| blob | 43105d20be35c8f9853b0f0ef647b899ffb66260 |
1 /* Fold a constant sub-tree into a single node for C-compiler
2 Copyright (C) 1987-2024 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
9 version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /*@@ This file should be rewritten to use an arbitrary precision
21 @@ representation for "struct tree_int_cst" and "struct tree_real_cst".
22 @@ Perhaps the routines could also be used for bc/dc, and made a lib.
23 @@ The routines that translate from the ap rep should
24 @@ warn if precision et. al. is lost.
25 @@ This would also make life easier when this technology is used
26 @@ for cross-compilers. */
28 /* The entry points in this file are fold, size_int_wide and size_binop.
30 fold takes a tree as argument and returns a simplified tree.
32 size_binop takes a tree code for an arithmetic operation
33 and two operands that are trees, and produces a tree for the
34 result, assuming the type comes from `sizetype'.
36 size_int takes an integer value, and creates a tree constant
37 with type from `sizetype'.
39 Note: Since the folders get called on non-gimple code as well as
40 gimple code, we need to handle GIMPLE tuples as well as their
41 corresponding tree equivalents. */
43 #define INCLUDE_ALGORITHM
90 /* Nonzero if we are folding constants inside an initializer or a C++
91 manifestly-constant-evaluated context; zero otherwise.
92 Should be used when folding in initializer enables additional
93 optimizations. */
96 /* Nonzero if we are folding C++ manifestly-constant-evaluated context; zero
97 otherwise.
98 Should be used when certain constructs shouldn't be optimized
99 during folding in that context. */
102 /* The following constants represent a bit based encoding of GCC's
103 comparison operators. This encoding simplifies transformations
104 on relational comparison operators, such as AND and OR. */
122 };
154 /* This is a helper function to detect min/max for some operands of COND_EXPR.
155 The form is "(EXP0 CMP EXP1) ? EXP2 : EXP3". */
156 tree_code
158 {
168 {
170 {
171 /* X <= Y - 1 equals to X < Y. */
174 /* X > Y - 1 equals to X >= Y. */
177 /* a != MIN_RANGE<a> ? a : MIN_RANGE<a>+1 -> MAX_EXPR<MIN_RANGE<a>+1, a> */
179 {
180 value_range r;
189 }
190 }
192 {
193 /* X < Y + 1 equals to X <= Y. */
196 /* X >= Y + 1 equals to X > Y. */
199 /* a != MAX_RANGE<a> ? a : MAX_RANGE<a>-1 -> MIN_EXPR<MIN_RANGE<a>-1, a> */
201 {
202 value_range r;
211 }
212 }
213 }
216 {
221 }
223 }
225 /* Return EXPR_LOCATION of T if it is not UNKNOWN_LOCATION.
226 Otherwise, return LOC. */
228 static location_t
230 {
233 }
235 /* Similar to protected_set_expr_location, but never modify x in place,
236 if location can and needs to be set, unshare it. */
238 tree
240 {
246 {
249 }
251 }
252 \f
253 /* If ARG2 divides ARG1 with zero remainder, carries out the exact
254 division and returns the quotient. Otherwise returns
255 NULL_TREE. */
257 tree
259 {
260 widest_int quo;
267 }
268 \f
269 /* This is nonzero if we should defer warnings about undefined
270 overflow. This facility exists because these warnings are a
271 special case. The code to estimate loop iterations does not want
272 to issue any warnings, since it works with expressions which do not
273 occur in user code. Various bits of cleanup code call fold(), but
274 only use the result if it has certain characteristics (e.g., is a
275 constant); that code only wants to issue a warning if the result is
276 used. */
280 /* If a warning about undefined overflow is deferred, this is the
281 warning. Note that this may cause us to turn two warnings into
282 one, but that is fine since it is sufficient to only give one
283 warning per expression. */
287 /* If a warning about undefined overflow is deferred, this is the
288 level at which the warning should be emitted. */
292 /* Start deferring overflow warnings. We could use a stack here to
293 permit nested calls, but at present it is not necessary. */
295 void
297 {
299 }
301 /* Stop deferring overflow warnings. If there is a pending warning,
302 and ISSUE is true, then issue the warning if appropriate. STMT is
303 the statement with which the warning should be associated (used for
304 location information); STMT may be NULL. CODE is the level of the
305 warning--a warn_strict_overflow_code value. This function will use
306 the smaller of CODE and the deferred code when deciding whether to
307 issue the warning. CODE may be zero to mean to always use the
308 deferred code. */
310 void
312 {
314 location_t locus;
319 {
325 }
336 /* Use the smallest code level when deciding to issue the
337 warning. */
346 else
349 }
351 /* Stop deferring overflow warnings, ignoring any deferred
352 warnings. */
354 void
356 {
358 }
360 /* Whether we are deferring overflow warnings. */
362 bool
364 {
366 }
368 /* This is called when we fold something based on the fact that signed
369 overflow is undefined. */
371 void
373 {
375 {
378 {
381 }
382 }
385 }
386 \f
387 /* Return true if the built-in mathematical function specified by CODE
388 is odd, i.e. -f(x) == f(-x). */
390 bool
392 {
394 {
395 CASE_CFN_ASIN:
396 CASE_CFN_ASIN_FN:
397 CASE_CFN_ASINH:
398 CASE_CFN_ASINH_FN:
399 CASE_CFN_ATAN:
400 CASE_CFN_ATAN_FN:
401 CASE_CFN_ATANH:
402 CASE_CFN_ATANH_FN:
403 CASE_CFN_CASIN:
404 CASE_CFN_CASIN_FN:
405 CASE_CFN_CASINH:
406 CASE_CFN_CASINH_FN:
407 CASE_CFN_CATAN:
408 CASE_CFN_CATAN_FN:
409 CASE_CFN_CATANH:
410 CASE_CFN_CATANH_FN:
411 CASE_CFN_CBRT:
412 CASE_CFN_CBRT_FN:
413 CASE_CFN_CPROJ:
414 CASE_CFN_CPROJ_FN:
415 CASE_CFN_CSIN:
416 CASE_CFN_CSIN_FN:
417 CASE_CFN_CSINH:
418 CASE_CFN_CSINH_FN:
419 CASE_CFN_CTAN:
420 CASE_CFN_CTAN_FN:
421 CASE_CFN_CTANH:
422 CASE_CFN_CTANH_FN:
423 CASE_CFN_ERF:
424 CASE_CFN_ERF_FN:
425 CASE_CFN_LLROUND:
426 CASE_CFN_LLROUND_FN:
427 CASE_CFN_LROUND:
428 CASE_CFN_LROUND_FN:
429 CASE_CFN_ROUND:
430 CASE_CFN_ROUNDEVEN:
431 CASE_CFN_ROUNDEVEN_FN:
432 CASE_CFN_SIN:
433 CASE_CFN_SIN_FN:
434 CASE_CFN_SINH:
435 CASE_CFN_SINH_FN:
436 CASE_CFN_TAN:
437 CASE_CFN_TAN_FN:
438 CASE_CFN_TANH:
439 CASE_CFN_TANH_FN:
440 CASE_CFN_TRUNC:
441 CASE_CFN_TRUNC_FN:
444 CASE_CFN_LLRINT:
445 CASE_CFN_LLRINT_FN:
446 CASE_CFN_LRINT:
447 CASE_CFN_LRINT_FN:
448 CASE_CFN_NEARBYINT:
449 CASE_CFN_NEARBYINT_FN:
450 CASE_CFN_RINT:
451 CASE_CFN_RINT_FN:
456 }
458 }
460 /* Check whether we may negate an integer constant T without causing
461 overflow. */
463 bool
465 {
466 tree type;
475 }
477 /* Determine whether an expression T can be cheaply negated using
478 the function negate_expr without introducing undefined overflow. */
480 static bool
482 {
483 tree type;
492 {
497 /* Check that -CST will not overflow type. */
510 /* We want to canonicalize to positive real constants. Pretend
511 that only negative ones can be easily negated. */
519 {
523 /* Steps don't prevent negation. */
530 }
545 /* -(A + B) -> (-B) - A. */
548 /* -(A + B) -> (-A) - B. */
552 /* We can't turn -(A-B) into B-A when we honor signed zeros. */
561 /* INT_MIN/n * n doesn't overflow while negating one operand it does
562 if n is a (negative) power of two. */
573 /* Fall through. */
586 /* In general we can't negate A in A / B, because if A is INT_MIN and
587 B is not 1 we change the sign of the result. */
591 /* In general we can't negate B in A / B, because if A is INT_MIN and
592 B is 1, we may turn this into INT_MIN / -1 which is undefined
593 and actually traps on some architectures. */
602 /* Negate -((double)float) as (double)(-float). */
604 {
608 }
612 /* Negate -f(x) as f(-x). */
618 /* Optimize -((int)x >> 31) into (unsigned)x >> 31 for int. */
620 {
624 }
629 }
631 }
633 /* Given T, an expression, return a folded tree for -T or NULL_TREE, if no
634 simplification is possible.
635 If negate_expr_p would return true for T, NULL_TREE will never be
636 returned. */
638 static tree
640 {
642 tree tem;
645 {
646 /* Convert - (~A) to A + 1. */
670 {
675 }
679 {
680 tree_vector_builder elts;
684 {
689 }
692 }
715 {
716 /* -(A + B) -> (-B) - A. */
718 {
722 }
724 /* -(A + B) -> (-A) - B. */
726 {
730 }
731 }
735 /* - (A - B) -> B - A */
746 /* Fall through. */
750 {
759 }
767 /* In general we can't negate A in A / B, because if A is INT_MIN and
768 B is not 1 we change the sign of the result. */
774 /* In general we can't negate B in A / B, because if A is INT_MIN and
775 B is 1, we may turn this into INT_MIN / -1 which is undefined
776 and actually traps on some architectures. */
788 /* Convert -((double)float) into (double)(-float). */
790 {
794 }
798 /* Negate -f(x) as f(-x). */
801 {
807 }
811 /* Optimize -((int)x >> 31) into (unsigned)x >> 31 for int. */
813 {
816 {
823 }
824 }
829 }
832 }
834 /* A wrapper for fold_negate_expr_1. */
836 static tree
838 {
845 }
847 /* Like fold_negate_expr, but return a NEGATE_EXPR tree, if T cannot be
848 negated in a simpler way. Also allow for T to be NULL_TREE, in which case
849 return NULL_TREE. */
851 static tree
853 {
855 location_t loc;
868 }
869 \f
870 /* Split a tree IN into a constant, literal and variable parts that could be
871 combined with CODE to make IN. "constant" means an expression with
872 TREE_CONSTANT but that isn't an actual constant. CODE must be a
873 commutative arithmetic operation. Store the constant part into *CONP,
874 the literal in *LITP and return the variable part. If a part isn't
875 present, set it to null. If the tree does not decompose in this way,
876 return the entire tree as the variable part and the other parts as null.
878 If CODE is PLUS_EXPR we also split trees that use MINUS_EXPR. In that
879 case, we negate an operand that was subtracted. Except if it is a
880 literal for which we use *MINUS_LITP instead.
882 If NEGATE_P is true, we are negating all of IN, again except a literal
883 for which we use *MINUS_LITP instead. If a variable part is of pointer
884 type, it is negated after converting to TYPE. This prevents us from
885 generating illegal MINUS pointer expression. LOC is the location of
886 the converted variable part.
888 If IN is itself a literal or constant, return it as appropriate.
890 Note that we do not guarantee that any of the three values will be the
891 same type as IN, but they will have the same signedness and mode. */
893 static tree
897 {
905 /* Strip any conversions that don't change the machine mode or signedness. */
914 /* We can associate addition and subtraction together (even
915 though the C standard doesn't say so) for integers because
916 the value is not affected. For reals, the value might be
917 affected, so we can't. */
923 {
929 /* First see if either of the operands is a literal, then a constant. */
942 /* If we haven't dealt with either operand, this is not a case we can
943 decompose. Otherwise, VAR is either of the ones remaining, if any. */
948 else
951 /* Now do any needed negations. */
958 }
963 {
964 /* -1 - X is folded to ~X, undo that here. Do _not_ do this
965 when IN is constant. */
968 }
969 else
973 {
986 }
996 }
998 /* Re-associate trees split by the above function. T1 and T2 are
999 either expressions to associate or null. Return the new
1000 expression, if any. LOC is the location of the new expression. If
1001 we build an operation, do it in TYPE and with CODE. */
1003 static tree
1005 {
1007 {
1010 }
1014 /* If either input is CODE, a PLUS_EXPR, or a MINUS_EXPR, don't
1015 try to fold this since we will have infinite recursion. But do
1016 deal with any NEGATE_EXPRs. */
1020 {
1022 {
1035 }
1037 {
1040 }
1044 }
1048 }
1049 \f
1050 /* Check whether TYPE1 and TYPE2 are equivalent integer types, suitable
1051 for use in int_const_binop, size_binop and size_diffop. */
1053 static bool
1055 {
1062 {
1071 }
1076 }
1078 /* Combine two wide ints ARG1 and ARG2 under operation CODE to produce
1079 a new constant in RES. Return FALSE if we don't know how to
1080 evaluate CODE at compile-time. */
1082 bool
1086 {
1087 wide_int tmp;
1090 {
1112 /* It's unclear from the C standard whether shifts can overflow.
1113 The following code ignores overflow; perhaps a C standard
1114 interpretation ruling is needed. */
1121 {
1125 else
1127 }
1128 else
1133 else
1212 }
1214 }
1216 /* Returns true if we know who is smaller or equal, ARG1 or ARG2, and set the
1217 min value to RES. */
1218 bool
1220 {
1222 {
1225 }
1227 {
1230 }
1233 }
1235 /* Combine two poly int's ARG1 and ARG2 under operation CODE to
1236 produce a new constant in RES. Return FALSE if we don't know how
1237 to evaluate CODE at compile-time. */
1239 static bool
1243 {
1247 {
1265 else
1272 else
1290 }
1292 }
1294 /* Combine two integer constants ARG1 and ARG2 under operation CODE to
1295 produce a new constant. Return NULL_TREE if we don't know how to
1296 evaluate CODE at compile-time. */
1298 tree
1301 {
1302 poly_wide_int poly_res;
1308 {
1314 }
1323 }
1325 /* Return true if binary operation OP distributes over addition in operand
1326 OPNO, with the other operand being held constant. OPNO counts from 1. */
1328 static bool
1330 {
1332 {
1343 }
1344 }
1346 /* OP is the INDEXth operand to CODE (counting from zero) and OTHER_OP
1347 is the other operand. Try to use the value of OP to simplify the
1348 operation in one step, without having to process individual elements. */
1349 static tree
1352 {
1353 /* AND, IOR as well as XOR with a zerop can be simplified directly. */
1355 {
1357 {
1362 }
1363 }
1366 }
1369 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1370 constant. We assume ARG1 and ARG2 have the same data type, or at least
1371 are the same kind of constant and the same machine mode. Return zero if
1372 combining the constants is not allowed in the current operating mode. */
1374 static tree
1376 {
1377 /* Sanity check for the recursive cases. */
1385 {
1391 }
1394 {
1395 machine_mode mode;
1396 REAL_VALUE_TYPE d1;
1397 REAL_VALUE_TYPE d2;
1398 REAL_VALUE_TYPE value;
1399 REAL_VALUE_TYPE result;
1403 /* The following codes are handled by real_arithmetic. */
1405 {
1416 }
1424 /* Don't perform operation if we honor signaling NaNs and
1425 either operand is a signaling NaN. */
1431 /* Don't perform operation if it would raise a division
1432 by zero exception. */
1438 /* If either operand is a NaN, just return it. Otherwise, set up
1439 for floating-point trap; we return an overflow. */
1441 {
1442 /* Make resulting NaN value to be qNaN when flag_signaling_nans
1443 is off. */
1447 }
1449 {
1450 /* Make resulting NaN value to be qNaN when flag_signaling_nans
1451 is off. */
1455 }
1460 /* Don't constant fold this floating point operation if
1461 both operands are not NaN but the result is NaN, and
1462 flag_trapping_math. Such operations should raise an
1463 invalid operation exception. */
1471 /* Don't constant fold this floating point operation if
1472 the result has overflowed and flag_trapping_math. */
1480 /* Don't constant fold this floating point operation if the
1481 result may dependent upon the run-time rounding mode and
1482 flag_rounding_math is set, or if GCC's software emulation
1483 is unable to accurately represent the result. */
1493 }
1496 {
1497 FIXED_VALUE_TYPE f1;
1498 FIXED_VALUE_TYPE f2;
1499 FIXED_VALUE_TYPE result;
1504 /* The following codes are handled by fixed_arithmetic. */
1506 {
1518 {
1525 }
1530 }
1537 /* Propagate overflow flags. */
1541 }
1544 {
1553 {
1564 mpc_mul);
1578 mpc_div);
1579 /* Fallthru. */
1585 {
1586 /* Keep this algorithm in sync with
1587 tree-complex.cc:expand_complex_div_straight().
1589 Expand complex division to scalars, straightforward algorithm.
1590 a / b = ((ar*br + ai*bi)/t) + i((ai*br - ar*bi)/t)
1591 t = br*br + bi*bi
1592 */
1593 tree magsquared
1597 tree t1
1601 tree t2
1608 }
1609 else
1610 {
1611 /* Keep this algorithm in sync with
1612 tree-complex.cc:expand_complex_div_wide().
1614 Expand complex division to scalars, modified algorithm to minimize
1615 overflow with wide input ranges. */
1621 {
1622 /* In the TRUE branch, we compute
1623 ratio = br/bi;
1624 div = (br * ratio) + bi;
1625 tr = (ar * ratio) + ai;
1626 ti = (ai * ratio) - ar;
1627 tr = tr / div;
1628 ti = ti / div; */
1639 }
1640 else
1641 {
1642 /* In the FALSE branch, we compute
1643 ratio = d/c;
1644 divisor = (d * ratio) + c;
1645 tr = (b * ratio) + a;
1646 ti = b - (a * ratio);
1647 tr = tr / div;
1648 ti = ti / div; */
1660 }
1661 }
1666 }
1670 }
1672 tree simplified;
1684 {
1689 /* We can operate directly on the encoding if:
1691 a3 - a2 == a2 - a1 && b3 - b2 == b2 - b1
1692 implies
1693 (a3 op b3) - (a2 op b2) == (a2 op b2) - (a1 op b1)
1695 Addition and subtraction are the supported operators
1696 for which this is true. */
1699 /* We can operate directly on stepped encodings if:
1701 a3 - a2 == a2 - a1
1702 implies:
1703 (a3 op c) - (a2 op c) == (a2 op c) - (a1 op c)
1705 which is true if (x -> x op c) distributes over addition. */
1707 else
1708 /* Similarly in reverse. */
1710 tree_vector_builder elts;
1715 {
1721 /* It is possible that const_binop cannot handle the given
1722 code and return NULL_TREE */
1726 }
1729 }
1731 /* Shifts allow a scalar offset for a vector. */
1734 {
1737 tree_vector_builder elts;
1742 {
1747 /* It is possible that const_binop cannot handle the given
1748 code and return NULL_TREE. */
1752 }
1755 }
1757 }
1759 /* Overload that adds a TYPE parameter to be able to dispatch
1760 to fold_relational_const. */
1762 tree
1764 {
1768 /* ??? Until we make the const_binop worker take the type of the
1769 result as argument put those cases that need it here. */
1771 {
1788 {
1793 }
1799 {
1815 {
1820 ? NOP_EXPR
1827 }
1830 }
1836 {
1859 {
1872 }
1875 }
1878 }
1883 /* Make sure type and arg0 have the same saturating flag. */
1888 }
1890 /* Compute CODE ARG1 with resulting type TYPE with ARG1 being constant.
1891 Return zero if computing the constants is not possible. */
1893 tree
1895 {
1896 /* Don't perform the operation, other than NEGATE and ABS, if
1897 flag_signaling_nans is on and the operand is a signaling NaN. */
1907 {
1908 CASE_CONVERT:
1915 /* If the source address is 0, and the source address space
1916 cannot have a valid object at 0, fold to dest type null. */
1927 {
1928 /* Can't call fold_negate_const directly here as that doesn't
1929 handle all cases and we might not be able to negate some
1930 constants. */
1935 }
1945 {
1949 }
1957 /* Perform BIT_NOT_EXPR on each element individually. */
1959 {
1960 tree elem;
1962 /* This can cope with stepped encodings because ~x == -1 - x. */
1963 tree_vector_builder elements;
1967 {
1973 }
1976 }
2000 {
2023 else
2028 {
2034 }
2037 }
2046 }
2049 }
2051 /* Create a sizetype INT_CST node with NUMBER sign extended. KIND
2052 indicates which particular sizetype to create. */
2054 tree
2056 {
2058 }
2059 \f
2060 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
2061 is a tree code. The type of the result is taken from the operands.
2062 Both must be equivalent integer types, ala int_binop_types_match_p.
2063 If the operands are constant, so is the result. */
2065 tree
2067 {
2076 /* Handle the special case of two poly_int constants faster. */
2078 {
2079 /* And some specific cases even faster than that. */
2081 {
2088 }
2090 {
2094 }
2096 {
2100 }
2102 /* Handle general case of two integer constants. For sizetype
2103 constant calculations we always want to know about overflow,
2104 even in the unsigned case. */
2108 }
2111 }
2113 /* Given two values, either both of sizetype or both of bitsizetype,
2114 compute the difference between the two values. Return the value
2115 in signed type corresponding to the type of the operands. */
2117 tree
2119 {
2121 tree ctype;
2126 /* If the type is already signed, just do the simple thing. */
2134 else
2137 /* If either operand is not a constant, do the conversions to the signed
2138 type and subtract. The hardware will do the right thing with any
2139 overflow in the subtraction. */
2145 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
2146 Otherwise, subtract the other way, convert to CTYPE (we know that can't
2147 overflow) and negate (which can't either). Special-case a result
2148 of zero while we're here. */
2154 else
2158 MINUS_EXPR,
2160 }
2161 \f
2162 /* A subroutine of fold_convert_const handling conversions of an
2163 INTEGER_CST to another integer type. */
2165 static tree
2167 {
2168 /* Given an integer constant, make new constant with new type,
2169 appropriately sign-extended or truncated. Use widest_int
2170 so that any extension is done according ARG1's type. */
2177 }
2179 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2180 to an integer type. */
2182 static tree
2184 {
2186 tree t;
2188 /* The following code implements the floating point to integer
2189 conversion rules required by the Java Language Specification,
2190 that IEEE NaNs are mapped to zero and values that overflow
2191 the target precision saturate, i.e. values greater than
2192 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
2193 are mapped to INT_MIN. These semantics are allowed by the
2194 C and C++ standards that simply state that the behavior of
2195 FP-to-integer conversion is unspecified upon overflow. */
2197 wide_int val;
2198 REAL_VALUE_TYPE r;
2202 {
2209 }
2211 /* If R is NaN, return zero and show we have an overflow. */
2213 {
2216 }
2218 /* See if R is less than the lower bound or greater than the
2219 upper bound. */
2222 {
2226 {
2229 }
2230 }
2233 {
2236 {
2239 {
2242 }
2243 }
2244 }
2251 }
2253 /* A subroutine of fold_convert_const handling conversions of a
2254 FIXED_CST to an integer type. */
2256 static tree
2258 {
2259 tree t;
2261 scalar_mode mode;
2263 /* Right shift FIXED_CST to temp by fbit. */
2267 {
2269 HOST_BITS_PER_DOUBLE_INT,
2272 /* Left shift temp to temp_trunc by fbit. */
2274 HOST_BITS_PER_DOUBLE_INT,
2276 }
2277 else
2278 {
2281 }
2283 /* If FIXED_CST is negative, we need to round the value toward 0.
2284 By checking if the fractional bits are not zero to add 1 to temp. */
2290 /* Given a fixed-point constant, make new constant with new type,
2291 appropriately sign-extended or truncated. */
2299 }
2301 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2302 to another floating point type. */
2304 static tree
2306 {
2307 REAL_VALUE_TYPE value;
2308 tree t;
2310 /* If the underlying modes are the same, simply treat it as
2311 copy and rebuild with TREE_REAL_CST information and the
2312 given type. */
2314 {
2317 }
2319 /* Don't perform the operation if flag_signaling_nans is on
2320 and the operand is a signaling NaN. */
2325 /* With flag_rounding_math we should respect the current rounding mode
2326 unless the conversion is exact. */
2334 /* If converting an infinity or NAN to a representation that doesn't
2335 have one, set the overflow bit so that we can produce some kind of
2336 error message at the appropriate point if necessary. It's not the
2337 most user-friendly message, but it's better than nothing. */
2344 /* Regular overflow, conversion produced an infinity in a mode that
2345 can't represent them. */
2350 else
2353 }
2355 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2356 to a floating point type. */
2358 static tree
2360 {
2361 REAL_VALUE_TYPE value;
2362 tree t;
2370 }
2372 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2373 to another fixed-point type. */
2375 static tree
2377 {
2378 FIXED_VALUE_TYPE value;
2379 tree t;
2386 /* Propagate overflow flags. */
2390 }
2392 /* A subroutine of fold_convert_const handling conversions an INTEGER_CST
2393 to a fixed-point type. */
2395 static tree
2397 {
2398 FIXED_VALUE_TYPE value;
2399 tree t;
2401 double_int di;
2408 else
2416 /* Propagate overflow flags. */
2420 }
2422 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2423 to a fixed-point type. */
2425 static tree
2427 {
2428 FIXED_VALUE_TYPE value;
2429 tree t;
2437 /* Propagate overflow flags. */
2441 }
2443 /* Attempt to fold type conversion operation CODE of expression ARG1 to
2444 type TYPE. If no simplification can be done return NULL_TREE. */
2446 static tree
2448 {
2453 /* We can't widen types, since the runtime value could overflow the
2454 original type before being extended to the new type. */
2465 {
2472 }
2474 {
2476 {
2478 /* Avoid the folding if flag_rounding_math is on and the
2479 conversion is not exact. */
2481 {
2487 }
2489 }
2494 }
2496 {
2503 }
2505 {
2508 {
2511 /* We can't handle steps directly when extending, since the
2512 values need to wrap at the original precision first. */
2513 bool step_ok_p
2517 tree_vector_builder v;
2522 {
2528 }
2530 }
2531 }
2533 }
2535 /* Construct a vector of zero elements of vector type TYPE. */
2537 static tree
2539 {
2540 tree t;
2544 }
2546 /* Returns true, if ARG is convertible to TYPE using a NOP_EXPR. */
2548 bool
2550 {
2565 {
2587 }
2588 }
2590 /* Convert expression ARG to type TYPE. Used by the middle-end for
2591 simple conversions in preference to calling the front-end's convert. */
2593 tree
2595 {
2597 tree tem;
2608 {
2611 /* Handle conversions between pointers to different address spaces. */
2616 /* fall through */
2621 {
2625 }
2639 {
2643 }
2645 {
2649 }
2651 {
2655 }
2658 {
2676 }
2681 {
2685 }
2688 {
2703 }
2707 {
2716 integer_zero_node));
2718 {
2722 {
2728 }
2736 }
2740 }
2758 }
2759 fold_convert_exit:
2762 }
2763 \f
2764 /* Return false if expr can be assumed not to be an lvalue, true
2765 otherwise. */
2767 static bool
2769 {
2770 /* We only need to wrap lvalue tree codes. */
2772 {
2805 /* Assume the worst for front-end tree codes. */
2809 }
2812 }
2814 /* Return an expr equal to X but certainly not valid as an lvalue. */
2816 tree
2818 {
2819 /* While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to
2820 us. */
2827 }
2828 \f
2829 /* Given a tree comparison code, return the code that is the logical inverse.
2830 It is generally not safe to do this for floating-point comparisons, except
2831 for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
2832 ERROR_MARK in this case. */
2834 enum tree_code
2836 {
2842 {
2873 }
2874 }
2876 /* Similar, but return the comparison that results if the operands are
2877 swapped. This is safe for floating-point. */
2879 enum tree_code
2881 {
2883 {
2909 }
2910 }
2913 /* Convert a comparison tree code from an enum tree_code representation
2914 into a compcode bit-based encoding. This function is the inverse of
2915 compcode_to_comparison. */
2917 static enum comparison_code
2919 {
2921 {
2952 }
2953 }
2955 /* Convert a compcode bit-based encoding of a comparison operator back
2956 to GCC's enum tree_code representation. This function is the
2957 inverse of comparison_to_compcode. */
2959 static enum tree_code
2961 {
2963 {
2994 }
2995 }
2997 /* Return true if COND1 tests the opposite condition of COND2. */
2999 bool
3001 {
3004 && (invert_tree_comparison
3011 }
3013 /* Return a tree for the comparison which is the combination of
3014 doing the AND or OR (depending on CODE) of the two operations LCODE
3015 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
3016 the possibility of trapping if the mode has NaNs, and return NULL_TREE
3017 if this makes the transformation invalid. */
3019 tree
3024 {
3031 {
3042 }
3045 {
3046 /* Eliminate unordered comparisons, as well as LTGT and ORD
3047 which are not used unless the mode has NaNs. */
3053 }
3055 {
3056 /* Check that the original operation and the optimized ones will trap
3057 under the same condition. */
3068 /* In a short-circuited boolean expression the LHS might be
3069 such that the RHS, if evaluated, will never trap. For
3070 example, in ORD (x, y) && (x < y), we evaluate the RHS only
3071 if neither x nor y is NaN. (This is a mixed blessing: for
3072 example, the expression above will never trap, hence
3073 optimizing it to x < y would be invalid). */
3078 /* If the comparison was short-circuited, and only the RHS
3079 trapped, we may now generate a spurious trap. */
3084 /* If we changed the conditions that cause a trap, we lose. */
3087 }
3093 else
3094 {
3099 }
3100 }
3101 \f
3102 /* Return nonzero if two operands (typically of the same tree node)
3103 are necessarily equal. FLAGS modifies behavior as follows:
3105 If OEP_ONLY_CONST is set, only return nonzero for constants.
3106 This function tests whether the operands are indistinguishable;
3107 it does not test whether they are equal using C's == operation.
3108 The distinction is important for IEEE floating point, because
3109 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
3110 (2) two NaNs may be indistinguishable, but NaN!=NaN.
3112 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
3113 even though it may hold multiple values during a function.
3114 This is because a GCC tree node guarantees that nothing else is
3115 executed between the evaluation of its "operands" (which may often
3116 be evaluated in arbitrary order). Hence if the operands themselves
3117 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
3118 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
3119 unset means assuming isochronic (or instantaneous) tree equivalence.
3120 Unless comparing arbitrary expression trees, such as from different
3121 statements, this flag can usually be left unset.
3123 If OEP_PURE_SAME is set, then pure functions with identical arguments
3124 are considered the same. It is used when the caller has other ways
3125 to ensure that global memory is unchanged in between.
3127 If OEP_ADDRESS_OF is set, we are actually comparing addresses of objects,
3128 not values of expressions.
3130 If OEP_LEXICOGRAPHIC is set, then also handle expressions with side-effects
3131 such as MODIFY_EXPR, RETURN_EXPR, as well as STATEMENT_LISTs.
3133 If OEP_BITWISE is set, then require the values to be bitwise identical
3134 rather than simply numerically equal. Do not take advantage of things
3135 like math-related flags or undefined behavior; only return true for
3136 values that are provably bitwise identical in all circumstances.
3138 Unless OEP_MATCH_SIDE_EFFECTS is set, the function returns false on
3139 any operand with side effect. This is unnecesarily conservative in the
3140 case we know that arg0 and arg1 are in disjoint code paths (such as in
3141 ?: operator). In addition OEP_MATCH_SIDE_EFFECTS is used when comparing
3142 addresses with TREE_CONSTANT flag set so we know that &var == &var
3143 even if var is volatile. */
3145 bool
3148 {
3156 /* If either is ERROR_MARK, they aren't equal. */
3162 /* Similar, if either does not have a type (like a template id),
3163 they aren't equal. */
3167 /* Bitwise identity makes no sense if the values have different layouts. */
3172 /* We cannot consider pointers to different address space equal. */
3179 /* Check equality of integer constants before bailing out due to
3180 precision differences. */
3182 {
3183 /* Address of INTEGER_CST is not defined; check that we did not forget
3184 to drop the OEP_ADDRESS_OF flags. */
3187 }
3190 {
3191 /* If both types don't have the same signedness, then we can't consider
3192 them equal. We must check this before the STRIP_NOPS calls
3193 because they may change the signedness of the arguments. As pointers
3194 strictly don't have a signedness, require either two pointers or
3195 two non-pointers as well. */
3201 /* If both types don't have the same precision, then it is not safe
3202 to strip NOPs. */
3209 }
3210 #if 0
3211 /* FIXME: Fortran FE currently produce ADDR_EXPR of NOP_EXPR. Enable the
3212 sanity check once the issue is solved. */
3213 else
3214 /* Addresses of conversions and SSA_NAMEs (and many other things)
3215 are not defined. Check that we did not forget to drop the
3216 OEP_ADDRESS_OF/OEP_CONSTANT_ADDRESS_OF flags. */
3219 #endif
3221 /* In case both args are comparisons but with different comparison
3222 code, try to swap the comparison operands of one arg to produce
3223 a match and compare that variant. */
3227 {
3235 }
3238 {
3239 /* NOP_EXPR and CONVERT_EXPR are considered equal. */
3241 ;
3243 {
3244 /* If we are interested in comparing addresses ignore
3245 MEM_REF wrappings of the base that can appear just for
3246 TBAA reasons. */
3260 }
3261 else
3263 }
3265 /* When not checking adddresses, this is needed for conversions and for
3266 COMPONENT_REF. Might as well play it safe and always test this. */
3273 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
3274 We don't care about side effects in that case because the SAVE_EXPR
3275 takes care of that for us. In all other cases, two expressions are
3276 equal if they have no side effects. If we have two identical
3277 expressions with side effects that should be treated the same due
3278 to the only side effects being identical SAVE_EXPR's, that will
3279 be detected in the recursive calls below.
3280 If we are taking an invariant address of two identical objects
3281 they are necessarily equal as well. */
3288 /* Next handle constant cases, those for which we can return 1 even
3289 if ONLY_CONST is set. */
3292 {
3305 {
3306 /* If we do not distinguish between signed and unsigned zero,
3307 consider them equal. */
3310 }
3314 {
3329 }
3333 flags)
3335 flags));
3346 flags | OEP_ADDRESS_OF
3349 {
3350 /* In GIMPLE empty constructors are allowed in initializers of
3351 aggregates. */
3355 /* See sem_variable::equals in ipa-icf for a similar approach. */
3362 {
3363 /* For arrays, check that the sizes all match. */
3369 }
3378 /* Address of CONSTRUCTOR is defined in GENERIC to mean the value
3379 of the CONSTRUCTOR referenced indirectly. */
3383 {
3387 /* Check that the values are the same... */
3392 /* ... and that they apply to the same field! */
3398 flags)
3401 flags)))
3403 }
3406 }
3410 }
3412 /* Don't handle more cases for OEP_BITWISE, since we can't guarantee that
3413 two instances of undefined behavior will give identical results. */
3417 /* Define macros to test an operand from arg0 and arg1 for equality and a
3418 variant that allows null and views null as being different from any
3419 non-null value. In the latter case, if either is null, the both
3420 must be; otherwise, do the normal comparison. */
3421 #define OP_SAME(N) operand_equal_p (TREE_OPERAND (arg0, N), \
3422 TREE_OPERAND (arg1, N), flags)
3424 #define OP_SAME_WITH_NULL(N) \
3425 ((!TREE_OPERAND (arg0, N) || !TREE_OPERAND (arg1, N)) \
3426 ? TREE_OPERAND (arg0, N) == TREE_OPERAND (arg1, N) : OP_SAME (N))
3429 {
3431 /* Two conversions are equal only if signedness and modes match. */
3433 {
3434 CASE_CONVERT:
3442 }
3452 /* For commutative ops, allow the other order. */
3460 /* If either of the pointer (or reference) expressions we are
3461 dereferencing contain a side effect, these cannot be equal,
3462 but their addresses can be. */
3469 {
3472 {
3476 /* Verify that the access types are compatible. */
3480 }
3485 /* Require the same offset. */
3491 /* Fallthru. */
3499 {
3500 /* Require equal access sizes */
3506 flags)))
3508 /* Verify that access happens in similar types. */
3511 /* Verify that accesses are TBAA compatible. */
3520 /* Verify that alignment is compatible. */
3524 }
3527 /* TARGET_MEM_REF require equal extra operands. */
3538 /* Compare the array index by value if it is constant first as we
3539 may have different types but same value here. */
3545 /* Compare low bound and element size as with OEP_ADDRESS_OF
3546 we have to account for the offset of the ref. */
3551 array_ref_low_bound
3555 array_ref_element_size
3557 flags))));
3560 /* Handle operand 2 the same as for ARRAY_REF. Operand 0
3561 may be NULL when we're called to compare MEM_EXPRs. */
3564 {
3567 /* Most of time we only need to compare FIELD_DECLs for equality.
3568 However when determining address look into actual offsets.
3569 These may match for unions and unshared record types. */
3572 {
3575 {
3579 /* Non-FIELD_DECL operands can appear in C++ templates. */
3586 flags))
3588 }
3589 else
3591 }
3592 }
3603 }
3607 {
3609 /* Be sure we pass right ADDRESS_OF flag. */
3626 /* The multiplcation operands are commutative. */
3627 /* FALLTHRU */
3635 /* Otherwise take into account this is a commutative operation. */
3648 /* BIT_INSERT_EXPR has an implict operand as the type precision
3649 of op1. Need to check to make sure they are the same. */
3655 /* FALLTHRU */
3680 /* Virtual table reference. */
3692 {
3697 }
3702 }
3706 {
3710 /* If not both CALL_EXPRs are either internal or normal function
3711 functions, then they are not equal. */
3714 {
3715 /* If the CALL_EXPRs call different internal functions, then they
3716 are not equal. */
3719 }
3720 else
3721 {
3722 /* If the CALL_EXPRs call different functions, then they are not
3723 equal. */
3725 flags))
3727 }
3729 /* FIXME: We could skip this test for OEP_MATCH_SIDE_EFFECTS. */
3730 {
3734 else
3738 }
3740 /* Now see if all the arguments are the same. */
3741 {
3752 /* If we get here and both argument lists are exhausted
3753 then the CALL_EXPRs are equal. */
3755 }
3758 }
3761 /* Consider __builtin_sqrt equal to sqrt. */
3771 {
3772 /* Consider decls with the same name equal. The caller needs
3773 to make sure they refer to the same entity (such as a function
3774 formal parameter). */
3780 }
3785 {
3789 /* In GIMPLE constructors are used only to build vectors from
3790 elements. Individual elements in the constructor must be
3791 indexed in increasing order and form an initial sequence.
3793 We make no effort to compare nonconstant ones in GENERIC. */
3798 /* Be sure that vectors constructed have the same representation.
3799 We only tested element precision and modes to match.
3800 Vectors may be BLKmode and thus also check that the number of
3801 parts match. */
3814 {
3819 /* In GIMPLE the indexes can be either NULL or matching i.
3820 Double check this so we won't get false
3821 positives for GENERIC. */
3829 }
3831 }
3834 {
3835 /* Compare the STATEMENT_LISTs. */
3841 {
3842 /* The lists don't have the same number of statements. */
3848 flags & (OEP_LEXICOGRAPHIC
3851 }
3852 }
3857 {
3868 }
3872 }
3874 #undef OP_SAME
3875 #undef OP_SAME_WITH_NULL
3876 }
3878 /* Generate a hash value for an expression. This can be used iteratively
3879 by passing a previous result as the HSTATE argument. */
3881 void
3884 {
3890 {
3893 }
3903 {
3904 /* Alas, constants aren't shared, so we can't rely on pointer
3905 identity. */
3915 {
3919 else
3923 }
3925 {
3929 }
3939 {
3946 }
3948 /* We can just compare by pointer. */
3952 /* The node itself doesn't matter. */
3956 /* Ignore. */
3959 /* A list of expressions, for a CALL_EXPR or as the elements of a
3960 VECTOR_CST. */
3965 {
3971 {
3972 /* In GIMPLE the indexes can be either NULL or matching i. */
3976 {
3979 }
3980 else
3983 }
3985 }
3987 {
3988 tree_stmt_iterator i;
3993 }
4002 /* When referring to a built-in FUNCTION_DECL, use the __builtin__ form.
4003 Otherwise nodes that compare equal according to operand_equal_p might
4004 get different hash codes. However, don't do this for machine specific
4005 or front end builtins, since the function code is overloaded in those
4006 cases. */
4009 {
4012 }
4013 /* FALL THROUGH */
4016 {
4020 }
4024 {
4025 /* DECL's have a unique ID */
4027 }
4029 {
4030 /* For comparisons that can be swapped, use the lower
4031 tree code. */
4038 }
4040 {
4041 /* NOP_EXPR and CONVERT_EXPR are considered equal by
4042 operand_equal_p. */
4046 /* Don't hash the type, that can lead to having nodes which
4047 compare equal according to operand_equal_p, but which
4048 have different hash codes. Make sure to include signedness
4049 in the hash computation. */
4052 }
4053 /* For OEP_ADDRESS_OF, hash MEM_EXPR[&decl, 0] the same as decl. */
4061 /* Don't ICE on FE specific trees, or their arguments etc.
4062 during operand_equal_p hash verification. */
4065 else
4066 {
4072 {
4088 {
4095 }
4109 {
4110 /* The multiplication operands are commutative. */
4117 }
4125 /* For TARGET_EXPR, just hash on the TARGET_EXPR_SLOT.
4126 Usually different TARGET_EXPRs just should use
4127 different temporaries in their slots. */
4132 /* Virtual table reference. */
4140 {
4142 /* We compute mangled names only when free_lang_data is run.
4143 In that case we can hash precisely. */
4146 hstate.add_object
4147 (IDENTIFIER_HASH_VALUE
4149 }
4153 }
4155 /* Don't hash the type, that can lead to having nodes which
4156 compare equal according to operand_equal_p, but which
4157 have different hash codes. */
4159 {
4160 /* Make sure to include signness in the hash computation. */
4163 }
4166 {
4167 /* It's a commutative expression. We want to hash it the same
4168 however it appears. We do this by first hashing both operands
4169 and then rehashing based on the order of their independent
4170 hashes. */
4175 }
4176 else
4180 }
4182 }
4183 }
4185 bool
4188 {
4189 /* When checking and unless comparing DECL names, verify that if
4190 the outermost operand_equal_p call returns non-zero then ARG0
4191 and ARG1 have the same hash value. */
4193 {
4195 {
4197 {
4204 }
4206 }
4207 else
4211 }
4214 }
4219 /* Conveinece wrapper around operand_compare class because usually we do
4220 not need to play with the valueizer. */
4222 bool
4224 {
4226 }
4228 namespace inchash
4229 {
4231 /* Generate a hash value for an expression. This can be used iteratively
4232 by passing a previous result as the HSTATE argument.
4234 This function is intended to produce the same hash for expressions which
4235 would compare equal using operand_equal_p. */
4236 void
4238 {
4240 }
4242 }
4243 \f
4244 /* Similar to operand_equal_p, but see if ARG0 might be a variant of ARG1
4245 with a different signedness or a narrower precision. */
4247 static bool
4249 {
4257 /* Discard any conversions that don't change the modes of ARG0 and ARG1
4258 and see if the inner values are the same. This removes any
4259 signedness comparison, which doesn't matter here. */
4267 /* Discard a single widening conversion from ARG1 and see if the inner
4268 value is the same as ARG0. */
4277 }
4278 \f
4279 /* See if ARG is an expression that is either a comparison or is performing
4280 arithmetic on comparisons. The comparisons must only be comparing
4281 two different values, which will be stored in *CVAL1 and *CVAL2; if
4282 they are nonzero it means that some operands have already been found.
4283 No variables may be used anywhere else in the expression except in the
4284 comparisons.
4286 If this is true, return 1. Otherwise, return zero. */
4288 static bool
4290 {
4294 /* We can handle some of the tcc_expression cases here. */
4303 {
4322 /* First see if we can handle the first operand, then the second. For
4323 the second operand, we know *CVAL1 can't be zero. It must be that
4324 one side of the comparison is each of the values; test for the
4325 case where this isn't true by failing if the two operands
4326 are the same. */
4335 ;
4339 ;
4340 else
4344 ;
4348 ;
4349 else
4356 }
4357 }
4358 \f
4359 /* ARG is a tree that is known to contain just arithmetic operations and
4360 comparisons. Evaluate the operations in the tree substituting NEW0 for
4361 any occurrence of OLD0 as an operand of a comparison and likewise for
4362 NEW1 and OLD1. */
4364 static tree
4367 {
4372 /* We can handle some of the tcc_expression cases here. */
4380 {
4395 {
4414 }
4415 /* Fall through - ??? */
4418 {
4422 /* We need to check both for exact equality and tree equality. The
4423 former will be true if the operand has a side-effect. In that
4424 case, we know the operand occurred exactly once. */
4437 }
4441 }
4442 }
4443 \f
4444 /* Return a tree for the case when the result of an expression is RESULT
4445 converted to TYPE and OMITTED was previously an operand of the expression
4446 but is now not needed (e.g., we folded OMITTED * 0).
4448 If OMITTED has side effects, we must evaluate it. Otherwise, just do
4449 the conversion of RESULT to TYPE. */
4451 tree
4453 {
4456 /* If the resulting operand is an empty statement, just return the omitted
4457 statement casted to void. */
4467 }
4469 /* Return a tree for the case when the result of an expression is RESULT
4470 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
4471 of the expression but are now not needed.
4473 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
4474 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
4475 evaluated before OMITTED2. Otherwise, if neither has side effects,
4476 just do the conversion of RESULT to TYPE. */
4478 tree
4481 {
4490 }
4492 \f
4493 /* Return a simplified tree node for the truth-negation of ARG. This
4494 never alters ARG itself. We assume that ARG is an operation that
4495 returns a truth value (0 or 1).
4497 FIXME: one would think we would fold the result, but it causes
4498 problems with the dominator optimizer. */
4500 static tree
4502 {
4507 /* If this is a comparison, we can simply invert it, except for
4508 floating-point non-equality comparisons, in which case we just
4509 enclose a TRUTH_NOT_EXPR around what we have. */
4512 {
4515 && flag_trapping_math
4528 }
4531 {
4550 /* Here we can invert either operand. We invert the first operand
4551 unless the second operand is a TRUTH_NOT_EXPR in which case our
4552 result is the XOR of the first operand with the inside of the
4553 negation of the second operand. */
4558 else
4581 {
4588 /* A COND_EXPR may have a throw as one operand, which
4589 then has void type. Just leave void operands
4590 as they are. */
4596 }
4608 CASE_CONVERT:
4612 /* fall through */
4634 }
4635 }
4637 /* Fold the truth-negation of ARG. This never alters ARG itself. We
4638 assume that ARG is an operation that returns a truth value (0 or 1
4639 for scalars, 0 or -1 for vectors). Return the folded expression if
4640 folding is successful. Otherwise, return NULL_TREE. */
4642 static tree
4644 {
4647 ? BIT_NOT_EXPR
4650 }
4652 /* Return a simplified tree node for the truth-negation of ARG. This
4653 never alters ARG itself. We assume that ARG is an operation that
4654 returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). */
4656 tree
4658 {
4664 ? BIT_NOT_EXPR
4667 }
4668 \f
4669 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
4670 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero
4671 and uses reverse storage order if REVERSEP is nonzero. ORIG_INNER
4672 is the original memory reference used to preserve the alias set of
4673 the access. */
4675 static tree
4679 {
4682 /* Attempt not to lose the access path if possible. */
4684 {
4686 machine_mode nmode;
4688 tree noffset;
4696 && !reversep
4697 && !nreversep
4699 {
4702 }
4703 }
4712 {
4719 }
4734 }
4736 /* Optimize a bit-field compare.
4738 There are two cases: First is a compare against a constant and the
4739 second is a comparison of two items where the fields are at the same
4740 bit position relative to the start of a chunk (byte, halfword, word)
4741 large enough to contain it. In these cases we can avoid the shift
4742 implicit in bitfield extractions.
4744 For constants, we emit a compare of the shifted constant with the
4745 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
4746 compared. For two fields at the same position, we do the ANDs with the
4747 similar mask and compare the result of the ANDs.
4749 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
4750 COMPARE_TYPE is the type of the comparison, and LHS and RHS
4751 are the left and right operands of the comparison, respectively.
4753 If the optimization described above can be done, we return the resulting
4754 tree. Otherwise we return zero. */
4756 static tree
4759 {
4763 tree unsigned_type;
4766 scalar_int_mode nmode;
4771 tree mask;
4772 tree offset;
4774 /* Get all the information about the extractions being done. If the bit size
4775 is the same as the size of the underlying object, we aren't doing an
4776 extraction at all and so can do nothing. We also don't want to
4777 do anything if the inner expression is a PLACEHOLDER_EXPR since we
4778 then will no longer be able to replace it. */
4793 else
4794 {
4795 /* If this is not a constant, we can only do something if bit positions,
4796 sizes, signedness and storage order are the same. */
4797 rinner
4810 }
4812 /* Honor the C++ memory model and mimic what RTL expansion does. */
4816 {
4820 }
4822 /* See if we can find a mode to refer to this field. We should be able to,
4823 but fail if we can't. */
4831 /* Set signed and unsigned types of the precision of this mode for the
4832 shifts below. */
4835 /* Compute the bit position and size for the new reference and our offset
4836 within it. If the new reference is the same size as the original, we
4837 won't optimize anything, so return zero. */
4847 /* Make the mask to be used against the extracted field. */
4854 {
4858 /* If not comparing with constant, just rework the comparison
4859 and return. */
4867 }
4869 /* Otherwise, we are handling the constant case. See if the constant is too
4870 big for the field. Warn and return a tree for 0 (false) if so. We do
4871 this not only for its own sake, but to avoid having to test for this
4872 error case below. If we didn't, we might generate wrong code.
4874 For unsigned fields, the constant shifted right by the field length should
4875 be all zero. For signed fields, the high-order bits should agree with
4876 the sign bit. */
4879 {
4881 {
4885 }
4886 }
4887 else
4888 {
4891 {
4895 }
4896 }
4901 /* Single-bit compares should always be against zero. */
4903 {
4906 }
4908 /* Make a new bitfield reference, shift the constant over the
4909 appropriate number of bits and mask it with the computed mask
4910 (in case this was a signed field). If we changed it, make a new one. */
4918 mask);
4923 }
4924 \f
4925 /* Subroutine for fold_truth_andor_1: decode a field reference.
4927 If EXP is a comparison reference, we return the innermost reference.
4929 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
4930 set to the starting bit number.
4932 If the innermost field can be completely contained in a mode-sized
4933 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
4935 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
4936 otherwise it is not changed.
4938 *PUNSIGNEDP is set to the signedness of the field.
4940 *PREVERSEP is set to the storage order of the field.
4942 *PMASK is set to the mask used. This is either contained in a
4943 BIT_AND_EXPR or derived from the width of the field.
4945 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
4947 Return 0 if this is not a component reference or is one that we can't
4948 do anything with. */
4950 static tree
4955 {
4960 tree unsigned_type;
4963 /* All the optimizations using this function assume integer fields.
4964 There are problems with FP fields since the type_for_size call
4965 below can fail for, e.g., XFmode. */
4969 /* We are interested in the bare arrangement of bits, so strip everything
4970 that doesn't affect the machine mode. However, record the type of the
4971 outermost expression if it may matter below. */
4978 {
4984 }
4995 /* Reject out-of-bound accesses (PR79731). */
5007 /* If the number of bits in the reference is the same as the bitsize of
5008 the outer type, then the outer type gives the signedness. Otherwise
5009 (in case of a small bitfield) the signedness is unchanged. */
5013 /* Compute the mask to access the bitfield. */
5021 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
5029 }
5031 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
5032 bit positions and MASK is SIGNED. */
5034 static bool
5036 {
5040 /* If this function returns true when the type of the mask is
5041 UNSIGNED, then there will be errors. In particular see
5042 gcc.c-torture/execute/990326-1.c. There does not appear to be
5043 any documentation paper trail as to why this is so. But the pre
5044 wide-int worked with that restriction and it has been preserved
5045 here. */
5050 }
5052 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
5053 represents the sign bit of EXP's type. If EXP represents a sign
5054 or zero extension, also test VAL against the unextended type.
5055 The return value is the (sub)expression whose sign bit is VAL,
5056 or NULL_TREE otherwise. */
5058 tree
5060 {
5062 tree t;
5064 /* Tree EXP must have an integral type. */
5069 /* Tree VAL must be an integer constant. */
5078 /* Handle extension from a narrower type. */
5084 }
5086 /* Subroutine for fold_truth_andor_1 and simple_condition_p: determine if an
5087 operand is simple enough to be evaluated unconditionally. */
5089 static bool
5091 {
5092 /* Strip any conversions that don't change the machine mode. */
5101 /* Don't regard global variables as simple. They may be
5102 allocated in ways unknown to the compiler (shared memory,
5103 #pragma weak, etc). */
5106 /* Weakrefs are not safe to be read, since they can be NULL.
5107 They are !TREE_PUBLIC && !DECL_EXTERNAL but still
5108 have DECL_WEAK flag set. */
5110 /* Loading a static variable is unduly expensive, but global
5111 registers aren't expensive. */
5113 }
5115 /* Determine if an operand is simple enough to be evaluated unconditionally.
5116 In addition to simple_operand_p, we assume that comparisons, conversions,
5117 and logic-not operations are simple, if their operands are simple, too. */
5119 bool
5121 {
5140 }
5142 \f
5143 /* The following functions are subroutines to fold_range_test and allow it to
5144 try to change a logical combination of comparisons into a range test.
5146 For example, both
5147 X == 2 || X == 3 || X == 4 || X == 5
5148 and
5149 X >= 2 && X <= 5
5150 are converted to
5151 (unsigned) (X - 2) <= 3
5153 We describe each set of comparisons as being either inside or outside
5154 a range, using a variable named like IN_P, and then describe the
5155 range with a lower and upper bound. If one of the bounds is omitted,
5156 it represents either the highest or lowest value of the type.
5158 In the comments below, we represent a range by two numbers in brackets
5159 preceded by a "+" to designate being inside that range, or a "-" to
5160 designate being outside that range, so the condition can be inverted by
5161 flipping the prefix. An omitted bound is represented by a "-". For
5162 example, "- [-, 10]" means being outside the range starting at the lowest
5163 possible value and ending at 10, in other words, being greater than 10.
5164 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
5165 always false.
5167 We set up things so that the missing bounds are handled in a consistent
5168 manner so neither a missing bound nor "true" and "false" need to be
5169 handled using a special case. */
5171 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
5172 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
5173 and UPPER1_P are nonzero if the respective argument is an upper bound
5174 and zero for a lower. TYPE, if nonzero, is the type of the result; it
5175 must be specified for a comparison. ARG1 will be converted to ARG0's
5176 type if both are specified. */
5178 static tree
5181 {
5182 tree tem;
5186 /* If neither arg represents infinity, do the normal operation.
5187 Else, if not a comparison, return infinity. Else handle the special
5188 comparison rules. Note that most of the cases below won't occur, but
5189 are handled for consistency. */
5192 {
5197 }
5202 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
5203 for neither. In real maths, we cannot assume open ended ranges are
5204 the same. But, this is computer arithmetic, where numbers are finite.
5205 We can therefore make the transformation of any unbounded range with
5206 the value Z, Z being greater than any representable number. This permits
5207 us to treat unbounded ranges as equal. */
5211 {
5232 }
5235 }
5236 \f
5237 /* Helper routine for make_range. Perform one step for it, return
5238 new expression if the loop should continue or NULL_TREE if it should
5239 stop. */
5241 tree
5245 {
5251 {
5253 /* We can only do something if the range is testing for zero. */
5262 /* We can only do something if the range is testing for zero
5263 and if the second operand is an integer constant. Note that
5264 saying something is "in" the range we make is done by
5265 complementing IN_P since it will set in the initial case of
5266 being not equal to zero; "out" is leaving it alone. */
5273 {
5294 }
5296 /* If this is an unsigned comparison, we also know that EXP is
5297 greater than or equal to zero. We base the range tests we make
5298 on that fact, so we record it here so we can parse existing
5299 range tests. We test arg0_type since often the return type
5300 of, e.g. EQ_EXPR, is boolean. */
5302 {
5306 NULL_TREE))
5311 /* If the high bound is missing, but we have a nonzero low
5312 bound, reverse the range so it goes from zero to the low bound
5313 minus 1. */
5315 {
5320 }
5321 }
5329 /* If flag_wrapv and ARG0_TYPE is signed, make sure
5330 low and high are non-NULL, then normalize will DTRT. */
5333 {
5338 }
5340 /* (-x) IN [a,b] -> x in [-b, -a] */
5352 /* ~ X -> -X - 1 */
5361 /* If flag_wrapv and ARG0_TYPE is signed, then we cannot
5362 move a constant to the other side. */
5367 /* If EXP is signed, any overflow in the computation is undefined,
5368 so we don't worry about it so long as our computations on
5369 the bounds don't overflow. For unsigned, overflow is defined
5370 and this is exactly the right thing. */
5382 normalize:
5383 /* Check for an unsigned range which has wrapped around the maximum
5384 value thus making n_high < n_low, and normalize it. */
5386 {
5392 /* If the range is of the form +/- [ x+1, x ], we won't
5393 be able to normalize it. But then, it represents the
5394 whole range or the empty set, so make it
5395 +/- [ -, - ]. */
5399 else
5401 }
5402 else
5410 CASE_CONVERT:
5428 /* If we're converting arg0 from an unsigned type, to exp,
5429 a signed type, we will be doing the comparison as unsigned.
5430 The tests above have already verified that LOW and HIGH
5431 are both positive.
5433 So we have to ensure that we will handle large unsigned
5434 values the same way that the current signed bounds treat
5435 negative values. */
5438 {
5439 tree high_positive;
5440 tree equiv_type;
5441 /* For fixed-point modes, we need to pass the saturating flag
5442 as the 2nd parameter. */
5444 equiv_type
5449 else
5450 equiv_type
5453 /* A range without an upper bound is, naturally, unbounded.
5454 Since convert would have cropped a very large value, use
5455 the max value for the destination type. */
5456 high_positive
5463 high_positive),
5466 /* If the low bound is specified, "and" the range with the
5467 range for which the original unsigned value will be
5468 positive. */
5470 {
5473 integer_zero_node),
5474 high_positive))
5478 }
5479 else
5480 {
5481 /* Otherwise, "or" the range with the range of the input
5482 that will be interpreted as negative. */
5485 integer_zero_node),
5486 high_positive))
5490 }
5491 }
5493 /* Otherwise, if we are converting arg0 from signed type, to exp,
5494 an unsigned type, we will do the comparison as signed. If
5495 high is non-NULL, we punt above if it doesn't fit in the signed
5496 type, so if we get through here, +[-, high] or +[low, high] are
5497 equivalent to +[-, n_high] or +[n_low, n_high]. Similarly,
5498 +[-, -] or -[-, -] are equivalent too. But if low is specified and
5499 high is not, the +[low, -] range is equivalent to union of
5500 +[n_low, -] and +[-, -1] ranges, so +[low, -] is equivalent to
5501 -[0, n_low-1] and similarly -[low, -] to +[0, n_low-1], except for
5502 low being 0, which should be treated as [-, -]. */
5505 && low
5507 {
5510 else
5511 {
5516 }
5517 }
5526 }
5527 }
5529 /* Given EXP, a logical expression, set the range it is testing into
5530 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
5531 actually being tested. *PLOW and *PHIGH will be made of the same
5532 type as the returned expression. If EXP is not a comparison, we
5533 will most likely not be returning a useful value and range. Set
5534 *STRICT_OVERFLOW_P to true if the return value is only valid
5535 because signed overflow is undefined; otherwise, do not change
5536 *STRICT_OVERFLOW_P. */
5538 tree
5541 {
5549 /* Start with simply saying "EXP != 0" and then look at the code of EXP
5550 and see if we can refine the range. Some of the cases below may not
5551 happen, but it doesn't seem worth worrying about this. We "continue"
5552 the outer loop when we've changed something; otherwise we "break"
5553 the switch, which will "break" the while. */
5559 {
5565 {
5573 }
5582 }
5584 /* If EXP is a constant, we can evaluate whether this is true or false. */
5586 {
5593 }
5597 }
5599 /* Returns TRUE if [LOW, HIGH] range check can be optimized to
5600 a bitwise check i.e. when
5601 LOW == 0xXX...X00...0
5602 HIGH == 0xXX...X11...1
5603 Return corresponding mask in MASK and stem in VALUE. */
5605 static bool
5608 {
5631 }
5632 \f
5633 /* Helper routine for build_range_check and match.pd. Return the type to
5634 perform the check or NULL if it shouldn't be optimized. */
5636 tree
5638 {
5639 /* First make sure that arithmetics in this type is valid, then make sure
5640 that it wraps around. */
5645 {
5648 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
5649 for the type in question, as we rely on this here. */
5659 else
5661 }
5664 /* Right now all BITINT_TYPEs satisfy
5665 (unsigned) max + 1 == (unsigned) min, so no need to verify
5666 that like for INTEGER_TYPEs. */
5670 }
5672 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
5673 type, TYPE, return an expression to test if EXP is in (or out of, depending
5674 on IN_P) the range. Return 0 if the test couldn't be created. */
5676 tree
5679 {
5682 /* Disable this optimization for function pointer expressions
5683 on targets that require function pointer canonicalization. */
5690 {
5696 }
5718 value);
5721 {
5723 {
5727 }
5729 }
5731 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
5733 {
5737 {
5739 {
5742 etype
5744 else
5747 }
5750 }
5751 }
5753 /* Optimize (c>=low) && (c<=high) into (c-low>=0) && (c-low<=high-low).
5754 This requires wrap-around arithmetics for the type of the expression. */
5771 }
5772 \f
5773 /* Return the predecessor of VAL in its type, handling the infinite case. */
5775 static tree
5777 {
5783 else
5786 }
5788 /* Return the successor of VAL in its type, handling the infinite case. */
5790 static tree
5792 {
5798 else
5801 }
5803 /* Given two ranges, see if we can merge them into one. Return 1 if we
5804 can, 0 if we can't. Set the output range into the specified parameters. */
5806 bool
5809 {
5813 tree tem;
5823 /* Make range 0 be the range that starts first, or ends last if they
5824 start at the same value. Swap them if it isn't. */
5827 || (lowequal
5830 {
5834 }
5836 /* If the second range is != high1 where high1 is the type maximum of
5837 the type, try first merging with < high1 range. */
5839 && high1
5846 {
5851 /* Similarly for the second range != low1 where low1 is the type minimum
5852 of the type, try first merging with > low1 range. */
5857 }
5859 /* Now flag two cases, whether the ranges are disjoint or whether the
5860 second range is totally subsumed in the first. Note that the tests
5861 below are simplified by the ones above. */
5867 /* We now have four cases, depending on whether we are including or
5868 excluding the two ranges. */
5870 {
5871 /* If they don't overlap, the result is false. If the second range
5872 is a subset it is the result. Otherwise, the range is from the start
5873 of the second to the end of the first. */
5878 else
5880 }
5883 {
5884 /* If they don't overlap, the result is the first range. If they are
5885 equal, the result is false. If the second range is a subset of the
5886 first, and the ranges begin at the same place, we go from just after
5887 the end of the second range to the end of the first. If the second
5888 range is not a subset of the first, or if it is a subset and both
5889 ranges end at the same place, the range starts at the start of the
5890 first range and ends just before the second range.
5891 Otherwise, we can't describe this as a single range. */
5897 {
5902 {
5903 /* We are in the weird situation where high0 > high1 but
5904 high1 has no successor. Punt. */
5906 }
5907 }
5909 {
5914 {
5915 /* low0 < low1 but low1 has no predecessor. Punt. */
5917 }
5918 }
5919 else
5921 }
5924 {
5925 /* If they don't overlap, the result is the second range. If the second
5926 is a subset of the first, the result is false. Otherwise,
5927 the range starts just after the first range and ends at the
5928 end of the second. */
5933 else
5934 {
5939 {
5940 /* high1 > high0 but high0 has no successor. Punt. */
5942 }
5943 }
5944 }
5946 else
5947 {
5948 /* The case where we are excluding both ranges. Here the complex case
5949 is if they don't overlap. In that case, the only time we have a
5950 range is if they are adjacent. If the second is a subset of the
5951 first, the result is the first. Otherwise, the range to exclude
5952 starts at the beginning of the first range and ends at the end of the
5953 second. */
5955 {
5960 else
5961 {
5962 /* Canonicalize - [min, x] into - [-, x]. */
5965 {
5968 GET_MODE_BITSIZE
5971 /* FALLTHROUGH */
5984 }
5986 /* Canonicalize - [x, max] into - [x, -]. */
5989 {
5992 GET_MODE_BITSIZE
5995 /* FALLTHROUGH */
6011 }
6013 /* The ranges might be also adjacent between the maximum and
6014 minimum values of the given type. For
6015 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
6016 return + [x + 1, y - 1]. */
6018 {
6025 }
6026 else
6028 }
6029 }
6032 else
6034 }
6038 }
6039 \f
6041 /* Subroutine of fold, looking inside expressions of the form
6042 A op B ? A : C, where (ARG00, COMP_CODE, ARG01), ARG1 and ARG2
6043 are the three operands of the COND_EXPR. This function is
6044 being used also to optimize A op B ? C : A, by reversing the
6045 comparison first.
6047 Return a folded expression whose code is not a COND_EXPR
6048 anymore, or NULL_TREE if no folding opportunity is found. */
6050 static tree
6054 {
6056 tree tem;
6061 /* If we have A op 0 ? A : -A, consider applying the following
6062 transformations:
6064 A == 0? A : -A same as -A
6065 A != 0? A : -A same as A
6066 A >= 0? A : -A same as abs (A)
6067 A > 0? A : -A same as abs (A)
6068 A <= 0? A : -A same as -abs (A)
6069 A < 0? A : -A same as -abs (A)
6071 None of these transformations work for modes with signed
6072 zeros. If A is +/-0, the first two transformations will
6073 change the sign of the result (from +0 to -0, or vice
6074 versa). The last four will fix the sign of the result,
6075 even though the original expressions could be positive or
6076 negative, depending on the sign of A.
6078 Note that all these transformations are correct if A is
6079 NaN, since the two alternatives (A and -A) are also NaNs. */
6086 /* In the case that A is of the form X-Y, '-A' (arg2) may
6087 have already been folded to Y-X, check for that. */
6095 {
6107 /* Fall through. */
6118 /* FALLTHRU */
6125 {
6126 /* A <= 0 ? A : -A for A INT_MIN is valid, but -abs(INT_MIN)
6127 is not, invokes UB both in abs and in the negation of it.
6128 So, use ABSU_EXPR instead. */
6133 }
6134 else
6135 {
6138 }
6142 }
6144 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
6145 A == 0 ? A : 0 is always 0 unless A is -0. Note that
6146 both transformations are correct when A is NaN: A != 0
6147 is then true, and A == 0 is false. */
6151 {
6156 }
6158 /* Try some transformations of A op B ? A : B.
6160 A == B? A : B same as B
6161 A != B? A : B same as A
6162 A >= B? A : B same as max (A, B)
6163 A > B? A : B same as max (B, A)
6164 A <= B? A : B same as min (A, B)
6165 A < B? A : B same as min (B, A)
6167 As above, these transformations don't work in the presence
6168 of signed zeros. For example, if A and B are zeros of
6169 opposite sign, the first two transformations will change
6170 the sign of the result. In the last four, the original
6171 expressions give different results for (A=+0, B=-0) and
6172 (A=-0, B=+0), but the transformed expressions do not.
6174 The first two transformations are correct if either A or B
6175 is a NaN. In the first transformation, the condition will
6176 be false, and B will indeed be chosen. In the case of the
6177 second transformation, the condition A != B will be true,
6178 and A will be chosen.
6180 The conversions to max() and min() are not correct if B is
6181 a number and A is not. The conditions in the original
6182 expressions will be false, so all four give B. The min()
6183 and max() versions would give a NaN instead. */
6186 /* Avoid these transformations if the COND_EXPR may be used
6187 as an lvalue in the C++ front-end. PR c++/19199. */
6188 && (in_gimple_form
6194 {
6200 {
6209 /* In C++ a ?: expression can be an lvalue, so put the
6210 operand which will be used if they are equal first
6211 so that we can convert this back to the
6212 corresponding COND_EXPR. */
6214 {
6222 }
6229 {
6237 }
6250 }
6251 }
6254 }
6257 \f
6258 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
6259 #define LOGICAL_OP_NON_SHORT_CIRCUIT \
6260 (BRANCH_COST (optimize_function_for_speed_p (cfun), \
6261 false) >= 2)
6262 #endif
6264 /* EXP is some logical combination of boolean tests. See if we can
6265 merge it into some range test. Return the new tree if so. */
6267 static tree
6270 {
6284 /* If op0 is known true or false and this is a short-circuiting
6285 operation we must not merge with op1 since that makes side-effects
6286 unconditional. So special-case this. */
6293 /* If this is an OR operation, invert both sides; we will invert
6294 again at the end. */
6298 /* If both expressions are the same, if we can merge the ranges, and we
6299 can build the range test, return it or it inverted. If one of the
6300 ranges is always true or always false, consider it to be the same
6301 expression as the other. */
6309 {
6313 }
6315 /* On machines where the branch cost is expensive, if this is a
6316 short-circuited branch and the underlying object on both sides
6317 is the same, make a non-short-circuit operation. */
6320 logical_op_non_short_circuit
6327 {
6328 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
6329 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
6330 which cases we can't do this. */
6338 {
6347 {
6350 WARN_STRICT_OVERFLOW_COMPARISON);
6354 }
6355 }
6356 }
6359 }
6360 \f
6361 /* Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P
6362 bit value. Arrange things so the extra bits will be set to zero if and
6363 only if C is signed-extended to its full width. If MASK is nonzero,
6364 it is an INTEGER_CST that should be AND'ed with the extra bits. */
6366 static tree
6368 {
6371 tree temp;
6376 /* We work by getting just the sign bit into the low-order bit, then
6377 into the high-order bit, then sign-extend. We then XOR that value
6378 with C. */
6382 /* We must use a signed type in order to get an arithmetic right shift.
6383 However, we must also avoid introducing accidental overflows, so that
6384 a subsequent call to integer_zerop will work. Hence we must
6385 do the type conversion here. At this point, the constant is either
6386 zero or one, and the conversion to a signed type can never overflow.
6387 We could get an overflow if this conversion is done anywhere else. */
6396 /* If necessary, convert the type back to match the type of C. */
6401 }
6402 \f
6403 /* For an expression that has the form
6404 (A && B) || ~B
6405 or
6406 (A || B) && ~B,
6407 we can drop one of the inner expressions and simplify to
6408 A || ~B
6409 or
6410 A && ~B
6411 LOC is the location of the resulting expression. OP is the inner
6412 logical operation; the left-hand side in the examples above, while CMPOP
6413 is the right-hand side. RHS_ONLY is used to prevent us from accidentally
6414 removing a condition that guards another, as in
6415 (A != NULL && A->...) || A == NULL
6416 which we must not transform. If RHS_ONLY is true, only eliminate the
6417 right-most operand of the inner logical operation. */
6419 static tree
6422 {
6440 {
6443 {
6446 }
6447 }
6449 {
6452 {
6455 }
6456 }
6471 }
6473 /* Find ways of folding logical expressions of LHS and RHS:
6474 Try to merge two comparisons to the same innermost item.
6475 Look for range tests like "ch >= '0' && ch <= '9'".
6476 Look for combinations of simple terms on machines with expensive branches
6477 and evaluate the RHS unconditionally.
6479 For example, if we have p->a == 2 && p->b == 4 and we can make an
6480 object large enough to span both A and B, we can do this with a comparison
6481 against the object ANDed with the a mask.
6483 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
6484 operations to do this with one comparison.
6486 We check for both normal comparisons and the BIT_AND_EXPRs made this by
6487 function and the one above.
6489 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
6490 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
6492 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
6493 two operands.
6495 We return the simplified tree or 0 if no optimization is possible. */
6497 static tree
6500 {
6501 /* If this is the "or" of two comparisons, we can do something if
6502 the comparisons are NE_EXPR. If this is the "and", we can do something
6503 if the comparisons are EQ_EXPR. I.e.,
6504 (a->b == 2 && a->c == 4) can become (a->new == NEW).
6506 WANTED_CODE is this operation code. For single bit fields, we can
6507 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
6508 comparison for one-bit fields. */
6529 /* Start by getting the comparison codes. Fail if anything is volatile.
6530 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
6531 it were surrounded with a NE_EXPR. */
6540 {
6544 }
6547 {
6551 }
6562 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
6565 {
6568 {
6573 }
6576 {
6582 }
6583 }
6588 /* If the RHS can be evaluated unconditionally and its operands are
6589 simple, it wins to evaluate the RHS unconditionally on machines
6590 with expensive branches. In this case, this isn't a comparison
6591 that can be merged. */
6598 {
6599 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
6610 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
6620 }
6622 /* See if the comparisons can be merged. Then get all the parameters for
6623 each side. */
6648 /* It must be true that the inner operation on the lhs of each
6649 comparison must be the same if we are to be able to do anything.
6650 Then see if we have constants. If not, the same must be true for
6651 the rhs's. */
6660 {
6663 }
6668 else
6671 /* If either comparison code is not correct for our logical operation,
6672 fail. However, we can convert a one-bit comparison against zero into
6673 the opposite comparison against that bit being set in the field. */
6677 {
6679 {
6680 /* Make the left operand unsigned, since we are only interested
6681 in the value of one bit. Otherwise we are doing the wrong
6682 thing below. */
6685 }
6686 else
6688 }
6690 /* This is analogous to the code for l_const above. */
6692 {
6694 {
6697 }
6698 else
6700 }
6702 /* See if we can find a mode that contains both fields being compared on
6703 the left. If we can't, fail. Otherwise, update all constants and masks
6704 to be relative to a field of that size. */
6718 {
6721 }
6731 {
6740 {
6744 }
6745 }
6747 {
6756 {
6760 }
6761 }
6763 /* If the right sides are not constant, do the same for it. Also,
6764 disallow this optimization if a size, signedness or storage order
6765 mismatch occurs between the left and right sides. */
6767 {
6771 /* Make sure the two fields on the right
6772 correspond to the left without being swapped. */
6789 {
6792 }
6803 /* Make a mask that corresponds to both fields being compared.
6804 Do this for both items being compared. If the operands are the
6805 same size and the bits being compared are in the same position
6806 then we can do this by masking both and comparing the masked
6807 results. */
6814 {
6828 }
6830 /* There is still another way we can do something: If both pairs of
6831 fields being compared are adjacent, we may be able to make a wider
6832 field containing them both.
6834 Note that we still must mask the lhs/rhs expressions. Furthermore,
6835 the mask must be shifted to account for the shift done by
6836 make_bit_field_ref. */
6845 {
6846 tree type;
6864 /* Convert to the smaller type before masking out unwanted bits. */
6867 {
6869 {
6873 }
6875 {
6879 }
6880 }
6889 }
6892 }
6894 /* Handle the case of comparisons with constants. If there is something in
6895 common between the masks, those bits of the constants must be the same.
6896 If not, the condition is always false. Test for this to avoid generating
6897 incorrect code below. */
6902 {
6904 {
6907 }
6908 else
6909 {
6912 }
6913 }
6918 /* Construct the expression we will return. First get the component
6919 reference we will make. Unless the mask is all ones the width of
6920 that field, perform the mask operation. Then compare with the
6921 merged constant. */
6932 }
6933 \f
6934 /* T is an integer expression that is being multiplied, divided, or taken a
6935 modulus (CODE says which and what kind of divide or modulus) by a
6936 constant C. See if we can eliminate that operation by folding it with
6937 other operations already in T. WIDE_TYPE, if non-null, is a type that
6938 should be used for the computation if wider than our type.
6940 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
6941 (X * 2) + (Y * 4). We must, however, be assured that either the original
6942 expression would not overflow or that overflow is undefined for the type
6943 in the language in question.
6945 If we return a non-null expression, it is an equivalent form of the
6946 original computation, but need not be in the original type.
6948 We set *STRICT_OVERFLOW_P to true if the return values depends on
6949 signed overflow being undefined. Otherwise we do not change
6950 *STRICT_OVERFLOW_P. */
6952 static tree
6955 {
6956 /* To avoid exponential search depth, refuse to allow recursion past
6957 three levels. Beyond that (1) it's highly unlikely that we'll find
6958 something interesting and (2) we've probably processed it before
6959 when we built the inner expression. */
6962 tree ret;
6967 depth++;
6969 depth--;
6972 }
6974 static tree
6977 {
6982 {
6985 {
6988 }
6992 }
6998 /* Don't deal with constants of zero here; they confuse the code below. */
7008 /* Note that we need not handle conditional operations here since fold
7009 already handles those cases. So just do arithmetic here. */
7011 {
7013 /* For a constant, we can always simplify if we are a multiply
7014 or (for divide and modulus) if it is a multiple of our constant. */
7018 {
7021 /* If the multiplication overflowed, we lost information on it.
7022 See PR68142 and PR69845. */
7026 }
7032 /* If op0 is an expression ... */
7038 /* ... and has wrapping overflow, and its type is smaller
7039 than ctype, then we cannot pass through as widening. */
7043 /* ... or this is a truncation (t is narrower than op0),
7044 then we cannot pass through this narrowing. */
7047 /* ... or signedness changes for division or modulus,
7048 then we cannot pass through this conversion. */
7052 /* ... or has undefined overflow while the converted to
7053 type has not, we cannot do the operation in the inner type
7054 as that would introduce undefined overflow. */
7059 /* Pass the constant down and see if we can make a simplification. If
7060 we can, replace this expression with the inner simplification for
7061 possible later conversion to our or some other type. */
7072 /* If widening the type changes it from signed to unsigned, then we
7073 must avoid building ABS_EXPR itself as unsigned. */
7075 {
7079 {
7082 }
7084 }
7085 /* If the constant is negative, we cannot simplify this. */
7088 /* FALLTHROUGH */
7090 /* For division and modulus, type can't be unsigned, as e.g.
7091 (-(x / 2U)) / 2U isn't equal to -((x / 2U) / 2U) for x >= 2.
7092 For signed types, even with wrapping overflow, this is fine. */
7101 /* If widening the type changes the signedness, then we can't perform
7102 this optimization as that changes the result. */
7106 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
7112 {
7119 }
7123 /* If the second operand is constant, this is a multiplication
7124 or floor division, by a power of two, so we can treat it that
7125 way unless the multiplier or divisor overflows. Signed
7126 left-shift overflow is implementation-defined rather than
7127 undefined in C90, so do not convert signed left shift into
7128 multiplication. */
7131 /* const_binop may not detect overflow correctly,
7132 so check for it explicitly here. */
7141 ctype,
7143 t1),
7148 /* See if we can eliminate the operation on both sides. If we can, we
7149 can return a new PLUS or MINUS. If we can't, the only remaining
7150 cases where we can do anything are if the second operand is a
7151 constant. */
7158 /* If not multiplication, we can only do this if both operands
7159 are divisible by c. */
7162 {
7167 }
7169 /* If this was a subtraction, negate OP1 and set it to be an addition.
7170 This simplifies the logic below. */
7172 {
7174 /* If OP1 was not easily negatable, the constant may be OP0. */
7176 {
7179 }
7180 }
7185 /* If either OP1 or C are negative, this optimization is not safe for
7186 some of the division and remainder types while for others we need
7187 to change the code. */
7189 {
7197 }
7199 /* If it's a multiply or a division/modulus operation of a multiple
7200 of our constant, do the operation and verify it doesn't overflow. */
7204 {
7207 /* We allow the constant to overflow with wrapping semantics. */
7211 }
7212 else
7215 /* If we have an unsigned type, we cannot widen the operation since it
7216 will change the result if the original computation overflowed. */
7220 /* The last case is if we are a multiply. In that case, we can
7221 apply the distributive law to commute the multiply and addition
7222 if the multiplication of the constants doesn't overflow
7223 and overflow is defined. With undefined overflow
7224 op0 * c might overflow, while (op0 + orig_op1) * c doesn't.
7225 But fold_plusminus_mult_expr would factor back any power-of-two
7226 value so do not distribute in the first place in this case. */
7234 op1);
7239 /* We have a special case here if we are doing something like
7240 (C * 8) % 4 since we know that's zero. */
7243 /* If the multiplication can overflow we cannot optimize this. */
7248 {
7251 }
7253 /* ... fall through ... */
7257 /* If we can extract our operation from the LHS, do so and return a
7258 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
7259 do something only if the second operand is a constant. */
7275 /* If these are the same operation types, we can associate them
7276 assuming no overflow. */
7278 {
7293 }
7295 /* If these operations "cancel" each other, we have the main
7296 optimizations of this pass, which occur when either constant is a
7297 multiple of the other, in which case we replace this with either an
7298 operation or CODE or TCODE.
7300 If we have an unsigned type, we cannot do this since it will change
7301 the result if the original computation overflowed. */
7309 {
7312 {
7318 }
7321 {
7327 }
7328 }
7333 }
7336 }
7337 \f
7338 /* Return a node which has the indicated constant VALUE (either 0 or
7339 1 for scalars or {-1,-1,..} or {0,0,...} for vectors),
7340 and is of the indicated TYPE. */
7342 tree
7344 {
7353 else
7355 }
7358 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
7359 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
7360 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
7361 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
7362 COND is the first argument to CODE; otherwise (as in the example
7363 given here), it is the second argument. TYPE is the type of the
7364 original expression. Return NULL_TREE if no simplification is
7365 possible. */
7367 static tree
7372 {
7380 /* Do not move possibly trapping operations into the conditional as this
7381 pessimizes code and causes gimplification issues when applied late. */
7389 {
7393 /* If this operand throws an expression, then it does not make
7394 sense to try to perform a logical or arithmetic operation
7395 involving it. */
7400 }
7403 {
7408 }
7409 else
7410 /* Detect the case of mixing vector and scalar types - bail out. */
7416 /* This transformation is only worthwhile if we don't have to wrap ARG
7417 in a SAVE_EXPR and the operation can be simplified without recursing
7418 on at least one of the branches once its pushed inside the COND_EXPR. */
7427 {
7431 else
7433 }
7435 {
7439 else
7441 }
7443 /* Check that we have simplified at least one of the branches. */
7448 }
7450 \f
7451 /* Subroutine of fold() that checks for the addition of ARG +/- 0.0.
7453 If !NEGATE, return true if ZERO_ARG is +/-0.0 and, for all ARG of
7454 type TYPE, ARG + ZERO_ARG is the same as ARG. If NEGATE, return true
7455 if ARG - ZERO_ARG is the same as X.
7457 If ARG is NULL, check for any value of type TYPE.
7459 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
7460 and finite. The problematic cases are when X is zero, and its mode
7461 has signed zeros. In the case of rounding towards -infinity,
7462 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
7463 modes, X + 0 is not the same as X because -0 + 0 is 0. */
7465 bool
7468 {
7472 /* Don't allow the fold with -fsignaling-nans. */
7476 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
7480 /* There is no case that is safe for all rounding modes. */
7484 /* In a vector or complex, we would need to check the sign of all zeros. */
7490 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
7494 /* The mode has signed zeros, and we have to honor their sign.
7495 In this situation, there are only two cases we can return true for.
7496 (i) X - 0 is the same as X with default rounding.
7497 (ii) X + 0 is X when X can't possibly be -0.0. */
7499 }
7501 /* Subroutine of match.pd that optimizes comparisons of a division by
7502 a nonzero integer constant against an integer constant, i.e.
7503 X/C1 op C2.
7505 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
7506 GE_EXPR or LE_EXPR. ARG01 and ARG1 must be a INTEGER_CST. */
7508 enum tree_code
7511 {
7516 /* We have to do this the hard way to detect unsigned overflow.
7517 prod = int_const_binop (MULT_EXPR, c1, c2); */
7523 {
7527 /* Likewise *hi = int_const_binop (PLUS_EXPR, prod, tmp). */
7530 }
7532 {
7535 {
7554 }
7555 }
7556 else
7557 {
7558 /* A negative divisor reverses the relational operators. */
7563 {
7582 }
7583 }
7596 }
7598 /* Test whether it is preferable to swap two operands, ARG0 and
7599 ARG1, for example because ARG0 is an integer constant and ARG1
7600 isn't. */
7602 bool
7604 {
7618 /* It is preferable to swap two SSA_NAME to ensure a canonical form
7619 for commutative and comparison operators. Ensuring a canonical
7620 form allows the optimizers to find additional redundancies without
7621 having to explicitly check for both orderings. */
7627 /* Put SSA_NAMEs last. */
7633 /* Put variables last. */
7640 }
7643 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
7644 means A >= Y && A != MAX, but in this case we know that
7645 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
7647 static tree
7649 {
7656 else
7665 {
7668 }
7670 {
7673 }
7674 else
7681 {
7682 /* Convert the pointer types into integer before taking the difference. */
7686 }
7687 else
7694 }
7696 /* Fold a sum or difference of at least one multiplication.
7697 Returns the folded tree or NULL if no simplification could be made. */
7699 static tree
7702 {
7706 /* (A * C) +- (B * C) -> (A+-B) * C.
7707 (A * C) +- A -> A * (C+-1).
7708 We are most concerned about the case where C is a constant,
7709 but other combinations show up during loop reduction. Since
7710 it is not difficult, try all four possibilities. */
7713 {
7716 }
7718 {
7721 }
7722 else
7723 {
7724 /* We cannot generate constant 1 for fract. */
7729 }
7731 {
7734 }
7736 {
7738 /* As we canonicalize A - 2 to A + -2 get rid of that sign for
7739 the purpose of this canonicalization. */
7743 {
7746 }
7747 else
7749 }
7750 else
7751 {
7752 /* We cannot generate constant 1 for fract. */
7757 }
7760 /* Prefer factoring a common non-constant. */
7770 /* No identical multiplicands; see if we can find a common
7771 power-of-two factor in non-power-of-two multiplies. This
7772 can help in multi-dimensional array access. */
7774 {
7777 HOST_WIDE_INT tmp;
7779 tree maybe_same;
7781 /* Move min of absolute values to int11. */
7783 {
7788 }
7789 else
7796 /* The remainder should not be a constant, otherwise we
7797 end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
7798 increased the number of multiplications necessary. */
7800 {
7808 }
7809 }
7816 /* We are neither factoring zero nor minus one. */
7824 /* Same may be zero and thus the operation 'code' may overflow. Likewise
7825 same may be minus one and thus the multiplication may overflow. Perform
7826 the sum operation in an unsigned type. */
7831 /* If the sum evaluated to a constant that is not -INF the multiplication
7832 cannot overflow. */
7839 /* Do not resort to unsigned multiplication because
7840 we lose the no-overflow property of the expression. */
7842 }
7844 /* Subroutine of native_encode_expr. Encode the INTEGER_CST
7845 specified by EXPR into the buffer PTR of length LEN bytes.
7846 Return the number of bytes placed in the buffer, or zero
7847 upon failure. */
7849 static int
7851 {
7855 {
7861 {
7863 /* More work is needed when adding _BitInt support to PDP endian
7864 if limb is smaller than word, or if _BitInt limb ordering doesn't
7865 match target endianity here. */
7870 }
7871 else
7873 }
7874 else
7885 /* Dry run. */
7891 {
7893 /* Extend EXPR according to TYPE_SIGN if the precision isn't a whole
7894 number of bytes. */
7898 {
7905 else
7907 }
7908 else
7912 }
7914 }
7917 /* Subroutine of native_encode_expr. Encode the FIXED_CST
7918 specified by EXPR into the buffer PTR of length LEN bytes.
7919 Return the number of bytes placed in the buffer, or zero
7920 upon failure. */
7922 static int
7924 {
7928 FIXED_VALUE_TYPE value;
7943 }
7946 /* Subroutine of native_encode_expr. Encode the REAL_CST
7947 specified by EXPR into the buffer PTR of length LEN bytes.
7948 Return the number of bytes placed in the buffer, or zero
7949 upon failure. */
7951 static int
7953 {
7959 /* There are always 32 bits in each long, no matter the size of
7960 the hosts long. We handle floating point representations with
7961 up to 192 bits. */
7970 /* Dry run. */
7979 {
7984 {
7991 else
7993 }
7994 else
7995 {
7998 {
7999 /* Reverse bytes within each long, or within the entire float
8000 if it's smaller than a long (for HFmode). */
8003 }
8004 }
8009 }
8011 }
8013 /* Subroutine of native_encode_expr. Encode the COMPLEX_CST
8014 specified by EXPR into the buffer PTR of length LEN bytes.
8015 Return the number of bytes placed in the buffer, or zero
8016 upon failure. */
8018 static int
8020 {
8022 tree part;
8036 }
8038 /* Like native_encode_vector, but only encode the first COUNT elements.
8039 The other arguments are as for native_encode_vector. */
8041 static int
8044 {
8048 {
8049 /* This is the only case in which elements can be smaller than a byte.
8050 Element 0 is always in the lsb of the containing byte. */
8059 /* Zero the buffer and then set bits later where necessary. */
8068 {
8074 {
8077 }
8078 }
8080 }
8085 {
8087 {
8090 }
8101 }
8103 }
8105 /* Subroutine of native_encode_expr. Encode the VECTOR_CST
8106 specified by EXPR into the buffer PTR of length LEN bytes.
8107 Return the number of bytes placed in the buffer, or zero
8108 upon failure. */
8110 static int
8112 {
8117 }
8120 /* Subroutine of native_encode_expr. Encode the STRING_CST
8121 specified by EXPR into the buffer PTR of length LEN bytes.
8122 Return the number of bytes placed in the buffer, or zero
8123 upon failure. */
8125 static int
8127 {
8130 /* Wide-char strings are encoded in target byte-order so native
8131 encoding them is trivial. */
8146 else
8147 {
8150 {
8153 }
8155 }
8157 }
8160 /* Subroutine of fold_view_convert_expr. Encode the INTEGER_CST, REAL_CST,
8161 FIXED_CST, COMPLEX_CST, STRING_CST, or VECTOR_CST specified by EXPR into
8162 the buffer PTR of size LEN bytes. If PTR is NULL, don't actually store
8163 anything, just do a dry run. Fail either if OFF is -1 and LEN isn't
8164 sufficient to encode the entire EXPR, or if OFF is out of bounds.
8165 Otherwise, start at byte offset OFF and encode at most LEN bytes.
8166 Return the number of bytes placed in the buffer, or zero upon failure. */
8168 int
8170 {
8171 /* We don't support starting at negative offset and -1 is special. */
8176 {
8197 }
8198 }
8200 /* Try to find a type whose byte size is smaller or equal to LEN bytes larger
8201 or equal to FIELDSIZE bytes, with underlying mode precision/size multiple
8202 of BITS_PER_UNIT. As native_{interpret,encode}_int works in term of
8203 machine modes, we can't just use build_nonstandard_integer_type. */
8205 tree
8207 {
8208 machine_mode mode;
8210 {
8217 {
8221 }
8222 }
8228 {
8236 }
8239 }
8241 /* Similar to native_encode_expr, but also handle CONSTRUCTORs, VCEs,
8242 NON_LVALUE_EXPRs and nops. If MASK is non-NULL (then PTR has
8243 to be non-NULL and OFF zero), then in addition to filling the
8244 bytes pointed by PTR with the value also clear any bits pointed
8245 by MASK that are known to be initialized, keep them as is for
8246 e.g. uninitialized padding bits or uninitialized fields. */
8248 int
8251 {
8254 /* We don't support starting at negative offset and -1 is special. */
8262 {
8266 mask);
8281 {
8282 tree min_index;
8300 {
8305 {
8308 }
8313 else
8317 {
8321 offset_int wpos
8329 offset_int wcount
8336 }
8338 {
8341 offset_int wpos
8349 }
8352 {
8354 {
8365 }
8367 {
8371 }
8372 }
8376 do
8377 {
8382 {
8384 {
8387 fieldsize);
8390 }
8392 ptr
8395 fieldsize,
8398 mask
8402 else
8403 {
8406 }
8407 }
8410 {
8411 /* Partial overlap. */
8417 {
8421 fieldsize);
8422 }
8423 else
8424 {
8428 }
8431 }
8433 }
8435 }
8437 }
8440 {
8450 {
8456 {
8459 }
8462 {
8471 }
8475 else
8479 {
8480 tree fld;
8482 {
8495 }
8497 {
8501 }
8504 {
8505 cnt--;
8511 }
8512 }
8517 {
8530 {
8535 }
8536 }
8537 else
8538 {
8543 }
8547 /* Prepare to deal with integral bit-fields and filter out other
8548 bit-fields that do not start and end on a byte boundary. */
8550 {
8555 {
8561 }
8567 }
8577 {
8578 /* FIXME: Handle PDP endian. */
8589 {
8592 }
8593 else
8594 {
8602 else
8603 {
8606 }
8607 }
8617 else
8630 /* If the bitfield does not start at byte boundary, handle
8631 the partial byte at the start. */
8634 {
8636 {
8641 {
8644 else
8646 }
8647 }
8648 else
8649 {
8654 {
8657 else
8661 }
8662 }
8663 }
8664 /* If the bitfield does not end at byte boundary, handle
8665 the partial byte at the end. */
8669 {
8671 {
8678 }
8679 else
8680 {
8687 }
8688 }
8692 {
8697 }
8698 else
8699 {
8700 /* Partial overlap. */
8704 {
8707 }
8711 }
8713 }
8718 {
8721 {
8724 {
8728 }
8731 }
8734 fldsize,
8744 }
8745 else
8746 {
8747 /* Partial overlap. */
8753 {
8757 fieldsize);
8758 }
8759 else
8760 {
8764 }
8767 }
8768 }
8770 }
8772 }
8773 }
8776 /* Subroutine of native_interpret_expr. Interpret the contents of
8777 the buffer PTR of length LEN as an INTEGER_CST of type TYPE.
8778 If the buffer cannot be interpreted, return NULL_TREE. */
8780 static tree
8782 {
8785 {
8791 {
8793 /* More work is needed when adding _BitInt support to PDP endian
8794 if limb is smaller than word, or if _BitInt limb ordering doesn't
8795 match target endianity here. */
8800 }
8801 else
8803 }
8804 else
8813 }
8816 /* Subroutine of native_interpret_expr. Interpret the contents of
8817 the buffer PTR of length LEN as a FIXED_CST of type TYPE.
8818 If the buffer cannot be interpreted, return NULL_TREE. */
8820 static tree
8822 {
8825 double_int result;
8826 FIXED_VALUE_TYPE fixed_value;
8836 }
8839 /* Subroutine of native_interpret_expr. Interpret the contents of
8840 the buffer PTR of length LEN as a REAL_CST of type TYPE.
8841 If the buffer cannot be interpreted, return NULL_TREE. */
8843 tree
8845 {
8849 /* There are always 32 bits in each long, no matter the size of
8850 the hosts long. We handle floating point representations with
8851 up to 192 bits. */
8852 REAL_VALUE_TYPE r;
8862 {
8863 /* Both OFFSET and BYTE index within a long;
8864 bitpos indexes the whole float. */
8867 {
8874 else
8876 }
8877 else
8878 {
8881 {
8882 /* Reverse bytes within each long, or within the entire float
8883 if it's smaller than a long (for HFmode). */
8886 }
8887 }
8891 }
8895 }
8898 /* Subroutine of native_interpret_expr. Interpret the contents of
8899 the buffer PTR of length LEN as a COMPLEX_CST of type TYPE.
8900 If the buffer cannot be interpreted, return NULL_TREE. */
8902 static tree
8904 {
8919 }
8921 /* Read a vector of type TYPE from the target memory image given by BYTES,
8922 which contains LEN bytes. The vector is known to be encodable using
8923 NPATTERNS interleaved patterns with NELTS_PER_PATTERN elements each.
8925 Return the vector on success, otherwise return null. */
8927 static tree
8931 {
8935 {
8936 /* This is the only case in which elements can be smaller than a byte.
8937 Element 0 is always in the lsb of the containing byte. */
8944 {
8951 }
8953 }
8961 {
8967 }
8969 }
8971 /* Subroutine of native_interpret_expr. Interpret the contents of
8972 the buffer PTR of length LEN as a VECTOR_CST of type TYPE.
8973 If the buffer cannot be interpreted, return NULL_TREE. */
8975 static tree
8977 {
8986 }
8989 /* Subroutine of fold_view_convert_expr. Interpret the contents of
8990 the buffer PTR of length LEN as a constant of type TYPE. For
8991 INTEGRAL_TYPE_P we return an INTEGER_CST, for SCALAR_FLOAT_TYPE_P
8992 we return a REAL_CST, etc... If the buffer cannot be interpreted,
8993 return NULL_TREE. */
8995 tree
8997 {
8999 {
9011 {
9012 /* For floating point values in composite modes, punt if this
9013 folding doesn't preserve bit representation. As the mode doesn't
9014 have fixed precision while GCC pretends it does, there could be
9015 valid values that GCC can't really represent accurately.
9016 See PR95450. Even for other modes, e.g. x86 XFmode can have some
9017 bit combinationations which GCC doesn't preserve. */
9027 }
9041 }
9042 }
9044 /* Returns true if we can interpret the contents of a native encoding
9045 as TYPE. */
9047 bool
9049 {
9051 {
9065 }
9066 }
9068 /* Attempt to interpret aggregate of TYPE from bytes encoded in target
9069 byte order at PTR + OFF with LEN bytes. Does not handle unions. */
9071 tree
9074 {
9077 {
9084 {
9088 }
9093 {
9098 {
9103 }
9107 eltsz);
9111 }
9113 }
9117 {
9126 {
9129 {
9130 poly_int64 bitoffset;
9135 else
9145 }
9146 else
9147 {
9165 HOST_WIDE_INT rpos;
9168 else
9169 {
9172 }
9181 }
9182 }
9185 {
9195 }
9199 {
9204 }
9211 {
9215 /* FIXME: Figure out how to handle PDP endian bitfields. */
9221 else
9225 }
9227 }
9229 }
9231 /* Routines for manipulation of native_encode_expr encoded data if the encoded
9232 or extracted constant positions and/or sizes aren't byte aligned. */
9234 /* Shift left the bytes in PTR of SZ elements by AMNT bits, carrying over the
9235 bits between adjacent elements. AMNT should be within
9236 [0, BITS_PER_UNIT).
9237 Example, AMNT = 2:
9238 00011111|11100000 << 2 = 01111111|10000000
9239 PTR[1] | PTR[0] PTR[1] | PTR[0]. */
9241 void
9244 {
9253 {
9259 {
9262 }
9263 }
9264 }
9266 /* Like shift_bytes_in_array_left but for big-endian.
9267 Shift right the bytes in PTR of SZ elements by AMNT bits, carrying over the
9268 bits between adjacent elements. AMNT should be within
9269 [0, BITS_PER_UNIT).
9270 Example, AMNT = 2:
9271 00011111|11100000 >> 2 = 00000111|11111000
9272 PTR[0] | PTR[1] PTR[0] | PTR[1]. */
9274 void
9277 {
9285 {
9292 }
9293 }
9295 /* Try to view-convert VECTOR_CST EXPR to VECTOR_TYPE TYPE by operating
9296 directly on the VECTOR_CST encoding, in a way that works for variable-
9297 length vectors. Return the resulting VECTOR_CST on success or null
9298 on failure. */
9300 static tree
9302 {
9314 /* We can only preserve the semantics of a stepped pattern if the new
9315 vector element is an integer of the same size. */
9320 /* The number of bits needed to encode one element from every pattern
9321 of the original vector. */
9322 unsigned int expr_sequence_bits
9325 /* The number of bits needed to encode one element from every pattern
9326 of the result. */
9327 unsigned int type_sequence_bits
9330 /* Don't try to read more bytes than are available, which can happen
9331 for constant-sized vectors if TYPE has larger elements than EXPR_TYPE.
9332 The general VIEW_CONVERT handling can cope with that case, so there's
9333 no point complicating things here. */
9336 BITS_PER_UNIT);
9341 /* Get enough bytes of EXPR to form the new encoding. */
9349 /* Reencode the bytes as TYPE. */
9353 }
9355 /* Fold a VIEW_CONVERT_EXPR of a constant expression EXPR to type
9356 TYPE at compile-time. If we're unable to perform the conversion
9357 return NULL_TREE. */
9359 static tree
9361 {
9365 HOST_WIDE_INT l;
9367 /* Check that the host and target are sane. */
9378 {
9381 }
9382 else
9383 {
9386 }
9392 }
9394 /* Build an expression for the address of T. Folds away INDIRECT_REF
9395 to avoid confusing the gimplify process. */
9397 tree
9399 {
9400 /* The size of the object is not relevant when talking about its address. */
9405 {
9410 }
9413 {
9418 }
9425 {
9430 }
9431 else
9435 }
9437 /* Build an expression for the address of T. */
9439 tree
9441 {
9445 }
9447 /* Fold a unary expression of code CODE and type TYPE with operand
9448 OP0. Return the folded expression if folding is successful.
9449 Otherwise, return NULL_TREE. */
9451 tree
9453 {
9454 tree tem;
9455 tree arg0;
9463 {
9466 {
9467 /* Don't use STRIP_NOPS, because signedness of argument type
9468 matters. */
9470 }
9471 else
9472 {
9473 /* Strip any conversions that don't change the mode. This
9474 is safe for every expression, except for a comparison
9475 expression because its signedness is derived from its
9476 operands.
9478 Note that this is done as an internal manipulation within
9479 the constant folder, in order to find the simplest
9480 representation of the arguments so that their form can be
9481 studied. In any cases, the appropriate type conversions
9482 should be put back in the tree that will get out of the
9483 constant folder. */
9485 }
9488 {
9491 {
9495 }
9496 }
9497 }
9504 {
9511 {
9525 /* If this was a conversion, and all we did was to move into
9526 inside the COND_EXPR, bring it back out. But leave it if
9527 it is a conversion from integer to integer and the
9528 result precision is no wider than a word since such a
9529 conversion is cheap and may be optimized away by combine,
9530 while it couldn't if it were outside the COND_EXPR. Then return
9531 so we don't get into an infinite recursion loop taking the
9532 conversion out and then back in. */
9544 && (INTEGRAL_TYPE_P
9557 }
9558 }
9561 {
9567 CASE_CONVERT:
9571 {
9572 /* If we have (type) (a CMP b) and type is an integral type, return
9573 new expression involving the new type. Canonicalize
9574 (type) (a CMP b) to (a CMP b) ? (type) true : (type) false for
9575 non-integral type.
9576 Do not fold the result as that would not simplify further, also
9577 folding again results in recursions. */
9587 }
9589 /* Handle (T *)&A.B.C for A being of type T and B and C
9590 living at offset zero. This occurs frequently in
9591 C++ upcasting and then accessing the base. */
9595 {
9597 tree offset;
9598 machine_mode mode;
9600 tree base
9604 /* If the reference was to a (constant) zero offset, we can use
9605 the address of the base if it has the same base type
9606 as the result type and the pointer type is unqualified. */
9614 }
9618 /* Detect assigning a bitfield. */
9620 && DECL_BIT_FIELD
9622 {
9623 /* Don't leave an assignment inside a conversion
9624 unless assigning a bitfield. */
9626 /* First do the assignment, then return converted constant. */
9631 }
9633 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
9634 constants (if x has signed type, the sign bit cannot be set
9635 in c). This folds extension into the BIT_AND_EXPR.
9636 ??? We don't do it for BOOLEAN_TYPE or ENUMERAL_TYPE because they
9637 very likely don't have maximal range for their precision and this
9638 transformation effectively doesn't preserve non-maximal ranges. */
9642 {
9653 <= HOST_BITS_PER_WIDE_INT
9655 {
9659 cst &= HOST_WIDE_INT_M1U
9663 && !flag_syntax_only
9666 {
9670 }
9671 }
9673 {
9683 }
9684 }
9686 /* Convert (T1)(X p+ Y) into ((T1)X p+ Y), for pointer type, when the new
9687 cast (T1)X will fold away. We assume that this happens when X itself
9688 is a cast. */
9692 {
9696 /* If -fsanitize=alignment, avoid this optimization in GENERIC
9697 when the pointed type needs higher alignment than
9698 the p+ first operand's pointed type. */
9705 /* Similarly, avoid this optimization in GENERIC for -fsanitize=null
9706 when type is a reference type and arg00's type is not,
9707 because arg00 could be validly nullptr and if arg01 doesn't return,
9708 we don't want false positive binding of reference to nullptr. */
9710 && !in_gimple_form
9717 }
9719 /* Convert (T1)(~(T2)X) into ~(T1)X if T1 and T2 are integral types
9720 of the same precision, and X is an integer type not narrower than
9721 types T1 or T2, i.e. the cast (T2)X isn't an extension. */
9727 {
9733 }
9735 /* Convert (T1)(X * Y) into (T1)X * (T1)Y if T1 is narrower than the
9736 type of X and Y (integer types only). */
9743 {
9744 /* Be careful not to introduce new overflows. */
9745 tree mult_type;
9748 else
9752 {
9759 }
9760 }
9766 {
9773 }
9784 /* Convert fabs((double)float) into (double)fabsf(float). */
9787 {
9793 targ0));
9794 }
9798 /* Convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
9817 /* Note that the operand of this must be an int
9818 and its values must be 0 or 1.
9819 ("true" is a fixed value perhaps depending on the language,
9820 but we don't handle values other than 1 correctly yet.) */
9827 /* Fold *&X to X if X is an lvalue. */
9829 {
9836 }
9842 }
9845 /* If the operation was a conversion do _not_ mark a resulting constant
9846 with TREE_OVERFLOW if the original constant was not. These conversions
9847 have implementation defined behavior and retaining the TREE_OVERFLOW
9848 flag here would confuse later passes such as VRP. */
9849 tree
9852 {
9861 }
9863 /* Fold a binary bitwise/truth expression of code CODE and type TYPE with
9864 operands OP0 and OP1. LOC is the location of the resulting expression.
9865 ARG0 and ARG1 are the NOP_STRIPed results of OP0 and OP1.
9866 Return the folded expression if folding is successful. Otherwise,
9867 return NULL_TREE. */
9868 static tree
9871 {
9872 tree tem;
9874 /* We only do these simplifications if we are optimizing. */
9878 /* Check for things like (A || B) && (A || C). We can convert this
9879 to A || (B && C). Note that either operator can be any of the four
9880 truth and/or operations and the transformation will still be
9881 valid. Also note that we only care about order for the
9882 ANDIF and ORIF operators. If B contains side effects, this
9883 might change the truth-value of A. */
9890 {
9910 /* This case if tricky because we must either have commutative
9911 operators or else A10 must not have side-effects. */
9917 a01);
9918 }
9920 /* See if we can build a range comparison. */
9926 {
9930 }
9934 {
9938 }
9940 /* Check for the possibility of merging component references. If our
9941 lhs is another similar operation, try to merge its rhs with our
9942 rhs. Then try to merge our lhs and rhs. */
9953 logical_op_non_short_circuit
9961 {
9968 /* Transform ((A AND-IF B) AND[-IF] C) into (A AND-IF (B AND C)),
9969 or ((A OR-IF B) OR[-IF] C) into (A OR-IF (B OR C))
9970 We don't want to pack more than two leafs to a non-IF AND/OR
9971 expression.
9972 If tree-code of left-hand operand isn't an AND/OR-IF code and not
9973 equal to IF-CODE, then we don't want to add right-hand operand.
9974 If the inner right-hand side of left-hand operand has
9975 side-effects, or isn't simple, then we can't add to it,
9976 as otherwise we might destroy if-sequence. */
9979 /* Needed for sequence points to handle trappings, and
9980 side-effects. */
9982 {
9984 arg1);
9986 tem);
9987 }
9988 /* Same as above but for (A AND[-IF] (B AND-IF C)) -> ((A AND B) AND-IF C),
9989 or (A OR[-IF] (B OR-IF C) -> ((A OR B) OR-IF C). */
9992 /* Needed for sequence points to handle trappings, and
9993 side-effects. */
9995 {
10000 }
10001 /* Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
10002 into (A OR B).
10003 For sequence point consistancy, we need to check for trapping,
10004 and side-effects. */
10008 }
10011 }
10013 /* Helper that tries to canonicalize the comparison ARG0 CODE ARG1
10014 by changing CODE to reduce the magnitude of constants involved in
10015 ARG0 of the comparison.
10016 Returns a canonicalized comparison tree if a simplification was
10017 possible, otherwise returns NULL_TREE.
10018 Set *STRICT_OVERFLOW_P to true if the canonicalization is only
10019 valid if signed overflow is undefined. */
10021 static tree
10025 {
10030 /* Match A +- CST code arg1. We can change this only if overflow
10031 is undefined. */
10034 /* In principle pointers also have undefined overflow behavior,
10035 but that causes problems elsewhere. */
10042 /* Identify the constant in arg0 and its sign. */
10046 /* Overflowed constants and zero will cause problems. */
10051 /* See if we can reduce the magnitude of the constant in
10052 arg0 by changing the comparison code. */
10053 /* A - CST < arg1 -> A - CST-1 <= arg1. */
10057 /* A + CST > arg1 -> A + CST-1 >= arg1. */
10061 /* A + CST <= arg1 -> A + CST-1 < arg1. */
10065 /* A - CST >= arg1 -> A - CST-1 > arg1. */
10069 else
10073 /* Now build the constant reduced in magnitude. But not if that
10074 would produce one outside of its types range. */
10090 }
10092 /* Canonicalize the comparison ARG0 CODE ARG1 with type TYPE with undefined
10093 overflow further. Try to decrease the magnitude of constants involved
10094 by changing LE_EXPR and GE_EXPR to LT_EXPR and GT_EXPR or vice versa
10095 and put sole constants at the second argument position.
10096 Returns the canonicalized tree if changed, otherwise NULL_TREE. */
10098 static tree
10101 {
10102 tree t;
10107 /* Try canonicalization by simplifying arg0. */
10112 {
10116 }
10118 /* Try canonicalization by simplifying arg1 using the swapped
10119 comparison. */
10127 }
10129 /* Return whether BASE + OFFSET + BITPOS may wrap around the address
10130 space. This is used to avoid issuing overflow warnings for
10131 expressions like &p->x which cannot wrap. */
10133 static bool
10135 {
10142 poly_wide_int wi_offset;
10148 else
10153 precision);
10168 /* We can do slightly better for SIZE if we have an ADDR_EXPR of an
10169 array. */
10178 }
10180 /* Return a positive integer when the symbol DECL is known to have
10181 a nonzero address, zero when it's known not to (e.g., it's a weak
10182 symbol), and a negative integer when the symbol is not yet in the
10183 symbol table and so whether or not its address is zero is unknown.
10184 For function local objects always return positive integer. */
10185 static int
10187 {
10188 /* Normally, don't do anything for variables and functions before symtab is
10189 built; it is quite possible that DECL will be declared weak later.
10190 But if folding_initializer, we need a constant answer now, so create
10191 the symtab entry and prevent later weak declaration. */
10194 = (folding_initializer
10199 /* Function local objects are never NULL. */
10207 }
10209 /* Subroutine of fold_binary. This routine performs all of the
10210 transformations that are common to the equality/inequality
10211 operators (EQ_EXPR and NE_EXPR) and the ordering operators
10212 (LT_EXPR, LE_EXPR, GE_EXPR and GT_EXPR). Callers other than
10213 fold_binary should call fold_binary. Fold a comparison with
10214 tree code CODE and type TYPE with operands OP0 and OP1. Return
10215 the folded comparison or NULL_TREE. */
10217 static tree
10220 {
10230 /* For comparisons of pointers we can decompose it to a compile time
10231 comparison of the base objects and the offsets into the object.
10232 This requires at least one operand being an ADDR_EXPR or a
10233 POINTER_PLUS_EXPR to do more than the operand_equal_p test below. */
10239 {
10242 machine_mode mode;
10246 /* Get base and offset for the access. Strip ADDR_EXPR for
10247 get_inner_reference, but put it back by stripping INDIRECT_REF
10248 off the base object if possible. indirect_baseN will be true
10249 if baseN is not an address but refers to the object itself. */
10252 {
10253 base0
10259 else
10261 }
10263 {
10267 {
10268 base0
10274 else
10276 }
10279 else
10283 {
10290 }
10291 }
10295 {
10296 base1
10302 else
10304 }
10306 {
10310 {
10311 base1
10317 else
10319 }
10322 else
10326 {
10333 }
10334 }
10336 /* If we have equivalent bases we might be able to simplify. */
10340 {
10341 /* We can fold this expression to a constant if the non-constant
10342 offset parts are equal. */
10346 && (equality_code
10347 || (indirect_base0
10350 {
10356 "occur when comparing P +- C1 with "
10358 WARN_STRICT_OVERFLOW_CONDITIONAL);
10361 {
10399 }
10400 }
10401 /* We can simplify the comparison to a comparison of the variable
10402 offset parts if the constant offset parts are equal.
10403 Be careful to use signed sizetype here because otherwise we
10404 mess with array offsets in the wrong way. This is possible
10405 because pointer arithmetic is restricted to retain within an
10406 object and overflow on pointer differences is undefined as of
10407 6.5.6/8 and /9 with respect to the signed ptrdiff_t. */
10409 && (equality_code
10410 || (indirect_base0
10413 {
10414 /* By converting to signed sizetype we cover middle-end pointer
10415 arithmetic which operates on unsigned pointer types of size
10416 type size and ARRAY_REF offsets which are properly sign or
10417 zero extended from their type in case it is narrower than
10418 sizetype. */
10421 else
10425 else
10432 "occur when comparing P +- C1 with "
10434 WARN_STRICT_OVERFLOW_COMPARISON);
10437 }
10438 }
10439 /* For equal offsets we can simplify to a comparison of the
10440 base addresses. */
10442 && (indirect_base0
10444 && (indirect_base1
10449 {
10455 }
10456 /* Comparison between an ordinary (non-weak) symbol and a null
10457 pointer can be eliminated since such symbols must have a non
10458 null address. In C, relational expressions between pointers
10459 to objects and null pointers are undefined. The results
10460 below follow the C++ rules with the additional property that
10461 every object pointer compares greater than a null pointer.
10462 */
10465 /* Avoid folding references to struct members at offset 0 to
10466 prevent tests like '&ptr->firstmember == 0' from getting
10467 eliminated. When ptr is null, although the -> expression
10468 is strictly speaking invalid, GCC retains it as a matter
10469 of QoI. See PR c/44555. */
10472 && indirect_base0
10473 /* The caller guarantees that when one of the arguments is
10474 constant (i.e., null in this case) it is second. */
10476 {
10478 {
10489 }
10490 }
10491 }
10493 /* Transform comparisons of the form X +- C1 CMP Y +- C2 to
10494 X CMP Y +- C2 +- C1 for signed X, Y. This is valid if
10495 the resulting offset is smaller in absolute value than the
10496 original one and has the same sign. */
10505 {
10510 tree cst;
10512 "occur when combining constants around "
10515 /* Put the constant on the side where it doesn't overflow and is
10516 of lower absolute value and of same sign than before. */
10523 {
10526 variable1,
10530 }
10538 {
10544 variable2);
10545 }
10546 }
10552 /* If we are comparing an expression that just has comparisons
10553 of two integer values, arithmetic expressions of those comparisons,
10554 and constants, we can simplify it. There are only three cases
10555 to check: the two values can either be equal, the first can be
10556 greater, or the second can be greater. Fold the expression for
10557 those three values. Since each value must be 0 or 1, we have
10558 eight possibilities, each of which corresponds to the constant 0
10559 or 1 or one of the six possible comparisons.
10561 This handles common cases like (a > b) == 0 but also handles
10562 expressions like ((x > y) - (y > x)) > 0, which supposedly
10563 occur in macroized code. */
10566 {
10570 /* Don't handle degenerate cases here; they should already
10571 have been handled anyway. */
10580 {
10584 /* We can't just pass T to eval_subst in case cval1 or cval2
10585 was the same as ARG1. */
10587 tree high_result
10591 arg1);
10592 tree equal_result
10596 arg1);
10597 tree low_result
10601 arg1);
10603 /* All three of these results should be 0 or 1. Confirm they are.
10604 Then use those values to select the proper code to use. */
10609 {
10610 /* Make a 3-bit mask with the high-order bit being the
10611 value for `>', the next for '=', and the low for '<'. */
10615 {
10617 /* Always false. */
10638 /* Always true. */
10640 }
10643 }
10644 }
10645 }
10648 }
10651 /* Subroutine of fold_binary. Optimize complex multiplications of the
10652 form z * conj(z), as pow(realpart(z),2) + pow(imagpart(z),2). The
10653 argument EXPR represents the expression "z" of type TYPE. */
10655 static tree
10657 {
10662 {
10665 }
10667 {
10670 }
10671 else
10672 {
10676 }
10685 }
10688 /* Helper function for fold_vec_perm. Store elements of VECTOR_CST or
10689 CONSTRUCTOR ARG into array ELTS, which has NELTS elements, and return
10690 true if successful. */
10692 static bool
10694 {
10699 {
10702 }
10704 {
10710 else
10712 }
10713 else
10719 }
10721 /* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
10722 mask for VLA vec_perm folding.
10723 REASON if specified, will contain the reason why SEL is not suitable.
10724 Used only for debugging and unit-testing. */
10726 static bool
10730 {
10737 {
10741 }
10743 /* We want to avoid cases where sel.length is not a multiple of npatterns.
10744 For eg: sel.length = 2 + 2x, and sel npatterns = 4. */
10745 poly_uint64 esel;
10747 {
10751 }
10757 {
10760 HOST_WIDE_INT step;
10762 {
10766 }
10767 // FIXME: Punt on step < 0 for now, revisit later.
10774 {
10778 }
10780 /* Ensure that stepped sequence of the pattern selects elements
10781 only from the same input vector. */
10790 {
10794 }
10796 /* Ensure that the stepped sequence always selects from the same
10797 input pattern. */
10802 {
10806 }
10808 /* If a1 chooses base element from arg, ensure that it's a natural
10809 stepped sequence, ie, (arg[2] - arg[1]) == (arg[1] - arg[0])
10810 to preserve arg's encoding. */
10813 {
10826 {
10830 }
10831 }
10832 }
10835 }
10837 /* Try to fold permutation of ARG0 and ARG1 with SEL selector when
10838 the input vectors are VECTOR_CST. Return NULL_TREE otherwise.
10839 REASON has same purpose as described in
10840 valid_mask_for_fold_vec_perm_cst_p. */
10842 static tree
10845 {
10849 /* First try to implement the fold in a VLA-friendly way.
10851 (1) If the selector is simply a duplication of N elements, the
10852 result is likewise a duplication of N elements.
10854 (2) If the selector is N elements followed by a duplication
10855 of N elements, the result is too.
10857 (3) If the selector is N elements followed by an interleaving
10858 of N linear series, the situation is more complex.
10860 valid_mask_for_fold_vec_perm_cst_p detects whether we
10861 can handle this case. If we can, then each of the N linear
10862 series either (a) selects the same element each time or
10863 (b) selects a linear series from one of the input patterns.
10865 If (b) holds for one of the linear series, the result
10866 will contain a linear series, and so the result will have
10867 the same shape as the selector. If (a) holds for all of
10868 the linear series, the result will be the same as (2) above.
10870 (b) can only hold if one of the input patterns has a
10871 stepped encoding. */
10874 {
10882 }
10884 {
10887 }
10888 else
10893 {
10896 poly_uint64 r;
10899 /* Punt if sel[i] /trunc_div len cannot be determined,
10900 because the input vector to be chosen will depend on
10901 runtime vector length.
10902 For example if len == 4 + 4x, and sel[i] == 4,
10903 If len at runtime equals 4, we choose arg1[0].
10904 For any other value of len > 4 at runtime, we choose arg0[4].
10905 which makes the element choice dependent on runtime vector length. */
10907 {
10911 }
10913 /* sel[i] % len will give the index of element in the chosen input
10914 vector. For example if sel[i] == 5 + 4x and len == 4 + 4x,
10915 we will choose arg1[1] since (5 + 4x) % (4 + 4x) == 1. */
10917 {
10921 }
10926 }
10929 }
10931 /* Attempt to fold vector permutation of ARG0 and ARG1 vectors using SEL
10932 selector. Return the folded VECTOR_CST or CONSTRUCTOR if successful,
10933 NULL_TREE otherwise. */
10935 tree
10937 {
10953 /* For fall back case, we want to ensure we have VLS vectors
10954 with equal length. */
10968 {
10969 HOST_WIDE_INT index;
10973 }
10975 }
10977 /* Try to fold a pointer difference of type TYPE two address expressions of
10978 array references AREF0 and AREF1 using location LOC. Return a
10979 simplified expression for the difference or NULL_TREE. */
10981 static tree
10985 {
10990 /* If the bases are array references as well, recurse. If the bases
10991 are pointer indirections compute the difference of the pointers.
10992 If the bases are equal, we are set. */
10995 && (base_offset
10997 use_pointer_diff)))
11000 && (base_offset
11001 = use_pointer_diff
11011 {
11017 base_offset,
11020 }
11022 }
11024 /* If the real or vector real constant CST of type TYPE has an exact
11025 inverse, return it, else return NULL. */
11027 tree
11029 {
11030 REAL_VALUE_TYPE r;
11031 tree unit_type;
11032 machine_mode mode;
11035 {
11045 {
11049 tree_vector_builder elts;
11054 {
11059 }
11062 }
11066 }
11067 }
11069 /* Mask out the tz least significant bits of X of type TYPE where
11070 tz is the number of trailing zeroes in Y. */
11071 static wide_int
11073 {
11078 }
11080 /* Return true when T is an address and is known to be nonzero.
11081 For floating point we further ensure that T is not denormal.
11082 Similar logic is present in nonzero_address in rtlanal.h.
11084 If the return value is based on the assumption that signed overflow
11085 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
11086 change *STRICT_OVERFLOW_P. */
11088 static bool
11090 {
11094 /* Doing something useful for floating point would need more work. */
11100 {
11103 strict_overflow_p);
11109 strict_overflow_p);
11117 }
11120 {
11123 strict_overflow_p);
11131 strict_overflow_p);
11145 strict_overflow_p);
11149 strict_overflow_p);
11152 {
11164 }
11168 }
11170 }
11172 /* Return true when T is an address and is known to be nonzero.
11173 Handle warnings about undefined signed overflow. */
11175 bool
11177 {
11184 "determining that expression is always "
11186 WARN_STRICT_OVERFLOW_MISC);
11188 }
11190 /* Return true if T is known not to be equal to an integer W. */
11192 bool
11194 {
11195 int_range_max vr;
11197 {
11208 /* If T has some known zero bits and W has any of those bits set,
11209 then T is known not to be equal to W. */
11217 }
11218 }
11220 /* Fold a binary expression of code CODE and type TYPE with operands
11221 OP0 and OP1. LOC is the location of the resulting expression.
11222 Return the folded expression if folding is successful. Otherwise,
11223 return NULL_TREE. */
11225 tree
11228 {
11243 /* Strip any conversions that don't change the mode. This is
11244 safe for every expression, except for a comparison expression
11245 because its signedness is derived from its operands. So, in
11246 the latter case, only strip conversions that don't change the
11247 signedness. MIN_EXPR/MAX_EXPR also need signedness of arguments
11248 preserved.
11250 Note that this is done as an internal manipulation within the
11251 constant folder, in order to find the simplest representation
11252 of the arguments so that their form can be studied. In any
11253 cases, the appropriate type conversions should be put back in
11254 the tree that will get out of the constant folder. */
11257 {
11260 }
11261 else
11262 {
11265 }
11267 /* Note that TREE_CONSTANT isn't enough: static var addresses are
11268 constant but we can't do arithmetic on them. */
11270 {
11273 {
11277 }
11278 }
11280 /* If this is a commutative operation, and ARG0 is a constant, move it
11281 to ARG1 to reduce the number of tests below. */
11286 /* Likewise if this is a comparison, and ARG0 is a constant, move it
11287 to ARG1 to reduce the number of tests below. */
11296 /* ARG0 is the first operand of EXPR, and ARG1 is the second operand.
11298 First check for cases where an arithmetic operation is applied to a
11299 compound, conditional, or comparison operation. Push the arithmetic
11300 operation inside the compound or conditional to see if any folding
11301 can then be done. Convert comparison to conditional for this purpose.
11302 The also optimizes non-constant cases that used to be done in
11303 expand_expr.
11305 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
11306 one of the operands is a comparison and the other is a comparison, a
11307 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
11308 code below would make the expression more complex. Change it to a
11309 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
11310 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
11323 {
11327 boolean_type_node,
11335 }
11339 {
11341 {
11346 tem);
11347 }
11349 {
11354 tem);
11355 }
11360 {
11366 }
11371 {
11377 }
11378 }
11381 {
11383 /* MEM[&MEM[p, CST1], CST2] -> MEM[p, CST1 + CST2]. */
11386 {
11392 }
11394 /* MEM[&a.b, CST2] -> MEM[&a, offsetof (a, b) + CST2]. */
11397 {
11398 tree base;
11399 poly_int64 coffset;
11408 }
11413 /* INT +p INT -> (PTR)(INT + INT). Stripping types allows for this. */
11419 arg1),
11421 arg0)));
11427 {
11428 /* X + (X / CST) * -CST is X % CST. */
11433 {
11442 cst0));
11443 }
11444 }
11446 /* Handle (A1 * C1) + (A2 * C2) with A1, A2 or C1, C2 being the same or
11447 one. Make sure the type is not saturating and has the signedness of
11448 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11449 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11456 {
11460 }
11463 {
11464 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
11465 (plus (plus (mult) (mult)) (foo)) so that we can
11466 take advantage of the factoring cases below. */
11475 {
11481 else
11494 parg0),
11496 marg)),
11500 return
11506 parg1)));
11507 }
11508 }
11509 else
11510 {
11511 /* Fold __complex__ ( x, 0 ) + __complex__ ( 0, y )
11512 to __complex__ ( x, y ). This is not the same for SNaNs or
11513 if signed zeros are involved. */
11517 {
11524 {
11528 {
11534 }
11536 {
11542 }
11543 }
11544 }
11546 /* Convert a + (b*c + d*e) into (a + b*c) + d*e.
11547 We associate floats only if the user has specified
11548 -fassociative-math. */
11552 {
11557 {
11558 tree tree0;
11561 }
11562 }
11563 /* Convert (b*c + d*e) + a into b*c + (d*e +a).
11564 We associate floats only if the user has specified
11565 -fassociative-math. */
11569 {
11574 {
11575 tree tree0;
11578 }
11579 }
11580 }
11582 bit_rotate:
11583 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
11584 is a rotate of A by C1 bits. */
11585 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
11586 is a rotate of A by B bits.
11587 Similarly for (A << B) | (A >> (-B & C3)) where C3 is Z-1,
11588 though in this case CODE must be | and not + or ^, otherwise
11589 it doesn't return A when B is 0. */
11590 {
11592 tree rtype;
11601 /* Only create rotates in complete modes. Other cases are not
11602 expanded properly. */
11605 {
11618 {
11623 }
11628 {
11631 code0 == LSHIFT_EXPR
11634 }
11636 {
11646 {
11650 orig_tree01);
11652 }
11653 }
11657 {
11668 {
11672 orig_tree01);
11674 }
11675 }
11676 }
11677 }
11679 associate:
11680 /* In most languages, can't associate operations on floats through
11681 parentheses. Rather than remember where the parentheses were, we
11682 don't associate floats at all, unless the user has specified
11683 -fassociative-math.
11684 And, we need to make sure type is not saturating. */
11689 {
11695 /* Split both trees into variables, constants, and literals. Then
11696 associate each group together, the constants with literals,
11697 then the result with variables. This increases the chances of
11698 literals being recombined later and of generating relocatable
11699 expressions for the sum of a constant and literal. */
11707 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
11711 /* With undefined overflow prefer doing association in a type
11712 which wraps on overflow, if that is one of the operand types. */
11715 {
11723 }
11725 /* With undefined overflow we can only associate constants with one
11726 variable, and constants whose association doesn't overflow. */
11729 {
11731 {
11732 /* ??? If split_tree would handle NEGATE_EXPR we could
11733 simply reject these cases and the allowed cases would
11734 be the var0/minus_var1 ones. */
11740 {
11743 }
11750 {
11753 }
11759 /* The only case we can still associate with two variables
11760 is if they cancel out. */
11764 }
11770 }
11772 /* Only do something if we found more than two objects. Otherwise,
11773 nothing has changed and we risk infinite recursion. */
11781 {
11783 int minus_var0_origin
11786 int minus_con0_origin
11789 int minus_lit0_origin
11802 {
11808 }
11810 {
11816 }
11818 /* Preserve the MINUS_EXPR if the negative part of the literal is
11819 greater than the positive part. Otherwise, the multiplicative
11820 folding code (i.e extract_muldiv) may be fooled in case
11821 unsigned constants are subtracted, like in the following
11822 example: ((X*2 + 4) - 8U)/2. */
11824 {
11828 /* But avoid ending up with only negated parts. */
11830 {
11836 }
11837 else
11838 {
11844 }
11845 }
11847 /* Don't introduce overflows through reassociation. */
11852 /* Eliminate lit0 and minus_lit0 to con0 and minus_con0. */
11859 /* Eliminate minus_con0. */
11861 {
11863 {
11867 }
11869 {
11873 }
11874 else
11876 }
11878 /* Eliminate minus_var0. */
11880 {
11882 {
11886 }
11887 else
11889 }
11891 /* Reassociate only if there has been any actual association
11892 between subtrees from op0 and subtrees from op1 in at
11893 least one of the operands, otherwise we risk infinite
11894 recursion. See PR114084. */
11898 return
11901 }
11902 }
11908 /* Fold &a[i] - &a[j] to i-j. */
11913 {
11917 code
11921 }
11923 /* Further transformations are not for pointers. */
11927 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
11930 /* If arg0 is e.g. unsigned int and type is int, then this could
11931 introduce UB, because if A is INT_MIN at runtime, the original
11932 expression can be well defined while the latter is not.
11933 See PR83269. */
11942 /* Fold __complex__ ( x, 0 ) - __complex__ ( 0, y ) to
11943 __complex__ ( x, -y ). This is not the same for SNaNs or if
11944 signed zeros are involved. */
11948 {
11955 {
11959 {
11961 arg1r ? arg1r
11966 }
11968 {
11972 arg1i ? arg1i
11975 }
11976 }
11977 }
11979 /* A - B -> A + (-B) if B is easily negatable. */
11983 /* Avoid this transformation if B is a positive REAL_CST. */
11991 /* Handle (A1 * C1) - (A2 * C2) with A1, A2 or C1, C2 being the same or
11992 one. Make sure the type is not saturating and has the signedness of
11993 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11994 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
12001 {
12005 }
12011 {
12012 /* Transform x * -C into -x * C if x is easily negatable. */
12027 {
12030 "occur when simplifying "
12032 WARN_STRICT_OVERFLOW_MISC);
12034 }
12036 /* Optimize z * conj(z) for integer complex numbers. */
12043 }
12044 else
12045 {
12046 /* Fold z * +-I to __complex__ (-+__imag z, +-__real z).
12047 This is not the same for NaNs or if signed zeros are
12048 involved. */
12054 {
12057 return
12063 return
12068 }
12070 /* Optimize z * conj(z) for floating point complex numbers.
12071 Guarded by flag_unsafe_math_optimizations as non-finite
12072 imaginary components don't produce scalar results. */
12081 }
12085 /* Canonicalize (X & C1) | C2. */
12089 {
12094 /* If (C1&C2) == C1, then (X&C1)|C2 becomes (X,C2). */
12102 /* If (C1|C2) == ~0 then (X&C1)|C2 becomes X|C2. */
12104 {
12107 }
12109 /* Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2,
12110 unless (C1 & ~C2) | (C2 & C3) for some C3 is a mask of some
12111 mode which allows further optimizations. */
12116 {
12121 {
12124 }
12125 }
12128 {
12133 }
12134 }
12136 /* See if this can be simplified into a rotate first. If that
12137 is unsuccessful continue in the association code. */
12141 /* Fold (X & 1) ^ 1 as (X & 1) == 0. */
12149 /* See if this can be simplified into a rotate first. If that
12150 is unsuccessful continue in the association code. */
12154 /* Fold !X & 1 as X == 0. */
12157 {
12161 }
12163 /* Fold (X * Y) & -(1 << CST) to X * Y if Y is a constant
12164 multiple of 1 << CST. */
12166 {
12173 }
12175 /* Fold (X * CST1) & CST2 to zero if we can, or drop known zero
12176 bits from CST2. */
12180 {
12182 wide_int masked
12189 {
12190 /* Avoid the transform if arg1 is a mask of some
12191 mode which allows further optimizations. */
12198 }
12199 }
12201 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
12204 {
12209 return
12211 }
12216 /* Don't touch a floating-point divide by zero unless the mode
12217 of the constant can represent infinity. */
12223 /* (-A) / (-B) -> A / B */
12235 /* Fall through */
12238 /* Simplify A / (B << N) where A and B are positive and B is
12239 a power of 2, to A >> (N + log2(B)). */
12244 {
12247 {
12255 WARN_STRICT_OVERFLOW_MISC);
12261 }
12262 }
12264 /* Fall through */
12272 /* Convert -A / -B to A / B when the type is signed and overflow is
12273 undefined. */
12277 {
12280 "when distributing negation across "
12282 WARN_STRICT_OVERFLOW_MISC);
12287 }
12291 {
12294 "when distributing negation across "
12296 WARN_STRICT_OVERFLOW_MISC);
12301 }
12303 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
12304 operation, EXACT_DIV_EXPR.
12306 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
12307 At one time others generated faster code, it's not clear if they do
12308 after the last round to changes to the DIV code in expmed.cc. */
12319 {
12323 WARN_STRICT_OVERFLOW_MISC);
12325 }
12337 {
12341 WARN_STRICT_OVERFLOW_MISC);
12343 }
12351 /* Since negative shift count is not well-defined,
12352 don't try to compute it in the compiler. */
12358 /* If we have a rotate of a bit operation with the rotate count and
12359 the second operand of the bit operation both constant,
12360 permute the two operations. */
12366 {
12374 }
12376 /* Two consecutive rotates adding up to the some integer
12377 multiple of the precision of the type can be ignored. */
12393 /* Note that the operands of this must be ints
12394 and their values must be 0 or 1.
12395 ("true" is a fixed value perhaps depending on the language.) */
12396 /* If first arg is constant zero, return it. */
12399 /* FALLTHRU */
12401 /* If either arg is constant true, drop it. */
12405 /* Preserve sequence points. */
12408 /* If second arg is constant zero, result is zero, but first arg
12409 must be evaluated. */
12412 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
12413 case will be handled here. */
12417 /* !X && X is always false. */
12421 /* X && !X is always false. */
12426 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
12427 means A >= Y && A != MAX, but in this case we know that
12428 A < X <= MAX. */
12432 {
12444 }
12453 /* Note that the operands of this must be ints
12454 and their values must be 0 or true.
12455 ("true" is a fixed value perhaps depending on the language.) */
12456 /* If first arg is constant true, return it. */
12459 /* FALLTHRU */
12461 /* If either arg is constant zero, drop it. */
12465 /* Preserve sequence points. */
12468 /* If second arg is constant true, result is true, but we must
12469 evaluate first arg. */
12472 /* Likewise for first arg, but note this only occurs here for
12473 TRUTH_OR_EXPR. */
12477 /* !X || X is always true. */
12481 /* X || !X is always true. */
12486 /* (X && !Y) || (!X && Y) is X ^ Y */
12489 {
12506 }
12515 /* If the second arg is constant zero, drop it. */
12518 /* If the second arg is constant true, this is a logical inversion. */
12520 {
12523 }
12524 /* Identical arguments cancel to zero. */
12528 /* !X ^ X is always true. */
12533 /* X ^ !X is always true. */
12549 /* bool_var != 1 becomes !bool_var. */
12556 /* bool_var == 0 becomes !bool_var. */
12563 /* !exp != 0 becomes !exp */
12568 /* If this is an EQ or NE comparison with zero and ARG0 is
12569 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
12570 two operations, but the latter can be done in one less insn
12571 on machines that have only two-operand insns or on which a
12572 constant cannot be the first operand. */
12575 {
12580 {
12588 }
12591 {
12599 }
12600 }
12602 /* If this is a comparison of a field, we may be able to simplify it. */
12605 /* Handle the constant case even without -O
12606 to make sure the warnings are given. */
12608 {
12612 }
12614 /* Optimize comparisons of strlen vs zero to a compare of the
12615 first character of the string vs zero. To wit,
12616 strlen(ptr) == 0 => *ptr == 0
12617 strlen(ptr) != 0 => *ptr != 0
12618 Other cases should reduce to one of these two (or a constant)
12619 due to the return value of strlen being unsigned. */
12621 {
12629 {
12630 tree ptrtype
12632 TYPE_QUAL_CONST));
12638 }
12639 }
12641 /* Fold (X >> C) != 0 into X < 0 if C is one less than the width
12642 of X. Similarly fold (X >> C) == 0 into X >= 0. */
12646 {
12651 {
12653 {
12656 }
12659 }
12660 }
12662 /* Fold (~X & C) == 0 into (X & C) != 0 and (~X & C) != 0 into
12663 (X & C) == 0 when C is a single bit. */
12668 {
12675 arg1));
12676 }
12678 /* Fold ((X & C) ^ C) eq/ne 0 into (X & C) ne/eq 0, when the
12679 constant C is a power of two, i.e. a single bit. */
12686 {
12690 }
12692 /* Likewise, fold ((X ^ C) & C) eq/ne 0 into (X & C) ne/eq 0,
12693 when is C is a power of two, i.e. a single bit. */
12700 {
12706 }
12710 {
12717 /* Optimize (X ^ Z) op (Y ^ Z) as X op Y, and symmetries.
12718 operand_equal_p guarantees no side-effects so we don't need
12719 to use omit_one_operand on Z. */
12723 arg10));
12727 arg11));
12731 arg10));
12735 arg11));
12737 /* Optimize (X ^ C1) op (Y ^ C2) as (X ^ (C1 ^ C2)) op Y. */
12740 {
12746 }
12747 }
12749 /* Attempt to simplify equality/inequality comparisons of complex
12750 values. Only lower the comparison if the result is known or
12751 can be simplified to a single scalar comparison. */
12756 {
12761 {
12764 }
12765 else
12766 {
12769 }
12772 {
12775 }
12776 else
12777 {
12780 }
12784 {
12786 {
12791 }
12792 else
12793 {
12798 }
12799 }
12803 {
12805 {
12810 }
12811 else
12812 {
12817 }
12818 }
12819 }
12831 /* Transform comparisons of the form X +- C CMP X. */
12836 {
12841 /* (X - c) > X becomes false. */
12847 /* Likewise (X + c) < X becomes false. */
12853 /* Convert (X - c) <= X to true. */
12860 /* Convert (X + c) >= X to true. */
12866 }
12868 /* If we are comparing an ABS_EXPR with a constant, we can
12869 convert all the cases into explicit comparisons, but they may
12870 well not be faster than doing the ABS and one comparison.
12871 But ABS (X) <= C is a range comparison, which becomes a subtraction
12872 and a comparison, and is probably faster. */
12886 /* Convert ABS_EXPR<x> >= 0 to true. */
12893 {
12896 "when simplifying comparison of "
12898 WARN_STRICT_OVERFLOW_CONDITIONAL);
12901 arg0);
12902 }
12904 /* Convert ABS_EXPR<x> < 0 to false. */
12909 {
12912 "when simplifying comparison of "
12914 WARN_STRICT_OVERFLOW_CONDITIONAL);
12917 arg0);
12918 }
12920 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
12921 and similarly for >= into !=. */
12931 /* Similarly for X < (cast) (1 << Y). But cast can't be narrowing,
12932 otherwise Y might be >= # of bits in X's type and thus e.g.
12933 (unsigned char) (1 << Y) for Y 15 might be 0.
12934 If the cast is widening, then 1 << Y should have unsigned type,
12935 otherwise if Y is number of bits in the signed shift type minus 1,
12936 we can't optimize this. E.g. (unsigned long long) (1 << Y) for Y
12937 31 might be 0xffffffff80000000. */
12950 {
12956 }
12968 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
12969 {
12982 }
12987 /* When pedantic, a compound expression can be neither an lvalue
12988 nor an integer constant expression. */
12991 /* Don't let (0, 0) be null pointer constant. */
12999 }
13001 /* For constants M and N, if M == (1LL << cst) - 1 && (N & M) == M,
13002 ((A & N) + B) & M -> (A + B) & M
13003 Similarly if (N & M) == 0,
13004 ((A | N) + B) & M -> (A + B) & M
13005 and for - instead of + (or unary - instead of +)
13006 and/or ^ instead of |.
13007 If B is constant and (B & M) == 0, fold into A & M.
13009 This function is a helper for match.pd patterns. Return non-NULL
13010 type in which the simplified operation should be performed only
13011 if any optimization is possible.
13013 ARG1 is M above, ARG00 is left operand of +/-, if CODE00 is BIT_*_EXPR,
13014 then ARG00{0,1} are operands of that bitop, otherwise CODE00 is ERROR_MARK.
13015 Similarly for ARG01, CODE01 and ARG01{0,1}, just for the right operand of
13016 +/-. */
13017 tree
13022 {
13037 wide_int cst0;
13039 /* Now we know that arg0 is (C + D) or (C - D) or -C and
13040 arg1 (M) is == (1LL << cst) - 1.
13041 Store C into PMOP[0] and D into PMOP[1]. */
13048 {
13055 {
13058 }
13061 /* If C or D is of the form (A & N) where
13062 (N & M) == M, or of the form (A | N) or
13063 (A ^ N) where (N & M) == 0, replace it with A. */
13069 /* If C or D is a N where (N & M) == 0, it can be
13070 omitted (replaced with 0). */
13075 /* Similarly, with C - N where (-N & M) == 0. */
13083 }
13085 /* Only build anything new if we optimized one or both arguments above. */
13091 else
13093 }
13095 /* Used by contains_label_[p1]. */
13097 struct contains_label_data
13098 {
13101 };
13103 /* Callback for walk_tree, looking for LABEL_EXPR. Return *TP if it is
13104 a LABEL_EXPR or CASE_LABEL_EXPR not inside of another SWITCH_EXPR; otherwise
13105 return NULL_TREE. Do not check the subtrees of GOTO_EXPR. */
13107 static tree
13109 {
13112 {
13123 {
13131 }
13140 }
13141 }
13143 /* Return whether the sub-tree ST contains a label which is accessible from
13144 outside the sub-tree. */
13146 static bool
13148 {
13152 }
13154 /* Fold a ternary expression of code CODE and type TYPE with operands
13155 OP0, OP1, and OP2. Return the folded expression if folding is
13156 successful. Otherwise, return NULL_TREE. */
13158 tree
13161 {
13162 tree tem;
13169 /* If this is a commutative operation, and OP0 is a constant, move it
13170 to OP1 to reduce the number of tests below. */
13179 /* Strip any conversions that don't change the mode. This is safe
13180 for every expression, except for a comparison expression because
13181 its signedness is derived from its operands. So, in the latter
13182 case, only strip conversions that don't change the signedness.
13184 Note that this is done as an internal manipulation within the
13185 constant folder, in order to find the simplest representation of
13186 the arguments so that their form can be studied. In any cases,
13187 the appropriate type conversions should be put back in the tree
13188 that will get out of the constant folder. */
13190 {
13193 }
13196 {
13199 }
13202 {
13205 }
13208 {
13212 {
13218 }
13223 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
13224 so all simple results must be passed through pedantic_non_lvalue. */
13226 {
13229 /* Only optimize constant conditions when the selected branch
13230 has the same type as the COND_EXPR. This avoids optimizing
13231 away "c ? x : throw", where the throw has a void type.
13232 Avoid throwing away that operand which contains label. */
13239 }
13241 {
13248 {
13251 {
13259 }
13264 }
13265 }
13267 /* If we have A op B ? A : C, we may be able to convert this to a
13268 simpler expression, depending on the operation and the values
13269 of B and C. Signed zeros prevent all of these transformations,
13270 for reasons given above each one.
13272 Also try swapping the arguments and inverting the conditional. */
13276 {
13283 }
13288 {
13295 arg00,
13296 arg01,
13300 }
13302 /* If the second operand is simpler than the third, swap them
13303 since that produces better jump optimization results. */
13306 {
13308 /* See if this can be inverted. If it can't, possibly because
13309 it was a floating-point inequality comparison, don't do
13310 anything. */
13314 }
13316 /* Convert A ? 1 : 0 to simply A. */
13321 /* If we try to convert OP0 to our type, the
13322 call to fold will try to move the conversion inside
13323 a COND, which will recurse. In that case, the COND_EXPR
13324 is probably the best choice, so leave it alone. */
13328 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
13329 over COND_EXPR in cases such as floating point comparisons. */
13338 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
13343 {
13344 /* sign_bit_p looks through both zero and sign extensions,
13345 but for this optimization only sign extensions are
13346 usable. */
13349 {
13352 {
13355 }
13357 }
13358 /* sign_bit_p only checks ARG1 bits within A's precision.
13359 If <sign bit of A> has wider type than A, bits outside
13360 of A's precision in <sign bit of A> need to be checked.
13361 If they are all 0, this optimization needs to be done
13362 in unsigned A's type, if they are all 1 in signed A's type,
13363 otherwise this can't be done. */
13369 {
13371 tree tem_type;
13384 {
13387 }
13389 {
13392 }
13393 else
13395 }
13398 return
13404 arg1)));
13405 }
13407 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
13408 already handled above. */
13413 {
13423 op1);
13424 }
13426 /* A & N ? N : 0 is simply A & N if N is a power of two. This
13427 is probably obsolete because the first operand should be a
13428 truth value (that's why we have the two cases above), but let's
13429 leave it in until we can confirm this for all front-ends. */
13437 /* operand_equal_p compares just value, not precision, so e.g.
13438 arg1 could be 8-bit -128 and be power of two, but BIT_AND_EXPR
13439 second operand 32-bit -128, which is not a power of two (or vice
13440 versa. */
13444 /* Disable the transformations below for vectors, since
13445 fold_binary_op_with_conditional_arg may undo them immediately,
13446 yielding an infinite loop. */
13450 /* Convert A ? B : 0 into A && B if A and B are truth values. */
13459 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
13464 {
13466 /* Only perform transformation if ARG0 is easily inverted. */
13470 ? BIT_IOR_EXPR
13473 op1);
13474 }
13476 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
13481 {
13483 /* Only perform transformation if ARG0 is easily inverted. */
13489 op2);
13490 }
13492 /* Convert A ? 1 : B into A || B if A and B are truth values. */
13504 /* CALL_EXPRs used to be ternary exprs. Catch any mistaken uses
13505 of fold_ternary on them. */
13515 {
13517 unsigned HOST_WIDE_INT width
13528 {
13533 {
13535 {
13540 }
13546 }
13547 }
13548 }
13550 /* On constants we can use native encode/interpret to constant
13551 fold (nearly) all BIT_FIELD_REFs. */
13557 {
13560 /* Limit us to a reasonable amount of work. To relax the
13561 other limitations we need bit-shifting of the buffer
13562 and rounding up the size. */
13566 {
13568 unsigned HOST_WIDE_INT len
13573 {
13578 }
13579 }
13580 }
13585 /* Perform constant folding of BIT_INSERT_EXPR. */
13589 {
13590 /* Build a vector of integers from the tree mask. */
13591 vec_perm_builder builder;
13595 /* Create a vec_perm_indices for the integer vector. */
13600 }
13604 /* Perform (partial) constant folding of BIT_INSERT_EXPR. */
13607 {
13613 wide_int tem2
13617 }
13622 {
13624 unsigned HOST_WIDE_INT elsize
13627 {
13633 {
13639 }
13640 }
13641 }
13647 }
13649 /* Gets the element ACCESS_INDEX from CTOR, which must be a CONSTRUCTOR
13650 of an array (or vector). *CTOR_IDX if non-NULL is updated with the
13651 constructor element index of the value returned. If the element is
13652 not found NULL_TREE is returned and *CTOR_IDX is updated to
13653 the index of the element after the ACCESS_INDEX position (which
13654 may be outside of the CTOR array). */
13656 tree
13659 {
13665 {
13668 {
13669 /* Static constructors for variably sized objects makes no sense. */
13672 /* ??? When it is obvious that the range is signed, treat it so. */
13677 {
13679 low_bound
13681 SIGNED);
13682 }
13683 else
13684 {
13687 }
13688 }
13689 }
13693 index_sgn);
13705 {
13706 /* Array constructor might explicitly set index, or specify a range,
13707 or leave index NULL meaning that it is next index after previous
13708 one. */
13710 {
13712 max_index = index
13714 else
13715 {
13718 index_sgn);
13719 max_index
13721 index_sgn);
13723 }
13724 }
13726 {
13732 }
13733 else
13736 /* Do we have match? */
13738 {
13740 {
13744 }
13745 }
13747 /* We're past the element we search for. Note during parsing
13748 the elements might not be sorted.
13749 ??? We should use a binary search and a flag on the
13750 CONSTRUCTOR as to whether elements are sorted in declaration
13751 order. */
13753 }
13757 }
13759 /* Perform constant folding and related simplification of EXPR.
13760 The related simplifications include x*1 => x, x*0 => 0, etc.,
13761 and application of the associative law.
13762 NOP_EXPR conversions may be removed freely (as long as we
13763 are careful not to change the type of the overall expression).
13764 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
13765 but we can constant-fold them if they have constant operands. */
13767 #ifdef ENABLE_FOLD_CHECKING
13768 # define fold(x) fold_1 (x)
13770 static
13771 #endif
13772 tree
13774 {
13778 tree tem;
13781 /* Return right away if a constant. */
13785 /* CALL_EXPR-like objects with variable numbers of operands are
13786 treated specially. */
13788 {
13790 {
13793 }
13795 }
13798 {
13803 {
13821 }
13822 }
13825 {
13827 {
13834 {
13839 }
13842 }
13844 /* Return a VECTOR_CST if possible. */
13846 {
13852 tree val;
13858 }
13866 }
13868 #ifdef ENABLE_FOLD_CHECKING
13869 #undef fold
13876 /* When --enable-checking=fold, compute a digest of expr before
13877 and after actual fold call to see if fold did not accidentally
13878 change original expr. */
13880 tree
13882 {
13883 tree ret;
13903 }
13905 void
13907 {
13918 }
13920 static void
13922 {
13924 }
13926 static void
13929 {
13935 recursive_label:
13945 {
13946 /* Allow DECL_ASSEMBLER_NAME and symtab_node to be modified. */
13954 }
13962 {
13963 /* Allow these fields to be modified. */
13964 tree tmp;
13975 {
13978 }
13979 }
13981 {
13982 /* Allow the no-warning bit to be set. Perhaps we shouldn't allow
13983 that and change builtins.cc etc. instead - see PR89543. */
13989 }
14000 {
14003 {
14019 }
14023 {
14036 }
14053 {
14059 }
14062 {
14064 {
14067 }
14069 }
14080 {
14083 }
14091 }
14092 }
14094 /* Helper function for outputting the checksum of a tree T. When
14095 debugging with gdb, you can "define mynext" to be "next" followed
14096 by "call debug_fold_checksum (op0)", then just trace down till the
14097 outputs differ. */
14099 DEBUG_FUNCTION void
14101 {
14116 }
14118 #endif
14120 /* Fold a unary tree expression with code CODE of type TYPE with an
14121 operand OP0. LOC is the location of the resulting expression.
14122 Return a folded expression if successful. Otherwise, return a tree
14123 expression with code CODE of type TYPE with an operand OP0. */
14125 tree
14128 {
14129 tree tem;
14130 #ifdef ENABLE_FOLD_CHECKING
14139 #endif
14145 #ifdef ENABLE_FOLD_CHECKING
14152 #endif
14154 }
14156 /* Fold a binary tree expression with code CODE of type TYPE with
14157 operands OP0 and OP1. LOC is the location of the resulting
14158 expression. Return a folded expression if successful. Otherwise,
14159 return a tree expression with code CODE of type TYPE with operands
14160 OP0 and OP1. */
14162 tree
14165 MEM_STAT_DECL)
14166 {
14167 tree tem;
14168 #ifdef ENABLE_FOLD_CHECKING
14185 #endif
14191 #ifdef ENABLE_FOLD_CHECKING
14206 #endif
14208 }
14210 /* Fold a ternary tree expression with code CODE of type TYPE with
14211 operands OP0, OP1, and OP2. Return a folded expression if
14212 successful. Otherwise, return a tree expression with code CODE of
14213 type TYPE with operands OP0, OP1, and OP2. */
14215 tree
14218 {
14219 tree tem;
14220 #ifdef ENABLE_FOLD_CHECKING
14244 #endif
14251 #ifdef ENABLE_FOLD_CHECKING
14274 #endif
14276 }
14278 /* Fold a CALL_EXPR expression of type TYPE with operands FN and NARGS
14279 arguments in ARGARRAY, and a null static chain.
14280 Return a folded expression if successful. Otherwise, return a CALL_EXPR
14281 of type TYPE from the given operands as constructed by build_call_array. */
14283 tree
14286 {
14287 tree tem;
14288 #ifdef ENABLE_FOLD_CHECKING
14307 #endif
14313 #ifdef ENABLE_FOLD_CHECKING
14329 #endif
14331 }
14333 /* Perform constant folding and related simplification of initializer
14334 expression EXPR. These behave identically to "fold_buildN" but ignore
14335 potential run-time traps and exceptions that fold must preserve. */
14337 #define START_FOLD_INIT \
14338 int saved_signaling_nans = flag_signaling_nans;\
14339 int saved_trapping_math = flag_trapping_math;\
14340 int saved_rounding_math = flag_rounding_math;\
14341 int saved_trapv = flag_trapv;\
14342 int saved_folding_initializer = folding_initializer;\
14343 flag_signaling_nans = 0;\
14344 flag_trapping_math = 0;\
14345 flag_rounding_math = 0;\
14346 flag_trapv = 0;\
14347 folding_initializer = 1;
14349 #define END_FOLD_INIT \
14350 flag_signaling_nans = saved_signaling_nans;\
14351 flag_trapping_math = saved_trapping_math;\
14352 flag_rounding_math = saved_rounding_math;\
14353 flag_trapv = saved_trapv;\
14354 folding_initializer = saved_folding_initializer;
14356 tree
14358 {
14359 tree result;
14360 START_FOLD_INIT;
14364 END_FOLD_INIT;
14366 }
14368 tree
14371 {
14372 tree result;
14373 START_FOLD_INIT;
14377 END_FOLD_INIT;
14379 }
14381 tree
14384 {
14385 tree result;
14386 START_FOLD_INIT;
14390 END_FOLD_INIT;
14392 }
14394 tree
14397 {
14398 tree result;
14399 START_FOLD_INIT;
14403 END_FOLD_INIT;
14405 }
14407 tree
14410 {
14411 tree result;
14412 START_FOLD_INIT;
14416 END_FOLD_INIT;
14418 }
14420 #undef START_FOLD_INIT
14421 #undef END_FOLD_INIT
14423 /* Determine if first argument is a multiple of second argument. Return
14424 false if it is not, or we cannot easily determined it to be.
14426 An example of the sort of thing we care about (at this point; this routine
14427 could surely be made more general, and expanded to do what the *_DIV_EXPR's
14428 fold cases do now) is discovering that
14430 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14432 is a multiple of
14434 SAVE_EXPR (J * 8)
14436 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
14438 This code also handles discovering that
14440 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14442 is a multiple of 8 so we don't have to worry about dealing with a
14443 possible remainder.
14445 Note that we *look* inside a SAVE_EXPR only to determine how it was
14446 calculated; it is not safe for fold to do much of anything else with the
14447 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
14448 at run time. For example, the latter example above *cannot* be implemented
14449 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
14450 evaluation time of the original SAVE_EXPR is not necessarily the same at
14451 the time the new expression is evaluated. The only optimization of this
14452 sort that would be valid is changing
14454 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
14456 divided by 8 to
14458 SAVE_EXPR (I) * SAVE_EXPR (J)
14460 (where the same SAVE_EXPR (J) is used in the original and the
14461 transformed version).
14463 NOWRAP specifies whether all outer operations in TYPE should
14464 be considered not wrapping. Any type conversion within TOP acts
14465 as a barrier and we will fall back to NOWRAP being false.
14466 NOWRAP is mostly used to treat expressions in TYPE_SIZE and friends
14467 as not wrapping even though they are generally using unsigned arithmetic. */
14469 bool
14471 {
14482 {
14484 /* Bitwise and provides a power of two multiple. If the mask is
14485 a multiple of BOTTOM then TOP is a multiple of BOTTOM. */
14492 /* If the multiplication can wrap we cannot recurse further unless
14493 the bottom is a power of two which is where wrapping does not
14494 matter. */
14500 {
14506 {
14509 /* Handle multiple_of_p ((x * 2 + 2) * 4, 8). */
14511 {
14515 {
14518 }
14519 }
14521 }
14522 }
14527 /* Handle X << CST as X * (1 << CST) and only process the constant. */
14529 {
14532 {
14533 wide_int mul_op
14537 nowrap);
14538 }
14539 }
14544 /* If the addition or subtraction can wrap we cannot recurse further
14545 unless bottom is a power of two which is where wrapping does not
14546 matter. */
14552 /* Handle cases like op0 + 0xfffffffd as op0 - 3 if the expression has
14553 unsigned type. For example, (X / 3) + 0xfffffffd is multiple of 3,
14554 but 0xfffffffd is not. */
14557 && nowrap
14562 /* It is impossible to prove if op0 +- op1 is multiple of bottom
14563 precisely, so be conservative here checking if both op0 and op1
14564 are multiple of bottom. Note we check the second operand first
14565 since it's usually simpler. */
14569 CASE_CONVERT:
14570 /* Can't handle conversions from non-integral or wider integral type. */
14575 /* NOWRAP only extends to operations in the outermost type so
14576 make sure to strip it off here. */
14591 SIGNED);
14597 {
14600 /* Check for special cases to see if top is defined as multiple
14601 of bottom:
14603 top = (X & ~(bottom - 1) ; bottom is power of 2
14605 or
14607 Y = X % bottom
14608 top = X - Y. */
14627 }
14629 /* fall through */
14637 }
14638 }
14640 /* Return true if expression X cannot be (or contain) a NaN or infinity.
14641 This function returns true for integer expressions, and returns
14642 false if uncertain. */
14644 bool
14646 {
14651 {
14674 {
14675 CASE_CFN_FABS:
14676 CASE_CFN_FABS_FN:
14678 CASE_CFN_FMAX:
14679 CASE_CFN_FMAX_FN:
14680 CASE_CFN_FMIN:
14681 CASE_CFN_FMIN_FN:
14686 }
14690 }
14691 }
14693 /* Return true if expression X evaluates to an infinity.
14694 This function returns false for integer expressions. */
14696 bool
14698 {
14702 {
14715 }
14716 }
14718 /* Return true if expression X could evaluate to an infinity.
14719 This function returns false for integer expressions, and returns
14720 true if uncertain. */
14722 bool
14724 {
14728 {
14741 }
14742 }
14744 /* Return true if expression X evaluates to a signaling NaN.
14745 This function returns false for integer expressions. */
14747 bool
14749 {
14753 {
14764 }
14765 }
14767 /* Return true if expression X could evaluate to a signaling NaN.
14768 This function returns false for integer expressions, and returns
14769 true if uncertain. */
14771 bool
14773 {
14777 {
14797 {
14798 CASE_CFN_FABS:
14799 CASE_CFN_FABS_FN:
14801 CASE_CFN_FMAX:
14802 CASE_CFN_FMAX_FN:
14803 CASE_CFN_FMIN:
14804 CASE_CFN_FMIN_FN:
14809 }
14812 }
14813 }
14815 /* Return true if expression X evaluates to a NaN.
14816 This function returns false for integer expressions. */
14818 bool
14820 {
14824 {
14835 }
14836 }
14838 /* Return true if expression X could evaluate to a NaN.
14839 This function returns false for integer expressions, and returns
14840 true if uncertain. */
14842 bool
14844 {
14848 {
14873 {
14874 CASE_CFN_FABS:
14875 CASE_CFN_FABS_FN:
14877 CASE_CFN_FMAX:
14878 CASE_CFN_FMAX_FN:
14879 CASE_CFN_FMIN:
14880 CASE_CFN_FMIN_FN:
14885 }
14888 }
14889 }
14891 /* Return true if expression X could evaluate to -0.0.
14892 This function returns true if uncertain. */
14894 bool
14896 {
14900 {
14915 {
14916 CASE_CFN_FABS:
14917 CASE_CFN_FABS_FN:
14921 }
14924 }
14925 /* Ideally !(tree_expr_nonzero_p (X) || tree_expr_nonnegative_p (X))
14926 * but currently those predicates require tree and not const_tree. */
14928 }
14930 #define tree_expr_nonnegative_warnv_p(X, Y) \
14933 #define RECURSE(X) \
14934 ((tree_expr_nonnegative_warnv_p) (X, strict_overflow_p, depth + 1))
14936 /* Return true if CODE or TYPE is known to be non-negative. */
14938 static bool
14940 {
14944 /* Truth values evaluate to 0 or 1, which is nonnegative unless we
14945 have a signed:1 type (where the value is -1 and 0). */
14948 }
14950 /* Return true if (CODE OP0) is known to be non-negative. If the return
14951 value is based on the assumption that signed overflow is undefined,
14952 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14953 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14955 bool
14958 {
14963 {
14965 /* We can't return 1 if flag_wrapv is set because
14966 ABS_EXPR<INT_MIN> = INT_MIN. */
14970 {
14973 }
14981 CASE_CONVERT:
14982 {
14987 {
14991 {
14995 }
14996 }
14998 {
15004 }
15005 }
15010 }
15012 /* We don't know sign of `t', so be conservative and return false. */
15014 }
15016 /* Return true if (CODE OP0 OP1) is known to be non-negative. If the return
15017 value is based on the assumption that signed overflow is undefined,
15018 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15019 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15021 bool
15025 {
15030 {
15036 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
15037 both unsigned and at least 2 bits shorter than the result. */
15041 {
15046 {
15050 }
15051 }
15056 {
15057 /* x * x is always non-negative for floating point x
15058 or without overflow. */
15061 {
15066 }
15067 }
15069 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
15070 both unsigned and their total bits is shorter than the result. */
15074 {
15093 {
15103 }
15104 }
15111 /* Usually RECURSE (op0) || RECURSE (op1) but NaNs complicate
15112 things. */
15137 }
15139 /* We don't know sign of `t', so be conservative and return false. */
15141 }
15143 /* Return true if T is known to be non-negative. If the return
15144 value is based on the assumption that signed overflow is undefined,
15145 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15146 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15148 bool
15150 {
15155 {
15169 /* Limit the depth of recursion to avoid quadratic behavior.
15170 This is expected to catch almost all occurrences in practice.
15171 If this code misses important cases that unbounded recursion
15172 would not, passes that need this information could be revised
15173 to provide it through dataflow propagation. */
15181 }
15182 }
15184 /* Return true if T is known to be non-negative. If the return
15185 value is based on the assumption that signed overflow is undefined,
15186 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15187 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15189 bool
15192 {
15194 {
15195 CASE_CFN_ACOS:
15196 CASE_CFN_ACOS_FN:
15197 CASE_CFN_ACOSH:
15198 CASE_CFN_ACOSH_FN:
15199 CASE_CFN_CABS:
15200 CASE_CFN_CABS_FN:
15201 CASE_CFN_COSH:
15202 CASE_CFN_COSH_FN:
15203 CASE_CFN_ERFC:
15204 CASE_CFN_ERFC_FN:
15205 CASE_CFN_EXP:
15206 CASE_CFN_EXP_FN:
15207 CASE_CFN_EXP10:
15208 CASE_CFN_EXP2:
15209 CASE_CFN_EXP2_FN:
15210 CASE_CFN_FABS:
15211 CASE_CFN_FABS_FN:
15212 CASE_CFN_FDIM:
15213 CASE_CFN_FDIM_FN:
15214 CASE_CFN_HYPOT:
15215 CASE_CFN_HYPOT_FN:
15216 CASE_CFN_POW10:
15217 CASE_CFN_FFS:
15218 CASE_CFN_PARITY:
15219 CASE_CFN_POPCOUNT:
15220 CASE_CFN_CLZ:
15221 CASE_CFN_CLRSB:
15226 /* Always true. */
15229 CASE_CFN_SQRT:
15230 CASE_CFN_SQRT_FN:
15231 /* sqrt(-0.0) is -0.0. */
15236 CASE_CFN_ASINH:
15237 CASE_CFN_ASINH_FN:
15238 CASE_CFN_ATAN:
15239 CASE_CFN_ATAN_FN:
15240 CASE_CFN_ATANH:
15241 CASE_CFN_ATANH_FN:
15242 CASE_CFN_CBRT:
15243 CASE_CFN_CBRT_FN:
15244 CASE_CFN_CEIL:
15245 CASE_CFN_CEIL_FN:
15246 CASE_CFN_ERF:
15247 CASE_CFN_ERF_FN:
15248 CASE_CFN_EXPM1:
15249 CASE_CFN_EXPM1_FN:
15250 CASE_CFN_FLOOR:
15251 CASE_CFN_FLOOR_FN:
15252 CASE_CFN_FMOD:
15253 CASE_CFN_FMOD_FN:
15254 CASE_CFN_FREXP:
15255 CASE_CFN_FREXP_FN:
15256 CASE_CFN_ICEIL:
15257 CASE_CFN_IFLOOR:
15258 CASE_CFN_IRINT:
15259 CASE_CFN_IROUND:
15260 CASE_CFN_LCEIL:
15261 CASE_CFN_LDEXP:
15262 CASE_CFN_LFLOOR:
15263 CASE_CFN_LLCEIL:
15264 CASE_CFN_LLFLOOR:
15265 CASE_CFN_LLRINT:
15266 CASE_CFN_LLRINT_FN:
15267 CASE_CFN_LLROUND:
15268 CASE_CFN_LLROUND_FN:
15269 CASE_CFN_LRINT:
15270 CASE_CFN_LRINT_FN:
15271 CASE_CFN_LROUND:
15272 CASE_CFN_LROUND_FN:
15273 CASE_CFN_MODF:
15274 CASE_CFN_MODF_FN:
15275 CASE_CFN_NEARBYINT:
15276 CASE_CFN_NEARBYINT_FN:
15277 CASE_CFN_RINT:
15278 CASE_CFN_RINT_FN:
15279 CASE_CFN_ROUND:
15280 CASE_CFN_ROUND_FN:
15281 CASE_CFN_ROUNDEVEN:
15282 CASE_CFN_ROUNDEVEN_FN:
15283 CASE_CFN_SCALB:
15284 CASE_CFN_SCALBLN:
15285 CASE_CFN_SCALBLN_FN:
15286 CASE_CFN_SCALBN:
15287 CASE_CFN_SCALBN_FN:
15288 CASE_CFN_SIGNBIT:
15289 CASE_CFN_SIGNIFICAND:
15290 CASE_CFN_SINH:
15291 CASE_CFN_SINH_FN:
15292 CASE_CFN_TANH:
15293 CASE_CFN_TANH_FN:
15294 CASE_CFN_TRUNC:
15295 CASE_CFN_TRUNC_FN:
15296 /* True if the 1st argument is nonnegative. */
15299 CASE_CFN_FMAX:
15300 CASE_CFN_FMAX_FN:
15301 /* Usually RECURSE (arg0) || RECURSE (arg1) but NaNs complicate
15302 things. In the presence of sNaNs, we're only guaranteed to be
15303 non-negative if both operands are non-negative. In the presence
15304 of qNaNs, we're non-negative if either operand is non-negative
15305 and can't be a qNaN, or if both operands are non-negative. */
15314 CASE_CFN_FMIN:
15315 CASE_CFN_FMIN_FN:
15316 /* True if the 1st AND 2nd arguments are nonnegative. */
15319 CASE_CFN_COPYSIGN:
15320 CASE_CFN_COPYSIGN_FN:
15321 /* True if the 2nd argument is nonnegative. */
15324 CASE_CFN_POWI:
15325 /* True if the 1st argument is nonnegative or the second
15326 argument is an even integer. */
15332 CASE_CFN_POW:
15333 CASE_CFN_POW_FN:
15334 /* True if the 1st argument is nonnegative or the second
15335 argument is an even integer valued real. */
15337 {
15338 REAL_VALUE_TYPE c;
15339 HOST_WIDE_INT n;
15344 {
15345 REAL_VALUE_TYPE cint;
15349 }
15350 }
15355 }
15357 }
15359 /* Return true if T is known to be non-negative. If the return
15360 value is based on the assumption that signed overflow is undefined,
15361 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15362 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15364 static bool
15366 {
15372 {
15374 {
15378 /* If the initializer is non-void, then it's a normal expression
15379 that will be assigned to the slot. */
15383 /* Otherwise, the initializer sets the slot in some way. One common
15384 way is an assignment statement at the end of the initializer. */
15386 {
15394 else
15396 }
15402 }
15405 {
15411 arg0,
15412 arg1,
15414 }
15427 }
15428 }
15430 #undef RECURSE
15431 #undef tree_expr_nonnegative_warnv_p
15433 /* Return true if T is known to be non-negative. If the return
15434 value is based on the assumption that signed overflow is undefined,
15435 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15436 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15438 bool
15440 {
15447 {
15469 }
15472 {
15497 }
15498 }
15500 /* Return true if `t' is known to be non-negative. Handle warnings
15501 about undefined signed overflow. */
15503 bool
15505 {
15512 "determining that expression is always "
15514 WARN_STRICT_OVERFLOW_MISC);
15516 }
15519 /* Return true when (CODE OP0) is an address and is known to be nonzero.
15520 For floating point we further ensure that T is not denormal.
15521 Similar logic is present in nonzero_address in rtlanal.h.
15523 If the return value is based on the assumption that signed overflow
15524 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15525 change *STRICT_OVERFLOW_P. */
15527 bool
15530 {
15532 {
15535 strict_overflow_p);
15538 {
15544 strict_overflow_p));
15545 }
15550 strict_overflow_p);
15554 }
15557 }
15559 /* Return true when (CODE OP0 OP1) is an address and is known to be nonzero.
15560 For floating point we further ensure that T is not denormal.
15561 Similar logic is present in nonzero_address in rtlanal.h.
15563 If the return value is based on the assumption that signed overflow
15564 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15565 change *STRICT_OVERFLOW_P. */
15567 bool
15569 tree type,
15570 tree op0,
15572 {
15575 {
15579 {
15580 /* With the presence of negative values it is hard
15581 to say something. */
15588 /* One of operands must be positive and the other non-negative. */
15589 /* We don't set *STRICT_OVERFLOW_P here: even if this value
15590 overflows, on a twos-complement machine the sum of two
15591 nonnegative numbers can never be zero. */
15593 strict_overflow_p)
15595 strict_overflow_p));
15596 }
15601 {
15603 strict_overflow_p)
15605 strict_overflow_p))
15606 {
15609 }
15610 }
15619 {
15622 }
15629 {
15633 /* When both operands are nonzero, then MAX must be too. */
15635 strict_overflow_p))
15638 /* MAX where operand 0 is positive is positive. */
15640 strict_overflow_p);
15641 }
15642 /* MAX where operand 1 is positive is positive. */
15647 {
15651 }
15656 strict_overflow_p)
15658 strict_overflow_p));
15662 }
15665 }
15667 /* Return true when T is an address and is known to be nonzero.
15668 For floating point we further ensure that T is not denormal.
15669 Similar logic is present in nonzero_address in rtlanal.h.
15671 If the return value is based on the assumption that signed overflow
15672 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15673 change *STRICT_OVERFLOW_P. */
15675 bool
15677 {
15680 {
15685 {
15697 /* For objects in symbol table check if we know they are non-zero.
15698 Don't do anything for variables and functions before symtab is built;
15699 it is quite possible that they will be declared weak later. */
15704 /* Constants are never weak. */
15709 }
15717 {
15721 }
15731 }
15733 }
15735 #define integer_valued_real_p(X) \
15738 #define RECURSE(X) \
15739 ((integer_valued_real_p) (X, depth + 1))
15741 /* Return true if the floating point result of (CODE OP0) has an
15742 integer value. We also allow +Inf, -Inf and NaN to be considered
15743 integer values. Return false for signaling NaN.
15745 DEPTH is the current nesting depth of the query. */
15747 bool
15749 {
15751 {
15758 CASE_CONVERT:
15759 {
15766 }
15770 }
15772 }
15774 /* Return true if the floating point result of (CODE OP0 OP1) has an
15775 integer value. We also allow +Inf, -Inf and NaN to be considered
15776 integer values. Return false for signaling NaN.
15778 DEPTH is the current nesting depth of the query. */
15780 bool
15782 {
15784 {
15794 }
15796 }
15798 /* Return true if the floating point result of calling FNDECL with arguments
15799 ARG0 and ARG1 has an integer value. We also allow +Inf, -Inf and NaN to be
15800 considered integer values. Return false for signaling NaN. If FNDECL
15801 takes fewer than 2 arguments, the remaining ARGn are null.
15803 DEPTH is the current nesting depth of the query. */
15805 bool
15807 {
15809 {
15810 CASE_CFN_CEIL:
15811 CASE_CFN_CEIL_FN:
15812 CASE_CFN_FLOOR:
15813 CASE_CFN_FLOOR_FN:
15814 CASE_CFN_NEARBYINT:
15815 CASE_CFN_NEARBYINT_FN:
15816 CASE_CFN_RINT:
15817 CASE_CFN_RINT_FN:
15818 CASE_CFN_ROUND:
15819 CASE_CFN_ROUND_FN:
15820 CASE_CFN_ROUNDEVEN:
15821 CASE_CFN_ROUNDEVEN_FN:
15822 CASE_CFN_TRUNC:
15823 CASE_CFN_TRUNC_FN:
15826 CASE_CFN_FMIN:
15827 CASE_CFN_FMIN_FN:
15828 CASE_CFN_FMAX:
15829 CASE_CFN_FMAX_FN:
15834 }
15836 }
15838 /* Return true if the floating point expression T (a GIMPLE_SINGLE_RHS)
15839 has an integer value. We also allow +Inf, -Inf and NaN to be
15840 considered integer values. Return false for signaling NaN.
15842 DEPTH is the current nesting depth of the query. */
15844 bool
15846 {
15848 {
15856 /* Limit the depth of recursion to avoid quadratic behavior.
15857 This is expected to catch almost all occurrences in practice.
15858 If this code misses important cases that unbounded recursion
15859 would not, passes that need this information could be revised
15860 to provide it through dataflow propagation. */
15864 depth));
15868 }
15870 }
15872 /* Return true if the floating point expression T (a GIMPLE_INVALID_RHS)
15873 has an integer value. We also allow +Inf, -Inf and NaN to be
15874 considered integer values. Return false for signaling NaN.
15876 DEPTH is the current nesting depth of the query. */
15878 static bool
15880 {
15882 {
15893 }
15895 }
15897 #undef RECURSE
15898 #undef integer_valued_real_p
15900 /* Return true if the floating point expression T has an integer value.
15901 We also allow +Inf, -Inf and NaN to be considered integer values.
15902 Return false for signaling NaN.
15904 DEPTH is the current nesting depth of the query. */
15906 bool
15908 {
15916 {
15932 }
15935 {
15941 {
15950 }
15954 }
15955 }
15957 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
15958 attempt to fold the expression to a constant without modifying TYPE,
15959 OP0 or OP1.
15961 If the expression could be simplified to a constant, then return
15962 the constant. If the expression would not be simplified to a
15963 constant, then return NULL_TREE. */
15965 tree
15967 {
15970 }
15972 /* Given the components of a unary expression CODE, TYPE and OP0,
15973 attempt to fold the expression to a constant without modifying
15974 TYPE or OP0.
15976 If the expression could be simplified to a constant, then return
15977 the constant. If the expression would not be simplified to a
15978 constant, then return NULL_TREE. */
15980 tree
15982 {
15985 }
15987 /* If EXP represents referencing an element in a constant string
15988 (either via pointer arithmetic or array indexing), return the
15989 tree representing the value accessed, otherwise return NULL. */
15991 tree
15993 {
15997 {
15999 tree index;
16000 tree string;
16005 else
16006 {
16010 /* Optimize the special-case of a zero lower bound.
16012 We convert the low_bound to sizetype to avoid some problems
16013 with constant folding. (E.g. suppose the lower bound is 1,
16014 and its mode is QI. Without the conversion,l (ARRAY
16015 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
16016 +INDEX), which becomes (ARRAY+255+INDEX). Oops!) */
16022 }
16024 scalar_int_mode char_mode;
16036 }
16038 }
16040 /* Folds a read from vector element at IDX of vector ARG. */
16042 tree
16044 {
16049 {
16053 {
16060 }
16061 }
16063 }
16065 /* Return the tree for neg (ARG0) when ARG0 is known to be either
16066 an integer constant, real, or fixed-point constant.
16068 TYPE is the type of the result. */
16070 static tree
16072 {
16076 {
16082 {
16083 FIXED_VALUE_TYPE f;
16088 /* Propagate overflow flags. */
16092 }
16096 {
16103 }
16106 }
16109 }
16111 /* Return the tree for abs (ARG0) when ARG0 is known to be either
16112 an integer constant or real constant.
16114 TYPE is the type of the result. */
16116 tree
16118 {
16122 {
16124 {
16125 /* If the value is unsigned or non-negative, then the absolute value
16126 is the same as the ordinary value. */
16130 ;
16132 /* If the value is negative, then the absolute value is
16133 its negation. */
16134 else
16137 /* Force to the destination type, set TREE_OVERFLOW for signed
16138 TYPE only. */
16140 }
16146 else
16152 }
16155 }
16157 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
16158 constant. TYPE is the type of the result. */
16160 static tree
16162 {
16166 }
16168 /* Given CODE, a relational operator, the target type, TYPE and two
16169 constant operands OP0 and OP1, return the result of the
16170 relational operation. If the result is not a compile time
16171 constant, then return NULL_TREE. */
16173 static tree
16175 {
16178 /* From here on, the only cases we handle are when the result is
16179 known to be a constant. */
16182 {
16186 /* Handle the cases where either operand is a NaN. */
16188 {
16190 {
16218 }
16221 }
16224 }
16227 {
16231 }
16233 /* Handle equality/inequality of complex constants. */
16235 {
16246 else
16248 }
16251 {
16253 {
16254 /* Have vector comparison with scalar boolean result. */
16262 {
16270 }
16272 }
16273 tree_vector_builder elts;
16278 {
16291 }
16294 }
16296 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
16298 To compute GT, swap the arguments and do LT.
16299 To compute GE, do LT and invert the result.
16300 To compute LE, swap the arguments, do LT and invert the result.
16301 To compute NE, do EQ and invert the result.
16303 Therefore, the code below must handle only EQ and LT. */
16306 {
16309 }
16311 /* Note that it is safe to invert for real values here because we
16312 have already handled the one case that it matters. */
16316 {
16319 }
16321 /* Compute a result for LT or EQ if args permit;
16322 Otherwise return T. */
16324 {
16327 else
16329 }
16330 else
16336 }
16338 /* If necessary, return a CLEANUP_POINT_EXPR for EXPR with the
16339 indicated TYPE. If no CLEANUP_POINT_EXPR is necessary, return EXPR
16340 itself. */
16342 tree
16344 {
16345 /* If the expression does not have side effects then we don't have to wrap
16346 it with a cleanup point expression. */
16350 /* If the expression is a return, check to see if the expression inside the
16351 return has no side effects or the right hand side of the modify expression
16352 inside the return. If either don't have side effects set we don't need to
16353 wrap the expression in a cleanup point expression. Note we don't check the
16354 left hand side of the modify because it should always be a return decl. */
16356 {
16363 }
16366 }
16368 /* Given a pointer value OP0 and a type TYPE, return a simplified version
16369 of an indirection through OP0, or NULL_TREE if no simplification is
16370 possible. */
16372 tree
16374 {
16376 tree subtype;
16377 poly_uint64 const_op01;
16386 {
16390 /* *&CONST_DECL -> to the value of the const decl. */
16393 /* *&p => p; make sure to handle *&"str"[cst] here. */
16395 {
16399 else
16401 }
16402 /* *(foo *)&fooarray => fooarray[0] */
16405 && (!in_gimple_form
16407 {
16417 }
16418 /* *(foo *)&complexfoo => __real__ complexfoo */
16422 /* *(foo *)&vectorfoo => BIT_FIELD_REF<vectorfoo,...> */
16425 {
16429 index);
16430 }
16431 }
16435 {
16441 {
16442 tree op00type;
16446 /* ((foo*)&vectorfoo)[1] => BIT_FIELD_REF<vectorfoo,...> */
16449 /* POINTER_PLUS_EXPR second operand is sizetype, unsigned,
16450 but we want to treat offsets with MSB set as negative.
16451 For the code below negative offsets are invalid and
16452 TYPE_SIZE of the element is something unsigned, so
16453 check whether op01 fits into poly_int64, which implies
16454 it is from 0 to INTTYPE_MAXIMUM (HOST_WIDE_INT), and
16455 then just use poly_uint64 because we want to treat the
16456 value as unsigned. */
16458 {
16460 poly_uint64 max_offset
16464 {
16469 }
16470 }
16471 /* ((foo*)&complexfoo)[1] => __imag__ complexfoo */
16474 {
16476 const_op01))
16478 }
16479 /* ((foo *)&fooarray)[1] => fooarray[1] */
16482 {
16491 {
16496 }
16497 }
16498 }
16499 }
16501 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
16504 && (!in_gimple_form
16506 {
16507 tree type_domain;
16517 NULL_TREE);
16518 }
16521 }
16523 /* Builds an expression for an indirection through T, simplifying some
16524 cases. */
16526 tree
16528 {
16536 }
16538 /* Given an INDIRECT_REF T, return either T or a simplified version. */
16540 tree
16542 {
16547 else
16549 }
16551 /* Strip non-trapping, non-side-effecting tree nodes from an expression
16552 whose result is ignored. The type of the returned tree need not be
16553 the same as the original expression. */
16555 tree
16557 {
16563 {
16574 else
16580 {
16596 }
16601 }
16602 }
16604 /* Return the value of VALUE, rounded up to a multiple of DIVISOR. */
16606 tree
16608 {
16614 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16615 have to do anything. Only do this when we are not given a const,
16616 because in that case, this check is more expensive than just
16617 doing it. */
16619 {
16624 }
16626 /* If divisor is a power of two, simplify this to bit manipulation. */
16628 {
16630 {
16644 }
16645 else
16646 {
16647 tree t;
16653 }
16654 }
16655 else
16656 {
16661 }
16664 }
16666 /* Likewise, but round down. */
16668 tree
16670 {
16677 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16678 have to do anything. Only do this when we are not given a const,
16679 because in that case, this check is more expensive than just
16680 doing it. */
16682 {
16687 }
16689 /* If divisor is a power of two, simplify this to bit manipulation. */
16691 {
16692 tree t;
16696 }
16697 else
16698 {
16703 }
16706 }
16708 /* Returns the pointer to the base of the object addressed by EXP and
16709 extracts the information about the offset of the access, storing it
16710 to PBITPOS and POFFSET. */
16712 static tree
16715 {
16716 tree core;
16717 machine_mode mode;
16719 poly_int64 bitsize;
16728 {
16733 }
16735 {
16741 {
16742 poly_offset_int tem
16748 }
16749 }
16750 else
16751 {
16755 }
16758 }
16760 /* Returns true if addresses of E1 and E2 differ by a constant, false
16761 otherwise. If they do, E1 - E2 is stored in *DIFF. */
16763 bool
16765 {
16780 {
16790 }
16792 {
16793 /* If only one of the offsets is non-constant, the difference cannot
16794 be a constant. */
16796 }
16797 else
16802 }
16804 /* Return OFF converted to a pointer offset type suitable as offset for
16805 POINTER_PLUS_EXPR. Use location LOC for this conversion. */
16806 tree
16808 {
16812 }
16814 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16815 tree
16817 {
16820 }
16822 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16823 tree
16825 {
16828 }
16830 /* Return a pointer to a NUL-terminated string containing the sequence
16831 of bytes corresponding to the representation of the object referred to
16832 by SRC (or a subsequence of such bytes within it if SRC is a reference
16833 to an initialized constant array plus some constant offset).
16834 Set *STRSIZE the number of bytes in the constant sequence including
16835 the terminating NUL byte. *STRSIZE is equal to sizeof(A) - OFFSET
16836 where A is the array that stores the constant sequence that SRC points
16837 to and OFFSET is the byte offset of SRC from the beginning of A. SRC
16838 need not point to a string or even an array of characters but may point
16839 to an object of any type. */
16843 {
16844 /* The offset into the array A storing the string, and A's byte size. */
16845 tree offset_node;
16846 tree mem_size;
16853 else
16860 {
16863 else
16865 }
16870 /* ARRAY_SIZE is the byte size of the array the constant sequence
16871 is stored in and equal to sizeof A. INIT_BYTES is the number
16872 of bytes in the constant sequence used to initialize the array,
16873 including any embedded NULs as well as the terminating NUL (for
16874 strings), but not including any trailing zeros/NULs past
16875 the terminating one appended implicitly to a string literal to
16876 zero out the remainder of the array it's stored in. For example,
16877 given:
16878 const char a[7] = "abc\0d";
16879 n = strlen (a + 1);
16880 ARRAY_SIZE is 7, INIT_BYTES is 6, and OFFSET is 1. For a valid
16881 (i.e., nul-terminated) string with no embedded nuls, INIT_BYTES
16882 is equal to strlen (A) + 1. */
16887 /* Ideally this would turn into a gcc_checking_assert over time. */
16895 {
16896 /* Compute and store the number of characters from the beginning
16897 of the substring at OFFSET to the end, including the terminating
16898 nul. Offsets past the initial length refer to null strings. */
16901 else
16903 }
16904 else
16905 {
16907 /* Support only properly NUL-terminated single byte strings. */
16912 }
16915 }
16917 /* Return a pointer to a NUL-terminated string corresponding to
16918 the expression STR referencing a constant string, possibly
16919 involving a constant offset. Return null if STR either doesn't
16920 reference a constant string or if it involves a nonconstant
16921 offset. */
16925 {
16927 }
16929 /* Given a tree T, compute which bits in T may be nonzero. */
16931 wide_int
16933 {
16935 {
16953 CASE_CONVERT:
16959 {
16964 }
16968 {
16976 }
16980 {
16988 }
16992 }
16995 }
16997 /* Helper function for address compare simplifications in match.pd.
16998 OP0 and OP1 are ADDR_EXPR operands being compared by CODE.
16999 TYPE is the type of comparison operands.
17000 BASE0, BASE1, OFF0 and OFF1 are set by the function.
17001 GENERIC is true if GENERIC folding and false for GIMPLE folding.
17002 Returns 0 if OP0 is known to be unequal to OP1 regardless of OFF{0,1},
17003 1 if bases are known to be equal and OP0 cmp OP1 depends on OFF0 cmp OFF1,
17004 and 2 if unknown. */
17006 int
17010 {
17020 {
17023 }
17025 {
17028 }
17033 /* Punt in GENERIC on variables with value expressions;
17034 the value expressions might point to fields/elements
17035 of other vars etc. */
17041 {
17045 }
17053 /* Assume different STRING_CSTs with the same content will be
17054 merged. */
17063 {
17066 /* If the offsets are equal we can ignore overflow. */
17069 /* Or if we compare using pointers to decls or strings. */
17074 }
17080 /* At this point we know (or assume) the two pointers point at
17081 different objects. */
17085 /* Punt on non-zero offsets from functions. */
17089 /* Or if the bases are neither decls nor string literals. */
17094 /* For initializers, assume addresses of different functions are
17095 different. */
17101 /* Compute whether one address points to the start of one
17102 object and another one to the end of another one. */
17105 {
17108 else
17110 }
17113 else
17114 {
17118 else
17120 }
17122 {
17125 else
17127 }
17130 else
17131 {
17135 else
17137 }
17139 {
17140 /* If one offset is pointing (or could be) to the beginning of one
17141 object and the other is pointing to one past the last byte of the
17142 other object, punt. */
17147 /* If both offsets are the same, there are some cases we know that are
17148 ok. Either if we know they aren't zero, or if we know both sizes
17149 are no zero. */
17155 }
17157 /* At this point, equal is 2 if either one or both pointers are out of
17158 bounds of their object, or one points to start of its object and the
17159 other points to end of its object. This is unspecified behavior
17160 e.g. in C++. Otherwise equal is 0. */
17164 /* When both pointers point to string literals, even when equal is 0,
17165 due to tail merging of string literals the pointers might be the same. */
17167 {
17174 /* If the bytes in the string literals starting at the pointers
17175 differ, the pointers need to be different. */
17180 {
17185 /* If even the bytes in the string literal before the
17186 pointers are the same, the string literals could be
17187 tail merged. */
17189 }
17191 }
17196 /* If this is a pointer comparison, ignore for now even
17197 valid equalities where one pointer is the offset zero
17198 of one object and the other to one past end of another one. */
17202 /* Assume that string literals can't be adjacent to variables
17203 (automatic or global). */
17207 /* Assume that automatic variables can't be adjacent to global
17208 variables. */
17213 }
17215 /* Return the single non-zero element of a CONSTRUCTOR or NULL_TREE. */
17216 tree
17218 {
17227 {
17231 }
17233 }
17235 #if CHECKING_P
17239 /* Helper functions for writing tests of folding trees. */
17241 /* Verify that the binary op (LHS CODE RHS) folds to CONSTANT. */
17243 static void
17245 tree constant)
17246 {
17248 }
17250 /* Verify that the binary op (LHS CODE RHS) folds to an NON_LVALUE_EXPR
17251 wrapping WRAPPED_EXPR. */
17253 static void
17255 tree wrapped_expr)
17256 {
17261 }
17263 /* Verify that various arithmetic binary operations are folded
17264 correctly. */
17266 static void
17268 {
17274 /* Addition. */
17275 /* 1 <-- (0 + 1) */
17277 one);
17279 one);
17281 /* (nonlvalue)x <-- (x + 0) */
17283 x);
17285 /* Subtraction. */
17286 /* 0 <-- (x - x) */
17288 zero);
17290 x);
17292 /* Multiplication. */
17293 /* 0 <-- (x * 0) */
17295 zero);
17297 /* (nonlvalue)x <-- (x * 1) */
17299 x);
17300 }
17304 /* Build a VECTOR_CST corresponding to VMODE, and has
17305 encoding given by NPATTERNS, NELTS_PER_PATTERN and STEP.
17306 Fill it with randomized elements, using rand() % THRESHOLD. */
17308 static tree
17313 {
17318 // Fill a0 for each pattern
17325 // Fill a1 for each pattern
17327 {
17328 tree a1;
17330 {
17335 }
17336 else
17339 }
17344 {
17349 }
17352 }
17354 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17355 when result is VLA. */
17357 static void
17360 {
17361 /* Actual npatterns and encoded_elts in res may be less than expected due
17362 to canonicalization. */
17369 }
17371 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17372 when the result is VLS. */
17374 static void
17376 {
17380 }
17382 /* Helper routine to push multiple elements into BUILDER. */
17386 {
17389 }
17391 #define ARG0(index) vector_cst_elt (arg0, index)
17392 #define ARG1(index) vector_cst_elt (arg1, index)
17394 /* Test cases where result is VNx4SI and input vectors are V4SI. */
17396 static void
17398 {
17400 {
17401 /* Case 1:
17402 sel = { 0, 4, 1, 5, ... }
17403 res = { arg[0], arg1[0], arg0[1], arg1[1], ...} // (4, 1) */
17404 {
17408 tree inner_type
17422 }
17424 /* Case 2: Same as case 1, but contains an out of bounds access which
17425 should wrap around.
17426 sel = {0, 8, 4, 12, ...} (4, 1)
17427 res = { arg0[0], arg0[0], arg1[0], arg1[0], ... } (4, 1). */
17428 {
17432 tree inner_type
17446 }
17447 }
17448 }
17450 /* Test cases where result is V4SI and input vectors are VNx4SI. */
17452 static void
17454 {
17456 {
17457 /* Case 1:
17458 sel = { 0, 1, 2, 3}
17459 res = { arg0[0], arg0[1], arg0[2], arg0[3] }. */
17460 {
17464 tree inner_type
17478 }
17480 /* Case 2: Same as Case 1, but crossing input vector.
17481 sel = {0, 2, 4, 6}
17482 In this case,the index 4 is ambiguous since len = 4 + 4x.
17483 Since we cannot determine, which vector to choose from during
17484 compile time, should return NULL_TREE. */
17485 {
17489 tree inner_type
17504 }
17505 }
17506 }
17508 /* Test all input vectors. */
17510 static void
17512 {
17513 /* Test with 10 different inputs. */
17515 {
17520 /* Case 1: mask = {0, ...} // (1, 1)
17521 res = { arg0[0], ... } // (1, 1) */
17522 {
17529 }
17531 /* Case 2: mask = {len, ...} // (1, 1)
17532 res = { arg1[0], ... } // (1, 1) */
17533 {
17541 }
17542 }
17543 }
17545 /* Test all vectors which contain at-least 2 elements. */
17547 static void
17549 {
17551 {
17552 /* Case 1: mask = { 0, len, ... } // (2, 1)
17553 res = { arg0[0], arg1[0], ... } // (2, 1) */
17554 {
17568 }
17570 /* Case 2: mask = { 0, len, 1, len+1, ... } // (2, 2)
17571 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2) */
17572 {
17586 }
17588 /* Case 4: mask = {0, 0, 1, ...} // (1, 3)
17589 Test that the stepped sequence of the pattern selects from
17590 same input pattern. Since input vectors have npatterns = 2,
17591 and step (a2 - a1) = 1, step is not a multiple of npatterns
17592 in input vector. So return NULL_TREE. */
17593 {
17608 }
17610 /* Case 5: mask = {len, 0, 1, ...} // (1, 3)
17611 Test that stepped sequence of the pattern selects from arg0.
17612 res = { arg1[0], arg0[0], arg0[1], ... } // (1, 3) */
17613 {
17627 }
17629 /* Case 6: PR111648 - a1 chooses base element from input vector arg.
17630 In this case ensure that arg has a natural stepped sequence
17631 to preserve arg's encoding.
17633 As a concrete example, consider:
17634 arg0: { -16, -9, -10, ... } // (1, 3)
17635 arg1: { -12, -5, -6, ... } // (1, 3)
17636 sel = { 0, len, len + 1, ... } // (1, 3)
17638 This will create res with following encoding:
17639 res = { arg0[0], arg1[0], arg1[1], ... } // (1, 3)
17640 = { -16, -12, -5, ... }
17642 The step in above encoding would be: (-5) - (-12) = 7
17643 And hence res[3] would be computed as -5 + 7 = 2.
17644 instead of arg1[2], ie, -6.
17645 Ensure that valid_mask_for_fold_vec_perm_cst returns false
17646 for this case. */
17647 {
17658 /* FIXME: It may happen that build_vec_cst_rand may build a natural
17659 stepped pattern, even if we didn't explicitly tell it to. So folding
17660 may not always fail, but if it does, ensure that's because arg1 does
17661 not have a natural stepped sequence (and not due to other reason) */
17665 }
17667 /* Case 7: Same as Case 6, except that arg1 contains natural stepped
17668 sequence and thus folding should be valid for this case. */
17669 {
17683 }
17685 /* Case 8: Same as aarch64/sve/slp_3.c:
17686 arg0, arg1 are dup vectors.
17687 sel = { 0, len, 1, len+1, 2, len+2, ... } // (2, 3)
17688 So res = { arg0[0], arg1[0], ... } // (2, 1)
17690 In this case, since the input vectors are dup, only the first two
17691 elements per pattern in sel are considered significant. */
17692 {
17706 }
17707 }
17708 }
17710 /* Test all vectors which contain at-least 4 elements. */
17712 static void
17714 {
17716 {
17717 /* Case 1: mask = { 0, len, 1, len+1, ... } // (4, 1)
17718 res: { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (4, 1) */
17719 {
17733 }
17735 /* Case 2: sel = {0, 1, 2, ...} // (1, 3)
17736 res: { arg0[0], arg0[1], arg0[2], ... } // (1, 3) */
17737 {
17750 }
17752 /* Case 3: sel = {len, len+1, len+2, ...} // (1, 3)
17753 res: { arg1[0], arg1[1], arg1[2], ... } // (1, 3) */
17754 {
17767 }
17769 /* Case 4:
17770 sel = { len, 0, 2, ... } // (1, 3)
17771 This should return NULL because we cross the input vectors.
17772 Because,
17773 Let's assume len = C + Cx
17774 a1 = 0
17775 S = 2
17776 esel = arg0_len / sel_npatterns = C + Cx
17777 ae = 0 + (esel - 2) * S
17778 = 0 + (C + Cx - 2) * 2
17779 = 2(C-2) + 2Cx
17781 For C >= 4:
17782 Let q1 = a1 / arg0_len = 0 / (C + Cx) = 0
17783 Let qe = ae / arg0_len = (2(C-2) + 2Cx) / (C + Cx) = 1
17784 Since q1 != qe, we cross input vectors.
17785 So return NULL_TREE. */
17786 {
17800 }
17802 /* Case 5: npatterns(arg0) = 4 > npatterns(sel) = 2
17803 mask = { 0, len, 1, len + 1, ...} // (2, 2)
17804 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2)
17806 Note that fold_vec_perm_cst will set
17807 res_npatterns = max(4, max(4, 2)) = 4
17808 However after canonicalizing, we will end up with shape (2, 2). */
17809 {
17822 }
17824 /* Case 6: Test combination in sel, where one pattern is dup and other
17825 is stepped sequence.
17826 sel = { 0, 0, 0, 1, 0, 2, ... } // (2, 3)
17827 res = { arg0[0], arg0[0], arg0[0],
17828 arg0[1], arg0[0], arg0[2], ... } // (2, 3) */
17829 {
17844 }
17846 /* Case 7: PR111048: Check that we set arg_npatterns correctly,
17847 when arg0, arg1 and sel have different number of patterns.
17848 arg0 is of shape (1, 1)
17849 arg1 is of shape (4, 1)
17850 sel is of shape (2, 3) = {1, len, 2, len+1, 3, len+2, ...}
17852 In this case the pattern: {len, len+1, len+2, ...} chooses arg1.
17853 However,
17854 step = (len+2) - (len+1) = 1
17855 arg_npatterns = VECTOR_CST_NPATTERNS (arg1) = 4
17856 Since step is not a multiple of arg_npatterns,
17857 valid_mask_for_fold_vec_perm_cst should return false,
17858 and thus fold_vec_perm_cst should return NULL_TREE. */
17859 {
17874 }
17876 /* Case 8: PR111754: When input vector is not a stepped sequence,
17877 check that the result is not a stepped sequence either, even
17878 if sel has a stepped sequence. */
17879 {
17892 }
17894 /* Case 9: If sel doesn't contain a stepped sequence,
17895 check that the result has same encoding as sel, irrespective
17896 of shape of input vectors. */
17897 {
17912 }
17913 }
17914 }
17916 /* Test all vectors which contain at-least 8 elements. */
17918 static void
17920 {
17922 {
17923 /* Case 1: sel_npatterns (4) > input npatterns (2)
17924 sel: { 0, 0, 1, len, 2, 0, 3, len, 4, 0, 5, len, ...} // (4, 3)
17925 res: { arg0[0], arg0[0], arg0[0], arg1[0],
17926 arg0[2], arg0[0], arg0[3], arg1[0],
17927 arg0[4], arg0[0], arg0[5], arg1[0], ... } // (4, 3) */
17928 {
17945 }
17946 }
17947 }
17949 /* Test vectors for which nunits[0] <= 4. */
17951 static void
17953 {
17954 /* Case 1: mask = {0, 4, ...} // (1, 2)
17955 This should return NULL_TREE because the index 4 may choose
17956 from either arg0 or arg1 depending on vector length. */
17957 {
17972 }
17973 }
17975 #undef ARG0
17976 #undef ARG1
17978 /* Return true if SIZE is of the form C + Cx and C is power of 2. */
17980 static bool
17982 {
17990 }
17992 /* Execute fold_vec_perm_cst unit tests. */
17994 static void
17996 {
18000 machine_mode vmode;
18002 {
18003 /* Obtain modes corresponding to VNx4SI and V4SI,
18004 to call mixed mode tests below.
18005 FIXME: Is there a better way to do this ? */
18007 {
18014 }
18031 }
18036 {
18039 }
18040 }
18043 /* Verify that various binary operations on vectors are folded
18044 correctly. */
18046 static void
18048 {
18055 /* Verify equality tests that return a scalar boolean result. */
18068 }
18070 /* Verify folding of VEC_DUPLICATE_EXPRs. */
18072 static void
18074 {
18077 /* This will be 1 if VEC_MODE isn't a vector mode. */
18084 }
18086 /* Run all of the selftests within this file. */
18088 void
18090 {
18095 }