| blob | 4f8561509ffb2abf1ace960544151790f25bf9f8 |
1 /* Fold a constant sub-tree into a single node for C-compiler
2 Copyright (C) 1987-2023 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 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1347 constant. We assume ARG1 and ARG2 have the same data type, or at least
1348 are the same kind of constant and the same machine mode. Return zero if
1349 combining the constants is not allowed in the current operating mode. */
1351 static tree
1353 {
1354 /* Sanity check for the recursive cases. */
1362 {
1368 }
1371 {
1372 machine_mode mode;
1373 REAL_VALUE_TYPE d1;
1374 REAL_VALUE_TYPE d2;
1375 REAL_VALUE_TYPE value;
1376 REAL_VALUE_TYPE result;
1380 /* The following codes are handled by real_arithmetic. */
1382 {
1393 }
1401 /* Don't perform operation if we honor signaling NaNs and
1402 either operand is a signaling NaN. */
1408 /* Don't perform operation if it would raise a division
1409 by zero exception. */
1415 /* If either operand is a NaN, just return it. Otherwise, set up
1416 for floating-point trap; we return an overflow. */
1418 {
1419 /* Make resulting NaN value to be qNaN when flag_signaling_nans
1420 is off. */
1424 }
1426 {
1427 /* Make resulting NaN value to be qNaN when flag_signaling_nans
1428 is off. */
1432 }
1437 /* Don't constant fold this floating point operation if
1438 both operands are not NaN but the result is NaN, and
1439 flag_trapping_math. Such operations should raise an
1440 invalid operation exception. */
1448 /* Don't constant fold this floating point operation if
1449 the result has overflowed and flag_trapping_math. */
1457 /* Don't constant fold this floating point operation if the
1458 result may dependent upon the run-time rounding mode and
1459 flag_rounding_math is set, or if GCC's software emulation
1460 is unable to accurately represent the result. */
1470 }
1473 {
1474 FIXED_VALUE_TYPE f1;
1475 FIXED_VALUE_TYPE f2;
1476 FIXED_VALUE_TYPE result;
1481 /* The following codes are handled by fixed_arithmetic. */
1483 {
1495 {
1502 }
1507 }
1514 /* Propagate overflow flags. */
1518 }
1521 {
1530 {
1541 mpc_mul);
1555 mpc_div);
1556 /* Fallthru. */
1562 {
1563 /* Keep this algorithm in sync with
1564 tree-complex.cc:expand_complex_div_straight().
1566 Expand complex division to scalars, straightforward algorithm.
1567 a / b = ((ar*br + ai*bi)/t) + i((ai*br - ar*bi)/t)
1568 t = br*br + bi*bi
1569 */
1570 tree magsquared
1574 tree t1
1578 tree t2
1585 }
1586 else
1587 {
1588 /* Keep this algorithm in sync with
1589 tree-complex.cc:expand_complex_div_wide().
1591 Expand complex division to scalars, modified algorithm to minimize
1592 overflow with wide input ranges. */
1598 {
1599 /* In the TRUE branch, we compute
1600 ratio = br/bi;
1601 div = (br * ratio) + bi;
1602 tr = (ar * ratio) + ai;
1603 ti = (ai * ratio) - ar;
1604 tr = tr / div;
1605 ti = ti / div; */
1616 }
1617 else
1618 {
1619 /* In the FALSE branch, we compute
1620 ratio = d/c;
1621 divisor = (d * ratio) + c;
1622 tr = (b * ratio) + a;
1623 ti = b - (a * ratio);
1624 tr = tr / div;
1625 ti = ti / div; */
1637 }
1638 }
1643 }
1647 }
1653 {
1658 /* We can operate directly on the encoding if:
1660 a3 - a2 == a2 - a1 && b3 - b2 == b2 - b1
1661 implies
1662 (a3 op b3) - (a2 op b2) == (a2 op b2) - (a1 op b1)
1664 Addition and subtraction are the supported operators
1665 for which this is true. */
1668 /* We can operate directly on stepped encodings if:
1670 a3 - a2 == a2 - a1
1671 implies:
1672 (a3 op c) - (a2 op c) == (a2 op c) - (a1 op c)
1674 which is true if (x -> x op c) distributes over addition. */
1676 else
1677 /* Similarly in reverse. */
1679 tree_vector_builder elts;
1684 {
1690 /* It is possible that const_binop cannot handle the given
1691 code and return NULL_TREE */
1695 }
1698 }
1700 /* Shifts allow a scalar offset for a vector. */
1703 {
1706 tree_vector_builder elts;
1711 {
1716 /* It is possible that const_binop cannot handle the given
1717 code and return NULL_TREE. */
1721 }
1724 }
1726 }
1728 /* Overload that adds a TYPE parameter to be able to dispatch
1729 to fold_relational_const. */
1731 tree
1733 {
1737 /* ??? Until we make the const_binop worker take the type of the
1738 result as argument put those cases that need it here. */
1740 {
1757 {
1762 }
1768 {
1784 {
1789 ? NOP_EXPR
1796 }
1799 }
1805 {
1828 {
1841 }
1844 }
1847 }
1852 /* Make sure type and arg0 have the same saturating flag. */
1857 }
1859 /* Compute CODE ARG1 with resulting type TYPE with ARG1 being constant.
1860 Return zero if computing the constants is not possible. */
1862 tree
1864 {
1865 /* Don't perform the operation, other than NEGATE and ABS, if
1866 flag_signaling_nans is on and the operand is a signaling NaN. */
1876 {
1877 CASE_CONVERT:
1884 /* If the source address is 0, and the source address space
1885 cannot have a valid object at 0, fold to dest type null. */
1896 {
1897 /* Can't call fold_negate_const directly here as that doesn't
1898 handle all cases and we might not be able to negate some
1899 constants. */
1904 }
1914 {
1918 }
1926 /* Perform BIT_NOT_EXPR on each element individually. */
1928 {
1929 tree elem;
1931 /* This can cope with stepped encodings because ~x == -1 - x. */
1932 tree_vector_builder elements;
1936 {
1942 }
1945 }
1969 {
1992 else
1997 {
2003 }
2006 }
2015 }
2018 }
2020 /* Create a sizetype INT_CST node with NUMBER sign extended. KIND
2021 indicates which particular sizetype to create. */
2023 tree
2025 {
2027 }
2028 \f
2029 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
2030 is a tree code. The type of the result is taken from the operands.
2031 Both must be equivalent integer types, ala int_binop_types_match_p.
2032 If the operands are constant, so is the result. */
2034 tree
2036 {
2045 /* Handle the special case of two poly_int constants faster. */
2047 {
2048 /* And some specific cases even faster than that. */
2050 {
2057 }
2059 {
2063 }
2065 {
2069 }
2071 /* Handle general case of two integer constants. For sizetype
2072 constant calculations we always want to know about overflow,
2073 even in the unsigned case. */
2077 }
2080 }
2082 /* Given two values, either both of sizetype or both of bitsizetype,
2083 compute the difference between the two values. Return the value
2084 in signed type corresponding to the type of the operands. */
2086 tree
2088 {
2090 tree ctype;
2095 /* If the type is already signed, just do the simple thing. */
2103 else
2106 /* If either operand is not a constant, do the conversions to the signed
2107 type and subtract. The hardware will do the right thing with any
2108 overflow in the subtraction. */
2114 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
2115 Otherwise, subtract the other way, convert to CTYPE (we know that can't
2116 overflow) and negate (which can't either). Special-case a result
2117 of zero while we're here. */
2123 else
2127 MINUS_EXPR,
2129 }
2130 \f
2131 /* A subroutine of fold_convert_const handling conversions of an
2132 INTEGER_CST to another integer type. */
2134 static tree
2136 {
2137 /* Given an integer constant, make new constant with new type,
2138 appropriately sign-extended or truncated. Use widest_int
2139 so that any extension is done according ARG1's type. */
2143 }
2145 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2146 to an integer type. */
2148 static tree
2150 {
2152 tree t;
2154 /* The following code implements the floating point to integer
2155 conversion rules required by the Java Language Specification,
2156 that IEEE NaNs are mapped to zero and values that overflow
2157 the target precision saturate, i.e. values greater than
2158 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
2159 are mapped to INT_MIN. These semantics are allowed by the
2160 C and C++ standards that simply state that the behavior of
2161 FP-to-integer conversion is unspecified upon overflow. */
2163 wide_int val;
2164 REAL_VALUE_TYPE r;
2168 {
2175 }
2177 /* If R is NaN, return zero and show we have an overflow. */
2179 {
2182 }
2184 /* See if R is less than the lower bound or greater than the
2185 upper bound. */
2188 {
2192 {
2195 }
2196 }
2199 {
2202 {
2205 {
2208 }
2209 }
2210 }
2217 }
2219 /* A subroutine of fold_convert_const handling conversions of a
2220 FIXED_CST to an integer type. */
2222 static tree
2224 {
2225 tree t;
2227 scalar_mode mode;
2229 /* Right shift FIXED_CST to temp by fbit. */
2233 {
2235 HOST_BITS_PER_DOUBLE_INT,
2238 /* Left shift temp to temp_trunc by fbit. */
2240 HOST_BITS_PER_DOUBLE_INT,
2242 }
2243 else
2244 {
2247 }
2249 /* If FIXED_CST is negative, we need to round the value toward 0.
2250 By checking if the fractional bits are not zero to add 1 to temp. */
2256 /* Given a fixed-point constant, make new constant with new type,
2257 appropriately sign-extended or truncated. */
2265 }
2267 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2268 to another floating point type. */
2270 static tree
2272 {
2273 REAL_VALUE_TYPE value;
2274 tree t;
2276 /* If the underlying modes are the same, simply treat it as
2277 copy and rebuild with TREE_REAL_CST information and the
2278 given type. */
2280 {
2283 }
2285 /* Don't perform the operation if flag_signaling_nans is on
2286 and the operand is a signaling NaN. */
2291 /* With flag_rounding_math we should respect the current rounding mode
2292 unless the conversion is exact. */
2300 /* If converting an infinity or NAN to a representation that doesn't
2301 have one, set the overflow bit so that we can produce some kind of
2302 error message at the appropriate point if necessary. It's not the
2303 most user-friendly message, but it's better than nothing. */
2310 /* Regular overflow, conversion produced an infinity in a mode that
2311 can't represent them. */
2316 else
2319 }
2321 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2322 to a floating point type. */
2324 static tree
2326 {
2327 REAL_VALUE_TYPE value;
2328 tree t;
2336 }
2338 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2339 to another fixed-point type. */
2341 static tree
2343 {
2344 FIXED_VALUE_TYPE value;
2345 tree t;
2352 /* Propagate overflow flags. */
2356 }
2358 /* A subroutine of fold_convert_const handling conversions an INTEGER_CST
2359 to a fixed-point type. */
2361 static tree
2363 {
2364 FIXED_VALUE_TYPE value;
2365 tree t;
2367 double_int di;
2374 else
2382 /* Propagate overflow flags. */
2386 }
2388 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2389 to a fixed-point type. */
2391 static tree
2393 {
2394 FIXED_VALUE_TYPE value;
2395 tree t;
2403 /* Propagate overflow flags. */
2407 }
2409 /* Attempt to fold type conversion operation CODE of expression ARG1 to
2410 type TYPE. If no simplification can be done return NULL_TREE. */
2412 static tree
2414 {
2419 /* We can't widen types, since the runtime value could overflow the
2420 original type before being extended to the new type. */
2431 {
2438 }
2440 {
2442 {
2444 /* Avoid the folding if flag_rounding_math is on and the
2445 conversion is not exact. */
2447 {
2453 }
2455 }
2460 }
2462 {
2469 }
2471 {
2474 {
2477 /* We can't handle steps directly when extending, since the
2478 values need to wrap at the original precision first. */
2479 bool step_ok_p
2483 tree_vector_builder v;
2488 {
2494 }
2496 }
2497 }
2499 }
2501 /* Construct a vector of zero elements of vector type TYPE. */
2503 static tree
2505 {
2506 tree t;
2510 }
2512 /* Returns true, if ARG is convertible to TYPE using a NOP_EXPR. */
2514 bool
2516 {
2531 {
2553 }
2554 }
2556 /* Convert expression ARG to type TYPE. Used by the middle-end for
2557 simple conversions in preference to calling the front-end's convert. */
2559 tree
2561 {
2563 tree tem;
2574 {
2577 /* Handle conversions between pointers to different address spaces. */
2582 /* fall through */
2587 {
2591 }
2605 {
2609 }
2611 {
2615 }
2617 {
2621 }
2624 {
2642 }
2647 {
2651 }
2654 {
2669 }
2673 {
2682 integer_zero_node));
2684 {
2688 {
2694 }
2702 }
2706 }
2724 }
2725 fold_convert_exit:
2728 }
2729 \f
2730 /* Return false if expr can be assumed not to be an lvalue, true
2731 otherwise. */
2733 static bool
2735 {
2736 /* We only need to wrap lvalue tree codes. */
2738 {
2771 /* Assume the worst for front-end tree codes. */
2775 }
2778 }
2780 /* Return an expr equal to X but certainly not valid as an lvalue. */
2782 tree
2784 {
2785 /* While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to
2786 us. */
2793 }
2794 \f
2795 /* Given a tree comparison code, return the code that is the logical inverse.
2796 It is generally not safe to do this for floating-point comparisons, except
2797 for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
2798 ERROR_MARK in this case. */
2800 enum tree_code
2802 {
2808 {
2839 }
2840 }
2842 /* Similar, but return the comparison that results if the operands are
2843 swapped. This is safe for floating-point. */
2845 enum tree_code
2847 {
2849 {
2875 }
2876 }
2879 /* Convert a comparison tree code from an enum tree_code representation
2880 into a compcode bit-based encoding. This function is the inverse of
2881 compcode_to_comparison. */
2883 static enum comparison_code
2885 {
2887 {
2918 }
2919 }
2921 /* Convert a compcode bit-based encoding of a comparison operator back
2922 to GCC's enum tree_code representation. This function is the
2923 inverse of comparison_to_compcode. */
2925 static enum tree_code
2927 {
2929 {
2960 }
2961 }
2963 /* Return true if COND1 tests the opposite condition of COND2. */
2965 bool
2967 {
2970 && (invert_tree_comparison
2977 }
2979 /* Return a tree for the comparison which is the combination of
2980 doing the AND or OR (depending on CODE) of the two operations LCODE
2981 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2982 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2983 if this makes the transformation invalid. */
2985 tree
2990 {
2997 {
3008 }
3011 {
3012 /* Eliminate unordered comparisons, as well as LTGT and ORD
3013 which are not used unless the mode has NaNs. */
3019 }
3021 {
3022 /* Check that the original operation and the optimized ones will trap
3023 under the same condition. */
3034 /* In a short-circuited boolean expression the LHS might be
3035 such that the RHS, if evaluated, will never trap. For
3036 example, in ORD (x, y) && (x < y), we evaluate the RHS only
3037 if neither x nor y is NaN. (This is a mixed blessing: for
3038 example, the expression above will never trap, hence
3039 optimizing it to x < y would be invalid). */
3044 /* If the comparison was short-circuited, and only the RHS
3045 trapped, we may now generate a spurious trap. */
3050 /* If we changed the conditions that cause a trap, we lose. */
3053 }
3059 else
3060 {
3065 }
3066 }
3067 \f
3068 /* Return nonzero if two operands (typically of the same tree node)
3069 are necessarily equal. FLAGS modifies behavior as follows:
3071 If OEP_ONLY_CONST is set, only return nonzero for constants.
3072 This function tests whether the operands are indistinguishable;
3073 it does not test whether they are equal using C's == operation.
3074 The distinction is important for IEEE floating point, because
3075 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
3076 (2) two NaNs may be indistinguishable, but NaN!=NaN.
3078 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
3079 even though it may hold multiple values during a function.
3080 This is because a GCC tree node guarantees that nothing else is
3081 executed between the evaluation of its "operands" (which may often
3082 be evaluated in arbitrary order). Hence if the operands themselves
3083 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
3084 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
3085 unset means assuming isochronic (or instantaneous) tree equivalence.
3086 Unless comparing arbitrary expression trees, such as from different
3087 statements, this flag can usually be left unset.
3089 If OEP_PURE_SAME is set, then pure functions with identical arguments
3090 are considered the same. It is used when the caller has other ways
3091 to ensure that global memory is unchanged in between.
3093 If OEP_ADDRESS_OF is set, we are actually comparing addresses of objects,
3094 not values of expressions.
3096 If OEP_LEXICOGRAPHIC is set, then also handle expressions with side-effects
3097 such as MODIFY_EXPR, RETURN_EXPR, as well as STATEMENT_LISTs.
3099 If OEP_BITWISE is set, then require the values to be bitwise identical
3100 rather than simply numerically equal. Do not take advantage of things
3101 like math-related flags or undefined behavior; only return true for
3102 values that are provably bitwise identical in all circumstances.
3104 Unless OEP_MATCH_SIDE_EFFECTS is set, the function returns false on
3105 any operand with side effect. This is unnecesarily conservative in the
3106 case we know that arg0 and arg1 are in disjoint code paths (such as in
3107 ?: operator). In addition OEP_MATCH_SIDE_EFFECTS is used when comparing
3108 addresses with TREE_CONSTANT flag set so we know that &var == &var
3109 even if var is volatile. */
3111 bool
3114 {
3122 /* If either is ERROR_MARK, they aren't equal. */
3128 /* Similar, if either does not have a type (like a template id),
3129 they aren't equal. */
3133 /* Bitwise identity makes no sense if the values have different layouts. */
3138 /* We cannot consider pointers to different address space equal. */
3145 /* Check equality of integer constants before bailing out due to
3146 precision differences. */
3148 {
3149 /* Address of INTEGER_CST is not defined; check that we did not forget
3150 to drop the OEP_ADDRESS_OF flags. */
3153 }
3156 {
3157 /* If both types don't have the same signedness, then we can't consider
3158 them equal. We must check this before the STRIP_NOPS calls
3159 because they may change the signedness of the arguments. As pointers
3160 strictly don't have a signedness, require either two pointers or
3161 two non-pointers as well. */
3167 /* If both types don't have the same precision, then it is not safe
3168 to strip NOPs. */
3175 }
3176 #if 0
3177 /* FIXME: Fortran FE currently produce ADDR_EXPR of NOP_EXPR. Enable the
3178 sanity check once the issue is solved. */
3179 else
3180 /* Addresses of conversions and SSA_NAMEs (and many other things)
3181 are not defined. Check that we did not forget to drop the
3182 OEP_ADDRESS_OF/OEP_CONSTANT_ADDRESS_OF flags. */
3185 #endif
3187 /* In case both args are comparisons but with different comparison
3188 code, try to swap the comparison operands of one arg to produce
3189 a match and compare that variant. */
3193 {
3201 }
3204 {
3205 /* NOP_EXPR and CONVERT_EXPR are considered equal. */
3207 ;
3209 {
3210 /* If we are interested in comparing addresses ignore
3211 MEM_REF wrappings of the base that can appear just for
3212 TBAA reasons. */
3226 }
3227 else
3229 }
3231 /* When not checking adddresses, this is needed for conversions and for
3232 COMPONENT_REF. Might as well play it safe and always test this. */
3239 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
3240 We don't care about side effects in that case because the SAVE_EXPR
3241 takes care of that for us. In all other cases, two expressions are
3242 equal if they have no side effects. If we have two identical
3243 expressions with side effects that should be treated the same due
3244 to the only side effects being identical SAVE_EXPR's, that will
3245 be detected in the recursive calls below.
3246 If we are taking an invariant address of two identical objects
3247 they are necessarily equal as well. */
3254 /* Next handle constant cases, those for which we can return 1 even
3255 if ONLY_CONST is set. */
3258 {
3271 {
3272 /* If we do not distinguish between signed and unsigned zero,
3273 consider them equal. */
3276 }
3280 {
3295 }
3299 flags)
3301 flags));
3312 flags | OEP_ADDRESS_OF
3315 /* In GIMPLE empty constructors are allowed in initializers of
3316 aggregates. */
3320 }
3322 /* Don't handle more cases for OEP_BITWISE, since we can't guarantee that
3323 two instances of undefined behavior will give identical results. */
3327 /* Define macros to test an operand from arg0 and arg1 for equality and a
3328 variant that allows null and views null as being different from any
3329 non-null value. In the latter case, if either is null, the both
3330 must be; otherwise, do the normal comparison. */
3331 #define OP_SAME(N) operand_equal_p (TREE_OPERAND (arg0, N), \
3332 TREE_OPERAND (arg1, N), flags)
3334 #define OP_SAME_WITH_NULL(N) \
3335 ((!TREE_OPERAND (arg0, N) || !TREE_OPERAND (arg1, N)) \
3336 ? TREE_OPERAND (arg0, N) == TREE_OPERAND (arg1, N) : OP_SAME (N))
3339 {
3341 /* Two conversions are equal only if signedness and modes match. */
3343 {
3344 CASE_CONVERT:
3352 }
3362 /* For commutative ops, allow the other order. */
3370 /* If either of the pointer (or reference) expressions we are
3371 dereferencing contain a side effect, these cannot be equal,
3372 but their addresses can be. */
3379 {
3382 {
3386 /* Verify that the access types are compatible. */
3390 }
3395 /* Require the same offset. */
3401 /* Fallthru. */
3409 {
3410 /* Require equal access sizes */
3416 flags)))
3418 /* Verify that access happens in similar types. */
3421 /* Verify that accesses are TBAA compatible. */
3430 /* Verify that alignment is compatible. */
3434 }
3437 /* TARGET_MEM_REF require equal extra operands. */
3448 /* Compare the array index by value if it is constant first as we
3449 may have different types but same value here. */
3455 /* Compare low bound and element size as with OEP_ADDRESS_OF
3456 we have to account for the offset of the ref. */
3461 array_ref_low_bound
3465 array_ref_element_size
3467 flags))));
3470 /* Handle operand 2 the same as for ARRAY_REF. Operand 0
3471 may be NULL when we're called to compare MEM_EXPRs. */
3474 {
3477 /* Most of time we only need to compare FIELD_DECLs for equality.
3478 However when determining address look into actual offsets.
3479 These may match for unions and unshared record types. */
3482 {
3485 {
3489 /* Non-FIELD_DECL operands can appear in C++ templates. */
3496 flags))
3498 }
3499 else
3501 }
3502 }
3513 }
3517 {
3519 /* Be sure we pass right ADDRESS_OF flag. */
3536 /* The multiplcation operands are commutative. */
3537 /* FALLTHRU */
3545 /* Otherwise take into account this is a commutative operation. */
3558 /* BIT_INSERT_EXPR has an implict operand as the type precision
3559 of op1. Need to check to make sure they are the same. */
3565 /* FALLTHRU */
3590 /* Virtual table reference. */
3602 {
3607 }
3612 }
3616 {
3620 /* If not both CALL_EXPRs are either internal or normal function
3621 functions, then they are not equal. */
3624 {
3625 /* If the CALL_EXPRs call different internal functions, then they
3626 are not equal. */
3629 }
3630 else
3631 {
3632 /* If the CALL_EXPRs call different functions, then they are not
3633 equal. */
3635 flags))
3637 }
3639 /* FIXME: We could skip this test for OEP_MATCH_SIDE_EFFECTS. */
3640 {
3644 else
3648 }
3650 /* Now see if all the arguments are the same. */
3651 {
3662 /* If we get here and both argument lists are exhausted
3663 then the CALL_EXPRs are equal. */
3665 }
3668 }
3671 /* Consider __builtin_sqrt equal to sqrt. */
3681 {
3682 /* Consider decls with the same name equal. The caller needs
3683 to make sure they refer to the same entity (such as a function
3684 formal parameter). */
3690 }
3695 {
3699 /* In GIMPLE constructors are used only to build vectors from
3700 elements. Individual elements in the constructor must be
3701 indexed in increasing order and form an initial sequence.
3703 We make no effort to compare constructors in generic.
3704 (see sem_variable::equals in ipa-icf which can do so for
3705 constants). */
3710 /* Be sure that vectors constructed have the same representation.
3711 We only tested element precision and modes to match.
3712 Vectors may be BLKmode and thus also check that the number of
3713 parts match. */
3726 {
3731 /* In GIMPLE the indexes can be either NULL or matching i.
3732 Double check this so we won't get false
3733 positives for GENERIC. */
3741 }
3743 }
3746 {
3747 /* Compare the STATEMENT_LISTs. */
3753 {
3754 /* The lists don't have the same number of statements. */
3760 flags & (OEP_LEXICOGRAPHIC
3763 }
3764 }
3769 {
3780 }
3784 }
3786 #undef OP_SAME
3787 #undef OP_SAME_WITH_NULL
3788 }
3790 /* Generate a hash value for an expression. This can be used iteratively
3791 by passing a previous result as the HSTATE argument. */
3793 void
3796 {
3802 {
3805 }
3815 {
3816 /* Alas, constants aren't shared, so we can't rely on pointer
3817 identity. */
3827 {
3831 else
3835 }
3837 {
3841 }
3851 {
3858 }
3860 /* We can just compare by pointer. */
3864 /* The node itself doesn't matter. */
3868 /* Ignore. */
3871 /* A list of expressions, for a CALL_EXPR or as the elements of a
3872 VECTOR_CST. */
3877 {
3883 {
3884 /* In GIMPLE the indexes can be either NULL or matching i. */
3889 }
3891 }
3893 {
3894 tree_stmt_iterator i;
3899 }
3908 /* When referring to a built-in FUNCTION_DECL, use the __builtin__ form.
3909 Otherwise nodes that compare equal according to operand_equal_p might
3910 get different hash codes. However, don't do this for machine specific
3911 or front end builtins, since the function code is overloaded in those
3912 cases. */
3915 {
3918 }
3919 /* FALL THROUGH */
3922 {
3926 }
3930 {
3931 /* DECL's have a unique ID */
3933 }
3935 {
3936 /* For comparisons that can be swapped, use the lower
3937 tree code. */
3944 }
3946 {
3947 /* NOP_EXPR and CONVERT_EXPR are considered equal by
3948 operand_equal_p. */
3952 /* Don't hash the type, that can lead to having nodes which
3953 compare equal according to operand_equal_p, but which
3954 have different hash codes. Make sure to include signedness
3955 in the hash computation. */
3958 }
3959 /* For OEP_ADDRESS_OF, hash MEM_EXPR[&decl, 0] the same as decl. */
3967 /* Don't ICE on FE specific trees, or their arguments etc.
3968 during operand_equal_p hash verification. */
3971 else
3972 {
3978 {
3994 {
4001 }
4015 {
4016 /* The multiplication operands are commutative. */
4023 }
4031 /* For TARGET_EXPR, just hash on the TARGET_EXPR_SLOT.
4032 Usually different TARGET_EXPRs just should use
4033 different temporaries in their slots. */
4038 /* Virtual table reference. */
4046 {
4048 /* We compute mangled names only when free_lang_data is run.
4049 In that case we can hash precisely. */
4052 hstate.add_object
4053 (IDENTIFIER_HASH_VALUE
4055 }
4059 }
4061 /* Don't hash the type, that can lead to having nodes which
4062 compare equal according to operand_equal_p, but which
4063 have different hash codes. */
4065 {
4066 /* Make sure to include signness in the hash computation. */
4069 }
4072 {
4073 /* It's a commutative expression. We want to hash it the same
4074 however it appears. We do this by first hashing both operands
4075 and then rehashing based on the order of their independent
4076 hashes. */
4081 }
4082 else
4086 }
4088 }
4089 }
4091 bool
4094 {
4095 /* When checking and unless comparing DECL names, verify that if
4096 the outermost operand_equal_p call returns non-zero then ARG0
4097 and ARG1 have the same hash value. */
4099 {
4101 {
4103 {
4110 }
4112 }
4113 else
4117 }
4120 }
4125 /* Conveinece wrapper around operand_compare class because usually we do
4126 not need to play with the valueizer. */
4128 bool
4130 {
4132 }
4134 namespace inchash
4135 {
4137 /* Generate a hash value for an expression. This can be used iteratively
4138 by passing a previous result as the HSTATE argument.
4140 This function is intended to produce the same hash for expressions which
4141 would compare equal using operand_equal_p. */
4142 void
4144 {
4146 }
4148 }
4149 \f
4150 /* Similar to operand_equal_p, but see if ARG0 might be a variant of ARG1
4151 with a different signedness or a narrower precision. */
4153 static bool
4155 {
4163 /* Discard any conversions that don't change the modes of ARG0 and ARG1
4164 and see if the inner values are the same. This removes any
4165 signedness comparison, which doesn't matter here. */
4173 /* Discard a single widening conversion from ARG1 and see if the inner
4174 value is the same as ARG0. */
4183 }
4184 \f
4185 /* See if ARG is an expression that is either a comparison or is performing
4186 arithmetic on comparisons. The comparisons must only be comparing
4187 two different values, which will be stored in *CVAL1 and *CVAL2; if
4188 they are nonzero it means that some operands have already been found.
4189 No variables may be used anywhere else in the expression except in the
4190 comparisons.
4192 If this is true, return 1. Otherwise, return zero. */
4194 static bool
4196 {
4200 /* We can handle some of the tcc_expression cases here. */
4209 {
4228 /* First see if we can handle the first operand, then the second. For
4229 the second operand, we know *CVAL1 can't be zero. It must be that
4230 one side of the comparison is each of the values; test for the
4231 case where this isn't true by failing if the two operands
4232 are the same. */
4241 ;
4245 ;
4246 else
4250 ;
4254 ;
4255 else
4262 }
4263 }
4264 \f
4265 /* ARG is a tree that is known to contain just arithmetic operations and
4266 comparisons. Evaluate the operations in the tree substituting NEW0 for
4267 any occurrence of OLD0 as an operand of a comparison and likewise for
4268 NEW1 and OLD1. */
4270 static tree
4273 {
4278 /* We can handle some of the tcc_expression cases here. */
4286 {
4301 {
4320 }
4321 /* Fall through - ??? */
4324 {
4328 /* We need to check both for exact equality and tree equality. The
4329 former will be true if the operand has a side-effect. In that
4330 case, we know the operand occurred exactly once. */
4343 }
4347 }
4348 }
4349 \f
4350 /* Return a tree for the case when the result of an expression is RESULT
4351 converted to TYPE and OMITTED was previously an operand of the expression
4352 but is now not needed (e.g., we folded OMITTED * 0).
4354 If OMITTED has side effects, we must evaluate it. Otherwise, just do
4355 the conversion of RESULT to TYPE. */
4357 tree
4359 {
4362 /* If the resulting operand is an empty statement, just return the omitted
4363 statement casted to void. */
4373 }
4375 /* Return a tree for the case when the result of an expression is RESULT
4376 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
4377 of the expression but are now not needed.
4379 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
4380 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
4381 evaluated before OMITTED2. Otherwise, if neither has side effects,
4382 just do the conversion of RESULT to TYPE. */
4384 tree
4387 {
4396 }
4398 \f
4399 /* Return a simplified tree node for the truth-negation of ARG. This
4400 never alters ARG itself. We assume that ARG is an operation that
4401 returns a truth value (0 or 1).
4403 FIXME: one would think we would fold the result, but it causes
4404 problems with the dominator optimizer. */
4406 static tree
4408 {
4413 /* If this is a comparison, we can simply invert it, except for
4414 floating-point non-equality comparisons, in which case we just
4415 enclose a TRUTH_NOT_EXPR around what we have. */
4418 {
4421 && flag_trapping_math
4434 }
4437 {
4456 /* Here we can invert either operand. We invert the first operand
4457 unless the second operand is a TRUTH_NOT_EXPR in which case our
4458 result is the XOR of the first operand with the inside of the
4459 negation of the second operand. */
4464 else
4487 {
4494 /* A COND_EXPR may have a throw as one operand, which
4495 then has void type. Just leave void operands
4496 as they are. */
4502 }
4514 CASE_CONVERT:
4518 /* fall through */
4540 }
4541 }
4543 /* Fold the truth-negation of ARG. This never alters ARG itself. We
4544 assume that ARG is an operation that returns a truth value (0 or 1
4545 for scalars, 0 or -1 for vectors). Return the folded expression if
4546 folding is successful. Otherwise, return NULL_TREE. */
4548 static tree
4550 {
4553 ? BIT_NOT_EXPR
4556 }
4558 /* Return a simplified tree node for the truth-negation of ARG. This
4559 never alters ARG itself. We assume that ARG is an operation that
4560 returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). */
4562 tree
4564 {
4570 ? BIT_NOT_EXPR
4573 }
4574 \f
4575 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
4576 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero
4577 and uses reverse storage order if REVERSEP is nonzero. ORIG_INNER
4578 is the original memory reference used to preserve the alias set of
4579 the access. */
4581 static tree
4585 {
4588 /* Attempt not to lose the access path if possible. */
4590 {
4592 machine_mode nmode;
4594 tree noffset;
4602 && !reversep
4603 && !nreversep
4605 {
4608 }
4609 }
4618 {
4625 }
4640 }
4642 /* Optimize a bit-field compare.
4644 There are two cases: First is a compare against a constant and the
4645 second is a comparison of two items where the fields are at the same
4646 bit position relative to the start of a chunk (byte, halfword, word)
4647 large enough to contain it. In these cases we can avoid the shift
4648 implicit in bitfield extractions.
4650 For constants, we emit a compare of the shifted constant with the
4651 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
4652 compared. For two fields at the same position, we do the ANDs with the
4653 similar mask and compare the result of the ANDs.
4655 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
4656 COMPARE_TYPE is the type of the comparison, and LHS and RHS
4657 are the left and right operands of the comparison, respectively.
4659 If the optimization described above can be done, we return the resulting
4660 tree. Otherwise we return zero. */
4662 static tree
4665 {
4669 tree unsigned_type;
4672 scalar_int_mode nmode;
4677 tree mask;
4678 tree offset;
4680 /* Get all the information about the extractions being done. If the bit size
4681 is the same as the size of the underlying object, we aren't doing an
4682 extraction at all and so can do nothing. We also don't want to
4683 do anything if the inner expression is a PLACEHOLDER_EXPR since we
4684 then will no longer be able to replace it. */
4699 else
4700 {
4701 /* If this is not a constant, we can only do something if bit positions,
4702 sizes, signedness and storage order are the same. */
4703 rinner
4716 }
4718 /* Honor the C++ memory model and mimic what RTL expansion does. */
4722 {
4726 }
4728 /* See if we can find a mode to refer to this field. We should be able to,
4729 but fail if we can't. */
4737 /* Set signed and unsigned types of the precision of this mode for the
4738 shifts below. */
4741 /* Compute the bit position and size for the new reference and our offset
4742 within it. If the new reference is the same size as the original, we
4743 won't optimize anything, so return zero. */
4753 /* Make the mask to be used against the extracted field. */
4760 {
4764 /* If not comparing with constant, just rework the comparison
4765 and return. */
4773 }
4775 /* Otherwise, we are handling the constant case. See if the constant is too
4776 big for the field. Warn and return a tree for 0 (false) if so. We do
4777 this not only for its own sake, but to avoid having to test for this
4778 error case below. If we didn't, we might generate wrong code.
4780 For unsigned fields, the constant shifted right by the field length should
4781 be all zero. For signed fields, the high-order bits should agree with
4782 the sign bit. */
4785 {
4787 {
4791 }
4792 }
4793 else
4794 {
4797 {
4801 }
4802 }
4807 /* Single-bit compares should always be against zero. */
4809 {
4812 }
4814 /* Make a new bitfield reference, shift the constant over the
4815 appropriate number of bits and mask it with the computed mask
4816 (in case this was a signed field). If we changed it, make a new one. */
4824 mask);
4829 }
4830 \f
4831 /* Subroutine for fold_truth_andor_1: decode a field reference.
4833 If EXP is a comparison reference, we return the innermost reference.
4835 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
4836 set to the starting bit number.
4838 If the innermost field can be completely contained in a mode-sized
4839 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
4841 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
4842 otherwise it is not changed.
4844 *PUNSIGNEDP is set to the signedness of the field.
4846 *PREVERSEP is set to the storage order of the field.
4848 *PMASK is set to the mask used. This is either contained in a
4849 BIT_AND_EXPR or derived from the width of the field.
4851 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
4853 Return 0 if this is not a component reference or is one that we can't
4854 do anything with. */
4856 static tree
4861 {
4866 tree unsigned_type;
4869 /* All the optimizations using this function assume integer fields.
4870 There are problems with FP fields since the type_for_size call
4871 below can fail for, e.g., XFmode. */
4875 /* We are interested in the bare arrangement of bits, so strip everything
4876 that doesn't affect the machine mode. However, record the type of the
4877 outermost expression if it may matter below. */
4884 {
4890 }
4901 /* Reject out-of-bound accesses (PR79731). */
4913 /* If the number of bits in the reference is the same as the bitsize of
4914 the outer type, then the outer type gives the signedness. Otherwise
4915 (in case of a small bitfield) the signedness is unchanged. */
4919 /* Compute the mask to access the bitfield. */
4927 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
4935 }
4937 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
4938 bit positions and MASK is SIGNED. */
4940 static bool
4942 {
4946 /* If this function returns true when the type of the mask is
4947 UNSIGNED, then there will be errors. In particular see
4948 gcc.c-torture/execute/990326-1.c. There does not appear to be
4949 any documentation paper trail as to why this is so. But the pre
4950 wide-int worked with that restriction and it has been preserved
4951 here. */
4956 }
4958 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
4959 represents the sign bit of EXP's type. If EXP represents a sign
4960 or zero extension, also test VAL against the unextended type.
4961 The return value is the (sub)expression whose sign bit is VAL,
4962 or NULL_TREE otherwise. */
4964 tree
4966 {
4968 tree t;
4970 /* Tree EXP must have an integral type. */
4975 /* Tree VAL must be an integer constant. */
4984 /* Handle extension from a narrower type. */
4990 }
4992 /* Subroutine for fold_truth_andor_1 and simple_condition_p: determine if an
4993 operand is simple enough to be evaluated unconditionally. */
4995 static bool
4997 {
4998 /* Strip any conversions that don't change the machine mode. */
5007 /* Don't regard global variables as simple. They may be
5008 allocated in ways unknown to the compiler (shared memory,
5009 #pragma weak, etc). */
5012 /* Weakrefs are not safe to be read, since they can be NULL.
5013 They are !TREE_PUBLIC && !DECL_EXTERNAL but still
5014 have DECL_WEAK flag set. */
5016 /* Loading a static variable is unduly expensive, but global
5017 registers aren't expensive. */
5019 }
5021 /* Determine if an operand is simple enough to be evaluated unconditionally.
5022 In addition to simple_operand_p, we assume that comparisons, conversions,
5023 and logic-not operations are simple, if their operands are simple, too. */
5025 bool
5027 {
5046 }
5048 \f
5049 /* The following functions are subroutines to fold_range_test and allow it to
5050 try to change a logical combination of comparisons into a range test.
5052 For example, both
5053 X == 2 || X == 3 || X == 4 || X == 5
5054 and
5055 X >= 2 && X <= 5
5056 are converted to
5057 (unsigned) (X - 2) <= 3
5059 We describe each set of comparisons as being either inside or outside
5060 a range, using a variable named like IN_P, and then describe the
5061 range with a lower and upper bound. If one of the bounds is omitted,
5062 it represents either the highest or lowest value of the type.
5064 In the comments below, we represent a range by two numbers in brackets
5065 preceded by a "+" to designate being inside that range, or a "-" to
5066 designate being outside that range, so the condition can be inverted by
5067 flipping the prefix. An omitted bound is represented by a "-". For
5068 example, "- [-, 10]" means being outside the range starting at the lowest
5069 possible value and ending at 10, in other words, being greater than 10.
5070 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
5071 always false.
5073 We set up things so that the missing bounds are handled in a consistent
5074 manner so neither a missing bound nor "true" and "false" need to be
5075 handled using a special case. */
5077 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
5078 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
5079 and UPPER1_P are nonzero if the respective argument is an upper bound
5080 and zero for a lower. TYPE, if nonzero, is the type of the result; it
5081 must be specified for a comparison. ARG1 will be converted to ARG0's
5082 type if both are specified. */
5084 static tree
5087 {
5088 tree tem;
5092 /* If neither arg represents infinity, do the normal operation.
5093 Else, if not a comparison, return infinity. Else handle the special
5094 comparison rules. Note that most of the cases below won't occur, but
5095 are handled for consistency. */
5098 {
5103 }
5108 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
5109 for neither. In real maths, we cannot assume open ended ranges are
5110 the same. But, this is computer arithmetic, where numbers are finite.
5111 We can therefore make the transformation of any unbounded range with
5112 the value Z, Z being greater than any representable number. This permits
5113 us to treat unbounded ranges as equal. */
5117 {
5138 }
5141 }
5142 \f
5143 /* Helper routine for make_range. Perform one step for it, return
5144 new expression if the loop should continue or NULL_TREE if it should
5145 stop. */
5147 tree
5151 {
5157 {
5159 /* We can only do something if the range is testing for zero. */
5168 /* We can only do something if the range is testing for zero
5169 and if the second operand is an integer constant. Note that
5170 saying something is "in" the range we make is done by
5171 complementing IN_P since it will set in the initial case of
5172 being not equal to zero; "out" is leaving it alone. */
5179 {
5200 }
5202 /* If this is an unsigned comparison, we also know that EXP is
5203 greater than or equal to zero. We base the range tests we make
5204 on that fact, so we record it here so we can parse existing
5205 range tests. We test arg0_type since often the return type
5206 of, e.g. EQ_EXPR, is boolean. */
5208 {
5212 NULL_TREE))
5217 /* If the high bound is missing, but we have a nonzero low
5218 bound, reverse the range so it goes from zero to the low bound
5219 minus 1. */
5221 {
5226 }
5227 }
5235 /* If flag_wrapv and ARG0_TYPE is signed, make sure
5236 low and high are non-NULL, then normalize will DTRT. */
5239 {
5244 }
5246 /* (-x) IN [a,b] -> x in [-b, -a] */
5258 /* ~ X -> -X - 1 */
5267 /* If flag_wrapv and ARG0_TYPE is signed, then we cannot
5268 move a constant to the other side. */
5273 /* If EXP is signed, any overflow in the computation is undefined,
5274 so we don't worry about it so long as our computations on
5275 the bounds don't overflow. For unsigned, overflow is defined
5276 and this is exactly the right thing. */
5288 normalize:
5289 /* Check for an unsigned range which has wrapped around the maximum
5290 value thus making n_high < n_low, and normalize it. */
5292 {
5298 /* If the range is of the form +/- [ x+1, x ], we won't
5299 be able to normalize it. But then, it represents the
5300 whole range or the empty set, so make it
5301 +/- [ -, - ]. */
5305 else
5307 }
5308 else
5316 CASE_CONVERT:
5334 /* If we're converting arg0 from an unsigned type, to exp,
5335 a signed type, we will be doing the comparison as unsigned.
5336 The tests above have already verified that LOW and HIGH
5337 are both positive.
5339 So we have to ensure that we will handle large unsigned
5340 values the same way that the current signed bounds treat
5341 negative values. */
5344 {
5345 tree high_positive;
5346 tree equiv_type;
5347 /* For fixed-point modes, we need to pass the saturating flag
5348 as the 2nd parameter. */
5350 equiv_type
5355 else
5356 equiv_type
5359 /* A range without an upper bound is, naturally, unbounded.
5360 Since convert would have cropped a very large value, use
5361 the max value for the destination type. */
5362 high_positive
5369 high_positive),
5372 /* If the low bound is specified, "and" the range with the
5373 range for which the original unsigned value will be
5374 positive. */
5376 {
5379 integer_zero_node),
5380 high_positive))
5384 }
5385 else
5386 {
5387 /* Otherwise, "or" the range with the range of the input
5388 that will be interpreted as negative. */
5391 integer_zero_node),
5392 high_positive))
5396 }
5397 }
5399 /* Otherwise, if we are converting arg0 from signed type, to exp,
5400 an unsigned type, we will do the comparison as signed. If
5401 high is non-NULL, we punt above if it doesn't fit in the signed
5402 type, so if we get through here, +[-, high] or +[low, high] are
5403 equivalent to +[-, n_high] or +[n_low, n_high]. Similarly,
5404 +[-, -] or -[-, -] are equivalent too. But if low is specified and
5405 high is not, the +[low, -] range is equivalent to union of
5406 +[n_low, -] and +[-, -1] ranges, so +[low, -] is equivalent to
5407 -[0, n_low-1] and similarly -[low, -] to +[0, n_low-1], except for
5408 low being 0, which should be treated as [-, -]. */
5411 && low
5413 {
5416 else
5417 {
5422 }
5423 }
5432 }
5433 }
5435 /* Given EXP, a logical expression, set the range it is testing into
5436 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
5437 actually being tested. *PLOW and *PHIGH will be made of the same
5438 type as the returned expression. If EXP is not a comparison, we
5439 will most likely not be returning a useful value and range. Set
5440 *STRICT_OVERFLOW_P to true if the return value is only valid
5441 because signed overflow is undefined; otherwise, do not change
5442 *STRICT_OVERFLOW_P. */
5444 tree
5447 {
5455 /* Start with simply saying "EXP != 0" and then look at the code of EXP
5456 and see if we can refine the range. Some of the cases below may not
5457 happen, but it doesn't seem worth worrying about this. We "continue"
5458 the outer loop when we've changed something; otherwise we "break"
5459 the switch, which will "break" the while. */
5465 {
5471 {
5479 }
5488 }
5490 /* If EXP is a constant, we can evaluate whether this is true or false. */
5492 {
5499 }
5503 }
5505 /* Returns TRUE if [LOW, HIGH] range check can be optimized to
5506 a bitwise check i.e. when
5507 LOW == 0xXX...X00...0
5508 HIGH == 0xXX...X11...1
5509 Return corresponding mask in MASK and stem in VALUE. */
5511 static bool
5514 {
5537 }
5538 \f
5539 /* Helper routine for build_range_check and match.pd. Return the type to
5540 perform the check or NULL if it shouldn't be optimized. */
5542 tree
5544 {
5545 /* First make sure that arithmetics in this type is valid, then make sure
5546 that it wraps around. */
5551 {
5554 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
5555 for the type in question, as we rely on this here. */
5565 else
5567 }
5570 /* Right now all BITINT_TYPEs satisfy
5571 (unsigned) max + 1 == (unsigned) min, so no need to verify
5572 that like for INTEGER_TYPEs. */
5576 }
5578 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
5579 type, TYPE, return an expression to test if EXP is in (or out of, depending
5580 on IN_P) the range. Return 0 if the test couldn't be created. */
5582 tree
5585 {
5588 /* Disable this optimization for function pointer expressions
5589 on targets that require function pointer canonicalization. */
5596 {
5602 }
5624 value);
5627 {
5629 {
5633 }
5635 }
5637 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
5639 {
5643 {
5645 {
5648 etype
5650 else
5653 }
5656 }
5657 }
5659 /* Optimize (c>=low) && (c<=high) into (c-low>=0) && (c-low<=high-low).
5660 This requires wrap-around arithmetics for the type of the expression. */
5677 }
5678 \f
5679 /* Return the predecessor of VAL in its type, handling the infinite case. */
5681 static tree
5683 {
5689 else
5692 }
5694 /* Return the successor of VAL in its type, handling the infinite case. */
5696 static tree
5698 {
5704 else
5707 }
5709 /* Given two ranges, see if we can merge them into one. Return 1 if we
5710 can, 0 if we can't. Set the output range into the specified parameters. */
5712 bool
5715 {
5719 tree tem;
5729 /* Make range 0 be the range that starts first, or ends last if they
5730 start at the same value. Swap them if it isn't. */
5733 || (lowequal
5736 {
5740 }
5742 /* If the second range is != high1 where high1 is the type maximum of
5743 the type, try first merging with < high1 range. */
5745 && high1
5752 {
5757 /* Similarly for the second range != low1 where low1 is the type minimum
5758 of the type, try first merging with > low1 range. */
5763 }
5765 /* Now flag two cases, whether the ranges are disjoint or whether the
5766 second range is totally subsumed in the first. Note that the tests
5767 below are simplified by the ones above. */
5773 /* We now have four cases, depending on whether we are including or
5774 excluding the two ranges. */
5776 {
5777 /* If they don't overlap, the result is false. If the second range
5778 is a subset it is the result. Otherwise, the range is from the start
5779 of the second to the end of the first. */
5784 else
5786 }
5789 {
5790 /* If they don't overlap, the result is the first range. If they are
5791 equal, the result is false. If the second range is a subset of the
5792 first, and the ranges begin at the same place, we go from just after
5793 the end of the second range to the end of the first. If the second
5794 range is not a subset of the first, or if it is a subset and both
5795 ranges end at the same place, the range starts at the start of the
5796 first range and ends just before the second range.
5797 Otherwise, we can't describe this as a single range. */
5803 {
5808 {
5809 /* We are in the weird situation where high0 > high1 but
5810 high1 has no successor. Punt. */
5812 }
5813 }
5815 {
5820 {
5821 /* low0 < low1 but low1 has no predecessor. Punt. */
5823 }
5824 }
5825 else
5827 }
5830 {
5831 /* If they don't overlap, the result is the second range. If the second
5832 is a subset of the first, the result is false. Otherwise,
5833 the range starts just after the first range and ends at the
5834 end of the second. */
5839 else
5840 {
5845 {
5846 /* high1 > high0 but high0 has no successor. Punt. */
5848 }
5849 }
5850 }
5852 else
5853 {
5854 /* The case where we are excluding both ranges. Here the complex case
5855 is if they don't overlap. In that case, the only time we have a
5856 range is if they are adjacent. If the second is a subset of the
5857 first, the result is the first. Otherwise, the range to exclude
5858 starts at the beginning of the first range and ends at the end of the
5859 second. */
5861 {
5866 else
5867 {
5868 /* Canonicalize - [min, x] into - [-, x]. */
5871 {
5874 GET_MODE_BITSIZE
5877 /* FALLTHROUGH */
5890 }
5892 /* Canonicalize - [x, max] into - [x, -]. */
5895 {
5898 GET_MODE_BITSIZE
5901 /* FALLTHROUGH */
5917 }
5919 /* The ranges might be also adjacent between the maximum and
5920 minimum values of the given type. For
5921 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
5922 return + [x + 1, y - 1]. */
5924 {
5931 }
5932 else
5934 }
5935 }
5938 else
5940 }
5944 }
5945 \f
5947 /* Subroutine of fold, looking inside expressions of the form
5948 A op B ? A : C, where (ARG00, COMP_CODE, ARG01), ARG1 and ARG2
5949 are the three operands of the COND_EXPR. This function is
5950 being used also to optimize A op B ? C : A, by reversing the
5951 comparison first.
5953 Return a folded expression whose code is not a COND_EXPR
5954 anymore, or NULL_TREE if no folding opportunity is found. */
5956 static tree
5960 {
5962 tree tem;
5967 /* If we have A op 0 ? A : -A, consider applying the following
5968 transformations:
5970 A == 0? A : -A same as -A
5971 A != 0? A : -A same as A
5972 A >= 0? A : -A same as abs (A)
5973 A > 0? A : -A same as abs (A)
5974 A <= 0? A : -A same as -abs (A)
5975 A < 0? A : -A same as -abs (A)
5977 None of these transformations work for modes with signed
5978 zeros. If A is +/-0, the first two transformations will
5979 change the sign of the result (from +0 to -0, or vice
5980 versa). The last four will fix the sign of the result,
5981 even though the original expressions could be positive or
5982 negative, depending on the sign of A.
5984 Note that all these transformations are correct if A is
5985 NaN, since the two alternatives (A and -A) are also NaNs. */
5992 /* In the case that A is of the form X-Y, '-A' (arg2) may
5993 have already been folded to Y-X, check for that. */
6001 {
6013 /* Fall through. */
6024 /* FALLTHRU */
6031 {
6032 /* A <= 0 ? A : -A for A INT_MIN is valid, but -abs(INT_MIN)
6033 is not, invokes UB both in abs and in the negation of it.
6034 So, use ABSU_EXPR instead. */
6039 }
6040 else
6041 {
6044 }
6048 }
6050 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
6051 A == 0 ? A : 0 is always 0 unless A is -0. Note that
6052 both transformations are correct when A is NaN: A != 0
6053 is then true, and A == 0 is false. */
6057 {
6062 }
6064 /* Try some transformations of A op B ? A : B.
6066 A == B? A : B same as B
6067 A != B? A : B same as A
6068 A >= B? A : B same as max (A, B)
6069 A > B? A : B same as max (B, A)
6070 A <= B? A : B same as min (A, B)
6071 A < B? A : B same as min (B, A)
6073 As above, these transformations don't work in the presence
6074 of signed zeros. For example, if A and B are zeros of
6075 opposite sign, the first two transformations will change
6076 the sign of the result. In the last four, the original
6077 expressions give different results for (A=+0, B=-0) and
6078 (A=-0, B=+0), but the transformed expressions do not.
6080 The first two transformations are correct if either A or B
6081 is a NaN. In the first transformation, the condition will
6082 be false, and B will indeed be chosen. In the case of the
6083 second transformation, the condition A != B will be true,
6084 and A will be chosen.
6086 The conversions to max() and min() are not correct if B is
6087 a number and A is not. The conditions in the original
6088 expressions will be false, so all four give B. The min()
6089 and max() versions would give a NaN instead. */
6092 /* Avoid these transformations if the COND_EXPR may be used
6093 as an lvalue in the C++ front-end. PR c++/19199. */
6094 && (in_gimple_form
6100 {
6106 {
6115 /* In C++ a ?: expression can be an lvalue, so put the
6116 operand which will be used if they are equal first
6117 so that we can convert this back to the
6118 corresponding COND_EXPR. */
6120 {
6128 }
6135 {
6143 }
6156 }
6157 }
6160 }
6163 \f
6164 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
6165 #define LOGICAL_OP_NON_SHORT_CIRCUIT \
6166 (BRANCH_COST (optimize_function_for_speed_p (cfun), \
6167 false) >= 2)
6168 #endif
6170 /* EXP is some logical combination of boolean tests. See if we can
6171 merge it into some range test. Return the new tree if so. */
6173 static tree
6176 {
6190 /* If op0 is known true or false and this is a short-circuiting
6191 operation we must not merge with op1 since that makes side-effects
6192 unconditional. So special-case this. */
6199 /* If this is an OR operation, invert both sides; we will invert
6200 again at the end. */
6204 /* If both expressions are the same, if we can merge the ranges, and we
6205 can build the range test, return it or it inverted. If one of the
6206 ranges is always true or always false, consider it to be the same
6207 expression as the other. */
6215 {
6219 }
6221 /* On machines where the branch cost is expensive, if this is a
6222 short-circuited branch and the underlying object on both sides
6223 is the same, make a non-short-circuit operation. */
6226 logical_op_non_short_circuit
6233 {
6234 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
6235 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
6236 which cases we can't do this. */
6244 {
6253 {
6256 WARN_STRICT_OVERFLOW_COMPARISON);
6260 }
6261 }
6262 }
6265 }
6266 \f
6267 /* Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P
6268 bit value. Arrange things so the extra bits will be set to zero if and
6269 only if C is signed-extended to its full width. If MASK is nonzero,
6270 it is an INTEGER_CST that should be AND'ed with the extra bits. */
6272 static tree
6274 {
6277 tree temp;
6282 /* We work by getting just the sign bit into the low-order bit, then
6283 into the high-order bit, then sign-extend. We then XOR that value
6284 with C. */
6288 /* We must use a signed type in order to get an arithmetic right shift.
6289 However, we must also avoid introducing accidental overflows, so that
6290 a subsequent call to integer_zerop will work. Hence we must
6291 do the type conversion here. At this point, the constant is either
6292 zero or one, and the conversion to a signed type can never overflow.
6293 We could get an overflow if this conversion is done anywhere else. */
6302 /* If necessary, convert the type back to match the type of C. */
6307 }
6308 \f
6309 /* For an expression that has the form
6310 (A && B) || ~B
6311 or
6312 (A || B) && ~B,
6313 we can drop one of the inner expressions and simplify to
6314 A || ~B
6315 or
6316 A && ~B
6317 LOC is the location of the resulting expression. OP is the inner
6318 logical operation; the left-hand side in the examples above, while CMPOP
6319 is the right-hand side. RHS_ONLY is used to prevent us from accidentally
6320 removing a condition that guards another, as in
6321 (A != NULL && A->...) || A == NULL
6322 which we must not transform. If RHS_ONLY is true, only eliminate the
6323 right-most operand of the inner logical operation. */
6325 static tree
6328 {
6346 {
6349 {
6352 }
6353 }
6355 {
6358 {
6361 }
6362 }
6377 }
6379 /* Find ways of folding logical expressions of LHS and RHS:
6380 Try to merge two comparisons to the same innermost item.
6381 Look for range tests like "ch >= '0' && ch <= '9'".
6382 Look for combinations of simple terms on machines with expensive branches
6383 and evaluate the RHS unconditionally.
6385 For example, if we have p->a == 2 && p->b == 4 and we can make an
6386 object large enough to span both A and B, we can do this with a comparison
6387 against the object ANDed with the a mask.
6389 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
6390 operations to do this with one comparison.
6392 We check for both normal comparisons and the BIT_AND_EXPRs made this by
6393 function and the one above.
6395 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
6396 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
6398 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
6399 two operands.
6401 We return the simplified tree or 0 if no optimization is possible. */
6403 static tree
6406 {
6407 /* If this is the "or" of two comparisons, we can do something if
6408 the comparisons are NE_EXPR. If this is the "and", we can do something
6409 if the comparisons are EQ_EXPR. I.e.,
6410 (a->b == 2 && a->c == 4) can become (a->new == NEW).
6412 WANTED_CODE is this operation code. For single bit fields, we can
6413 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
6414 comparison for one-bit fields. */
6435 /* Start by getting the comparison codes. Fail if anything is volatile.
6436 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
6437 it were surrounded with a NE_EXPR. */
6446 {
6450 }
6453 {
6457 }
6468 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
6471 {
6474 {
6479 }
6482 {
6488 }
6489 }
6494 /* If the RHS can be evaluated unconditionally and its operands are
6495 simple, it wins to evaluate the RHS unconditionally on machines
6496 with expensive branches. In this case, this isn't a comparison
6497 that can be merged. */
6504 {
6505 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
6516 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
6526 }
6528 /* See if the comparisons can be merged. Then get all the parameters for
6529 each side. */
6554 /* It must be true that the inner operation on the lhs of each
6555 comparison must be the same if we are to be able to do anything.
6556 Then see if we have constants. If not, the same must be true for
6557 the rhs's. */
6566 {
6569 }
6574 else
6577 /* If either comparison code is not correct for our logical operation,
6578 fail. However, we can convert a one-bit comparison against zero into
6579 the opposite comparison against that bit being set in the field. */
6583 {
6585 {
6586 /* Make the left operand unsigned, since we are only interested
6587 in the value of one bit. Otherwise we are doing the wrong
6588 thing below. */
6591 }
6592 else
6594 }
6596 /* This is analogous to the code for l_const above. */
6598 {
6600 {
6603 }
6604 else
6606 }
6608 /* See if we can find a mode that contains both fields being compared on
6609 the left. If we can't, fail. Otherwise, update all constants and masks
6610 to be relative to a field of that size. */
6624 {
6627 }
6637 {
6646 {
6650 }
6651 }
6653 {
6662 {
6666 }
6667 }
6669 /* If the right sides are not constant, do the same for it. Also,
6670 disallow this optimization if a size, signedness or storage order
6671 mismatch occurs between the left and right sides. */
6673 {
6677 /* Make sure the two fields on the right
6678 correspond to the left without being swapped. */
6695 {
6698 }
6709 /* Make a mask that corresponds to both fields being compared.
6710 Do this for both items being compared. If the operands are the
6711 same size and the bits being compared are in the same position
6712 then we can do this by masking both and comparing the masked
6713 results. */
6720 {
6734 }
6736 /* There is still another way we can do something: If both pairs of
6737 fields being compared are adjacent, we may be able to make a wider
6738 field containing them both.
6740 Note that we still must mask the lhs/rhs expressions. Furthermore,
6741 the mask must be shifted to account for the shift done by
6742 make_bit_field_ref. */
6751 {
6752 tree type;
6770 /* Convert to the smaller type before masking out unwanted bits. */
6773 {
6775 {
6779 }
6781 {
6785 }
6786 }
6795 }
6798 }
6800 /* Handle the case of comparisons with constants. If there is something in
6801 common between the masks, those bits of the constants must be the same.
6802 If not, the condition is always false. Test for this to avoid generating
6803 incorrect code below. */
6808 {
6810 {
6813 }
6814 else
6815 {
6818 }
6819 }
6824 /* Construct the expression we will return. First get the component
6825 reference we will make. Unless the mask is all ones the width of
6826 that field, perform the mask operation. Then compare with the
6827 merged constant. */
6838 }
6839 \f
6840 /* T is an integer expression that is being multiplied, divided, or taken a
6841 modulus (CODE says which and what kind of divide or modulus) by a
6842 constant C. See if we can eliminate that operation by folding it with
6843 other operations already in T. WIDE_TYPE, if non-null, is a type that
6844 should be used for the computation if wider than our type.
6846 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
6847 (X * 2) + (Y * 4). We must, however, be assured that either the original
6848 expression would not overflow or that overflow is undefined for the type
6849 in the language in question.
6851 If we return a non-null expression, it is an equivalent form of the
6852 original computation, but need not be in the original type.
6854 We set *STRICT_OVERFLOW_P to true if the return values depends on
6855 signed overflow being undefined. Otherwise we do not change
6856 *STRICT_OVERFLOW_P. */
6858 static tree
6861 {
6862 /* To avoid exponential search depth, refuse to allow recursion past
6863 three levels. Beyond that (1) it's highly unlikely that we'll find
6864 something interesting and (2) we've probably processed it before
6865 when we built the inner expression. */
6868 tree ret;
6873 depth++;
6875 depth--;
6878 }
6880 static tree
6883 {
6888 {
6891 {
6894 }
6898 }
6904 /* Don't deal with constants of zero here; they confuse the code below. */
6914 /* Note that we need not handle conditional operations here since fold
6915 already handles those cases. So just do arithmetic here. */
6917 {
6919 /* For a constant, we can always simplify if we are a multiply
6920 or (for divide and modulus) if it is a multiple of our constant. */
6924 {
6927 /* If the multiplication overflowed, we lost information on it.
6928 See PR68142 and PR69845. */
6932 }
6938 /* If op0 is an expression ... */
6944 /* ... and has wrapping overflow, and its type is smaller
6945 than ctype, then we cannot pass through as widening. */
6949 /* ... or this is a truncation (t is narrower than op0),
6950 then we cannot pass through this narrowing. */
6953 /* ... or signedness changes for division or modulus,
6954 then we cannot pass through this conversion. */
6958 /* ... or has undefined overflow while the converted to
6959 type has not, we cannot do the operation in the inner type
6960 as that would introduce undefined overflow. */
6965 /* Pass the constant down and see if we can make a simplification. If
6966 we can, replace this expression with the inner simplification for
6967 possible later conversion to our or some other type. */
6978 /* If widening the type changes it from signed to unsigned, then we
6979 must avoid building ABS_EXPR itself as unsigned. */
6981 {
6985 {
6988 }
6990 }
6991 /* If the constant is negative, we cannot simplify this. */
6994 /* FALLTHROUGH */
6996 /* For division and modulus, type can't be unsigned, as e.g.
6997 (-(x / 2U)) / 2U isn't equal to -((x / 2U) / 2U) for x >= 2.
6998 For signed types, even with wrapping overflow, this is fine. */
7007 /* If widening the type changes the signedness, then we can't perform
7008 this optimization as that changes the result. */
7012 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
7018 {
7025 }
7029 /* If the second operand is constant, this is a multiplication
7030 or floor division, by a power of two, so we can treat it that
7031 way unless the multiplier or divisor overflows. Signed
7032 left-shift overflow is implementation-defined rather than
7033 undefined in C90, so do not convert signed left shift into
7034 multiplication. */
7037 /* const_binop may not detect overflow correctly,
7038 so check for it explicitly here. */
7047 ctype,
7049 t1),
7054 /* See if we can eliminate the operation on both sides. If we can, we
7055 can return a new PLUS or MINUS. If we can't, the only remaining
7056 cases where we can do anything are if the second operand is a
7057 constant. */
7064 /* If not multiplication, we can only do this if both operands
7065 are divisible by c. */
7068 {
7073 }
7075 /* If this was a subtraction, negate OP1 and set it to be an addition.
7076 This simplifies the logic below. */
7078 {
7080 /* If OP1 was not easily negatable, the constant may be OP0. */
7082 {
7085 }
7086 }
7091 /* If either OP1 or C are negative, this optimization is not safe for
7092 some of the division and remainder types while for others we need
7093 to change the code. */
7095 {
7103 }
7105 /* If it's a multiply or a division/modulus operation of a multiple
7106 of our constant, do the operation and verify it doesn't overflow. */
7110 {
7113 /* We allow the constant to overflow with wrapping semantics. */
7117 }
7118 else
7121 /* If we have an unsigned type, we cannot widen the operation since it
7122 will change the result if the original computation overflowed. */
7126 /* The last case is if we are a multiply. In that case, we can
7127 apply the distributive law to commute the multiply and addition
7128 if the multiplication of the constants doesn't overflow
7129 and overflow is defined. With undefined overflow
7130 op0 * c might overflow, while (op0 + orig_op1) * c doesn't.
7131 But fold_plusminus_mult_expr would factor back any power-of-two
7132 value so do not distribute in the first place in this case. */
7140 op1);
7145 /* We have a special case here if we are doing something like
7146 (C * 8) % 4 since we know that's zero. */
7149 /* If the multiplication can overflow we cannot optimize this. */
7154 {
7157 }
7159 /* ... fall through ... */
7163 /* If we can extract our operation from the LHS, do so and return a
7164 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
7165 do something only if the second operand is a constant. */
7181 /* If these are the same operation types, we can associate them
7182 assuming no overflow. */
7184 {
7199 }
7201 /* If these operations "cancel" each other, we have the main
7202 optimizations of this pass, which occur when either constant is a
7203 multiple of the other, in which case we replace this with either an
7204 operation or CODE or TCODE.
7206 If we have an unsigned type, we cannot do this since it will change
7207 the result if the original computation overflowed. */
7215 {
7218 {
7224 }
7227 {
7233 }
7234 }
7239 }
7242 }
7243 \f
7244 /* Return a node which has the indicated constant VALUE (either 0 or
7245 1 for scalars or {-1,-1,..} or {0,0,...} for vectors),
7246 and is of the indicated TYPE. */
7248 tree
7250 {
7259 else
7261 }
7264 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
7265 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
7266 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
7267 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
7268 COND is the first argument to CODE; otherwise (as in the example
7269 given here), it is the second argument. TYPE is the type of the
7270 original expression. Return NULL_TREE if no simplification is
7271 possible. */
7273 static tree
7278 {
7286 /* Do not move possibly trapping operations into the conditional as this
7287 pessimizes code and causes gimplification issues when applied late. */
7295 {
7299 /* If this operand throws an expression, then it does not make
7300 sense to try to perform a logical or arithmetic operation
7301 involving it. */
7306 }
7309 {
7314 }
7315 else
7316 /* Detect the case of mixing vector and scalar types - bail out. */
7322 /* This transformation is only worthwhile if we don't have to wrap ARG
7323 in a SAVE_EXPR and the operation can be simplified without recursing
7324 on at least one of the branches once its pushed inside the COND_EXPR. */
7333 {
7337 else
7339 }
7341 {
7345 else
7347 }
7349 /* Check that we have simplified at least one of the branches. */
7354 }
7356 \f
7357 /* Subroutine of fold() that checks for the addition of ARG +/- 0.0.
7359 If !NEGATE, return true if ZERO_ARG is +/-0.0 and, for all ARG of
7360 type TYPE, ARG + ZERO_ARG is the same as ARG. If NEGATE, return true
7361 if ARG - ZERO_ARG is the same as X.
7363 If ARG is NULL, check for any value of type TYPE.
7365 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
7366 and finite. The problematic cases are when X is zero, and its mode
7367 has signed zeros. In the case of rounding towards -infinity,
7368 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
7369 modes, X + 0 is not the same as X because -0 + 0 is 0. */
7371 bool
7374 {
7378 /* Don't allow the fold with -fsignaling-nans. */
7382 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
7386 /* There is no case that is safe for all rounding modes. */
7390 /* In a vector or complex, we would need to check the sign of all zeros. */
7396 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
7400 /* The mode has signed zeros, and we have to honor their sign.
7401 In this situation, there are only two cases we can return true for.
7402 (i) X - 0 is the same as X with default rounding.
7403 (ii) X + 0 is X when X can't possibly be -0.0. */
7405 }
7407 /* Subroutine of match.pd that optimizes comparisons of a division by
7408 a nonzero integer constant against an integer constant, i.e.
7409 X/C1 op C2.
7411 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
7412 GE_EXPR or LE_EXPR. ARG01 and ARG1 must be a INTEGER_CST. */
7414 enum tree_code
7417 {
7422 /* We have to do this the hard way to detect unsigned overflow.
7423 prod = int_const_binop (MULT_EXPR, c1, c2); */
7429 {
7433 /* Likewise *hi = int_const_binop (PLUS_EXPR, prod, tmp). */
7436 }
7438 {
7441 {
7460 }
7461 }
7462 else
7463 {
7464 /* A negative divisor reverses the relational operators. */
7469 {
7488 }
7489 }
7502 }
7504 /* Test whether it is preferable to swap two operands, ARG0 and
7505 ARG1, for example because ARG0 is an integer constant and ARG1
7506 isn't. */
7508 bool
7510 {
7524 /* It is preferable to swap two SSA_NAME to ensure a canonical form
7525 for commutative and comparison operators. Ensuring a canonical
7526 form allows the optimizers to find additional redundancies without
7527 having to explicitly check for both orderings. */
7533 /* Put SSA_NAMEs last. */
7539 /* Put variables last. */
7546 }
7549 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
7550 means A >= Y && A != MAX, but in this case we know that
7551 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
7553 static tree
7555 {
7562 else
7571 {
7574 }
7576 {
7579 }
7580 else
7587 {
7588 /* Convert the pointer types into integer before taking the difference. */
7592 }
7593 else
7600 }
7602 /* Fold a sum or difference of at least one multiplication.
7603 Returns the folded tree or NULL if no simplification could be made. */
7605 static tree
7608 {
7612 /* (A * C) +- (B * C) -> (A+-B) * C.
7613 (A * C) +- A -> A * (C+-1).
7614 We are most concerned about the case where C is a constant,
7615 but other combinations show up during loop reduction. Since
7616 it is not difficult, try all four possibilities. */
7619 {
7622 }
7624 {
7627 }
7628 else
7629 {
7630 /* We cannot generate constant 1 for fract. */
7635 }
7637 {
7640 }
7642 {
7644 /* As we canonicalize A - 2 to A + -2 get rid of that sign for
7645 the purpose of this canonicalization. */
7649 {
7652 }
7653 else
7655 }
7656 else
7657 {
7658 /* We cannot generate constant 1 for fract. */
7663 }
7666 /* Prefer factoring a common non-constant. */
7676 /* No identical multiplicands; see if we can find a common
7677 power-of-two factor in non-power-of-two multiplies. This
7678 can help in multi-dimensional array access. */
7680 {
7683 HOST_WIDE_INT tmp;
7685 tree maybe_same;
7687 /* Move min of absolute values to int11. */
7689 {
7694 }
7695 else
7702 /* The remainder should not be a constant, otherwise we
7703 end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
7704 increased the number of multiplications necessary. */
7706 {
7714 }
7715 }
7722 /* We are neither factoring zero nor minus one. */
7730 /* Same may be zero and thus the operation 'code' may overflow. Likewise
7731 same may be minus one and thus the multiplication may overflow. Perform
7732 the sum operation in an unsigned type. */
7737 /* If the sum evaluated to a constant that is not -INF the multiplication
7738 cannot overflow. */
7745 /* Do not resort to unsigned multiplication because
7746 we lose the no-overflow property of the expression. */
7748 }
7750 /* Subroutine of native_encode_expr. Encode the INTEGER_CST
7751 specified by EXPR into the buffer PTR of length LEN bytes.
7752 Return the number of bytes placed in the buffer, or zero
7753 upon failure. */
7755 static int
7757 {
7761 {
7767 {
7769 /* More work is needed when adding _BitInt support to PDP endian
7770 if limb is smaller than word, or if _BitInt limb ordering doesn't
7771 match target endianity here. */
7776 }
7777 else
7779 }
7780 else
7791 /* Dry run. */
7797 {
7799 /* Extend EXPR according to TYPE_SIGN if the precision isn't a whole
7800 number of bytes. */
7804 {
7811 else
7813 }
7814 else
7818 }
7820 }
7823 /* Subroutine of native_encode_expr. Encode the FIXED_CST
7824 specified by EXPR into the buffer PTR of length LEN bytes.
7825 Return the number of bytes placed in the buffer, or zero
7826 upon failure. */
7828 static int
7830 {
7834 FIXED_VALUE_TYPE value;
7849 }
7852 /* Subroutine of native_encode_expr. Encode the REAL_CST
7853 specified by EXPR into the buffer PTR of length LEN bytes.
7854 Return the number of bytes placed in the buffer, or zero
7855 upon failure. */
7857 static int
7859 {
7865 /* There are always 32 bits in each long, no matter the size of
7866 the hosts long. We handle floating point representations with
7867 up to 192 bits. */
7876 /* Dry run. */
7885 {
7890 {
7897 else
7899 }
7900 else
7901 {
7904 {
7905 /* Reverse bytes within each long, or within the entire float
7906 if it's smaller than a long (for HFmode). */
7909 }
7910 }
7915 }
7917 }
7919 /* Subroutine of native_encode_expr. Encode the COMPLEX_CST
7920 specified by EXPR into the buffer PTR of length LEN bytes.
7921 Return the number of bytes placed in the buffer, or zero
7922 upon failure. */
7924 static int
7926 {
7928 tree part;
7942 }
7944 /* Like native_encode_vector, but only encode the first COUNT elements.
7945 The other arguments are as for native_encode_vector. */
7947 static int
7950 {
7954 {
7955 /* This is the only case in which elements can be smaller than a byte.
7956 Element 0 is always in the lsb of the containing byte. */
7965 /* Zero the buffer and then set bits later where necessary. */
7974 {
7980 {
7983 }
7984 }
7986 }
7991 {
7993 {
7996 }
8007 }
8009 }
8011 /* Subroutine of native_encode_expr. Encode the VECTOR_CST
8012 specified by EXPR into the buffer PTR of length LEN bytes.
8013 Return the number of bytes placed in the buffer, or zero
8014 upon failure. */
8016 static int
8018 {
8023 }
8026 /* Subroutine of native_encode_expr. Encode the STRING_CST
8027 specified by EXPR into the buffer PTR of length LEN bytes.
8028 Return the number of bytes placed in the buffer, or zero
8029 upon failure. */
8031 static int
8033 {
8036 /* Wide-char strings are encoded in target byte-order so native
8037 encoding them is trivial. */
8052 else
8053 {
8056 {
8059 }
8061 }
8063 }
8066 /* Subroutine of fold_view_convert_expr. Encode the INTEGER_CST, REAL_CST,
8067 FIXED_CST, COMPLEX_CST, STRING_CST, or VECTOR_CST specified by EXPR into
8068 the buffer PTR of size LEN bytes. If PTR is NULL, don't actually store
8069 anything, just do a dry run. Fail either if OFF is -1 and LEN isn't
8070 sufficient to encode the entire EXPR, or if OFF is out of bounds.
8071 Otherwise, start at byte offset OFF and encode at most LEN bytes.
8072 Return the number of bytes placed in the buffer, or zero upon failure. */
8074 int
8076 {
8077 /* We don't support starting at negative offset and -1 is special. */
8082 {
8103 }
8104 }
8106 /* Try to find a type whose byte size is smaller or equal to LEN bytes larger
8107 or equal to FIELDSIZE bytes, with underlying mode precision/size multiple
8108 of BITS_PER_UNIT. As native_{interpret,encode}_int works in term of
8109 machine modes, we can't just use build_nonstandard_integer_type. */
8111 tree
8113 {
8114 machine_mode mode;
8116 {
8123 {
8127 }
8128 }
8134 {
8142 }
8145 }
8147 /* Similar to native_encode_expr, but also handle CONSTRUCTORs, VCEs,
8148 NON_LVALUE_EXPRs and nops. If MASK is non-NULL (then PTR has
8149 to be non-NULL and OFF zero), then in addition to filling the
8150 bytes pointed by PTR with the value also clear any bits pointed
8151 by MASK that are known to be initialized, keep them as is for
8152 e.g. uninitialized padding bits or uninitialized fields. */
8154 int
8157 {
8160 /* We don't support starting at negative offset and -1 is special. */
8168 {
8172 mask);
8187 {
8188 tree min_index;
8206 {
8211 {
8214 }
8219 else
8223 {
8227 offset_int wpos
8235 offset_int wcount
8242 }
8244 {
8247 offset_int wpos
8255 }
8258 {
8260 {
8271 }
8273 {
8277 }
8278 }
8282 do
8283 {
8288 {
8290 {
8293 fieldsize);
8296 }
8298 ptr
8301 fieldsize,
8304 mask
8308 else
8309 {
8312 }
8313 }
8316 {
8317 /* Partial overlap. */
8323 {
8327 fieldsize);
8328 }
8329 else
8330 {
8334 }
8337 }
8339 }
8341 }
8343 }
8346 {
8356 {
8362 {
8365 }
8368 {
8377 }
8381 else
8385 {
8386 tree fld;
8388 {
8401 }
8403 {
8407 }
8410 {
8411 cnt--;
8417 }
8418 }
8423 {
8436 {
8441 }
8442 }
8443 else
8444 {
8449 }
8453 /* Prepare to deal with integral bit-fields and filter out other
8454 bit-fields that do not start and end on a byte boundary. */
8456 {
8461 {
8467 }
8473 }
8483 {
8484 /* FIXME: Handle PDP endian. */
8495 {
8498 }
8499 else
8500 {
8508 else
8509 {
8512 }
8513 }
8523 else
8536 /* If the bitfield does not start at byte boundary, handle
8537 the partial byte at the start. */
8540 {
8542 {
8547 {
8550 else
8552 }
8553 }
8554 else
8555 {
8560 {
8563 else
8567 }
8568 }
8569 }
8570 /* If the bitfield does not end at byte boundary, handle
8571 the partial byte at the end. */
8575 {
8577 {
8584 }
8585 else
8586 {
8593 }
8594 }
8598 {
8603 }
8604 else
8605 {
8606 /* Partial overlap. */
8610 {
8613 }
8617 }
8619 }
8624 {
8627 {
8630 {
8634 }
8637 }
8640 fldsize,
8650 }
8651 else
8652 {
8653 /* Partial overlap. */
8659 {
8663 fieldsize);
8664 }
8665 else
8666 {
8670 }
8673 }
8674 }
8676 }
8678 }
8679 }
8682 /* Subroutine of native_interpret_expr. Interpret the contents of
8683 the buffer PTR of length LEN as an INTEGER_CST of type TYPE.
8684 If the buffer cannot be interpreted, return NULL_TREE. */
8686 static tree
8688 {
8691 {
8697 {
8699 /* More work is needed when adding _BitInt support to PDP endian
8700 if limb is smaller than word, or if _BitInt limb ordering doesn't
8701 match target endianity here. */
8706 }
8707 else
8709 }
8710 else
8720 }
8723 /* Subroutine of native_interpret_expr. Interpret the contents of
8724 the buffer PTR of length LEN as a FIXED_CST of type TYPE.
8725 If the buffer cannot be interpreted, return NULL_TREE. */
8727 static tree
8729 {
8732 double_int result;
8733 FIXED_VALUE_TYPE fixed_value;
8743 }
8746 /* Subroutine of native_interpret_expr. Interpret the contents of
8747 the buffer PTR of length LEN as a REAL_CST of type TYPE.
8748 If the buffer cannot be interpreted, return NULL_TREE. */
8750 tree
8752 {
8756 /* There are always 32 bits in each long, no matter the size of
8757 the hosts long. We handle floating point representations with
8758 up to 192 bits. */
8759 REAL_VALUE_TYPE r;
8769 {
8770 /* Both OFFSET and BYTE index within a long;
8771 bitpos indexes the whole float. */
8774 {
8781 else
8783 }
8784 else
8785 {
8788 {
8789 /* Reverse bytes within each long, or within the entire float
8790 if it's smaller than a long (for HFmode). */
8793 }
8794 }
8798 }
8802 }
8805 /* Subroutine of native_interpret_expr. Interpret the contents of
8806 the buffer PTR of length LEN as a COMPLEX_CST of type TYPE.
8807 If the buffer cannot be interpreted, return NULL_TREE. */
8809 static tree
8811 {
8826 }
8828 /* Read a vector of type TYPE from the target memory image given by BYTES,
8829 which contains LEN bytes. The vector is known to be encodable using
8830 NPATTERNS interleaved patterns with NELTS_PER_PATTERN elements each.
8832 Return the vector on success, otherwise return null. */
8834 static tree
8838 {
8842 {
8843 /* This is the only case in which elements can be smaller than a byte.
8844 Element 0 is always in the lsb of the containing byte. */
8851 {
8858 }
8860 }
8868 {
8874 }
8876 }
8878 /* Subroutine of native_interpret_expr. Interpret the contents of
8879 the buffer PTR of length LEN as a VECTOR_CST of type TYPE.
8880 If the buffer cannot be interpreted, return NULL_TREE. */
8882 static tree
8884 {
8893 }
8896 /* Subroutine of fold_view_convert_expr. Interpret the contents of
8897 the buffer PTR of length LEN as a constant of type TYPE. For
8898 INTEGRAL_TYPE_P we return an INTEGER_CST, for SCALAR_FLOAT_TYPE_P
8899 we return a REAL_CST, etc... If the buffer cannot be interpreted,
8900 return NULL_TREE. */
8902 tree
8904 {
8906 {
8918 {
8919 /* For floating point values in composite modes, punt if this
8920 folding doesn't preserve bit representation. As the mode doesn't
8921 have fixed precision while GCC pretends it does, there could be
8922 valid values that GCC can't really represent accurately.
8923 See PR95450. Even for other modes, e.g. x86 XFmode can have some
8924 bit combinationations which GCC doesn't preserve. */
8934 }
8948 }
8949 }
8951 /* Returns true if we can interpret the contents of a native encoding
8952 as TYPE. */
8954 bool
8956 {
8958 {
8972 }
8973 }
8975 /* Attempt to interpret aggregate of TYPE from bytes encoded in target
8976 byte order at PTR + OFF with LEN bytes. Does not handle unions. */
8978 tree
8981 {
8984 {
8991 {
8995 }
9000 {
9005 {
9010 }
9014 eltsz);
9018 }
9020 }
9024 {
9033 {
9036 {
9037 poly_int64 bitoffset;
9042 else
9052 }
9053 else
9054 {
9072 HOST_WIDE_INT rpos;
9075 else
9076 {
9079 }
9088 }
9089 }
9092 {
9102 }
9106 {
9111 }
9118 {
9122 /* FIXME: Figure out how to handle PDP endian bitfields. */
9128 else
9132 }
9134 }
9136 }
9138 /* Routines for manipulation of native_encode_expr encoded data if the encoded
9139 or extracted constant positions and/or sizes aren't byte aligned. */
9141 /* Shift left the bytes in PTR of SZ elements by AMNT bits, carrying over the
9142 bits between adjacent elements. AMNT should be within
9143 [0, BITS_PER_UNIT).
9144 Example, AMNT = 2:
9145 00011111|11100000 << 2 = 01111111|10000000
9146 PTR[1] | PTR[0] PTR[1] | PTR[0]. */
9148 void
9151 {
9160 {
9166 {
9169 }
9170 }
9171 }
9173 /* Like shift_bytes_in_array_left but for big-endian.
9174 Shift right the bytes in PTR of SZ elements by AMNT bits, carrying over the
9175 bits between adjacent elements. AMNT should be within
9176 [0, BITS_PER_UNIT).
9177 Example, AMNT = 2:
9178 00011111|11100000 >> 2 = 00000111|11111000
9179 PTR[0] | PTR[1] PTR[0] | PTR[1]. */
9181 void
9184 {
9192 {
9199 }
9200 }
9202 /* Try to view-convert VECTOR_CST EXPR to VECTOR_TYPE TYPE by operating
9203 directly on the VECTOR_CST encoding, in a way that works for variable-
9204 length vectors. Return the resulting VECTOR_CST on success or null
9205 on failure. */
9207 static tree
9209 {
9221 /* We can only preserve the semantics of a stepped pattern if the new
9222 vector element is an integer of the same size. */
9227 /* The number of bits needed to encode one element from every pattern
9228 of the original vector. */
9229 unsigned int expr_sequence_bits
9232 /* The number of bits needed to encode one element from every pattern
9233 of the result. */
9234 unsigned int type_sequence_bits
9237 /* Don't try to read more bytes than are available, which can happen
9238 for constant-sized vectors if TYPE has larger elements than EXPR_TYPE.
9239 The general VIEW_CONVERT handling can cope with that case, so there's
9240 no point complicating things here. */
9243 BITS_PER_UNIT);
9248 /* Get enough bytes of EXPR to form the new encoding. */
9256 /* Reencode the bytes as TYPE. */
9260 }
9262 /* Fold a VIEW_CONVERT_EXPR of a constant expression EXPR to type
9263 TYPE at compile-time. If we're unable to perform the conversion
9264 return NULL_TREE. */
9266 static tree
9268 {
9269 /* We support up to 512-bit values (for V8DFmode). */
9273 /* Check that the host and target are sane. */
9286 }
9288 /* Build an expression for the address of T. Folds away INDIRECT_REF
9289 to avoid confusing the gimplify process. */
9291 tree
9293 {
9294 /* The size of the object is not relevant when talking about its address. */
9299 {
9304 }
9307 {
9312 }
9319 {
9324 }
9325 else
9329 }
9331 /* Build an expression for the address of T. */
9333 tree
9335 {
9339 }
9341 /* Fold a unary expression of code CODE and type TYPE with operand
9342 OP0. Return the folded expression if folding is successful.
9343 Otherwise, return NULL_TREE. */
9345 tree
9347 {
9348 tree tem;
9349 tree arg0;
9357 {
9360 {
9361 /* Don't use STRIP_NOPS, because signedness of argument type
9362 matters. */
9364 }
9365 else
9366 {
9367 /* Strip any conversions that don't change the mode. This
9368 is safe for every expression, except for a comparison
9369 expression because its signedness is derived from its
9370 operands.
9372 Note that this is done as an internal manipulation within
9373 the constant folder, in order to find the simplest
9374 representation of the arguments so that their form can be
9375 studied. In any cases, the appropriate type conversions
9376 should be put back in the tree that will get out of the
9377 constant folder. */
9379 }
9382 {
9385 {
9389 }
9390 }
9391 }
9398 {
9405 {
9419 /* If this was a conversion, and all we did was to move into
9420 inside the COND_EXPR, bring it back out. But leave it if
9421 it is a conversion from integer to integer and the
9422 result precision is no wider than a word since such a
9423 conversion is cheap and may be optimized away by combine,
9424 while it couldn't if it were outside the COND_EXPR. Then return
9425 so we don't get into an infinite recursion loop taking the
9426 conversion out and then back in. */
9438 && (INTEGRAL_TYPE_P
9451 }
9452 }
9455 {
9461 CASE_CONVERT:
9465 {
9466 /* If we have (type) (a CMP b) and type is an integral type, return
9467 new expression involving the new type. Canonicalize
9468 (type) (a CMP b) to (a CMP b) ? (type) true : (type) false for
9469 non-integral type.
9470 Do not fold the result as that would not simplify further, also
9471 folding again results in recursions. */
9481 }
9483 /* Handle (T *)&A.B.C for A being of type T and B and C
9484 living at offset zero. This occurs frequently in
9485 C++ upcasting and then accessing the base. */
9489 {
9491 tree offset;
9492 machine_mode mode;
9494 tree base
9498 /* If the reference was to a (constant) zero offset, we can use
9499 the address of the base if it has the same base type
9500 as the result type and the pointer type is unqualified. */
9508 }
9512 /* Detect assigning a bitfield. */
9514 && DECL_BIT_FIELD
9516 {
9517 /* Don't leave an assignment inside a conversion
9518 unless assigning a bitfield. */
9520 /* First do the assignment, then return converted constant. */
9525 }
9527 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
9528 constants (if x has signed type, the sign bit cannot be set
9529 in c). This folds extension into the BIT_AND_EXPR.
9530 ??? We don't do it for BOOLEAN_TYPE or ENUMERAL_TYPE because they
9531 very likely don't have maximal range for their precision and this
9532 transformation effectively doesn't preserve non-maximal ranges. */
9536 {
9547 <= HOST_BITS_PER_WIDE_INT
9549 {
9553 cst &= HOST_WIDE_INT_M1U
9557 && !flag_syntax_only
9560 {
9564 }
9565 }
9567 {
9572 }
9573 }
9575 /* Convert (T1)(X p+ Y) into ((T1)X p+ Y), for pointer type, when the new
9576 cast (T1)X will fold away. We assume that this happens when X itself
9577 is a cast. */
9581 {
9585 /* If -fsanitize=alignment, avoid this optimization in GENERIC
9586 when the pointed type needs higher alignment than
9587 the p+ first operand's pointed type. */
9594 /* Similarly, avoid this optimization in GENERIC for -fsanitize=null
9595 when type is a reference type and arg00's type is not,
9596 because arg00 could be validly nullptr and if arg01 doesn't return,
9597 we don't want false positive binding of reference to nullptr. */
9599 && !in_gimple_form
9606 }
9608 /* Convert (T1)(~(T2)X) into ~(T1)X if T1 and T2 are integral types
9609 of the same precision, and X is an integer type not narrower than
9610 types T1 or T2, i.e. the cast (T2)X isn't an extension. */
9616 {
9622 }
9624 /* Convert (T1)(X * Y) into (T1)X * (T1)Y if T1 is narrower than the
9625 type of X and Y (integer types only). */
9632 {
9633 /* Be careful not to introduce new overflows. */
9634 tree mult_type;
9637 else
9641 {
9648 }
9649 }
9655 {
9662 }
9673 /* Convert fabs((double)float) into (double)fabsf(float). */
9676 {
9682 targ0));
9683 }
9687 /* Convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
9706 /* Note that the operand of this must be an int
9707 and its values must be 0 or 1.
9708 ("true" is a fixed value perhaps depending on the language,
9709 but we don't handle values other than 1 correctly yet.) */
9716 /* Fold *&X to X if X is an lvalue. */
9718 {
9725 }
9731 }
9734 /* If the operation was a conversion do _not_ mark a resulting constant
9735 with TREE_OVERFLOW if the original constant was not. These conversions
9736 have implementation defined behavior and retaining the TREE_OVERFLOW
9737 flag here would confuse later passes such as VRP. */
9738 tree
9741 {
9750 }
9752 /* Fold a binary bitwise/truth expression of code CODE and type TYPE with
9753 operands OP0 and OP1. LOC is the location of the resulting expression.
9754 ARG0 and ARG1 are the NOP_STRIPed results of OP0 and OP1.
9755 Return the folded expression if folding is successful. Otherwise,
9756 return NULL_TREE. */
9757 static tree
9760 {
9761 tree tem;
9763 /* We only do these simplifications if we are optimizing. */
9767 /* Check for things like (A || B) && (A || C). We can convert this
9768 to A || (B && C). Note that either operator can be any of the four
9769 truth and/or operations and the transformation will still be
9770 valid. Also note that we only care about order for the
9771 ANDIF and ORIF operators. If B contains side effects, this
9772 might change the truth-value of A. */
9779 {
9799 /* This case if tricky because we must either have commutative
9800 operators or else A10 must not have side-effects. */
9806 a01);
9807 }
9809 /* See if we can build a range comparison. */
9815 {
9819 }
9823 {
9827 }
9829 /* Check for the possibility of merging component references. If our
9830 lhs is another similar operation, try to merge its rhs with our
9831 rhs. Then try to merge our lhs and rhs. */
9842 logical_op_non_short_circuit
9850 {
9857 /* Transform ((A AND-IF B) AND[-IF] C) into (A AND-IF (B AND C)),
9858 or ((A OR-IF B) OR[-IF] C) into (A OR-IF (B OR C))
9859 We don't want to pack more than two leafs to a non-IF AND/OR
9860 expression.
9861 If tree-code of left-hand operand isn't an AND/OR-IF code and not
9862 equal to IF-CODE, then we don't want to add right-hand operand.
9863 If the inner right-hand side of left-hand operand has
9864 side-effects, or isn't simple, then we can't add to it,
9865 as otherwise we might destroy if-sequence. */
9868 /* Needed for sequence points to handle trappings, and
9869 side-effects. */
9871 {
9873 arg1);
9875 tem);
9876 }
9877 /* Same as above but for (A AND[-IF] (B AND-IF C)) -> ((A AND B) AND-IF C),
9878 or (A OR[-IF] (B OR-IF C) -> ((A OR B) OR-IF C). */
9881 /* Needed for sequence points to handle trappings, and
9882 side-effects. */
9884 {
9889 }
9890 /* Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
9891 into (A OR B).
9892 For sequence point consistancy, we need to check for trapping,
9893 and side-effects. */
9897 }
9900 }
9902 /* Helper that tries to canonicalize the comparison ARG0 CODE ARG1
9903 by changing CODE to reduce the magnitude of constants involved in
9904 ARG0 of the comparison.
9905 Returns a canonicalized comparison tree if a simplification was
9906 possible, otherwise returns NULL_TREE.
9907 Set *STRICT_OVERFLOW_P to true if the canonicalization is only
9908 valid if signed overflow is undefined. */
9910 static tree
9914 {
9919 /* Match A +- CST code arg1. We can change this only if overflow
9920 is undefined. */
9923 /* In principle pointers also have undefined overflow behavior,
9924 but that causes problems elsewhere. */
9931 /* Identify the constant in arg0 and its sign. */
9935 /* Overflowed constants and zero will cause problems. */
9940 /* See if we can reduce the magnitude of the constant in
9941 arg0 by changing the comparison code. */
9942 /* A - CST < arg1 -> A - CST-1 <= arg1. */
9946 /* A + CST > arg1 -> A + CST-1 >= arg1. */
9950 /* A + CST <= arg1 -> A + CST-1 < arg1. */
9954 /* A - CST >= arg1 -> A - CST-1 > arg1. */
9958 else
9962 /* Now build the constant reduced in magnitude. But not if that
9963 would produce one outside of its types range. */
9979 }
9981 /* Canonicalize the comparison ARG0 CODE ARG1 with type TYPE with undefined
9982 overflow further. Try to decrease the magnitude of constants involved
9983 by changing LE_EXPR and GE_EXPR to LT_EXPR and GT_EXPR or vice versa
9984 and put sole constants at the second argument position.
9985 Returns the canonicalized tree if changed, otherwise NULL_TREE. */
9987 static tree
9990 {
9991 tree t;
9996 /* Try canonicalization by simplifying arg0. */
10001 {
10005 }
10007 /* Try canonicalization by simplifying arg1 using the swapped
10008 comparison. */
10016 }
10018 /* Return whether BASE + OFFSET + BITPOS may wrap around the address
10019 space. This is used to avoid issuing overflow warnings for
10020 expressions like &p->x which cannot wrap. */
10022 static bool
10024 {
10031 poly_wide_int wi_offset;
10037 else
10042 precision);
10057 /* We can do slightly better for SIZE if we have an ADDR_EXPR of an
10058 array. */
10067 }
10069 /* Return a positive integer when the symbol DECL is known to have
10070 a nonzero address, zero when it's known not to (e.g., it's a weak
10071 symbol), and a negative integer when the symbol is not yet in the
10072 symbol table and so whether or not its address is zero is unknown.
10073 For function local objects always return positive integer. */
10074 static int
10076 {
10077 /* Normally, don't do anything for variables and functions before symtab is
10078 built; it is quite possible that DECL will be declared weak later.
10079 But if folding_initializer, we need a constant answer now, so create
10080 the symtab entry and prevent later weak declaration. */
10083 = (folding_initializer
10088 /* Function local objects are never NULL. */
10096 }
10098 /* Subroutine of fold_binary. This routine performs all of the
10099 transformations that are common to the equality/inequality
10100 operators (EQ_EXPR and NE_EXPR) and the ordering operators
10101 (LT_EXPR, LE_EXPR, GE_EXPR and GT_EXPR). Callers other than
10102 fold_binary should call fold_binary. Fold a comparison with
10103 tree code CODE and type TYPE with operands OP0 and OP1. Return
10104 the folded comparison or NULL_TREE. */
10106 static tree
10109 {
10119 /* For comparisons of pointers we can decompose it to a compile time
10120 comparison of the base objects and the offsets into the object.
10121 This requires at least one operand being an ADDR_EXPR or a
10122 POINTER_PLUS_EXPR to do more than the operand_equal_p test below. */
10128 {
10131 machine_mode mode;
10135 /* Get base and offset for the access. Strip ADDR_EXPR for
10136 get_inner_reference, but put it back by stripping INDIRECT_REF
10137 off the base object if possible. indirect_baseN will be true
10138 if baseN is not an address but refers to the object itself. */
10141 {
10142 base0
10148 else
10150 }
10152 {
10156 {
10157 base0
10163 else
10165 }
10168 else
10172 {
10179 }
10180 }
10184 {
10185 base1
10191 else
10193 }
10195 {
10199 {
10200 base1
10206 else
10208 }
10211 else
10215 {
10222 }
10223 }
10225 /* If we have equivalent bases we might be able to simplify. */
10229 {
10230 /* We can fold this expression to a constant if the non-constant
10231 offset parts are equal. */
10235 && (equality_code
10236 || (indirect_base0
10239 {
10245 "occur when comparing P +- C1 with "
10247 WARN_STRICT_OVERFLOW_CONDITIONAL);
10250 {
10288 }
10289 }
10290 /* We can simplify the comparison to a comparison of the variable
10291 offset parts if the constant offset parts are equal.
10292 Be careful to use signed sizetype here because otherwise we
10293 mess with array offsets in the wrong way. This is possible
10294 because pointer arithmetic is restricted to retain within an
10295 object and overflow on pointer differences is undefined as of
10296 6.5.6/8 and /9 with respect to the signed ptrdiff_t. */
10298 && (equality_code
10299 || (indirect_base0
10302 {
10303 /* By converting to signed sizetype we cover middle-end pointer
10304 arithmetic which operates on unsigned pointer types of size
10305 type size and ARRAY_REF offsets which are properly sign or
10306 zero extended from their type in case it is narrower than
10307 sizetype. */
10310 else
10314 else
10321 "occur when comparing P +- C1 with "
10323 WARN_STRICT_OVERFLOW_COMPARISON);
10326 }
10327 }
10328 /* For equal offsets we can simplify to a comparison of the
10329 base addresses. */
10331 && (indirect_base0
10333 && (indirect_base1
10338 {
10344 }
10345 /* Comparison between an ordinary (non-weak) symbol and a null
10346 pointer can be eliminated since such symbols must have a non
10347 null address. In C, relational expressions between pointers
10348 to objects and null pointers are undefined. The results
10349 below follow the C++ rules with the additional property that
10350 every object pointer compares greater than a null pointer.
10351 */
10354 /* Avoid folding references to struct members at offset 0 to
10355 prevent tests like '&ptr->firstmember == 0' from getting
10356 eliminated. When ptr is null, although the -> expression
10357 is strictly speaking invalid, GCC retains it as a matter
10358 of QoI. See PR c/44555. */
10361 && indirect_base0
10362 /* The caller guarantees that when one of the arguments is
10363 constant (i.e., null in this case) it is second. */
10365 {
10367 {
10378 }
10379 }
10380 }
10382 /* Transform comparisons of the form X +- C1 CMP Y +- C2 to
10383 X CMP Y +- C2 +- C1 for signed X, Y. This is valid if
10384 the resulting offset is smaller in absolute value than the
10385 original one and has the same sign. */
10394 {
10399 tree cst;
10401 "occur when combining constants around "
10404 /* Put the constant on the side where it doesn't overflow and is
10405 of lower absolute value and of same sign than before. */
10412 {
10415 variable1,
10419 }
10427 {
10433 variable2);
10434 }
10435 }
10441 /* If we are comparing an expression that just has comparisons
10442 of two integer values, arithmetic expressions of those comparisons,
10443 and constants, we can simplify it. There are only three cases
10444 to check: the two values can either be equal, the first can be
10445 greater, or the second can be greater. Fold the expression for
10446 those three values. Since each value must be 0 or 1, we have
10447 eight possibilities, each of which corresponds to the constant 0
10448 or 1 or one of the six possible comparisons.
10450 This handles common cases like (a > b) == 0 but also handles
10451 expressions like ((x > y) - (y > x)) > 0, which supposedly
10452 occur in macroized code. */
10455 {
10459 /* Don't handle degenerate cases here; they should already
10460 have been handled anyway. */
10469 {
10473 /* We can't just pass T to eval_subst in case cval1 or cval2
10474 was the same as ARG1. */
10476 tree high_result
10480 arg1);
10481 tree equal_result
10485 arg1);
10486 tree low_result
10490 arg1);
10492 /* All three of these results should be 0 or 1. Confirm they are.
10493 Then use those values to select the proper code to use. */
10498 {
10499 /* Make a 3-bit mask with the high-order bit being the
10500 value for `>', the next for '=', and the low for '<'. */
10504 {
10506 /* Always false. */
10527 /* Always true. */
10529 }
10532 }
10533 }
10534 }
10537 }
10540 /* Subroutine of fold_binary. Optimize complex multiplications of the
10541 form z * conj(z), as pow(realpart(z),2) + pow(imagpart(z),2). The
10542 argument EXPR represents the expression "z" of type TYPE. */
10544 static tree
10546 {
10551 {
10554 }
10556 {
10559 }
10560 else
10561 {
10565 }
10574 }
10577 /* Helper function for fold_vec_perm. Store elements of VECTOR_CST or
10578 CONSTRUCTOR ARG into array ELTS, which has NELTS elements, and return
10579 true if successful. */
10581 static bool
10583 {
10588 {
10591 }
10593 {
10599 else
10601 }
10602 else
10608 }
10610 /* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
10611 mask for VLA vec_perm folding.
10612 REASON if specified, will contain the reason why SEL is not suitable.
10613 Used only for debugging and unit-testing. */
10615 static bool
10619 {
10626 {
10630 }
10632 /* We want to avoid cases where sel.length is not a multiple of npatterns.
10633 For eg: sel.length = 2 + 2x, and sel npatterns = 4. */
10634 poly_uint64 esel;
10636 {
10640 }
10646 {
10649 HOST_WIDE_INT step;
10651 {
10655 }
10656 // FIXME: Punt on step < 0 for now, revisit later.
10663 {
10667 }
10669 /* Ensure that stepped sequence of the pattern selects elements
10670 only from the same input vector. */
10679 {
10683 }
10685 /* Ensure that the stepped sequence always selects from the same
10686 input pattern. */
10687 unsigned arg_npatterns
10692 {
10696 }
10697 }
10700 }
10702 /* Try to fold permutation of ARG0 and ARG1 with SEL selector when
10703 the input vectors are VECTOR_CST. Return NULL_TREE otherwise.
10704 REASON has same purpose as described in
10705 valid_mask_for_fold_vec_perm_cst_p. */
10707 static tree
10710 {
10714 /* (1) If SEL is a suitable mask as determined by
10715 valid_mask_for_fold_vec_perm_cst_p, then:
10716 res_npatterns = max of npatterns between ARG0, ARG1, and SEL
10717 res_nelts_per_pattern = max of nelts_per_pattern between
10718 ARG0, ARG1 and SEL.
10719 (2) If SEL is not a suitable mask, and TYPE is VLS then:
10720 res_npatterns = nelts in result vector.
10721 res_nelts_per_pattern = 1.
10722 This exception is made so that VLS ARG0, ARG1 and SEL work as before. */
10724 {
10725 res_npatterns
10730 res_nelts_per_pattern
10736 }
10738 {
10741 }
10742 else
10747 {
10750 poly_uint64 r;
10753 /* Punt if sel[i] /trunc_div len cannot be determined,
10754 because the input vector to be chosen will depend on
10755 runtime vector length.
10756 For example if len == 4 + 4x, and sel[i] == 4,
10757 If len at runtime equals 4, we choose arg1[0].
10758 For any other value of len > 4 at runtime, we choose arg0[4].
10759 which makes the element choice dependent on runtime vector length. */
10761 {
10765 }
10767 /* sel[i] % len will give the index of element in the chosen input
10768 vector. For example if sel[i] == 5 + 4x and len == 4 + 4x,
10769 we will choose arg1[1] since (5 + 4x) % (4 + 4x) == 1. */
10771 {
10775 }
10780 }
10783 }
10785 /* Attempt to fold vector permutation of ARG0 and ARG1 vectors using SEL
10786 selector. Return the folded VECTOR_CST or CONSTRUCTOR if successful,
10787 NULL_TREE otherwise. */
10789 tree
10791 {
10807 /* For fall back case, we want to ensure we have VLS vectors
10808 with equal length. */
10822 {
10823 HOST_WIDE_INT index;
10827 }
10829 }
10831 /* Try to fold a pointer difference of type TYPE two address expressions of
10832 array references AREF0 and AREF1 using location LOC. Return a
10833 simplified expression for the difference or NULL_TREE. */
10835 static tree
10839 {
10844 /* If the bases are array references as well, recurse. If the bases
10845 are pointer indirections compute the difference of the pointers.
10846 If the bases are equal, we are set. */
10849 && (base_offset
10851 use_pointer_diff)))
10854 && (base_offset
10855 = use_pointer_diff
10865 {
10871 base_offset,
10874 }
10876 }
10878 /* If the real or vector real constant CST of type TYPE has an exact
10879 inverse, return it, else return NULL. */
10881 tree
10883 {
10884 REAL_VALUE_TYPE r;
10885 tree unit_type;
10886 machine_mode mode;
10889 {
10899 {
10903 tree_vector_builder elts;
10908 {
10913 }
10916 }
10920 }
10921 }
10923 /* Mask out the tz least significant bits of X of type TYPE where
10924 tz is the number of trailing zeroes in Y. */
10925 static wide_int
10927 {
10932 }
10934 /* Return true when T is an address and is known to be nonzero.
10935 For floating point we further ensure that T is not denormal.
10936 Similar logic is present in nonzero_address in rtlanal.h.
10938 If the return value is based on the assumption that signed overflow
10939 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
10940 change *STRICT_OVERFLOW_P. */
10942 static bool
10944 {
10948 /* Doing something useful for floating point would need more work. */
10954 {
10957 strict_overflow_p);
10963 strict_overflow_p);
10971 }
10974 {
10977 strict_overflow_p);
10985 strict_overflow_p);
10999 strict_overflow_p);
11003 strict_overflow_p);
11006 {
11018 }
11022 }
11024 }
11026 /* Return true when T is an address and is known to be nonzero.
11027 Handle warnings about undefined signed overflow. */
11029 bool
11031 {
11038 "determining that expression is always "
11040 WARN_STRICT_OVERFLOW_MISC);
11042 }
11044 /* Return true if T is known not to be equal to an integer W. */
11046 bool
11048 {
11049 int_range_max vr;
11051 {
11061 else
11066 /* If T has some known zero bits and W has any of those bits set,
11067 then T is known not to be equal to W. */
11075 }
11076 }
11078 /* Fold a binary expression of code CODE and type TYPE with operands
11079 OP0 and OP1. LOC is the location of the resulting expression.
11080 Return the folded expression if folding is successful. Otherwise,
11081 return NULL_TREE. */
11083 tree
11086 {
11101 /* Strip any conversions that don't change the mode. This is
11102 safe for every expression, except for a comparison expression
11103 because its signedness is derived from its operands. So, in
11104 the latter case, only strip conversions that don't change the
11105 signedness. MIN_EXPR/MAX_EXPR also need signedness of arguments
11106 preserved.
11108 Note that this is done as an internal manipulation within the
11109 constant folder, in order to find the simplest representation
11110 of the arguments so that their form can be studied. In any
11111 cases, the appropriate type conversions should be put back in
11112 the tree that will get out of the constant folder. */
11115 {
11118 }
11119 else
11120 {
11123 }
11125 /* Note that TREE_CONSTANT isn't enough: static var addresses are
11126 constant but we can't do arithmetic on them. */
11128 {
11131 {
11135 }
11136 }
11138 /* If this is a commutative operation, and ARG0 is a constant, move it
11139 to ARG1 to reduce the number of tests below. */
11144 /* Likewise if this is a comparison, and ARG0 is a constant, move it
11145 to ARG1 to reduce the number of tests below. */
11154 /* ARG0 is the first operand of EXPR, and ARG1 is the second operand.
11156 First check for cases where an arithmetic operation is applied to a
11157 compound, conditional, or comparison operation. Push the arithmetic
11158 operation inside the compound or conditional to see if any folding
11159 can then be done. Convert comparison to conditional for this purpose.
11160 The also optimizes non-constant cases that used to be done in
11161 expand_expr.
11163 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
11164 one of the operands is a comparison and the other is a comparison, a
11165 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
11166 code below would make the expression more complex. Change it to a
11167 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
11168 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
11181 {
11185 boolean_type_node,
11193 }
11197 {
11199 {
11204 tem);
11205 }
11207 {
11212 tem);
11213 }
11218 {
11224 }
11229 {
11235 }
11236 }
11239 {
11241 /* MEM[&MEM[p, CST1], CST2] -> MEM[p, CST1 + CST2]. */
11244 {
11250 }
11252 /* MEM[&a.b, CST2] -> MEM[&a, offsetof (a, b) + CST2]. */
11255 {
11256 tree base;
11257 poly_int64 coffset;
11266 }
11271 /* INT +p INT -> (PTR)(INT + INT). Stripping types allows for this. */
11277 arg1),
11279 arg0)));
11285 {
11286 /* X + (X / CST) * -CST is X % CST. */
11291 {
11300 cst0));
11301 }
11302 }
11304 /* Handle (A1 * C1) + (A2 * C2) with A1, A2 or C1, C2 being the same or
11305 one. Make sure the type is not saturating and has the signedness of
11306 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11307 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11314 {
11318 }
11321 {
11322 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
11323 (plus (plus (mult) (mult)) (foo)) so that we can
11324 take advantage of the factoring cases below. */
11333 {
11339 else
11352 parg0),
11354 marg)),
11358 return
11364 parg1)));
11365 }
11366 }
11367 else
11368 {
11369 /* Fold __complex__ ( x, 0 ) + __complex__ ( 0, y )
11370 to __complex__ ( x, y ). This is not the same for SNaNs or
11371 if signed zeros are involved. */
11375 {
11382 {
11386 {
11392 }
11394 {
11400 }
11401 }
11402 }
11404 /* Convert a + (b*c + d*e) into (a + b*c) + d*e.
11405 We associate floats only if the user has specified
11406 -fassociative-math. */
11410 {
11415 {
11416 tree tree0;
11419 }
11420 }
11421 /* Convert (b*c + d*e) + a into b*c + (d*e +a).
11422 We associate floats only if the user has specified
11423 -fassociative-math. */
11427 {
11432 {
11433 tree tree0;
11436 }
11437 }
11438 }
11440 bit_rotate:
11441 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
11442 is a rotate of A by C1 bits. */
11443 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
11444 is a rotate of A by B bits.
11445 Similarly for (A << B) | (A >> (-B & C3)) where C3 is Z-1,
11446 though in this case CODE must be | and not + or ^, otherwise
11447 it doesn't return A when B is 0. */
11448 {
11450 tree rtype;
11459 /* Only create rotates in complete modes. Other cases are not
11460 expanded properly. */
11463 {
11476 {
11481 }
11486 {
11489 code0 == LSHIFT_EXPR
11492 }
11494 {
11504 {
11508 orig_tree01);
11510 }
11511 }
11515 {
11526 {
11530 orig_tree01);
11532 }
11533 }
11534 }
11535 }
11537 associate:
11538 /* In most languages, can't associate operations on floats through
11539 parentheses. Rather than remember where the parentheses were, we
11540 don't associate floats at all, unless the user has specified
11541 -fassociative-math.
11542 And, we need to make sure type is not saturating. */
11547 {
11553 /* Split both trees into variables, constants, and literals. Then
11554 associate each group together, the constants with literals,
11555 then the result with variables. This increases the chances of
11556 literals being recombined later and of generating relocatable
11557 expressions for the sum of a constant and literal. */
11565 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
11569 /* With undefined overflow prefer doing association in a type
11570 which wraps on overflow, if that is one of the operand types. */
11573 {
11581 }
11583 /* With undefined overflow we can only associate constants with one
11584 variable, and constants whose association doesn't overflow. */
11587 {
11589 {
11590 /* ??? If split_tree would handle NEGATE_EXPR we could
11591 simply reject these cases and the allowed cases would
11592 be the var0/minus_var1 ones. */
11598 {
11601 }
11608 {
11611 }
11617 /* The only case we can still associate with two variables
11618 is if they cancel out. */
11622 }
11628 }
11630 /* Only do something if we found more than two objects. Otherwise,
11631 nothing has changed and we risk infinite recursion. */
11639 {
11651 {
11655 }
11657 {
11661 }
11663 /* Preserve the MINUS_EXPR if the negative part of the literal is
11664 greater than the positive part. Otherwise, the multiplicative
11665 folding code (i.e extract_muldiv) may be fooled in case
11666 unsigned constants are subtracted, like in the following
11667 example: ((X*2 + 4) - 8U)/2. */
11669 {
11673 /* But avoid ending up with only negated parts. */
11675 {
11679 }
11680 else
11681 {
11685 }
11686 }
11688 /* Don't introduce overflows through reassociation. */
11693 /* Eliminate lit0 and minus_lit0 to con0 and minus_con0. */
11700 /* Eliminate minus_con0. */
11702 {
11709 else
11712 }
11714 /* Eliminate minus_var0. */
11716 {
11720 else
11723 }
11725 return
11728 }
11729 }
11735 /* Fold &a[i] - &a[j] to i-j. */
11740 {
11744 code
11748 }
11750 /* Further transformations are not for pointers. */
11754 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
11757 /* If arg0 is e.g. unsigned int and type is int, then this could
11758 introduce UB, because if A is INT_MIN at runtime, the original
11759 expression can be well defined while the latter is not.
11760 See PR83269. */
11769 /* Fold __complex__ ( x, 0 ) - __complex__ ( 0, y ) to
11770 __complex__ ( x, -y ). This is not the same for SNaNs or if
11771 signed zeros are involved. */
11775 {
11782 {
11786 {
11788 arg1r ? arg1r
11793 }
11795 {
11799 arg1i ? arg1i
11802 }
11803 }
11804 }
11806 /* A - B -> A + (-B) if B is easily negatable. */
11810 /* Avoid this transformation if B is a positive REAL_CST. */
11818 /* Handle (A1 * C1) - (A2 * C2) with A1, A2 or C1, C2 being the same or
11819 one. Make sure the type is not saturating and has the signedness of
11820 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11821 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11828 {
11832 }
11838 {
11839 /* Transform x * -C into -x * C if x is easily negatable. */
11854 {
11857 "occur when simplifying "
11859 WARN_STRICT_OVERFLOW_MISC);
11861 }
11863 /* Optimize z * conj(z) for integer complex numbers. */
11870 }
11871 else
11872 {
11873 /* Fold z * +-I to __complex__ (-+__imag z, +-__real z).
11874 This is not the same for NaNs or if signed zeros are
11875 involved. */
11881 {
11884 return
11890 return
11895 }
11897 /* Optimize z * conj(z) for floating point complex numbers.
11898 Guarded by flag_unsafe_math_optimizations as non-finite
11899 imaginary components don't produce scalar results. */
11908 }
11912 /* Canonicalize (X & C1) | C2. */
11916 {
11921 /* If (C1&C2) == C1, then (X&C1)|C2 becomes (X,C2). */
11929 /* If (C1|C2) == ~0 then (X&C1)|C2 becomes X|C2. */
11931 {
11934 }
11936 /* Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2,
11937 unless (C1 & ~C2) | (C2 & C3) for some C3 is a mask of some
11938 mode which allows further optimizations. */
11943 {
11948 {
11951 }
11952 }
11955 {
11960 }
11961 }
11963 /* See if this can be simplified into a rotate first. If that
11964 is unsuccessful continue in the association code. */
11968 /* Fold (X & 1) ^ 1 as (X & 1) == 0. */
11976 /* See if this can be simplified into a rotate first. If that
11977 is unsuccessful continue in the association code. */
11981 /* Fold (X ^ 1) & 1 as (X & 1) == 0. */
11986 {
11987 tree tem2;
11994 }
11995 /* Fold ~X & 1 as (X & 1) == 0. */
11999 {
12000 tree tem2;
12007 }
12008 /* Fold !X & 1 as X == 0. */
12011 {
12015 }
12017 /* Fold (X * Y) & -(1 << CST) to X * Y if Y is a constant
12018 multiple of 1 << CST. */
12020 {
12027 }
12029 /* Fold (X * CST1) & CST2 to zero if we can, or drop known zero
12030 bits from CST2. */
12034 {
12036 wide_int masked
12043 {
12044 /* Avoid the transform if arg1 is a mask of some
12045 mode which allows further optimizations. */
12052 }
12053 }
12055 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
12058 {
12063 return
12065 }
12070 /* Don't touch a floating-point divide by zero unless the mode
12071 of the constant can represent infinity. */
12077 /* (-A) / (-B) -> A / B */
12089 /* Fall through */
12092 /* Simplify A / (B << N) where A and B are positive and B is
12093 a power of 2, to A >> (N + log2(B)). */
12098 {
12101 {
12109 WARN_STRICT_OVERFLOW_MISC);
12115 }
12116 }
12118 /* Fall through */
12126 /* Convert -A / -B to A / B when the type is signed and overflow is
12127 undefined. */
12131 {
12134 "when distributing negation across "
12136 WARN_STRICT_OVERFLOW_MISC);
12141 }
12145 {
12148 "when distributing negation across "
12150 WARN_STRICT_OVERFLOW_MISC);
12155 }
12157 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
12158 operation, EXACT_DIV_EXPR.
12160 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
12161 At one time others generated faster code, it's not clear if they do
12162 after the last round to changes to the DIV code in expmed.cc. */
12173 {
12177 WARN_STRICT_OVERFLOW_MISC);
12179 }
12191 {
12195 WARN_STRICT_OVERFLOW_MISC);
12197 }
12205 /* Since negative shift count is not well-defined,
12206 don't try to compute it in the compiler. */
12212 /* If we have a rotate of a bit operation with the rotate count and
12213 the second operand of the bit operation both constant,
12214 permute the two operations. */
12220 {
12228 }
12230 /* Two consecutive rotates adding up to the some integer
12231 multiple of the precision of the type can be ignored. */
12247 /* Note that the operands of this must be ints
12248 and their values must be 0 or 1.
12249 ("true" is a fixed value perhaps depending on the language.) */
12250 /* If first arg is constant zero, return it. */
12253 /* FALLTHRU */
12255 /* If either arg is constant true, drop it. */
12259 /* Preserve sequence points. */
12262 /* If second arg is constant zero, result is zero, but first arg
12263 must be evaluated. */
12266 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
12267 case will be handled here. */
12271 /* !X && X is always false. */
12275 /* X && !X is always false. */
12280 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
12281 means A >= Y && A != MAX, but in this case we know that
12282 A < X <= MAX. */
12286 {
12298 }
12307 /* Note that the operands of this must be ints
12308 and their values must be 0 or true.
12309 ("true" is a fixed value perhaps depending on the language.) */
12310 /* If first arg is constant true, return it. */
12313 /* FALLTHRU */
12315 /* If either arg is constant zero, drop it. */
12319 /* Preserve sequence points. */
12322 /* If second arg is constant true, result is true, but we must
12323 evaluate first arg. */
12326 /* Likewise for first arg, but note this only occurs here for
12327 TRUTH_OR_EXPR. */
12331 /* !X || X is always true. */
12335 /* X || !X is always true. */
12340 /* (X && !Y) || (!X && Y) is X ^ Y */
12343 {
12360 }
12369 /* If the second arg is constant zero, drop it. */
12372 /* If the second arg is constant true, this is a logical inversion. */
12374 {
12377 }
12378 /* Identical arguments cancel to zero. */
12382 /* !X ^ X is always true. */
12387 /* X ^ !X is always true. */
12403 /* bool_var != 1 becomes !bool_var. */
12410 /* bool_var == 0 becomes !bool_var. */
12417 /* !exp != 0 becomes !exp */
12422 /* If this is an EQ or NE comparison with zero and ARG0 is
12423 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
12424 two operations, but the latter can be done in one less insn
12425 on machines that have only two-operand insns or on which a
12426 constant cannot be the first operand. */
12429 {
12434 {
12442 }
12445 {
12453 }
12454 }
12456 /* If this is a comparison of a field, we may be able to simplify it. */
12459 /* Handle the constant case even without -O
12460 to make sure the warnings are given. */
12462 {
12466 }
12468 /* Optimize comparisons of strlen vs zero to a compare of the
12469 first character of the string vs zero. To wit,
12470 strlen(ptr) == 0 => *ptr == 0
12471 strlen(ptr) != 0 => *ptr != 0
12472 Other cases should reduce to one of these two (or a constant)
12473 due to the return value of strlen being unsigned. */
12475 {
12483 {
12484 tree ptrtype
12486 TYPE_QUAL_CONST));
12492 }
12493 }
12495 /* Fold (X >> C) != 0 into X < 0 if C is one less than the width
12496 of X. Similarly fold (X >> C) == 0 into X >= 0. */
12500 {
12505 {
12507 {
12510 }
12513 }
12514 }
12516 /* Fold (~X & C) == 0 into (X & C) != 0 and (~X & C) != 0 into
12517 (X & C) == 0 when C is a single bit. */
12522 {
12529 arg1));
12530 }
12532 /* Fold ((X & C) ^ C) eq/ne 0 into (X & C) ne/eq 0, when the
12533 constant C is a power of two, i.e. a single bit. */
12540 {
12544 }
12546 /* Likewise, fold ((X ^ C) & C) eq/ne 0 into (X & C) ne/eq 0,
12547 when is C is a power of two, i.e. a single bit. */
12554 {
12560 }
12564 {
12571 /* Optimize (X ^ Z) op (Y ^ Z) as X op Y, and symmetries.
12572 operand_equal_p guarantees no side-effects so we don't need
12573 to use omit_one_operand on Z. */
12577 arg10));
12581 arg11));
12585 arg10));
12589 arg11));
12591 /* Optimize (X ^ C1) op (Y ^ C2) as (X ^ (C1 ^ C2)) op Y. */
12594 {
12600 }
12601 }
12603 /* Attempt to simplify equality/inequality comparisons of complex
12604 values. Only lower the comparison if the result is known or
12605 can be simplified to a single scalar comparison. */
12610 {
12615 {
12618 }
12619 else
12620 {
12623 }
12626 {
12629 }
12630 else
12631 {
12634 }
12638 {
12640 {
12645 }
12646 else
12647 {
12652 }
12653 }
12657 {
12659 {
12664 }
12665 else
12666 {
12671 }
12672 }
12673 }
12685 /* Transform comparisons of the form X +- C CMP X. */
12690 {
12695 /* (X - c) > X becomes false. */
12701 /* Likewise (X + c) < X becomes false. */
12707 /* Convert (X - c) <= X to true. */
12714 /* Convert (X + c) >= X to true. */
12720 }
12722 /* If we are comparing an ABS_EXPR with a constant, we can
12723 convert all the cases into explicit comparisons, but they may
12724 well not be faster than doing the ABS and one comparison.
12725 But ABS (X) <= C is a range comparison, which becomes a subtraction
12726 and a comparison, and is probably faster. */
12740 /* Convert ABS_EXPR<x> >= 0 to true. */
12747 {
12750 "when simplifying comparison of "
12752 WARN_STRICT_OVERFLOW_CONDITIONAL);
12755 arg0);
12756 }
12758 /* Convert ABS_EXPR<x> < 0 to false. */
12763 {
12766 "when simplifying comparison of "
12768 WARN_STRICT_OVERFLOW_CONDITIONAL);
12771 arg0);
12772 }
12774 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
12775 and similarly for >= into !=. */
12785 /* Similarly for X < (cast) (1 << Y). But cast can't be narrowing,
12786 otherwise Y might be >= # of bits in X's type and thus e.g.
12787 (unsigned char) (1 << Y) for Y 15 might be 0.
12788 If the cast is widening, then 1 << Y should have unsigned type,
12789 otherwise if Y is number of bits in the signed shift type minus 1,
12790 we can't optimize this. E.g. (unsigned long long) (1 << Y) for Y
12791 31 might be 0xffffffff80000000. */
12804 {
12810 }
12822 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
12823 {
12835 }
12840 /* When pedantic, a compound expression can be neither an lvalue
12841 nor an integer constant expression. */
12844 /* Don't let (0, 0) be null pointer constant. */
12852 }
12854 /* For constants M and N, if M == (1LL << cst) - 1 && (N & M) == M,
12855 ((A & N) + B) & M -> (A + B) & M
12856 Similarly if (N & M) == 0,
12857 ((A | N) + B) & M -> (A + B) & M
12858 and for - instead of + (or unary - instead of +)
12859 and/or ^ instead of |.
12860 If B is constant and (B & M) == 0, fold into A & M.
12862 This function is a helper for match.pd patterns. Return non-NULL
12863 type in which the simplified operation should be performed only
12864 if any optimization is possible.
12866 ARG1 is M above, ARG00 is left operand of +/-, if CODE00 is BIT_*_EXPR,
12867 then ARG00{0,1} are operands of that bitop, otherwise CODE00 is ERROR_MARK.
12868 Similarly for ARG01, CODE01 and ARG01{0,1}, just for the right operand of
12869 +/-. */
12870 tree
12875 {
12890 wide_int cst0;
12892 /* Now we know that arg0 is (C + D) or (C - D) or -C and
12893 arg1 (M) is == (1LL << cst) - 1.
12894 Store C into PMOP[0] and D into PMOP[1]. */
12901 {
12908 {
12911 }
12914 /* If C or D is of the form (A & N) where
12915 (N & M) == M, or of the form (A | N) or
12916 (A ^ N) where (N & M) == 0, replace it with A. */
12922 /* If C or D is a N where (N & M) == 0, it can be
12923 omitted (replaced with 0). */
12928 /* Similarly, with C - N where (-N & M) == 0. */
12936 }
12938 /* Only build anything new if we optimized one or both arguments above. */
12944 else
12946 }
12948 /* Used by contains_label_[p1]. */
12950 struct contains_label_data
12951 {
12954 };
12956 /* Callback for walk_tree, looking for LABEL_EXPR. Return *TP if it is
12957 a LABEL_EXPR or CASE_LABEL_EXPR not inside of another SWITCH_EXPR; otherwise
12958 return NULL_TREE. Do not check the subtrees of GOTO_EXPR. */
12960 static tree
12962 {
12965 {
12976 {
12984 }
12993 }
12994 }
12996 /* Return whether the sub-tree ST contains a label which is accessible from
12997 outside the sub-tree. */
12999 static bool
13001 {
13005 }
13007 /* Fold a ternary expression of code CODE and type TYPE with operands
13008 OP0, OP1, and OP2. Return the folded expression if folding is
13009 successful. Otherwise, return NULL_TREE. */
13011 tree
13014 {
13015 tree tem;
13022 /* If this is a commutative operation, and OP0 is a constant, move it
13023 to OP1 to reduce the number of tests below. */
13032 /* Strip any conversions that don't change the mode. This is safe
13033 for every expression, except for a comparison expression because
13034 its signedness is derived from its operands. So, in the latter
13035 case, only strip conversions that don't change the signedness.
13037 Note that this is done as an internal manipulation within the
13038 constant folder, in order to find the simplest representation of
13039 the arguments so that their form can be studied. In any cases,
13040 the appropriate type conversions should be put back in the tree
13041 that will get out of the constant folder. */
13043 {
13046 }
13049 {
13052 }
13055 {
13058 }
13061 {
13065 {
13071 }
13076 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
13077 so all simple results must be passed through pedantic_non_lvalue. */
13079 {
13082 /* Only optimize constant conditions when the selected branch
13083 has the same type as the COND_EXPR. This avoids optimizing
13084 away "c ? x : throw", where the throw has a void type.
13085 Avoid throwing away that operand which contains label. */
13092 }
13094 {
13101 {
13104 {
13112 }
13117 }
13118 }
13120 /* If we have A op B ? A : C, we may be able to convert this to a
13121 simpler expression, depending on the operation and the values
13122 of B and C. Signed zeros prevent all of these transformations,
13123 for reasons given above each one.
13125 Also try swapping the arguments and inverting the conditional. */
13129 {
13136 }
13141 {
13148 arg00,
13149 arg01,
13153 }
13155 /* If the second operand is simpler than the third, swap them
13156 since that produces better jump optimization results. */
13159 {
13161 /* See if this can be inverted. If it can't, possibly because
13162 it was a floating-point inequality comparison, don't do
13163 anything. */
13167 }
13169 /* Convert A ? 1 : 0 to simply A. */
13174 /* If we try to convert OP0 to our type, the
13175 call to fold will try to move the conversion inside
13176 a COND, which will recurse. In that case, the COND_EXPR
13177 is probably the best choice, so leave it alone. */
13181 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
13182 over COND_EXPR in cases such as floating point comparisons. */
13191 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
13196 {
13197 /* sign_bit_p looks through both zero and sign extensions,
13198 but for this optimization only sign extensions are
13199 usable. */
13202 {
13205 {
13208 }
13210 }
13211 /* sign_bit_p only checks ARG1 bits within A's precision.
13212 If <sign bit of A> has wider type than A, bits outside
13213 of A's precision in <sign bit of A> need to be checked.
13214 If they are all 0, this optimization needs to be done
13215 in unsigned A's type, if they are all 1 in signed A's type,
13216 otherwise this can't be done. */
13222 {
13224 tree tem_type;
13237 {
13240 }
13242 {
13245 }
13246 else
13248 }
13251 return
13257 arg1)));
13258 }
13260 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
13261 already handled above. */
13266 {
13276 op1);
13277 }
13279 /* A & N ? N : 0 is simply A & N if N is a power of two. This
13280 is probably obsolete because the first operand should be a
13281 truth value (that's why we have the two cases above), but let's
13282 leave it in until we can confirm this for all front-ends. */
13290 /* operand_equal_p compares just value, not precision, so e.g.
13291 arg1 could be 8-bit -128 and be power of two, but BIT_AND_EXPR
13292 second operand 32-bit -128, which is not a power of two (or vice
13293 versa. */
13297 /* Disable the transformations below for vectors, since
13298 fold_binary_op_with_conditional_arg may undo them immediately,
13299 yielding an infinite loop. */
13303 /* Convert A ? B : 0 into A && B if A and B are truth values. */
13312 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
13317 {
13319 /* Only perform transformation if ARG0 is easily inverted. */
13323 ? BIT_IOR_EXPR
13326 op1);
13327 }
13329 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
13334 {
13336 /* Only perform transformation if ARG0 is easily inverted. */
13342 op2);
13343 }
13345 /* Convert A ? 1 : B into A || B if A and B are truth values. */
13357 /* CALL_EXPRs used to be ternary exprs. Catch any mistaken uses
13358 of fold_ternary on them. */
13368 {
13370 unsigned HOST_WIDE_INT width
13381 {
13386 {
13388 {
13393 }
13399 }
13400 }
13401 }
13403 /* On constants we can use native encode/interpret to constant
13404 fold (nearly) all BIT_FIELD_REFs. */
13410 {
13413 /* Limit us to a reasonable amount of work. To relax the
13414 other limitations we need bit-shifting of the buffer
13415 and rounding up the size. */
13419 {
13421 unsigned HOST_WIDE_INT len
13426 {
13431 }
13432 }
13433 }
13438 /* Perform constant folding of BIT_INSERT_EXPR. */
13442 {
13443 /* Build a vector of integers from the tree mask. */
13444 vec_perm_builder builder;
13448 /* Create a vec_perm_indices for the integer vector. */
13453 }
13457 /* Perform (partial) constant folding of BIT_INSERT_EXPR. */
13460 {
13466 wide_int tem2
13470 }
13475 {
13477 unsigned HOST_WIDE_INT elsize
13480 {
13486 {
13492 }
13493 }
13494 }
13500 }
13502 /* Gets the element ACCESS_INDEX from CTOR, which must be a CONSTRUCTOR
13503 of an array (or vector). *CTOR_IDX if non-NULL is updated with the
13504 constructor element index of the value returned. If the element is
13505 not found NULL_TREE is returned and *CTOR_IDX is updated to
13506 the index of the element after the ACCESS_INDEX position (which
13507 may be outside of the CTOR array). */
13509 tree
13512 {
13518 {
13521 {
13522 /* Static constructors for variably sized objects makes no sense. */
13525 /* ??? When it is obvious that the range is signed, treat it so. */
13530 {
13532 low_bound
13534 SIGNED);
13535 }
13536 else
13537 {
13540 }
13541 }
13542 }
13546 index_sgn);
13558 {
13559 /* Array constructor might explicitly set index, or specify a range,
13560 or leave index NULL meaning that it is next index after previous
13561 one. */
13563 {
13565 max_index = index
13567 else
13568 {
13571 index_sgn);
13572 max_index
13574 index_sgn);
13576 }
13577 }
13579 {
13585 }
13586 else
13589 /* Do we have match? */
13591 {
13593 {
13597 }
13598 }
13600 /* We're past the element we search for. Note during parsing
13601 the elements might not be sorted.
13602 ??? We should use a binary search and a flag on the
13603 CONSTRUCTOR as to whether elements are sorted in declaration
13604 order. */
13606 }
13610 }
13612 /* Perform constant folding and related simplification of EXPR.
13613 The related simplifications include x*1 => x, x*0 => 0, etc.,
13614 and application of the associative law.
13615 NOP_EXPR conversions may be removed freely (as long as we
13616 are careful not to change the type of the overall expression).
13617 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
13618 but we can constant-fold them if they have constant operands. */
13620 #ifdef ENABLE_FOLD_CHECKING
13621 # define fold(x) fold_1 (x)
13623 static
13624 #endif
13625 tree
13627 {
13631 tree tem;
13634 /* Return right away if a constant. */
13638 /* CALL_EXPR-like objects with variable numbers of operands are
13639 treated specially. */
13641 {
13643 {
13646 }
13648 }
13651 {
13656 {
13674 }
13675 }
13678 {
13680 {
13687 {
13692 }
13695 }
13697 /* Return a VECTOR_CST if possible. */
13699 {
13705 tree val;
13711 }
13719 }
13721 #ifdef ENABLE_FOLD_CHECKING
13722 #undef fold
13729 /* When --enable-checking=fold, compute a digest of expr before
13730 and after actual fold call to see if fold did not accidentally
13731 change original expr. */
13733 tree
13735 {
13736 tree ret;
13756 }
13758 void
13760 {
13771 }
13773 static void
13775 {
13777 }
13779 static void
13782 {
13788 recursive_label:
13798 {
13799 /* Allow DECL_ASSEMBLER_NAME and symtab_node to be modified. */
13807 }
13815 {
13816 /* Allow these fields to be modified. */
13817 tree tmp;
13828 {
13831 }
13832 }
13834 {
13835 /* Allow the no-warning bit to be set. Perhaps we shouldn't allow
13836 that and change builtins.cc etc. instead - see PR89543. */
13842 }
13853 {
13856 {
13872 }
13876 {
13889 }
13906 {
13912 }
13915 {
13917 {
13920 }
13922 }
13933 {
13936 }
13944 }
13945 }
13947 /* Helper function for outputting the checksum of a tree T. When
13948 debugging with gdb, you can "define mynext" to be "next" followed
13949 by "call debug_fold_checksum (op0)", then just trace down till the
13950 outputs differ. */
13952 DEBUG_FUNCTION void
13954 {
13969 }
13971 #endif
13973 /* Fold a unary tree expression with code CODE of type TYPE with an
13974 operand OP0. LOC is the location of the resulting expression.
13975 Return a folded expression if successful. Otherwise, return a tree
13976 expression with code CODE of type TYPE with an operand OP0. */
13978 tree
13981 {
13982 tree tem;
13983 #ifdef ENABLE_FOLD_CHECKING
13992 #endif
13998 #ifdef ENABLE_FOLD_CHECKING
14005 #endif
14007 }
14009 /* Fold a binary tree expression with code CODE of type TYPE with
14010 operands OP0 and OP1. LOC is the location of the resulting
14011 expression. Return a folded expression if successful. Otherwise,
14012 return a tree expression with code CODE of type TYPE with operands
14013 OP0 and OP1. */
14015 tree
14018 MEM_STAT_DECL)
14019 {
14020 tree tem;
14021 #ifdef ENABLE_FOLD_CHECKING
14038 #endif
14044 #ifdef ENABLE_FOLD_CHECKING
14059 #endif
14061 }
14063 /* Fold a ternary tree expression with code CODE of type TYPE with
14064 operands OP0, OP1, and OP2. Return a folded expression if
14065 successful. Otherwise, return a tree expression with code CODE of
14066 type TYPE with operands OP0, OP1, and OP2. */
14068 tree
14071 {
14072 tree tem;
14073 #ifdef ENABLE_FOLD_CHECKING
14097 #endif
14104 #ifdef ENABLE_FOLD_CHECKING
14127 #endif
14129 }
14131 /* Fold a CALL_EXPR expression of type TYPE with operands FN and NARGS
14132 arguments in ARGARRAY, and a null static chain.
14133 Return a folded expression if successful. Otherwise, return a CALL_EXPR
14134 of type TYPE from the given operands as constructed by build_call_array. */
14136 tree
14139 {
14140 tree tem;
14141 #ifdef ENABLE_FOLD_CHECKING
14160 #endif
14166 #ifdef ENABLE_FOLD_CHECKING
14182 #endif
14184 }
14186 /* Perform constant folding and related simplification of initializer
14187 expression EXPR. These behave identically to "fold_buildN" but ignore
14188 potential run-time traps and exceptions that fold must preserve. */
14190 #define START_FOLD_INIT \
14191 int saved_signaling_nans = flag_signaling_nans;\
14192 int saved_trapping_math = flag_trapping_math;\
14193 int saved_rounding_math = flag_rounding_math;\
14194 int saved_trapv = flag_trapv;\
14195 int saved_folding_initializer = folding_initializer;\
14196 flag_signaling_nans = 0;\
14197 flag_trapping_math = 0;\
14198 flag_rounding_math = 0;\
14199 flag_trapv = 0;\
14200 folding_initializer = 1;
14202 #define END_FOLD_INIT \
14203 flag_signaling_nans = saved_signaling_nans;\
14204 flag_trapping_math = saved_trapping_math;\
14205 flag_rounding_math = saved_rounding_math;\
14206 flag_trapv = saved_trapv;\
14207 folding_initializer = saved_folding_initializer;
14209 tree
14211 {
14212 tree result;
14213 START_FOLD_INIT;
14217 END_FOLD_INIT;
14219 }
14221 tree
14224 {
14225 tree result;
14226 START_FOLD_INIT;
14230 END_FOLD_INIT;
14232 }
14234 tree
14237 {
14238 tree result;
14239 START_FOLD_INIT;
14243 END_FOLD_INIT;
14245 }
14247 tree
14250 {
14251 tree result;
14252 START_FOLD_INIT;
14256 END_FOLD_INIT;
14258 }
14260 tree
14263 {
14264 tree result;
14265 START_FOLD_INIT;
14269 END_FOLD_INIT;
14271 }
14273 #undef START_FOLD_INIT
14274 #undef END_FOLD_INIT
14276 /* Determine if first argument is a multiple of second argument. Return
14277 false if it is not, or we cannot easily determined it to be.
14279 An example of the sort of thing we care about (at this point; this routine
14280 could surely be made more general, and expanded to do what the *_DIV_EXPR's
14281 fold cases do now) is discovering that
14283 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14285 is a multiple of
14287 SAVE_EXPR (J * 8)
14289 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
14291 This code also handles discovering that
14293 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14295 is a multiple of 8 so we don't have to worry about dealing with a
14296 possible remainder.
14298 Note that we *look* inside a SAVE_EXPR only to determine how it was
14299 calculated; it is not safe for fold to do much of anything else with the
14300 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
14301 at run time. For example, the latter example above *cannot* be implemented
14302 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
14303 evaluation time of the original SAVE_EXPR is not necessarily the same at
14304 the time the new expression is evaluated. The only optimization of this
14305 sort that would be valid is changing
14307 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
14309 divided by 8 to
14311 SAVE_EXPR (I) * SAVE_EXPR (J)
14313 (where the same SAVE_EXPR (J) is used in the original and the
14314 transformed version).
14316 NOWRAP specifies whether all outer operations in TYPE should
14317 be considered not wrapping. Any type conversion within TOP acts
14318 as a barrier and we will fall back to NOWRAP being false.
14319 NOWRAP is mostly used to treat expressions in TYPE_SIZE and friends
14320 as not wrapping even though they are generally using unsigned arithmetic. */
14322 bool
14324 {
14335 {
14337 /* Bitwise and provides a power of two multiple. If the mask is
14338 a multiple of BOTTOM then TOP is a multiple of BOTTOM. */
14345 /* If the multiplication can wrap we cannot recurse further unless
14346 the bottom is a power of two which is where wrapping does not
14347 matter. */
14353 {
14359 {
14362 /* Handle multiple_of_p ((x * 2 + 2) * 4, 8). */
14364 {
14368 {
14371 }
14372 }
14374 }
14375 }
14380 /* Handle X << CST as X * (1 << CST) and only process the constant. */
14382 {
14385 {
14386 wide_int mul_op
14390 nowrap);
14391 }
14392 }
14397 /* If the addition or subtraction can wrap we cannot recurse further
14398 unless bottom is a power of two which is where wrapping does not
14399 matter. */
14405 /* Handle cases like op0 + 0xfffffffd as op0 - 3 if the expression has
14406 unsigned type. For example, (X / 3) + 0xfffffffd is multiple of 3,
14407 but 0xfffffffd is not. */
14410 && nowrap
14415 /* It is impossible to prove if op0 +- op1 is multiple of bottom
14416 precisely, so be conservative here checking if both op0 and op1
14417 are multiple of bottom. Note we check the second operand first
14418 since it's usually simpler. */
14422 CASE_CONVERT:
14423 /* Can't handle conversions from non-integral or wider integral type. */
14428 /* NOWRAP only extends to operations in the outermost type so
14429 make sure to strip it off here. */
14444 SIGNED);
14450 {
14453 /* Check for special cases to see if top is defined as multiple
14454 of bottom:
14456 top = (X & ~(bottom - 1) ; bottom is power of 2
14458 or
14460 Y = X % bottom
14461 top = X - Y. */
14480 }
14482 /* fall through */
14490 }
14491 }
14493 /* Return true if expression X cannot be (or contain) a NaN or infinity.
14494 This function returns true for integer expressions, and returns
14495 false if uncertain. */
14497 bool
14499 {
14504 {
14527 {
14528 CASE_CFN_FABS:
14529 CASE_CFN_FABS_FN:
14531 CASE_CFN_FMAX:
14532 CASE_CFN_FMAX_FN:
14533 CASE_CFN_FMIN:
14534 CASE_CFN_FMIN_FN:
14539 }
14543 }
14544 }
14546 /* Return true if expression X evaluates to an infinity.
14547 This function returns false for integer expressions. */
14549 bool
14551 {
14555 {
14568 }
14569 }
14571 /* Return true if expression X could evaluate to an infinity.
14572 This function returns false for integer expressions, and returns
14573 true if uncertain. */
14575 bool
14577 {
14581 {
14594 }
14595 }
14597 /* Return true if expression X evaluates to a signaling NaN.
14598 This function returns false for integer expressions. */
14600 bool
14602 {
14606 {
14617 }
14618 }
14620 /* Return true if expression X could evaluate to a signaling NaN.
14621 This function returns false for integer expressions, and returns
14622 true if uncertain. */
14624 bool
14626 {
14630 {
14650 {
14651 CASE_CFN_FABS:
14652 CASE_CFN_FABS_FN:
14654 CASE_CFN_FMAX:
14655 CASE_CFN_FMAX_FN:
14656 CASE_CFN_FMIN:
14657 CASE_CFN_FMIN_FN:
14662 }
14665 }
14666 }
14668 /* Return true if expression X evaluates to a NaN.
14669 This function returns false for integer expressions. */
14671 bool
14673 {
14677 {
14688 }
14689 }
14691 /* Return true if expression X could evaluate to a NaN.
14692 This function returns false for integer expressions, and returns
14693 true if uncertain. */
14695 bool
14697 {
14701 {
14726 {
14727 CASE_CFN_FABS:
14728 CASE_CFN_FABS_FN:
14730 CASE_CFN_FMAX:
14731 CASE_CFN_FMAX_FN:
14732 CASE_CFN_FMIN:
14733 CASE_CFN_FMIN_FN:
14738 }
14741 }
14742 }
14744 /* Return true if expression X could evaluate to -0.0.
14745 This function returns true if uncertain. */
14747 bool
14749 {
14753 {
14768 {
14769 CASE_CFN_FABS:
14770 CASE_CFN_FABS_FN:
14774 }
14777 }
14778 /* Ideally !(tree_expr_nonzero_p (X) || tree_expr_nonnegative_p (X))
14779 * but currently those predicates require tree and not const_tree. */
14781 }
14783 #define tree_expr_nonnegative_warnv_p(X, Y) \
14786 #define RECURSE(X) \
14787 ((tree_expr_nonnegative_warnv_p) (X, strict_overflow_p, depth + 1))
14789 /* Return true if CODE or TYPE is known to be non-negative. */
14791 static bool
14793 {
14797 /* Truth values evaluate to 0 or 1, which is nonnegative unless we
14798 have a signed:1 type (where the value is -1 and 0). */
14801 }
14803 /* Return true if (CODE OP0) is known to be non-negative. If the return
14804 value is based on the assumption that signed overflow is undefined,
14805 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14806 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14808 bool
14811 {
14816 {
14818 /* We can't return 1 if flag_wrapv is set because
14819 ABS_EXPR<INT_MIN> = INT_MIN. */
14823 {
14826 }
14834 CASE_CONVERT:
14835 {
14840 {
14844 {
14848 }
14849 }
14851 {
14857 }
14858 }
14863 }
14865 /* We don't know sign of `t', so be conservative and return false. */
14867 }
14869 /* Return true if (CODE OP0 OP1) is known to be non-negative. If the return
14870 value is based on the assumption that signed overflow is undefined,
14871 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14872 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14874 bool
14878 {
14883 {
14889 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
14890 both unsigned and at least 2 bits shorter than the result. */
14894 {
14899 {
14903 }
14904 }
14909 {
14910 /* x * x is always non-negative for floating point x
14911 or without overflow. */
14914 {
14919 }
14920 }
14922 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
14923 both unsigned and their total bits is shorter than the result. */
14927 {
14946 {
14956 }
14957 }
14964 /* Usually RECURSE (op0) || RECURSE (op1) but NaNs complicate
14965 things. */
14990 }
14992 /* We don't know sign of `t', so be conservative and return false. */
14994 }
14996 /* Return true if T is known to be non-negative. If the return
14997 value is based on the assumption that signed overflow is undefined,
14998 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14999 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15001 bool
15003 {
15008 {
15022 /* Limit the depth of recursion to avoid quadratic behavior.
15023 This is expected to catch almost all occurrences in practice.
15024 If this code misses important cases that unbounded recursion
15025 would not, passes that need this information could be revised
15026 to provide it through dataflow propagation. */
15034 }
15035 }
15037 /* Return true if T is known to be non-negative. If the return
15038 value is based on the assumption that signed overflow is undefined,
15039 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15040 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15042 bool
15045 {
15047 {
15048 CASE_CFN_ACOS:
15049 CASE_CFN_ACOS_FN:
15050 CASE_CFN_ACOSH:
15051 CASE_CFN_ACOSH_FN:
15052 CASE_CFN_CABS:
15053 CASE_CFN_CABS_FN:
15054 CASE_CFN_COSH:
15055 CASE_CFN_COSH_FN:
15056 CASE_CFN_ERFC:
15057 CASE_CFN_ERFC_FN:
15058 CASE_CFN_EXP:
15059 CASE_CFN_EXP_FN:
15060 CASE_CFN_EXP10:
15061 CASE_CFN_EXP2:
15062 CASE_CFN_EXP2_FN:
15063 CASE_CFN_FABS:
15064 CASE_CFN_FABS_FN:
15065 CASE_CFN_FDIM:
15066 CASE_CFN_FDIM_FN:
15067 CASE_CFN_HYPOT:
15068 CASE_CFN_HYPOT_FN:
15069 CASE_CFN_POW10:
15070 CASE_CFN_FFS:
15071 CASE_CFN_PARITY:
15072 CASE_CFN_POPCOUNT:
15073 CASE_CFN_CLZ:
15074 CASE_CFN_CLRSB:
15079 /* Always true. */
15082 CASE_CFN_SQRT:
15083 CASE_CFN_SQRT_FN:
15084 /* sqrt(-0.0) is -0.0. */
15089 CASE_CFN_ASINH:
15090 CASE_CFN_ASINH_FN:
15091 CASE_CFN_ATAN:
15092 CASE_CFN_ATAN_FN:
15093 CASE_CFN_ATANH:
15094 CASE_CFN_ATANH_FN:
15095 CASE_CFN_CBRT:
15096 CASE_CFN_CBRT_FN:
15097 CASE_CFN_CEIL:
15098 CASE_CFN_CEIL_FN:
15099 CASE_CFN_ERF:
15100 CASE_CFN_ERF_FN:
15101 CASE_CFN_EXPM1:
15102 CASE_CFN_EXPM1_FN:
15103 CASE_CFN_FLOOR:
15104 CASE_CFN_FLOOR_FN:
15105 CASE_CFN_FMOD:
15106 CASE_CFN_FMOD_FN:
15107 CASE_CFN_FREXP:
15108 CASE_CFN_FREXP_FN:
15109 CASE_CFN_ICEIL:
15110 CASE_CFN_IFLOOR:
15111 CASE_CFN_IRINT:
15112 CASE_CFN_IROUND:
15113 CASE_CFN_LCEIL:
15114 CASE_CFN_LDEXP:
15115 CASE_CFN_LFLOOR:
15116 CASE_CFN_LLCEIL:
15117 CASE_CFN_LLFLOOR:
15118 CASE_CFN_LLRINT:
15119 CASE_CFN_LLRINT_FN:
15120 CASE_CFN_LLROUND:
15121 CASE_CFN_LLROUND_FN:
15122 CASE_CFN_LRINT:
15123 CASE_CFN_LRINT_FN:
15124 CASE_CFN_LROUND:
15125 CASE_CFN_LROUND_FN:
15126 CASE_CFN_MODF:
15127 CASE_CFN_MODF_FN:
15128 CASE_CFN_NEARBYINT:
15129 CASE_CFN_NEARBYINT_FN:
15130 CASE_CFN_RINT:
15131 CASE_CFN_RINT_FN:
15132 CASE_CFN_ROUND:
15133 CASE_CFN_ROUND_FN:
15134 CASE_CFN_ROUNDEVEN:
15135 CASE_CFN_ROUNDEVEN_FN:
15136 CASE_CFN_SCALB:
15137 CASE_CFN_SCALBLN:
15138 CASE_CFN_SCALBLN_FN:
15139 CASE_CFN_SCALBN:
15140 CASE_CFN_SCALBN_FN:
15141 CASE_CFN_SIGNBIT:
15142 CASE_CFN_SIGNIFICAND:
15143 CASE_CFN_SINH:
15144 CASE_CFN_SINH_FN:
15145 CASE_CFN_TANH:
15146 CASE_CFN_TANH_FN:
15147 CASE_CFN_TRUNC:
15148 CASE_CFN_TRUNC_FN:
15149 /* True if the 1st argument is nonnegative. */
15152 CASE_CFN_FMAX:
15153 CASE_CFN_FMAX_FN:
15154 /* Usually RECURSE (arg0) || RECURSE (arg1) but NaNs complicate
15155 things. In the presence of sNaNs, we're only guaranteed to be
15156 non-negative if both operands are non-negative. In the presence
15157 of qNaNs, we're non-negative if either operand is non-negative
15158 and can't be a qNaN, or if both operands are non-negative. */
15167 CASE_CFN_FMIN:
15168 CASE_CFN_FMIN_FN:
15169 /* True if the 1st AND 2nd arguments are nonnegative. */
15172 CASE_CFN_COPYSIGN:
15173 CASE_CFN_COPYSIGN_FN:
15174 /* True if the 2nd argument is nonnegative. */
15177 CASE_CFN_POWI:
15178 /* True if the 1st argument is nonnegative or the second
15179 argument is an even integer. */
15185 CASE_CFN_POW:
15186 CASE_CFN_POW_FN:
15187 /* True if the 1st argument is nonnegative or the second
15188 argument is an even integer valued real. */
15190 {
15191 REAL_VALUE_TYPE c;
15192 HOST_WIDE_INT n;
15197 {
15198 REAL_VALUE_TYPE cint;
15202 }
15203 }
15208 }
15210 }
15212 /* Return true if T is known to be non-negative. If the return
15213 value is based on the assumption that signed overflow is undefined,
15214 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15215 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15217 static bool
15219 {
15225 {
15227 {
15231 /* If the initializer is non-void, then it's a normal expression
15232 that will be assigned to the slot. */
15236 /* Otherwise, the initializer sets the slot in some way. One common
15237 way is an assignment statement at the end of the initializer. */
15239 {
15247 else
15249 }
15255 }
15258 {
15264 arg0,
15265 arg1,
15267 }
15280 }
15281 }
15283 #undef RECURSE
15284 #undef tree_expr_nonnegative_warnv_p
15286 /* Return true if T is known to be non-negative. If the return
15287 value is based on the assumption that signed overflow is undefined,
15288 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15289 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15291 bool
15293 {
15300 {
15322 }
15325 {
15350 }
15351 }
15353 /* Return true if `t' is known to be non-negative. Handle warnings
15354 about undefined signed overflow. */
15356 bool
15358 {
15365 "determining that expression is always "
15367 WARN_STRICT_OVERFLOW_MISC);
15369 }
15372 /* Return true when (CODE OP0) is an address and is known to be nonzero.
15373 For floating point we further ensure that T is not denormal.
15374 Similar logic is present in nonzero_address in rtlanal.h.
15376 If the return value is based on the assumption that signed overflow
15377 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15378 change *STRICT_OVERFLOW_P. */
15380 bool
15383 {
15385 {
15388 strict_overflow_p);
15391 {
15397 strict_overflow_p));
15398 }
15403 strict_overflow_p);
15407 }
15410 }
15412 /* Return true when (CODE OP0 OP1) is an address and is known to be nonzero.
15413 For floating point we further ensure that T is not denormal.
15414 Similar logic is present in nonzero_address in rtlanal.h.
15416 If the return value is based on the assumption that signed overflow
15417 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15418 change *STRICT_OVERFLOW_P. */
15420 bool
15422 tree type,
15423 tree op0,
15425 {
15428 {
15432 {
15433 /* With the presence of negative values it is hard
15434 to say something. */
15441 /* One of operands must be positive and the other non-negative. */
15442 /* We don't set *STRICT_OVERFLOW_P here: even if this value
15443 overflows, on a twos-complement machine the sum of two
15444 nonnegative numbers can never be zero. */
15446 strict_overflow_p)
15448 strict_overflow_p));
15449 }
15454 {
15456 strict_overflow_p)
15458 strict_overflow_p))
15459 {
15462 }
15463 }
15472 {
15475 }
15482 {
15486 /* When both operands are nonzero, then MAX must be too. */
15488 strict_overflow_p))
15491 /* MAX where operand 0 is positive is positive. */
15493 strict_overflow_p);
15494 }
15495 /* MAX where operand 1 is positive is positive. */
15500 {
15504 }
15509 strict_overflow_p)
15511 strict_overflow_p));
15515 }
15518 }
15520 /* Return true when T is an address and is known to be nonzero.
15521 For floating point we further ensure that T is not denormal.
15522 Similar logic is present in nonzero_address in rtlanal.h.
15524 If the return value is based on the assumption that signed overflow
15525 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15526 change *STRICT_OVERFLOW_P. */
15528 bool
15530 {
15533 {
15538 {
15550 /* For objects in symbol table check if we know they are non-zero.
15551 Don't do anything for variables and functions before symtab is built;
15552 it is quite possible that they will be declared weak later. */
15557 /* Constants are never weak. */
15562 }
15570 {
15574 }
15584 }
15586 }
15588 #define integer_valued_real_p(X) \
15591 #define RECURSE(X) \
15592 ((integer_valued_real_p) (X, depth + 1))
15594 /* Return true if the floating point result of (CODE OP0) has an
15595 integer value. We also allow +Inf, -Inf and NaN to be considered
15596 integer values. Return false for signaling NaN.
15598 DEPTH is the current nesting depth of the query. */
15600 bool
15602 {
15604 {
15611 CASE_CONVERT:
15612 {
15619 }
15623 }
15625 }
15627 /* Return true if the floating point result of (CODE OP0 OP1) has an
15628 integer value. We also allow +Inf, -Inf and NaN to be considered
15629 integer values. Return false for signaling NaN.
15631 DEPTH is the current nesting depth of the query. */
15633 bool
15635 {
15637 {
15647 }
15649 }
15651 /* Return true if the floating point result of calling FNDECL with arguments
15652 ARG0 and ARG1 has an integer value. We also allow +Inf, -Inf and NaN to be
15653 considered integer values. Return false for signaling NaN. If FNDECL
15654 takes fewer than 2 arguments, the remaining ARGn are null.
15656 DEPTH is the current nesting depth of the query. */
15658 bool
15660 {
15662 {
15663 CASE_CFN_CEIL:
15664 CASE_CFN_CEIL_FN:
15665 CASE_CFN_FLOOR:
15666 CASE_CFN_FLOOR_FN:
15667 CASE_CFN_NEARBYINT:
15668 CASE_CFN_NEARBYINT_FN:
15669 CASE_CFN_RINT:
15670 CASE_CFN_RINT_FN:
15671 CASE_CFN_ROUND:
15672 CASE_CFN_ROUND_FN:
15673 CASE_CFN_ROUNDEVEN:
15674 CASE_CFN_ROUNDEVEN_FN:
15675 CASE_CFN_TRUNC:
15676 CASE_CFN_TRUNC_FN:
15679 CASE_CFN_FMIN:
15680 CASE_CFN_FMIN_FN:
15681 CASE_CFN_FMAX:
15682 CASE_CFN_FMAX_FN:
15687 }
15689 }
15691 /* Return true if the floating point expression T (a GIMPLE_SINGLE_RHS)
15692 has an integer value. We also allow +Inf, -Inf and NaN to be
15693 considered integer values. Return false for signaling NaN.
15695 DEPTH is the current nesting depth of the query. */
15697 bool
15699 {
15701 {
15709 /* Limit the depth of recursion to avoid quadratic behavior.
15710 This is expected to catch almost all occurrences in practice.
15711 If this code misses important cases that unbounded recursion
15712 would not, passes that need this information could be revised
15713 to provide it through dataflow propagation. */
15717 depth));
15721 }
15723 }
15725 /* Return true if the floating point expression T (a GIMPLE_INVALID_RHS)
15726 has an integer value. We also allow +Inf, -Inf and NaN to be
15727 considered integer values. Return false for signaling NaN.
15729 DEPTH is the current nesting depth of the query. */
15731 static bool
15733 {
15735 {
15746 }
15748 }
15750 #undef RECURSE
15751 #undef integer_valued_real_p
15753 /* Return true if the floating point expression T has an integer value.
15754 We also allow +Inf, -Inf and NaN to be considered integer values.
15755 Return false for signaling NaN.
15757 DEPTH is the current nesting depth of the query. */
15759 bool
15761 {
15769 {
15785 }
15788 {
15794 {
15803 }
15807 }
15808 }
15810 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
15811 attempt to fold the expression to a constant without modifying TYPE,
15812 OP0 or OP1.
15814 If the expression could be simplified to a constant, then return
15815 the constant. If the expression would not be simplified to a
15816 constant, then return NULL_TREE. */
15818 tree
15820 {
15823 }
15825 /* Given the components of a unary expression CODE, TYPE and OP0,
15826 attempt to fold the expression to a constant without modifying
15827 TYPE or OP0.
15829 If the expression could be simplified to a constant, then return
15830 the constant. If the expression would not be simplified to a
15831 constant, then return NULL_TREE. */
15833 tree
15835 {
15838 }
15840 /* If EXP represents referencing an element in a constant string
15841 (either via pointer arithmetic or array indexing), return the
15842 tree representing the value accessed, otherwise return NULL. */
15844 tree
15846 {
15850 {
15852 tree index;
15853 tree string;
15858 else
15859 {
15863 /* Optimize the special-case of a zero lower bound.
15865 We convert the low_bound to sizetype to avoid some problems
15866 with constant folding. (E.g. suppose the lower bound is 1,
15867 and its mode is QI. Without the conversion,l (ARRAY
15868 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
15869 +INDEX), which becomes (ARRAY+255+INDEX). Oops!) */
15875 }
15877 scalar_int_mode char_mode;
15889 }
15891 }
15893 /* Folds a read from vector element at IDX of vector ARG. */
15895 tree
15897 {
15902 {
15906 {
15913 }
15914 }
15916 }
15918 /* Return the tree for neg (ARG0) when ARG0 is known to be either
15919 an integer constant, real, or fixed-point constant.
15921 TYPE is the type of the result. */
15923 static tree
15925 {
15929 {
15935 {
15936 FIXED_VALUE_TYPE f;
15941 /* Propagate overflow flags. */
15945 }
15949 {
15956 }
15959 }
15962 }
15964 /* Return the tree for abs (ARG0) when ARG0 is known to be either
15965 an integer constant or real constant.
15967 TYPE is the type of the result. */
15969 tree
15971 {
15975 {
15977 {
15978 /* If the value is unsigned or non-negative, then the absolute value
15979 is the same as the ordinary value. */
15983 ;
15985 /* If the value is negative, then the absolute value is
15986 its negation. */
15987 else
15990 /* Force to the destination type, set TREE_OVERFLOW for signed
15991 TYPE only. */
15993 }
15999 else
16005 }
16008 }
16010 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
16011 constant. TYPE is the type of the result. */
16013 static tree
16015 {
16019 }
16021 /* Given CODE, a relational operator, the target type, TYPE and two
16022 constant operands OP0 and OP1, return the result of the
16023 relational operation. If the result is not a compile time
16024 constant, then return NULL_TREE. */
16026 static tree
16028 {
16031 /* From here on, the only cases we handle are when the result is
16032 known to be a constant. */
16035 {
16039 /* Handle the cases where either operand is a NaN. */
16041 {
16043 {
16071 }
16074 }
16077 }
16080 {
16084 }
16086 /* Handle equality/inequality of complex constants. */
16088 {
16099 else
16101 }
16104 {
16106 {
16107 /* Have vector comparison with scalar boolean result. */
16115 {
16123 }
16125 }
16126 tree_vector_builder elts;
16131 {
16144 }
16147 }
16149 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
16151 To compute GT, swap the arguments and do LT.
16152 To compute GE, do LT and invert the result.
16153 To compute LE, swap the arguments, do LT and invert the result.
16154 To compute NE, do EQ and invert the result.
16156 Therefore, the code below must handle only EQ and LT. */
16159 {
16162 }
16164 /* Note that it is safe to invert for real values here because we
16165 have already handled the one case that it matters. */
16169 {
16172 }
16174 /* Compute a result for LT or EQ if args permit;
16175 Otherwise return T. */
16177 {
16180 else
16182 }
16183 else
16189 }
16191 /* If necessary, return a CLEANUP_POINT_EXPR for EXPR with the
16192 indicated TYPE. If no CLEANUP_POINT_EXPR is necessary, return EXPR
16193 itself. */
16195 tree
16197 {
16198 /* If the expression does not have side effects then we don't have to wrap
16199 it with a cleanup point expression. */
16203 /* If the expression is a return, check to see if the expression inside the
16204 return has no side effects or the right hand side of the modify expression
16205 inside the return. If either don't have side effects set we don't need to
16206 wrap the expression in a cleanup point expression. Note we don't check the
16207 left hand side of the modify because it should always be a return decl. */
16209 {
16216 }
16219 }
16221 /* Given a pointer value OP0 and a type TYPE, return a simplified version
16222 of an indirection through OP0, or NULL_TREE if no simplification is
16223 possible. */
16225 tree
16227 {
16229 tree subtype;
16230 poly_uint64 const_op01;
16239 {
16243 /* *&CONST_DECL -> to the value of the const decl. */
16246 /* *&p => p; make sure to handle *&"str"[cst] here. */
16248 {
16252 else
16254 }
16255 /* *(foo *)&fooarray => fooarray[0] */
16258 && (!in_gimple_form
16260 {
16270 }
16271 /* *(foo *)&complexfoo => __real__ complexfoo */
16275 /* *(foo *)&vectorfoo => BIT_FIELD_REF<vectorfoo,...> */
16278 {
16282 index);
16283 }
16284 }
16288 {
16294 {
16295 tree op00type;
16299 /* ((foo*)&vectorfoo)[1] => BIT_FIELD_REF<vectorfoo,...> */
16302 /* POINTER_PLUS_EXPR second operand is sizetype, unsigned,
16303 but we want to treat offsets with MSB set as negative.
16304 For the code below negative offsets are invalid and
16305 TYPE_SIZE of the element is something unsigned, so
16306 check whether op01 fits into poly_int64, which implies
16307 it is from 0 to INTTYPE_MAXIMUM (HOST_WIDE_INT), and
16308 then just use poly_uint64 because we want to treat the
16309 value as unsigned. */
16311 {
16313 poly_uint64 max_offset
16317 {
16322 }
16323 }
16324 /* ((foo*)&complexfoo)[1] => __imag__ complexfoo */
16327 {
16329 const_op01))
16331 }
16332 /* ((foo *)&fooarray)[1] => fooarray[1] */
16335 {
16344 {
16349 }
16350 }
16351 }
16352 }
16354 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
16357 && (!in_gimple_form
16359 {
16360 tree type_domain;
16370 NULL_TREE);
16371 }
16374 }
16376 /* Builds an expression for an indirection through T, simplifying some
16377 cases. */
16379 tree
16381 {
16389 }
16391 /* Given an INDIRECT_REF T, return either T or a simplified version. */
16393 tree
16395 {
16400 else
16402 }
16404 /* Strip non-trapping, non-side-effecting tree nodes from an expression
16405 whose result is ignored. The type of the returned tree need not be
16406 the same as the original expression. */
16408 tree
16410 {
16416 {
16427 else
16433 {
16449 }
16454 }
16455 }
16457 /* Return the value of VALUE, rounded up to a multiple of DIVISOR. */
16459 tree
16461 {
16467 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16468 have to do anything. Only do this when we are not given a const,
16469 because in that case, this check is more expensive than just
16470 doing it. */
16472 {
16477 }
16479 /* If divisor is a power of two, simplify this to bit manipulation. */
16481 {
16483 {
16497 }
16498 else
16499 {
16500 tree t;
16506 }
16507 }
16508 else
16509 {
16514 }
16517 }
16519 /* Likewise, but round down. */
16521 tree
16523 {
16530 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16531 have to do anything. Only do this when we are not given a const,
16532 because in that case, this check is more expensive than just
16533 doing it. */
16535 {
16540 }
16542 /* If divisor is a power of two, simplify this to bit manipulation. */
16544 {
16545 tree t;
16549 }
16550 else
16551 {
16556 }
16559 }
16561 /* Returns the pointer to the base of the object addressed by EXP and
16562 extracts the information about the offset of the access, storing it
16563 to PBITPOS and POFFSET. */
16565 static tree
16568 {
16569 tree core;
16570 machine_mode mode;
16572 poly_int64 bitsize;
16581 {
16586 }
16588 {
16594 {
16595 poly_offset_int tem
16601 }
16602 }
16603 else
16604 {
16608 }
16611 }
16613 /* Returns true if addresses of E1 and E2 differ by a constant, false
16614 otherwise. If they do, E1 - E2 is stored in *DIFF. */
16616 bool
16618 {
16633 {
16643 }
16645 {
16646 /* If only one of the offsets is non-constant, the difference cannot
16647 be a constant. */
16649 }
16650 else
16655 }
16657 /* Return OFF converted to a pointer offset type suitable as offset for
16658 POINTER_PLUS_EXPR. Use location LOC for this conversion. */
16659 tree
16661 {
16665 }
16667 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16668 tree
16670 {
16673 }
16675 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16676 tree
16678 {
16681 }
16683 /* Return a pointer to a NUL-terminated string containing the sequence
16684 of bytes corresponding to the representation of the object referred to
16685 by SRC (or a subsequence of such bytes within it if SRC is a reference
16686 to an initialized constant array plus some constant offset).
16687 Set *STRSIZE the number of bytes in the constant sequence including
16688 the terminating NUL byte. *STRSIZE is equal to sizeof(A) - OFFSET
16689 where A is the array that stores the constant sequence that SRC points
16690 to and OFFSET is the byte offset of SRC from the beginning of A. SRC
16691 need not point to a string or even an array of characters but may point
16692 to an object of any type. */
16696 {
16697 /* The offset into the array A storing the string, and A's byte size. */
16698 tree offset_node;
16699 tree mem_size;
16706 else
16713 {
16716 else
16718 }
16723 /* ARRAY_SIZE is the byte size of the array the constant sequence
16724 is stored in and equal to sizeof A. INIT_BYTES is the number
16725 of bytes in the constant sequence used to initialize the array,
16726 including any embedded NULs as well as the terminating NUL (for
16727 strings), but not including any trailing zeros/NULs past
16728 the terminating one appended implicitly to a string literal to
16729 zero out the remainder of the array it's stored in. For example,
16730 given:
16731 const char a[7] = "abc\0d";
16732 n = strlen (a + 1);
16733 ARRAY_SIZE is 7, INIT_BYTES is 6, and OFFSET is 1. For a valid
16734 (i.e., nul-terminated) string with no embedded nuls, INIT_BYTES
16735 is equal to strlen (A) + 1. */
16740 /* Ideally this would turn into a gcc_checking_assert over time. */
16748 {
16749 /* Compute and store the number of characters from the beginning
16750 of the substring at OFFSET to the end, including the terminating
16751 nul. Offsets past the initial length refer to null strings. */
16754 else
16756 }
16757 else
16758 {
16760 /* Support only properly NUL-terminated single byte strings. */
16765 }
16768 }
16770 /* Return a pointer to a NUL-terminated string corresponding to
16771 the expression STR referencing a constant string, possibly
16772 involving a constant offset. Return null if STR either doesn't
16773 reference a constant string or if it involves a nonconstant
16774 offset. */
16778 {
16780 }
16782 /* Given a tree T, compute which bits in T may be nonzero. */
16784 wide_int
16786 {
16788 {
16806 CASE_CONVERT:
16812 {
16817 }
16821 {
16829 }
16833 {
16841 }
16845 }
16848 }
16850 /* Helper function for address compare simplifications in match.pd.
16851 OP0 and OP1 are ADDR_EXPR operands being compared by CODE.
16852 TYPE is the type of comparison operands.
16853 BASE0, BASE1, OFF0 and OFF1 are set by the function.
16854 GENERIC is true if GENERIC folding and false for GIMPLE folding.
16855 Returns 0 if OP0 is known to be unequal to OP1 regardless of OFF{0,1},
16856 1 if bases are known to be equal and OP0 cmp OP1 depends on OFF0 cmp OFF1,
16857 and 2 if unknown. */
16859 int
16863 {
16873 {
16876 }
16878 {
16881 }
16886 /* Punt in GENERIC on variables with value expressions;
16887 the value expressions might point to fields/elements
16888 of other vars etc. */
16894 {
16898 }
16906 /* Assume different STRING_CSTs with the same content will be
16907 merged. */
16916 {
16919 /* If the offsets are equal we can ignore overflow. */
16922 /* Or if we compare using pointers to decls or strings. */
16927 }
16933 /* At this point we know (or assume) the two pointers point at
16934 different objects. */
16938 /* Punt on non-zero offsets from functions. */
16942 /* Or if the bases are neither decls nor string literals. */
16947 /* For initializers, assume addresses of different functions are
16948 different. */
16954 /* Compute whether one address points to the start of one
16955 object and another one to the end of another one. */
16958 {
16961 else
16963 }
16966 else
16967 {
16971 else
16973 }
16975 {
16978 else
16980 }
16983 else
16984 {
16988 else
16990 }
16992 {
16993 /* If one offset is pointing (or could be) to the beginning of one
16994 object and the other is pointing to one past the last byte of the
16995 other object, punt. */
17000 /* If both offsets are the same, there are some cases we know that are
17001 ok. Either if we know they aren't zero, or if we know both sizes
17002 are no zero. */
17008 }
17010 /* At this point, equal is 2 if either one or both pointers are out of
17011 bounds of their object, or one points to start of its object and the
17012 other points to end of its object. This is unspecified behavior
17013 e.g. in C++. Otherwise equal is 0. */
17017 /* When both pointers point to string literals, even when equal is 0,
17018 due to tail merging of string literals the pointers might be the same. */
17020 {
17027 /* If the bytes in the string literals starting at the pointers
17028 differ, the pointers need to be different. */
17033 {
17038 /* If even the bytes in the string literal before the
17039 pointers are the same, the string literals could be
17040 tail merged. */
17042 }
17044 }
17049 /* If this is a pointer comparison, ignore for now even
17050 valid equalities where one pointer is the offset zero
17051 of one object and the other to one past end of another one. */
17055 /* Assume that string literals can't be adjacent to variables
17056 (automatic or global). */
17060 /* Assume that automatic variables can't be adjacent to global
17061 variables. */
17066 }
17068 /* Return the single non-zero element of a CONSTRUCTOR or NULL_TREE. */
17069 tree
17071 {
17080 {
17084 }
17086 }
17088 #if CHECKING_P
17092 /* Helper functions for writing tests of folding trees. */
17094 /* Verify that the binary op (LHS CODE RHS) folds to CONSTANT. */
17096 static void
17098 tree constant)
17099 {
17101 }
17103 /* Verify that the binary op (LHS CODE RHS) folds to an NON_LVALUE_EXPR
17104 wrapping WRAPPED_EXPR. */
17106 static void
17108 tree wrapped_expr)
17109 {
17114 }
17116 /* Verify that various arithmetic binary operations are folded
17117 correctly. */
17119 static void
17121 {
17127 /* Addition. */
17128 /* 1 <-- (0 + 1) */
17130 one);
17132 one);
17134 /* (nonlvalue)x <-- (x + 0) */
17136 x);
17138 /* Subtraction. */
17139 /* 0 <-- (x - x) */
17141 zero);
17143 x);
17145 /* Multiplication. */
17146 /* 0 <-- (x * 0) */
17148 zero);
17150 /* (nonlvalue)x <-- (x * 1) */
17152 x);
17153 }
17157 /* Build a VECTOR_CST corresponding to VMODE, and has
17158 encoding given by NPATTERNS, NELTS_PER_PATTERN and STEP.
17159 Fill it with randomized elements, using rand() % THRESHOLD. */
17161 static tree
17165 {
17170 // Fill a0 for each pattern
17177 // Fill a1 for each pattern
17185 {
17190 }
17193 }
17195 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17196 when result is VLA. */
17198 static void
17201 {
17202 /* Actual npatterns and encoded_elts in res may be less than expected due
17203 to canonicalization. */
17210 }
17212 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17213 when the result is VLS. */
17215 static void
17217 {
17221 }
17223 /* Helper routine to push multiple elements into BUILDER. */
17227 {
17230 }
17232 #define ARG0(index) vector_cst_elt (arg0, index)
17233 #define ARG1(index) vector_cst_elt (arg1, index)
17235 /* Test cases where result is VNx4SI and input vectors are V4SI. */
17237 static void
17239 {
17241 {
17242 /* Case 1:
17243 sel = { 0, 4, 1, 5, ... }
17244 res = { arg[0], arg1[0], arg0[1], arg1[1], ...} // (4, 1) */
17245 {
17249 tree inner_type
17263 }
17265 /* Case 2: Same as case 1, but contains an out of bounds access which
17266 should wrap around.
17267 sel = {0, 8, 4, 12, ...} (4, 1)
17268 res = { arg0[0], arg0[0], arg1[0], arg1[0], ... } (4, 1). */
17269 {
17273 tree inner_type
17287 }
17288 }
17289 }
17291 /* Test cases where result is V4SI and input vectors are VNx4SI. */
17293 static void
17295 {
17297 {
17298 /* Case 1:
17299 sel = { 0, 1, 2, 3}
17300 res = { arg0[0], arg0[1], arg0[2], arg0[3] }. */
17301 {
17305 tree inner_type
17319 }
17321 /* Case 2: Same as Case 1, but crossing input vector.
17322 sel = {0, 2, 4, 6}
17323 In this case,the index 4 is ambiguous since len = 4 + 4x.
17324 Since we cannot determine, which vector to choose from during
17325 compile time, should return NULL_TREE. */
17326 {
17330 tree inner_type
17345 }
17346 }
17347 }
17349 /* Test all input vectors. */
17351 static void
17353 {
17354 /* Test with 10 different inputs. */
17356 {
17361 /* Case 1: mask = {0, ...} // (1, 1)
17362 res = { arg0[0], ... } // (1, 1) */
17363 {
17370 }
17372 /* Case 2: mask = {len, ...} // (1, 1)
17373 res = { arg1[0], ... } // (1, 1) */
17374 {
17382 }
17383 }
17384 }
17386 /* Test all vectors which contain at-least 2 elements. */
17388 static void
17390 {
17392 {
17393 /* Case 1: mask = { 0, len, ... } // (2, 1)
17394 res = { arg0[0], arg1[0], ... } // (2, 1) */
17395 {
17409 }
17411 /* Case 2: mask = { 0, len, 1, len+1, ... } // (2, 2)
17412 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2) */
17413 {
17427 }
17429 /* Case 4: mask = {0, 0, 1, ...} // (1, 3)
17430 Test that the stepped sequence of the pattern selects from
17431 same input pattern. Since input vectors have npatterns = 2,
17432 and step (a2 - a1) = 1, step is not a multiple of npatterns
17433 in input vector. So return NULL_TREE. */
17434 {
17449 }
17451 /* Case 5: mask = {len, 0, 1, ...} // (1, 3)
17452 Test that stepped sequence of the pattern selects from arg0.
17453 res = { arg1[0], arg0[0], arg0[1], ... } // (1, 3) */
17454 {
17468 }
17469 }
17470 }
17472 /* Test all vectors which contain at-least 4 elements. */
17474 static void
17476 {
17478 {
17479 /* Case 1: mask = { 0, len, 1, len+1, ... } // (4, 1)
17480 res: { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (4, 1) */
17481 {
17495 }
17497 /* Case 2: sel = {0, 1, 2, ...} // (1, 3)
17498 res: { arg0[0], arg0[1], arg0[2], ... } // (1, 3) */
17499 {
17512 }
17514 /* Case 3: sel = {len, len+1, len+2, ...} // (1, 3)
17515 res: { arg1[0], arg1[1], arg1[2], ... } // (1, 3) */
17516 {
17529 }
17531 /* Case 4:
17532 sel = { len, 0, 2, ... } // (1, 3)
17533 This should return NULL because we cross the input vectors.
17534 Because,
17535 Let's assume len = C + Cx
17536 a1 = 0
17537 S = 2
17538 esel = arg0_len / sel_npatterns = C + Cx
17539 ae = 0 + (esel - 2) * S
17540 = 0 + (C + Cx - 2) * 2
17541 = 2(C-2) + 2Cx
17543 For C >= 4:
17544 Let q1 = a1 / arg0_len = 0 / (C + Cx) = 0
17545 Let qe = ae / arg0_len = (2(C-2) + 2Cx) / (C + Cx) = 1
17546 Since q1 != qe, we cross input vectors.
17547 So return NULL_TREE. */
17548 {
17562 }
17564 /* Case 5: npatterns(arg0) = 4 > npatterns(sel) = 2
17565 mask = { 0, len, 1, len + 1, ...} // (2, 2)
17566 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2)
17568 Note that fold_vec_perm_cst will set
17569 res_npatterns = max(4, max(4, 2)) = 4
17570 However after canonicalizing, we will end up with shape (2, 2). */
17571 {
17584 }
17586 /* Case 6: Test combination in sel, where one pattern is dup and other
17587 is stepped sequence.
17588 sel = { 0, 0, 0, 1, 0, 2, ... } // (2, 3)
17589 res = { arg0[0], arg0[0], arg0[0],
17590 arg0[1], arg0[0], arg0[2], ... } // (2, 3) */
17591 {
17606 }
17608 /* Case 7: PR111048: Check that we set arg_npatterns correctly,
17609 when arg0, arg1 and sel have different number of patterns.
17610 arg0 is of shape (1, 1)
17611 arg1 is of shape (4, 1)
17612 sel is of shape (2, 3) = {1, len, 2, len+1, 3, len+2, ...}
17614 In this case the pattern: {len, len+1, len+2, ...} chooses arg1.
17615 However,
17616 step = (len+2) - (len+1) = 1
17617 arg_npatterns = VECTOR_CST_NPATTERNS (arg1) = 4
17618 Since step is not a multiple of arg_npatterns,
17619 valid_mask_for_fold_vec_perm_cst should return false,
17620 and thus fold_vec_perm_cst should return NULL_TREE. */
17621 {
17636 }
17637 }
17638 }
17640 /* Test all vectors which contain at-least 8 elements. */
17642 static void
17644 {
17646 {
17647 /* Case 1: sel_npatterns (4) > input npatterns (2)
17648 sel: { 0, 0, 1, len, 2, 0, 3, len, 4, 0, 5, len, ...} // (4, 3)
17649 res: { arg0[0], arg0[0], arg0[0], arg1[0],
17650 arg0[2], arg0[0], arg0[3], arg1[0],
17651 arg0[4], arg0[0], arg0[5], arg1[0], ... } // (4, 3) */
17652 {
17669 }
17670 }
17671 }
17673 /* Test vectors for which nunits[0] <= 4. */
17675 static void
17677 {
17678 /* Case 1: mask = {0, 4, ...} // (1, 2)
17679 This should return NULL_TREE because the index 4 may choose
17680 from either arg0 or arg1 depending on vector length. */
17681 {
17696 }
17697 }
17699 #undef ARG0
17700 #undef ARG1
17702 /* Return true if SIZE is of the form C + Cx and C is power of 2. */
17704 static bool
17706 {
17714 }
17716 /* Execute fold_vec_perm_cst unit tests. */
17718 static void
17720 {
17724 machine_mode vmode;
17726 {
17727 /* Obtain modes corresponding to VNx4SI and V4SI,
17728 to call mixed mode tests below.
17729 FIXME: Is there a better way to do this ? */
17731 {
17738 }
17755 }
17760 {
17763 }
17764 }
17767 /* Verify that various binary operations on vectors are folded
17768 correctly. */
17770 static void
17772 {
17779 /* Verify equality tests that return a scalar boolean result. */
17792 }
17794 /* Verify folding of VEC_DUPLICATE_EXPRs. */
17796 static void
17798 {
17801 /* This will be 1 if VEC_MODE isn't a vector mode. */
17808 }
17810 /* Run all of the selftests within this file. */
17812 void
17814 {
17819 }