| blob | f5d68ac323a1f287e69a47d27e5d759ed7d39c26 |
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. */
2146 }
2148 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2149 to an integer type. */
2151 static tree
2153 {
2155 tree t;
2157 /* The following code implements the floating point to integer
2158 conversion rules required by the Java Language Specification,
2159 that IEEE NaNs are mapped to zero and values that overflow
2160 the target precision saturate, i.e. values greater than
2161 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
2162 are mapped to INT_MIN. These semantics are allowed by the
2163 C and C++ standards that simply state that the behavior of
2164 FP-to-integer conversion is unspecified upon overflow. */
2166 wide_int val;
2167 REAL_VALUE_TYPE r;
2171 {
2178 }
2180 /* If R is NaN, return zero and show we have an overflow. */
2182 {
2185 }
2187 /* See if R is less than the lower bound or greater than the
2188 upper bound. */
2191 {
2195 {
2198 }
2199 }
2202 {
2205 {
2208 {
2211 }
2212 }
2213 }
2220 }
2222 /* A subroutine of fold_convert_const handling conversions of a
2223 FIXED_CST to an integer type. */
2225 static tree
2227 {
2228 tree t;
2230 scalar_mode mode;
2232 /* Right shift FIXED_CST to temp by fbit. */
2236 {
2238 HOST_BITS_PER_DOUBLE_INT,
2241 /* Left shift temp to temp_trunc by fbit. */
2243 HOST_BITS_PER_DOUBLE_INT,
2245 }
2246 else
2247 {
2250 }
2252 /* If FIXED_CST is negative, we need to round the value toward 0.
2253 By checking if the fractional bits are not zero to add 1 to temp. */
2259 /* Given a fixed-point constant, make new constant with new type,
2260 appropriately sign-extended or truncated. */
2268 }
2270 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2271 to another floating point type. */
2273 static tree
2275 {
2276 REAL_VALUE_TYPE value;
2277 tree t;
2279 /* If the underlying modes are the same, simply treat it as
2280 copy and rebuild with TREE_REAL_CST information and the
2281 given type. */
2283 {
2286 }
2288 /* Don't perform the operation if flag_signaling_nans is on
2289 and the operand is a signaling NaN. */
2294 /* With flag_rounding_math we should respect the current rounding mode
2295 unless the conversion is exact. */
2303 /* If converting an infinity or NAN to a representation that doesn't
2304 have one, set the overflow bit so that we can produce some kind of
2305 error message at the appropriate point if necessary. It's not the
2306 most user-friendly message, but it's better than nothing. */
2313 /* Regular overflow, conversion produced an infinity in a mode that
2314 can't represent them. */
2319 else
2322 }
2324 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2325 to a floating point type. */
2327 static tree
2329 {
2330 REAL_VALUE_TYPE value;
2331 tree t;
2339 }
2341 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2342 to another fixed-point type. */
2344 static tree
2346 {
2347 FIXED_VALUE_TYPE value;
2348 tree t;
2355 /* Propagate overflow flags. */
2359 }
2361 /* A subroutine of fold_convert_const handling conversions an INTEGER_CST
2362 to a fixed-point type. */
2364 static tree
2366 {
2367 FIXED_VALUE_TYPE value;
2368 tree t;
2370 double_int di;
2377 else
2385 /* Propagate overflow flags. */
2389 }
2391 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2392 to a fixed-point type. */
2394 static tree
2396 {
2397 FIXED_VALUE_TYPE value;
2398 tree t;
2406 /* Propagate overflow flags. */
2410 }
2412 /* Attempt to fold type conversion operation CODE of expression ARG1 to
2413 type TYPE. If no simplification can be done return NULL_TREE. */
2415 static tree
2417 {
2422 /* We can't widen types, since the runtime value could overflow the
2423 original type before being extended to the new type. */
2434 {
2441 }
2443 {
2445 {
2447 /* Avoid the folding if flag_rounding_math is on and the
2448 conversion is not exact. */
2450 {
2456 }
2458 }
2463 }
2465 {
2472 }
2474 {
2477 {
2480 /* We can't handle steps directly when extending, since the
2481 values need to wrap at the original precision first. */
2482 bool step_ok_p
2486 tree_vector_builder v;
2491 {
2497 }
2499 }
2500 }
2502 }
2504 /* Construct a vector of zero elements of vector type TYPE. */
2506 static tree
2508 {
2509 tree t;
2513 }
2515 /* Returns true, if ARG is convertible to TYPE using a NOP_EXPR. */
2517 bool
2519 {
2534 {
2556 }
2557 }
2559 /* Convert expression ARG to type TYPE. Used by the middle-end for
2560 simple conversions in preference to calling the front-end's convert. */
2562 tree
2564 {
2566 tree tem;
2577 {
2580 /* Handle conversions between pointers to different address spaces. */
2585 /* fall through */
2590 {
2594 }
2608 {
2612 }
2614 {
2618 }
2620 {
2624 }
2627 {
2645 }
2650 {
2654 }
2657 {
2672 }
2676 {
2685 integer_zero_node));
2687 {
2691 {
2697 }
2705 }
2709 }
2727 }
2728 fold_convert_exit:
2731 }
2732 \f
2733 /* Return false if expr can be assumed not to be an lvalue, true
2734 otherwise. */
2736 static bool
2738 {
2739 /* We only need to wrap lvalue tree codes. */
2741 {
2774 /* Assume the worst for front-end tree codes. */
2778 }
2781 }
2783 /* Return an expr equal to X but certainly not valid as an lvalue. */
2785 tree
2787 {
2788 /* While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to
2789 us. */
2796 }
2797 \f
2798 /* Given a tree comparison code, return the code that is the logical inverse.
2799 It is generally not safe to do this for floating-point comparisons, except
2800 for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
2801 ERROR_MARK in this case. */
2803 enum tree_code
2805 {
2811 {
2842 }
2843 }
2845 /* Similar, but return the comparison that results if the operands are
2846 swapped. This is safe for floating-point. */
2848 enum tree_code
2850 {
2852 {
2878 }
2879 }
2882 /* Convert a comparison tree code from an enum tree_code representation
2883 into a compcode bit-based encoding. This function is the inverse of
2884 compcode_to_comparison. */
2886 static enum comparison_code
2888 {
2890 {
2921 }
2922 }
2924 /* Convert a compcode bit-based encoding of a comparison operator back
2925 to GCC's enum tree_code representation. This function is the
2926 inverse of comparison_to_compcode. */
2928 static enum tree_code
2930 {
2932 {
2963 }
2964 }
2966 /* Return true if COND1 tests the opposite condition of COND2. */
2968 bool
2970 {
2973 && (invert_tree_comparison
2980 }
2982 /* Return a tree for the comparison which is the combination of
2983 doing the AND or OR (depending on CODE) of the two operations LCODE
2984 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2985 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2986 if this makes the transformation invalid. */
2988 tree
2993 {
3000 {
3011 }
3014 {
3015 /* Eliminate unordered comparisons, as well as LTGT and ORD
3016 which are not used unless the mode has NaNs. */
3022 }
3024 {
3025 /* Check that the original operation and the optimized ones will trap
3026 under the same condition. */
3037 /* In a short-circuited boolean expression the LHS might be
3038 such that the RHS, if evaluated, will never trap. For
3039 example, in ORD (x, y) && (x < y), we evaluate the RHS only
3040 if neither x nor y is NaN. (This is a mixed blessing: for
3041 example, the expression above will never trap, hence
3042 optimizing it to x < y would be invalid). */
3047 /* If the comparison was short-circuited, and only the RHS
3048 trapped, we may now generate a spurious trap. */
3053 /* If we changed the conditions that cause a trap, we lose. */
3056 }
3062 else
3063 {
3068 }
3069 }
3070 \f
3071 /* Return nonzero if two operands (typically of the same tree node)
3072 are necessarily equal. FLAGS modifies behavior as follows:
3074 If OEP_ONLY_CONST is set, only return nonzero for constants.
3075 This function tests whether the operands are indistinguishable;
3076 it does not test whether they are equal using C's == operation.
3077 The distinction is important for IEEE floating point, because
3078 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
3079 (2) two NaNs may be indistinguishable, but NaN!=NaN.
3081 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
3082 even though it may hold multiple values during a function.
3083 This is because a GCC tree node guarantees that nothing else is
3084 executed between the evaluation of its "operands" (which may often
3085 be evaluated in arbitrary order). Hence if the operands themselves
3086 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
3087 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
3088 unset means assuming isochronic (or instantaneous) tree equivalence.
3089 Unless comparing arbitrary expression trees, such as from different
3090 statements, this flag can usually be left unset.
3092 If OEP_PURE_SAME is set, then pure functions with identical arguments
3093 are considered the same. It is used when the caller has other ways
3094 to ensure that global memory is unchanged in between.
3096 If OEP_ADDRESS_OF is set, we are actually comparing addresses of objects,
3097 not values of expressions.
3099 If OEP_LEXICOGRAPHIC is set, then also handle expressions with side-effects
3100 such as MODIFY_EXPR, RETURN_EXPR, as well as STATEMENT_LISTs.
3102 If OEP_BITWISE is set, then require the values to be bitwise identical
3103 rather than simply numerically equal. Do not take advantage of things
3104 like math-related flags or undefined behavior; only return true for
3105 values that are provably bitwise identical in all circumstances.
3107 Unless OEP_MATCH_SIDE_EFFECTS is set, the function returns false on
3108 any operand with side effect. This is unnecesarily conservative in the
3109 case we know that arg0 and arg1 are in disjoint code paths (such as in
3110 ?: operator). In addition OEP_MATCH_SIDE_EFFECTS is used when comparing
3111 addresses with TREE_CONSTANT flag set so we know that &var == &var
3112 even if var is volatile. */
3114 bool
3117 {
3125 /* If either is ERROR_MARK, they aren't equal. */
3131 /* Similar, if either does not have a type (like a template id),
3132 they aren't equal. */
3136 /* Bitwise identity makes no sense if the values have different layouts. */
3141 /* We cannot consider pointers to different address space equal. */
3148 /* Check equality of integer constants before bailing out due to
3149 precision differences. */
3151 {
3152 /* Address of INTEGER_CST is not defined; check that we did not forget
3153 to drop the OEP_ADDRESS_OF flags. */
3156 }
3159 {
3160 /* If both types don't have the same signedness, then we can't consider
3161 them equal. We must check this before the STRIP_NOPS calls
3162 because they may change the signedness of the arguments. As pointers
3163 strictly don't have a signedness, require either two pointers or
3164 two non-pointers as well. */
3170 /* If both types don't have the same precision, then it is not safe
3171 to strip NOPs. */
3178 }
3179 #if 0
3180 /* FIXME: Fortran FE currently produce ADDR_EXPR of NOP_EXPR. Enable the
3181 sanity check once the issue is solved. */
3182 else
3183 /* Addresses of conversions and SSA_NAMEs (and many other things)
3184 are not defined. Check that we did not forget to drop the
3185 OEP_ADDRESS_OF/OEP_CONSTANT_ADDRESS_OF flags. */
3188 #endif
3190 /* In case both args are comparisons but with different comparison
3191 code, try to swap the comparison operands of one arg to produce
3192 a match and compare that variant. */
3196 {
3204 }
3207 {
3208 /* NOP_EXPR and CONVERT_EXPR are considered equal. */
3210 ;
3212 {
3213 /* If we are interested in comparing addresses ignore
3214 MEM_REF wrappings of the base that can appear just for
3215 TBAA reasons. */
3229 }
3230 else
3232 }
3234 /* When not checking adddresses, this is needed for conversions and for
3235 COMPONENT_REF. Might as well play it safe and always test this. */
3242 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
3243 We don't care about side effects in that case because the SAVE_EXPR
3244 takes care of that for us. In all other cases, two expressions are
3245 equal if they have no side effects. If we have two identical
3246 expressions with side effects that should be treated the same due
3247 to the only side effects being identical SAVE_EXPR's, that will
3248 be detected in the recursive calls below.
3249 If we are taking an invariant address of two identical objects
3250 they are necessarily equal as well. */
3257 /* Next handle constant cases, those for which we can return 1 even
3258 if ONLY_CONST is set. */
3261 {
3274 {
3275 /* If we do not distinguish between signed and unsigned zero,
3276 consider them equal. */
3279 }
3283 {
3298 }
3302 flags)
3304 flags));
3315 flags | OEP_ADDRESS_OF
3318 {
3319 /* In GIMPLE empty constructors are allowed in initializers of
3320 aggregates. */
3324 /* See sem_variable::equals in ipa-icf for a similar approach. */
3331 {
3332 /* For arrays, check that the sizes all match. */
3338 }
3347 /* Address of CONSTRUCTOR is defined in GENERIC to mean the value
3348 of the CONSTRUCTOR referenced indirectly. */
3352 {
3356 /* Check that the values are the same... */
3361 /* ... and that they apply to the same field! */
3367 flags)
3370 flags)))
3372 }
3375 }
3379 }
3381 /* Don't handle more cases for OEP_BITWISE, since we can't guarantee that
3382 two instances of undefined behavior will give identical results. */
3386 /* Define macros to test an operand from arg0 and arg1 for equality and a
3387 variant that allows null and views null as being different from any
3388 non-null value. In the latter case, if either is null, the both
3389 must be; otherwise, do the normal comparison. */
3390 #define OP_SAME(N) operand_equal_p (TREE_OPERAND (arg0, N), \
3391 TREE_OPERAND (arg1, N), flags)
3393 #define OP_SAME_WITH_NULL(N) \
3394 ((!TREE_OPERAND (arg0, N) || !TREE_OPERAND (arg1, N)) \
3395 ? TREE_OPERAND (arg0, N) == TREE_OPERAND (arg1, N) : OP_SAME (N))
3398 {
3400 /* Two conversions are equal only if signedness and modes match. */
3402 {
3403 CASE_CONVERT:
3411 }
3421 /* For commutative ops, allow the other order. */
3429 /* If either of the pointer (or reference) expressions we are
3430 dereferencing contain a side effect, these cannot be equal,
3431 but their addresses can be. */
3438 {
3441 {
3445 /* Verify that the access types are compatible. */
3449 }
3454 /* Require the same offset. */
3460 /* Fallthru. */
3468 {
3469 /* Require equal access sizes */
3475 flags)))
3477 /* Verify that access happens in similar types. */
3480 /* Verify that accesses are TBAA compatible. */
3489 /* Verify that alignment is compatible. */
3493 }
3496 /* TARGET_MEM_REF require equal extra operands. */
3507 /* Compare the array index by value if it is constant first as we
3508 may have different types but same value here. */
3514 /* Compare low bound and element size as with OEP_ADDRESS_OF
3515 we have to account for the offset of the ref. */
3520 array_ref_low_bound
3524 array_ref_element_size
3526 flags))));
3529 /* Handle operand 2 the same as for ARRAY_REF. Operand 0
3530 may be NULL when we're called to compare MEM_EXPRs. */
3533 {
3536 /* Most of time we only need to compare FIELD_DECLs for equality.
3537 However when determining address look into actual offsets.
3538 These may match for unions and unshared record types. */
3541 {
3544 {
3548 /* Non-FIELD_DECL operands can appear in C++ templates. */
3555 flags))
3557 }
3558 else
3560 }
3561 }
3572 }
3576 {
3578 /* Be sure we pass right ADDRESS_OF flag. */
3595 /* The multiplcation operands are commutative. */
3596 /* FALLTHRU */
3604 /* Otherwise take into account this is a commutative operation. */
3617 /* BIT_INSERT_EXPR has an implict operand as the type precision
3618 of op1. Need to check to make sure they are the same. */
3624 /* FALLTHRU */
3649 /* Virtual table reference. */
3661 {
3666 }
3671 }
3675 {
3679 /* If not both CALL_EXPRs are either internal or normal function
3680 functions, then they are not equal. */
3683 {
3684 /* If the CALL_EXPRs call different internal functions, then they
3685 are not equal. */
3688 }
3689 else
3690 {
3691 /* If the CALL_EXPRs call different functions, then they are not
3692 equal. */
3694 flags))
3696 }
3698 /* FIXME: We could skip this test for OEP_MATCH_SIDE_EFFECTS. */
3699 {
3703 else
3707 }
3709 /* Now see if all the arguments are the same. */
3710 {
3721 /* If we get here and both argument lists are exhausted
3722 then the CALL_EXPRs are equal. */
3724 }
3727 }
3730 /* Consider __builtin_sqrt equal to sqrt. */
3740 {
3741 /* Consider decls with the same name equal. The caller needs
3742 to make sure they refer to the same entity (such as a function
3743 formal parameter). */
3749 }
3754 {
3758 /* In GIMPLE constructors are used only to build vectors from
3759 elements. Individual elements in the constructor must be
3760 indexed in increasing order and form an initial sequence.
3762 We make no effort to compare nonconstant ones in GENERIC. */
3767 /* Be sure that vectors constructed have the same representation.
3768 We only tested element precision and modes to match.
3769 Vectors may be BLKmode and thus also check that the number of
3770 parts match. */
3783 {
3788 /* In GIMPLE the indexes can be either NULL or matching i.
3789 Double check this so we won't get false
3790 positives for GENERIC. */
3798 }
3800 }
3803 {
3804 /* Compare the STATEMENT_LISTs. */
3810 {
3811 /* The lists don't have the same number of statements. */
3817 flags & (OEP_LEXICOGRAPHIC
3820 }
3821 }
3826 {
3837 }
3841 }
3843 #undef OP_SAME
3844 #undef OP_SAME_WITH_NULL
3845 }
3847 /* Generate a hash value for an expression. This can be used iteratively
3848 by passing a previous result as the HSTATE argument. */
3850 void
3853 {
3859 {
3862 }
3872 {
3873 /* Alas, constants aren't shared, so we can't rely on pointer
3874 identity. */
3884 {
3888 else
3892 }
3894 {
3898 }
3908 {
3915 }
3917 /* We can just compare by pointer. */
3921 /* The node itself doesn't matter. */
3925 /* Ignore. */
3928 /* A list of expressions, for a CALL_EXPR or as the elements of a
3929 VECTOR_CST. */
3934 {
3940 {
3941 /* In GIMPLE the indexes can be either NULL or matching i. */
3945 {
3948 }
3949 else
3952 }
3954 }
3956 {
3957 tree_stmt_iterator i;
3962 }
3971 /* When referring to a built-in FUNCTION_DECL, use the __builtin__ form.
3972 Otherwise nodes that compare equal according to operand_equal_p might
3973 get different hash codes. However, don't do this for machine specific
3974 or front end builtins, since the function code is overloaded in those
3975 cases. */
3978 {
3981 }
3982 /* FALL THROUGH */
3985 {
3989 }
3993 {
3994 /* DECL's have a unique ID */
3996 }
3998 {
3999 /* For comparisons that can be swapped, use the lower
4000 tree code. */
4007 }
4009 {
4010 /* NOP_EXPR and CONVERT_EXPR are considered equal by
4011 operand_equal_p. */
4015 /* Don't hash the type, that can lead to having nodes which
4016 compare equal according to operand_equal_p, but which
4017 have different hash codes. Make sure to include signedness
4018 in the hash computation. */
4021 }
4022 /* For OEP_ADDRESS_OF, hash MEM_EXPR[&decl, 0] the same as decl. */
4030 /* Don't ICE on FE specific trees, or their arguments etc.
4031 during operand_equal_p hash verification. */
4034 else
4035 {
4041 {
4057 {
4064 }
4078 {
4079 /* The multiplication operands are commutative. */
4086 }
4094 /* For TARGET_EXPR, just hash on the TARGET_EXPR_SLOT.
4095 Usually different TARGET_EXPRs just should use
4096 different temporaries in their slots. */
4101 /* Virtual table reference. */
4109 {
4111 /* We compute mangled names only when free_lang_data is run.
4112 In that case we can hash precisely. */
4115 hstate.add_object
4116 (IDENTIFIER_HASH_VALUE
4118 }
4122 }
4124 /* Don't hash the type, that can lead to having nodes which
4125 compare equal according to operand_equal_p, but which
4126 have different hash codes. */
4128 {
4129 /* Make sure to include signness in the hash computation. */
4132 }
4135 {
4136 /* It's a commutative expression. We want to hash it the same
4137 however it appears. We do this by first hashing both operands
4138 and then rehashing based on the order of their independent
4139 hashes. */
4144 }
4145 else
4149 }
4151 }
4152 }
4154 bool
4157 {
4158 /* When checking and unless comparing DECL names, verify that if
4159 the outermost operand_equal_p call returns non-zero then ARG0
4160 and ARG1 have the same hash value. */
4162 {
4164 {
4166 {
4173 }
4175 }
4176 else
4180 }
4183 }
4188 /* Conveinece wrapper around operand_compare class because usually we do
4189 not need to play with the valueizer. */
4191 bool
4193 {
4195 }
4197 namespace inchash
4198 {
4200 /* Generate a hash value for an expression. This can be used iteratively
4201 by passing a previous result as the HSTATE argument.
4203 This function is intended to produce the same hash for expressions which
4204 would compare equal using operand_equal_p. */
4205 void
4207 {
4209 }
4211 }
4212 \f
4213 /* Similar to operand_equal_p, but see if ARG0 might be a variant of ARG1
4214 with a different signedness or a narrower precision. */
4216 static bool
4218 {
4226 /* Discard any conversions that don't change the modes of ARG0 and ARG1
4227 and see if the inner values are the same. This removes any
4228 signedness comparison, which doesn't matter here. */
4236 /* Discard a single widening conversion from ARG1 and see if the inner
4237 value is the same as ARG0. */
4246 }
4247 \f
4248 /* See if ARG is an expression that is either a comparison or is performing
4249 arithmetic on comparisons. The comparisons must only be comparing
4250 two different values, which will be stored in *CVAL1 and *CVAL2; if
4251 they are nonzero it means that some operands have already been found.
4252 No variables may be used anywhere else in the expression except in the
4253 comparisons.
4255 If this is true, return 1. Otherwise, return zero. */
4257 static bool
4259 {
4263 /* We can handle some of the tcc_expression cases here. */
4272 {
4291 /* First see if we can handle the first operand, then the second. For
4292 the second operand, we know *CVAL1 can't be zero. It must be that
4293 one side of the comparison is each of the values; test for the
4294 case where this isn't true by failing if the two operands
4295 are the same. */
4304 ;
4308 ;
4309 else
4313 ;
4317 ;
4318 else
4325 }
4326 }
4327 \f
4328 /* ARG is a tree that is known to contain just arithmetic operations and
4329 comparisons. Evaluate the operations in the tree substituting NEW0 for
4330 any occurrence of OLD0 as an operand of a comparison and likewise for
4331 NEW1 and OLD1. */
4333 static tree
4336 {
4341 /* We can handle some of the tcc_expression cases here. */
4349 {
4364 {
4383 }
4384 /* Fall through - ??? */
4387 {
4391 /* We need to check both for exact equality and tree equality. The
4392 former will be true if the operand has a side-effect. In that
4393 case, we know the operand occurred exactly once. */
4406 }
4410 }
4411 }
4412 \f
4413 /* Return a tree for the case when the result of an expression is RESULT
4414 converted to TYPE and OMITTED was previously an operand of the expression
4415 but is now not needed (e.g., we folded OMITTED * 0).
4417 If OMITTED has side effects, we must evaluate it. Otherwise, just do
4418 the conversion of RESULT to TYPE. */
4420 tree
4422 {
4425 /* If the resulting operand is an empty statement, just return the omitted
4426 statement casted to void. */
4436 }
4438 /* Return a tree for the case when the result of an expression is RESULT
4439 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
4440 of the expression but are now not needed.
4442 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
4443 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
4444 evaluated before OMITTED2. Otherwise, if neither has side effects,
4445 just do the conversion of RESULT to TYPE. */
4447 tree
4450 {
4459 }
4461 \f
4462 /* Return a simplified tree node for the truth-negation of ARG. This
4463 never alters ARG itself. We assume that ARG is an operation that
4464 returns a truth value (0 or 1).
4466 FIXME: one would think we would fold the result, but it causes
4467 problems with the dominator optimizer. */
4469 static tree
4471 {
4476 /* If this is a comparison, we can simply invert it, except for
4477 floating-point non-equality comparisons, in which case we just
4478 enclose a TRUTH_NOT_EXPR around what we have. */
4481 {
4484 && flag_trapping_math
4497 }
4500 {
4519 /* Here we can invert either operand. We invert the first operand
4520 unless the second operand is a TRUTH_NOT_EXPR in which case our
4521 result is the XOR of the first operand with the inside of the
4522 negation of the second operand. */
4527 else
4550 {
4557 /* A COND_EXPR may have a throw as one operand, which
4558 then has void type. Just leave void operands
4559 as they are. */
4565 }
4577 CASE_CONVERT:
4581 /* fall through */
4603 }
4604 }
4606 /* Fold the truth-negation of ARG. This never alters ARG itself. We
4607 assume that ARG is an operation that returns a truth value (0 or 1
4608 for scalars, 0 or -1 for vectors). Return the folded expression if
4609 folding is successful. Otherwise, return NULL_TREE. */
4611 static tree
4613 {
4616 ? BIT_NOT_EXPR
4619 }
4621 /* Return a simplified tree node for the truth-negation of ARG. This
4622 never alters ARG itself. We assume that ARG is an operation that
4623 returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). */
4625 tree
4627 {
4633 ? BIT_NOT_EXPR
4636 }
4637 \f
4638 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
4639 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero
4640 and uses reverse storage order if REVERSEP is nonzero. ORIG_INNER
4641 is the original memory reference used to preserve the alias set of
4642 the access. */
4644 static tree
4648 {
4651 /* Attempt not to lose the access path if possible. */
4653 {
4655 machine_mode nmode;
4657 tree noffset;
4665 && !reversep
4666 && !nreversep
4668 {
4671 }
4672 }
4681 {
4688 }
4703 }
4705 /* Optimize a bit-field compare.
4707 There are two cases: First is a compare against a constant and the
4708 second is a comparison of two items where the fields are at the same
4709 bit position relative to the start of a chunk (byte, halfword, word)
4710 large enough to contain it. In these cases we can avoid the shift
4711 implicit in bitfield extractions.
4713 For constants, we emit a compare of the shifted constant with the
4714 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
4715 compared. For two fields at the same position, we do the ANDs with the
4716 similar mask and compare the result of the ANDs.
4718 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
4719 COMPARE_TYPE is the type of the comparison, and LHS and RHS
4720 are the left and right operands of the comparison, respectively.
4722 If the optimization described above can be done, we return the resulting
4723 tree. Otherwise we return zero. */
4725 static tree
4728 {
4732 tree unsigned_type;
4735 scalar_int_mode nmode;
4740 tree mask;
4741 tree offset;
4743 /* Get all the information about the extractions being done. If the bit size
4744 is the same as the size of the underlying object, we aren't doing an
4745 extraction at all and so can do nothing. We also don't want to
4746 do anything if the inner expression is a PLACEHOLDER_EXPR since we
4747 then will no longer be able to replace it. */
4762 else
4763 {
4764 /* If this is not a constant, we can only do something if bit positions,
4765 sizes, signedness and storage order are the same. */
4766 rinner
4779 }
4781 /* Honor the C++ memory model and mimic what RTL expansion does. */
4785 {
4789 }
4791 /* See if we can find a mode to refer to this field. We should be able to,
4792 but fail if we can't. */
4800 /* Set signed and unsigned types of the precision of this mode for the
4801 shifts below. */
4804 /* Compute the bit position and size for the new reference and our offset
4805 within it. If the new reference is the same size as the original, we
4806 won't optimize anything, so return zero. */
4816 /* Make the mask to be used against the extracted field. */
4823 {
4827 /* If not comparing with constant, just rework the comparison
4828 and return. */
4836 }
4838 /* Otherwise, we are handling the constant case. See if the constant is too
4839 big for the field. Warn and return a tree for 0 (false) if so. We do
4840 this not only for its own sake, but to avoid having to test for this
4841 error case below. If we didn't, we might generate wrong code.
4843 For unsigned fields, the constant shifted right by the field length should
4844 be all zero. For signed fields, the high-order bits should agree with
4845 the sign bit. */
4848 {
4850 {
4854 }
4855 }
4856 else
4857 {
4860 {
4864 }
4865 }
4870 /* Single-bit compares should always be against zero. */
4872 {
4875 }
4877 /* Make a new bitfield reference, shift the constant over the
4878 appropriate number of bits and mask it with the computed mask
4879 (in case this was a signed field). If we changed it, make a new one. */
4887 mask);
4892 }
4893 \f
4894 /* Subroutine for fold_truth_andor_1: decode a field reference.
4896 If EXP is a comparison reference, we return the innermost reference.
4898 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
4899 set to the starting bit number.
4901 If the innermost field can be completely contained in a mode-sized
4902 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
4904 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
4905 otherwise it is not changed.
4907 *PUNSIGNEDP is set to the signedness of the field.
4909 *PREVERSEP is set to the storage order of the field.
4911 *PMASK is set to the mask used. This is either contained in a
4912 BIT_AND_EXPR or derived from the width of the field.
4914 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
4916 Return 0 if this is not a component reference or is one that we can't
4917 do anything with. */
4919 static tree
4924 {
4929 tree unsigned_type;
4932 /* All the optimizations using this function assume integer fields.
4933 There are problems with FP fields since the type_for_size call
4934 below can fail for, e.g., XFmode. */
4938 /* We are interested in the bare arrangement of bits, so strip everything
4939 that doesn't affect the machine mode. However, record the type of the
4940 outermost expression if it may matter below. */
4947 {
4953 }
4964 /* Reject out-of-bound accesses (PR79731). */
4976 /* If the number of bits in the reference is the same as the bitsize of
4977 the outer type, then the outer type gives the signedness. Otherwise
4978 (in case of a small bitfield) the signedness is unchanged. */
4982 /* Compute the mask to access the bitfield. */
4990 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
4998 }
5000 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
5001 bit positions and MASK is SIGNED. */
5003 static bool
5005 {
5009 /* If this function returns true when the type of the mask is
5010 UNSIGNED, then there will be errors. In particular see
5011 gcc.c-torture/execute/990326-1.c. There does not appear to be
5012 any documentation paper trail as to why this is so. But the pre
5013 wide-int worked with that restriction and it has been preserved
5014 here. */
5019 }
5021 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
5022 represents the sign bit of EXP's type. If EXP represents a sign
5023 or zero extension, also test VAL against the unextended type.
5024 The return value is the (sub)expression whose sign bit is VAL,
5025 or NULL_TREE otherwise. */
5027 tree
5029 {
5031 tree t;
5033 /* Tree EXP must have an integral type. */
5038 /* Tree VAL must be an integer constant. */
5047 /* Handle extension from a narrower type. */
5053 }
5055 /* Subroutine for fold_truth_andor_1 and simple_condition_p: determine if an
5056 operand is simple enough to be evaluated unconditionally. */
5058 static bool
5060 {
5061 /* Strip any conversions that don't change the machine mode. */
5070 /* Don't regard global variables as simple. They may be
5071 allocated in ways unknown to the compiler (shared memory,
5072 #pragma weak, etc). */
5075 /* Weakrefs are not safe to be read, since they can be NULL.
5076 They are !TREE_PUBLIC && !DECL_EXTERNAL but still
5077 have DECL_WEAK flag set. */
5079 /* Loading a static variable is unduly expensive, but global
5080 registers aren't expensive. */
5082 }
5084 /* Determine if an operand is simple enough to be evaluated unconditionally.
5085 In addition to simple_operand_p, we assume that comparisons, conversions,
5086 and logic-not operations are simple, if their operands are simple, too. */
5088 bool
5090 {
5109 }
5111 \f
5112 /* The following functions are subroutines to fold_range_test and allow it to
5113 try to change a logical combination of comparisons into a range test.
5115 For example, both
5116 X == 2 || X == 3 || X == 4 || X == 5
5117 and
5118 X >= 2 && X <= 5
5119 are converted to
5120 (unsigned) (X - 2) <= 3
5122 We describe each set of comparisons as being either inside or outside
5123 a range, using a variable named like IN_P, and then describe the
5124 range with a lower and upper bound. If one of the bounds is omitted,
5125 it represents either the highest or lowest value of the type.
5127 In the comments below, we represent a range by two numbers in brackets
5128 preceded by a "+" to designate being inside that range, or a "-" to
5129 designate being outside that range, so the condition can be inverted by
5130 flipping the prefix. An omitted bound is represented by a "-". For
5131 example, "- [-, 10]" means being outside the range starting at the lowest
5132 possible value and ending at 10, in other words, being greater than 10.
5133 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
5134 always false.
5136 We set up things so that the missing bounds are handled in a consistent
5137 manner so neither a missing bound nor "true" and "false" need to be
5138 handled using a special case. */
5140 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
5141 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
5142 and UPPER1_P are nonzero if the respective argument is an upper bound
5143 and zero for a lower. TYPE, if nonzero, is the type of the result; it
5144 must be specified for a comparison. ARG1 will be converted to ARG0's
5145 type if both are specified. */
5147 static tree
5150 {
5151 tree tem;
5155 /* If neither arg represents infinity, do the normal operation.
5156 Else, if not a comparison, return infinity. Else handle the special
5157 comparison rules. Note that most of the cases below won't occur, but
5158 are handled for consistency. */
5161 {
5166 }
5171 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
5172 for neither. In real maths, we cannot assume open ended ranges are
5173 the same. But, this is computer arithmetic, where numbers are finite.
5174 We can therefore make the transformation of any unbounded range with
5175 the value Z, Z being greater than any representable number. This permits
5176 us to treat unbounded ranges as equal. */
5180 {
5201 }
5204 }
5205 \f
5206 /* Helper routine for make_range. Perform one step for it, return
5207 new expression if the loop should continue or NULL_TREE if it should
5208 stop. */
5210 tree
5214 {
5220 {
5222 /* We can only do something if the range is testing for zero. */
5231 /* We can only do something if the range is testing for zero
5232 and if the second operand is an integer constant. Note that
5233 saying something is "in" the range we make is done by
5234 complementing IN_P since it will set in the initial case of
5235 being not equal to zero; "out" is leaving it alone. */
5242 {
5263 }
5265 /* If this is an unsigned comparison, we also know that EXP is
5266 greater than or equal to zero. We base the range tests we make
5267 on that fact, so we record it here so we can parse existing
5268 range tests. We test arg0_type since often the return type
5269 of, e.g. EQ_EXPR, is boolean. */
5271 {
5275 NULL_TREE))
5280 /* If the high bound is missing, but we have a nonzero low
5281 bound, reverse the range so it goes from zero to the low bound
5282 minus 1. */
5284 {
5289 }
5290 }
5298 /* If flag_wrapv and ARG0_TYPE is signed, make sure
5299 low and high are non-NULL, then normalize will DTRT. */
5302 {
5307 }
5309 /* (-x) IN [a,b] -> x in [-b, -a] */
5321 /* ~ X -> -X - 1 */
5330 /* If flag_wrapv and ARG0_TYPE is signed, then we cannot
5331 move a constant to the other side. */
5336 /* If EXP is signed, any overflow in the computation is undefined,
5337 so we don't worry about it so long as our computations on
5338 the bounds don't overflow. For unsigned, overflow is defined
5339 and this is exactly the right thing. */
5351 normalize:
5352 /* Check for an unsigned range which has wrapped around the maximum
5353 value thus making n_high < n_low, and normalize it. */
5355 {
5361 /* If the range is of the form +/- [ x+1, x ], we won't
5362 be able to normalize it. But then, it represents the
5363 whole range or the empty set, so make it
5364 +/- [ -, - ]. */
5368 else
5370 }
5371 else
5379 CASE_CONVERT:
5397 /* If we're converting arg0 from an unsigned type, to exp,
5398 a signed type, we will be doing the comparison as unsigned.
5399 The tests above have already verified that LOW and HIGH
5400 are both positive.
5402 So we have to ensure that we will handle large unsigned
5403 values the same way that the current signed bounds treat
5404 negative values. */
5407 {
5408 tree high_positive;
5409 tree equiv_type;
5410 /* For fixed-point modes, we need to pass the saturating flag
5411 as the 2nd parameter. */
5413 equiv_type
5418 else
5419 equiv_type
5422 /* A range without an upper bound is, naturally, unbounded.
5423 Since convert would have cropped a very large value, use
5424 the max value for the destination type. */
5425 high_positive
5432 high_positive),
5435 /* If the low bound is specified, "and" the range with the
5436 range for which the original unsigned value will be
5437 positive. */
5439 {
5442 integer_zero_node),
5443 high_positive))
5447 }
5448 else
5449 {
5450 /* Otherwise, "or" the range with the range of the input
5451 that will be interpreted as negative. */
5454 integer_zero_node),
5455 high_positive))
5459 }
5460 }
5462 /* Otherwise, if we are converting arg0 from signed type, to exp,
5463 an unsigned type, we will do the comparison as signed. If
5464 high is non-NULL, we punt above if it doesn't fit in the signed
5465 type, so if we get through here, +[-, high] or +[low, high] are
5466 equivalent to +[-, n_high] or +[n_low, n_high]. Similarly,
5467 +[-, -] or -[-, -] are equivalent too. But if low is specified and
5468 high is not, the +[low, -] range is equivalent to union of
5469 +[n_low, -] and +[-, -1] ranges, so +[low, -] is equivalent to
5470 -[0, n_low-1] and similarly -[low, -] to +[0, n_low-1], except for
5471 low being 0, which should be treated as [-, -]. */
5474 && low
5476 {
5479 else
5480 {
5485 }
5486 }
5495 }
5496 }
5498 /* Given EXP, a logical expression, set the range it is testing into
5499 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
5500 actually being tested. *PLOW and *PHIGH will be made of the same
5501 type as the returned expression. If EXP is not a comparison, we
5502 will most likely not be returning a useful value and range. Set
5503 *STRICT_OVERFLOW_P to true if the return value is only valid
5504 because signed overflow is undefined; otherwise, do not change
5505 *STRICT_OVERFLOW_P. */
5507 tree
5510 {
5518 /* Start with simply saying "EXP != 0" and then look at the code of EXP
5519 and see if we can refine the range. Some of the cases below may not
5520 happen, but it doesn't seem worth worrying about this. We "continue"
5521 the outer loop when we've changed something; otherwise we "break"
5522 the switch, which will "break" the while. */
5528 {
5534 {
5542 }
5551 }
5553 /* If EXP is a constant, we can evaluate whether this is true or false. */
5555 {
5562 }
5566 }
5568 /* Returns TRUE if [LOW, HIGH] range check can be optimized to
5569 a bitwise check i.e. when
5570 LOW == 0xXX...X00...0
5571 HIGH == 0xXX...X11...1
5572 Return corresponding mask in MASK and stem in VALUE. */
5574 static bool
5577 {
5600 }
5601 \f
5602 /* Helper routine for build_range_check and match.pd. Return the type to
5603 perform the check or NULL if it shouldn't be optimized. */
5605 tree
5607 {
5608 /* First make sure that arithmetics in this type is valid, then make sure
5609 that it wraps around. */
5614 {
5617 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
5618 for the type in question, as we rely on this here. */
5628 else
5630 }
5633 /* Right now all BITINT_TYPEs satisfy
5634 (unsigned) max + 1 == (unsigned) min, so no need to verify
5635 that like for INTEGER_TYPEs. */
5639 }
5641 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
5642 type, TYPE, return an expression to test if EXP is in (or out of, depending
5643 on IN_P) the range. Return 0 if the test couldn't be created. */
5645 tree
5648 {
5651 /* Disable this optimization for function pointer expressions
5652 on targets that require function pointer canonicalization. */
5659 {
5665 }
5687 value);
5690 {
5692 {
5696 }
5698 }
5700 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
5702 {
5706 {
5708 {
5711 etype
5713 else
5716 }
5719 }
5720 }
5722 /* Optimize (c>=low) && (c<=high) into (c-low>=0) && (c-low<=high-low).
5723 This requires wrap-around arithmetics for the type of the expression. */
5740 }
5741 \f
5742 /* Return the predecessor of VAL in its type, handling the infinite case. */
5744 static tree
5746 {
5752 else
5755 }
5757 /* Return the successor of VAL in its type, handling the infinite case. */
5759 static tree
5761 {
5767 else
5770 }
5772 /* Given two ranges, see if we can merge them into one. Return 1 if we
5773 can, 0 if we can't. Set the output range into the specified parameters. */
5775 bool
5778 {
5782 tree tem;
5792 /* Make range 0 be the range that starts first, or ends last if they
5793 start at the same value. Swap them if it isn't. */
5796 || (lowequal
5799 {
5803 }
5805 /* If the second range is != high1 where high1 is the type maximum of
5806 the type, try first merging with < high1 range. */
5808 && high1
5815 {
5820 /* Similarly for the second range != low1 where low1 is the type minimum
5821 of the type, try first merging with > low1 range. */
5826 }
5828 /* Now flag two cases, whether the ranges are disjoint or whether the
5829 second range is totally subsumed in the first. Note that the tests
5830 below are simplified by the ones above. */
5836 /* We now have four cases, depending on whether we are including or
5837 excluding the two ranges. */
5839 {
5840 /* If they don't overlap, the result is false. If the second range
5841 is a subset it is the result. Otherwise, the range is from the start
5842 of the second to the end of the first. */
5847 else
5849 }
5852 {
5853 /* If they don't overlap, the result is the first range. If they are
5854 equal, the result is false. If the second range is a subset of the
5855 first, and the ranges begin at the same place, we go from just after
5856 the end of the second range to the end of the first. If the second
5857 range is not a subset of the first, or if it is a subset and both
5858 ranges end at the same place, the range starts at the start of the
5859 first range and ends just before the second range.
5860 Otherwise, we can't describe this as a single range. */
5866 {
5871 {
5872 /* We are in the weird situation where high0 > high1 but
5873 high1 has no successor. Punt. */
5875 }
5876 }
5878 {
5883 {
5884 /* low0 < low1 but low1 has no predecessor. Punt. */
5886 }
5887 }
5888 else
5890 }
5893 {
5894 /* If they don't overlap, the result is the second range. If the second
5895 is a subset of the first, the result is false. Otherwise,
5896 the range starts just after the first range and ends at the
5897 end of the second. */
5902 else
5903 {
5908 {
5909 /* high1 > high0 but high0 has no successor. Punt. */
5911 }
5912 }
5913 }
5915 else
5916 {
5917 /* The case where we are excluding both ranges. Here the complex case
5918 is if they don't overlap. In that case, the only time we have a
5919 range is if they are adjacent. If the second is a subset of the
5920 first, the result is the first. Otherwise, the range to exclude
5921 starts at the beginning of the first range and ends at the end of the
5922 second. */
5924 {
5929 else
5930 {
5931 /* Canonicalize - [min, x] into - [-, x]. */
5934 {
5937 GET_MODE_BITSIZE
5940 /* FALLTHROUGH */
5953 }
5955 /* Canonicalize - [x, max] into - [x, -]. */
5958 {
5961 GET_MODE_BITSIZE
5964 /* FALLTHROUGH */
5980 }
5982 /* The ranges might be also adjacent between the maximum and
5983 minimum values of the given type. For
5984 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
5985 return + [x + 1, y - 1]. */
5987 {
5994 }
5995 else
5997 }
5998 }
6001 else
6003 }
6007 }
6008 \f
6010 /* Subroutine of fold, looking inside expressions of the form
6011 A op B ? A : C, where (ARG00, COMP_CODE, ARG01), ARG1 and ARG2
6012 are the three operands of the COND_EXPR. This function is
6013 being used also to optimize A op B ? C : A, by reversing the
6014 comparison first.
6016 Return a folded expression whose code is not a COND_EXPR
6017 anymore, or NULL_TREE if no folding opportunity is found. */
6019 static tree
6023 {
6025 tree tem;
6030 /* If we have A op 0 ? A : -A, consider applying the following
6031 transformations:
6033 A == 0? A : -A same as -A
6034 A != 0? A : -A same as A
6035 A >= 0? A : -A same as abs (A)
6036 A > 0? A : -A same as abs (A)
6037 A <= 0? A : -A same as -abs (A)
6038 A < 0? A : -A same as -abs (A)
6040 None of these transformations work for modes with signed
6041 zeros. If A is +/-0, the first two transformations will
6042 change the sign of the result (from +0 to -0, or vice
6043 versa). The last four will fix the sign of the result,
6044 even though the original expressions could be positive or
6045 negative, depending on the sign of A.
6047 Note that all these transformations are correct if A is
6048 NaN, since the two alternatives (A and -A) are also NaNs. */
6055 /* In the case that A is of the form X-Y, '-A' (arg2) may
6056 have already been folded to Y-X, check for that. */
6064 {
6076 /* Fall through. */
6087 /* FALLTHRU */
6094 {
6095 /* A <= 0 ? A : -A for A INT_MIN is valid, but -abs(INT_MIN)
6096 is not, invokes UB both in abs and in the negation of it.
6097 So, use ABSU_EXPR instead. */
6102 }
6103 else
6104 {
6107 }
6111 }
6113 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
6114 A == 0 ? A : 0 is always 0 unless A is -0. Note that
6115 both transformations are correct when A is NaN: A != 0
6116 is then true, and A == 0 is false. */
6120 {
6125 }
6127 /* Try some transformations of A op B ? A : B.
6129 A == B? A : B same as B
6130 A != B? A : B same as A
6131 A >= B? A : B same as max (A, B)
6132 A > B? A : B same as max (B, A)
6133 A <= B? A : B same as min (A, B)
6134 A < B? A : B same as min (B, A)
6136 As above, these transformations don't work in the presence
6137 of signed zeros. For example, if A and B are zeros of
6138 opposite sign, the first two transformations will change
6139 the sign of the result. In the last four, the original
6140 expressions give different results for (A=+0, B=-0) and
6141 (A=-0, B=+0), but the transformed expressions do not.
6143 The first two transformations are correct if either A or B
6144 is a NaN. In the first transformation, the condition will
6145 be false, and B will indeed be chosen. In the case of the
6146 second transformation, the condition A != B will be true,
6147 and A will be chosen.
6149 The conversions to max() and min() are not correct if B is
6150 a number and A is not. The conditions in the original
6151 expressions will be false, so all four give B. The min()
6152 and max() versions would give a NaN instead. */
6155 /* Avoid these transformations if the COND_EXPR may be used
6156 as an lvalue in the C++ front-end. PR c++/19199. */
6157 && (in_gimple_form
6163 {
6169 {
6178 /* In C++ a ?: expression can be an lvalue, so put the
6179 operand which will be used if they are equal first
6180 so that we can convert this back to the
6181 corresponding COND_EXPR. */
6183 {
6191 }
6198 {
6206 }
6219 }
6220 }
6223 }
6226 \f
6227 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
6228 #define LOGICAL_OP_NON_SHORT_CIRCUIT \
6229 (BRANCH_COST (optimize_function_for_speed_p (cfun), \
6230 false) >= 2)
6231 #endif
6233 /* EXP is some logical combination of boolean tests. See if we can
6234 merge it into some range test. Return the new tree if so. */
6236 static tree
6239 {
6253 /* If op0 is known true or false and this is a short-circuiting
6254 operation we must not merge with op1 since that makes side-effects
6255 unconditional. So special-case this. */
6262 /* If this is an OR operation, invert both sides; we will invert
6263 again at the end. */
6267 /* If both expressions are the same, if we can merge the ranges, and we
6268 can build the range test, return it or it inverted. If one of the
6269 ranges is always true or always false, consider it to be the same
6270 expression as the other. */
6278 {
6282 }
6284 /* On machines where the branch cost is expensive, if this is a
6285 short-circuited branch and the underlying object on both sides
6286 is the same, make a non-short-circuit operation. */
6289 logical_op_non_short_circuit
6296 {
6297 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
6298 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
6299 which cases we can't do this. */
6307 {
6316 {
6319 WARN_STRICT_OVERFLOW_COMPARISON);
6323 }
6324 }
6325 }
6328 }
6329 \f
6330 /* Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P
6331 bit value. Arrange things so the extra bits will be set to zero if and
6332 only if C is signed-extended to its full width. If MASK is nonzero,
6333 it is an INTEGER_CST that should be AND'ed with the extra bits. */
6335 static tree
6337 {
6340 tree temp;
6345 /* We work by getting just the sign bit into the low-order bit, then
6346 into the high-order bit, then sign-extend. We then XOR that value
6347 with C. */
6351 /* We must use a signed type in order to get an arithmetic right shift.
6352 However, we must also avoid introducing accidental overflows, so that
6353 a subsequent call to integer_zerop will work. Hence we must
6354 do the type conversion here. At this point, the constant is either
6355 zero or one, and the conversion to a signed type can never overflow.
6356 We could get an overflow if this conversion is done anywhere else. */
6365 /* If necessary, convert the type back to match the type of C. */
6370 }
6371 \f
6372 /* For an expression that has the form
6373 (A && B) || ~B
6374 or
6375 (A || B) && ~B,
6376 we can drop one of the inner expressions and simplify to
6377 A || ~B
6378 or
6379 A && ~B
6380 LOC is the location of the resulting expression. OP is the inner
6381 logical operation; the left-hand side in the examples above, while CMPOP
6382 is the right-hand side. RHS_ONLY is used to prevent us from accidentally
6383 removing a condition that guards another, as in
6384 (A != NULL && A->...) || A == NULL
6385 which we must not transform. If RHS_ONLY is true, only eliminate the
6386 right-most operand of the inner logical operation. */
6388 static tree
6391 {
6409 {
6412 {
6415 }
6416 }
6418 {
6421 {
6424 }
6425 }
6440 }
6442 /* Find ways of folding logical expressions of LHS and RHS:
6443 Try to merge two comparisons to the same innermost item.
6444 Look for range tests like "ch >= '0' && ch <= '9'".
6445 Look for combinations of simple terms on machines with expensive branches
6446 and evaluate the RHS unconditionally.
6448 For example, if we have p->a == 2 && p->b == 4 and we can make an
6449 object large enough to span both A and B, we can do this with a comparison
6450 against the object ANDed with the a mask.
6452 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
6453 operations to do this with one comparison.
6455 We check for both normal comparisons and the BIT_AND_EXPRs made this by
6456 function and the one above.
6458 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
6459 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
6461 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
6462 two operands.
6464 We return the simplified tree or 0 if no optimization is possible. */
6466 static tree
6469 {
6470 /* If this is the "or" of two comparisons, we can do something if
6471 the comparisons are NE_EXPR. If this is the "and", we can do something
6472 if the comparisons are EQ_EXPR. I.e.,
6473 (a->b == 2 && a->c == 4) can become (a->new == NEW).
6475 WANTED_CODE is this operation code. For single bit fields, we can
6476 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
6477 comparison for one-bit fields. */
6498 /* Start by getting the comparison codes. Fail if anything is volatile.
6499 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
6500 it were surrounded with a NE_EXPR. */
6509 {
6513 }
6516 {
6520 }
6531 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
6534 {
6537 {
6542 }
6545 {
6551 }
6552 }
6557 /* If the RHS can be evaluated unconditionally and its operands are
6558 simple, it wins to evaluate the RHS unconditionally on machines
6559 with expensive branches. In this case, this isn't a comparison
6560 that can be merged. */
6567 {
6568 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
6579 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
6589 }
6591 /* See if the comparisons can be merged. Then get all the parameters for
6592 each side. */
6617 /* It must be true that the inner operation on the lhs of each
6618 comparison must be the same if we are to be able to do anything.
6619 Then see if we have constants. If not, the same must be true for
6620 the rhs's. */
6629 {
6632 }
6637 else
6640 /* If either comparison code is not correct for our logical operation,
6641 fail. However, we can convert a one-bit comparison against zero into
6642 the opposite comparison against that bit being set in the field. */
6646 {
6648 {
6649 /* Make the left operand unsigned, since we are only interested
6650 in the value of one bit. Otherwise we are doing the wrong
6651 thing below. */
6654 }
6655 else
6657 }
6659 /* This is analogous to the code for l_const above. */
6661 {
6663 {
6666 }
6667 else
6669 }
6671 /* See if we can find a mode that contains both fields being compared on
6672 the left. If we can't, fail. Otherwise, update all constants and masks
6673 to be relative to a field of that size. */
6687 {
6690 }
6700 {
6709 {
6713 }
6714 }
6716 {
6725 {
6729 }
6730 }
6732 /* If the right sides are not constant, do the same for it. Also,
6733 disallow this optimization if a size, signedness or storage order
6734 mismatch occurs between the left and right sides. */
6736 {
6740 /* Make sure the two fields on the right
6741 correspond to the left without being swapped. */
6758 {
6761 }
6772 /* Make a mask that corresponds to both fields being compared.
6773 Do this for both items being compared. If the operands are the
6774 same size and the bits being compared are in the same position
6775 then we can do this by masking both and comparing the masked
6776 results. */
6783 {
6797 }
6799 /* There is still another way we can do something: If both pairs of
6800 fields being compared are adjacent, we may be able to make a wider
6801 field containing them both.
6803 Note that we still must mask the lhs/rhs expressions. Furthermore,
6804 the mask must be shifted to account for the shift done by
6805 make_bit_field_ref. */
6814 {
6815 tree type;
6833 /* Convert to the smaller type before masking out unwanted bits. */
6836 {
6838 {
6842 }
6844 {
6848 }
6849 }
6858 }
6861 }
6863 /* Handle the case of comparisons with constants. If there is something in
6864 common between the masks, those bits of the constants must be the same.
6865 If not, the condition is always false. Test for this to avoid generating
6866 incorrect code below. */
6871 {
6873 {
6876 }
6877 else
6878 {
6881 }
6882 }
6887 /* Construct the expression we will return. First get the component
6888 reference we will make. Unless the mask is all ones the width of
6889 that field, perform the mask operation. Then compare with the
6890 merged constant. */
6901 }
6902 \f
6903 /* T is an integer expression that is being multiplied, divided, or taken a
6904 modulus (CODE says which and what kind of divide or modulus) by a
6905 constant C. See if we can eliminate that operation by folding it with
6906 other operations already in T. WIDE_TYPE, if non-null, is a type that
6907 should be used for the computation if wider than our type.
6909 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
6910 (X * 2) + (Y * 4). We must, however, be assured that either the original
6911 expression would not overflow or that overflow is undefined for the type
6912 in the language in question.
6914 If we return a non-null expression, it is an equivalent form of the
6915 original computation, but need not be in the original type.
6917 We set *STRICT_OVERFLOW_P to true if the return values depends on
6918 signed overflow being undefined. Otherwise we do not change
6919 *STRICT_OVERFLOW_P. */
6921 static tree
6924 {
6925 /* To avoid exponential search depth, refuse to allow recursion past
6926 three levels. Beyond that (1) it's highly unlikely that we'll find
6927 something interesting and (2) we've probably processed it before
6928 when we built the inner expression. */
6931 tree ret;
6936 depth++;
6938 depth--;
6941 }
6943 static tree
6946 {
6951 {
6954 {
6957 }
6961 }
6967 /* Don't deal with constants of zero here; they confuse the code below. */
6977 /* Note that we need not handle conditional operations here since fold
6978 already handles those cases. So just do arithmetic here. */
6980 {
6982 /* For a constant, we can always simplify if we are a multiply
6983 or (for divide and modulus) if it is a multiple of our constant. */
6987 {
6990 /* If the multiplication overflowed, we lost information on it.
6991 See PR68142 and PR69845. */
6995 }
7001 /* If op0 is an expression ... */
7007 /* ... and has wrapping overflow, and its type is smaller
7008 than ctype, then we cannot pass through as widening. */
7012 /* ... or this is a truncation (t is narrower than op0),
7013 then we cannot pass through this narrowing. */
7016 /* ... or signedness changes for division or modulus,
7017 then we cannot pass through this conversion. */
7021 /* ... or has undefined overflow while the converted to
7022 type has not, we cannot do the operation in the inner type
7023 as that would introduce undefined overflow. */
7028 /* Pass the constant down and see if we can make a simplification. If
7029 we can, replace this expression with the inner simplification for
7030 possible later conversion to our or some other type. */
7041 /* If widening the type changes it from signed to unsigned, then we
7042 must avoid building ABS_EXPR itself as unsigned. */
7044 {
7048 {
7051 }
7053 }
7054 /* If the constant is negative, we cannot simplify this. */
7057 /* FALLTHROUGH */
7059 /* For division and modulus, type can't be unsigned, as e.g.
7060 (-(x / 2U)) / 2U isn't equal to -((x / 2U) / 2U) for x >= 2.
7061 For signed types, even with wrapping overflow, this is fine. */
7070 /* If widening the type changes the signedness, then we can't perform
7071 this optimization as that changes the result. */
7075 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
7081 {
7088 }
7092 /* If the second operand is constant, this is a multiplication
7093 or floor division, by a power of two, so we can treat it that
7094 way unless the multiplier or divisor overflows. Signed
7095 left-shift overflow is implementation-defined rather than
7096 undefined in C90, so do not convert signed left shift into
7097 multiplication. */
7100 /* const_binop may not detect overflow correctly,
7101 so check for it explicitly here. */
7110 ctype,
7112 t1),
7117 /* See if we can eliminate the operation on both sides. If we can, we
7118 can return a new PLUS or MINUS. If we can't, the only remaining
7119 cases where we can do anything are if the second operand is a
7120 constant. */
7127 /* If not multiplication, we can only do this if both operands
7128 are divisible by c. */
7131 {
7136 }
7138 /* If this was a subtraction, negate OP1 and set it to be an addition.
7139 This simplifies the logic below. */
7141 {
7143 /* If OP1 was not easily negatable, the constant may be OP0. */
7145 {
7148 }
7149 }
7154 /* If either OP1 or C are negative, this optimization is not safe for
7155 some of the division and remainder types while for others we need
7156 to change the code. */
7158 {
7166 }
7168 /* If it's a multiply or a division/modulus operation of a multiple
7169 of our constant, do the operation and verify it doesn't overflow. */
7173 {
7176 /* We allow the constant to overflow with wrapping semantics. */
7180 }
7181 else
7184 /* If we have an unsigned type, we cannot widen the operation since it
7185 will change the result if the original computation overflowed. */
7189 /* The last case is if we are a multiply. In that case, we can
7190 apply the distributive law to commute the multiply and addition
7191 if the multiplication of the constants doesn't overflow
7192 and overflow is defined. With undefined overflow
7193 op0 * c might overflow, while (op0 + orig_op1) * c doesn't.
7194 But fold_plusminus_mult_expr would factor back any power-of-two
7195 value so do not distribute in the first place in this case. */
7203 op1);
7208 /* We have a special case here if we are doing something like
7209 (C * 8) % 4 since we know that's zero. */
7212 /* If the multiplication can overflow we cannot optimize this. */
7217 {
7220 }
7222 /* ... fall through ... */
7226 /* If we can extract our operation from the LHS, do so and return a
7227 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
7228 do something only if the second operand is a constant. */
7244 /* If these are the same operation types, we can associate them
7245 assuming no overflow. */
7247 {
7262 }
7264 /* If these operations "cancel" each other, we have the main
7265 optimizations of this pass, which occur when either constant is a
7266 multiple of the other, in which case we replace this with either an
7267 operation or CODE or TCODE.
7269 If we have an unsigned type, we cannot do this since it will change
7270 the result if the original computation overflowed. */
7278 {
7281 {
7287 }
7290 {
7296 }
7297 }
7302 }
7305 }
7306 \f
7307 /* Return a node which has the indicated constant VALUE (either 0 or
7308 1 for scalars or {-1,-1,..} or {0,0,...} for vectors),
7309 and is of the indicated TYPE. */
7311 tree
7313 {
7322 else
7324 }
7327 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
7328 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
7329 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
7330 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
7331 COND is the first argument to CODE; otherwise (as in the example
7332 given here), it is the second argument. TYPE is the type of the
7333 original expression. Return NULL_TREE if no simplification is
7334 possible. */
7336 static tree
7341 {
7349 /* Do not move possibly trapping operations into the conditional as this
7350 pessimizes code and causes gimplification issues when applied late. */
7358 {
7362 /* If this operand throws an expression, then it does not make
7363 sense to try to perform a logical or arithmetic operation
7364 involving it. */
7369 }
7372 {
7377 }
7378 else
7379 /* Detect the case of mixing vector and scalar types - bail out. */
7385 /* This transformation is only worthwhile if we don't have to wrap ARG
7386 in a SAVE_EXPR and the operation can be simplified without recursing
7387 on at least one of the branches once its pushed inside the COND_EXPR. */
7396 {
7400 else
7402 }
7404 {
7408 else
7410 }
7412 /* Check that we have simplified at least one of the branches. */
7417 }
7419 \f
7420 /* Subroutine of fold() that checks for the addition of ARG +/- 0.0.
7422 If !NEGATE, return true if ZERO_ARG is +/-0.0 and, for all ARG of
7423 type TYPE, ARG + ZERO_ARG is the same as ARG. If NEGATE, return true
7424 if ARG - ZERO_ARG is the same as X.
7426 If ARG is NULL, check for any value of type TYPE.
7428 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
7429 and finite. The problematic cases are when X is zero, and its mode
7430 has signed zeros. In the case of rounding towards -infinity,
7431 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
7432 modes, X + 0 is not the same as X because -0 + 0 is 0. */
7434 bool
7437 {
7441 /* Don't allow the fold with -fsignaling-nans. */
7445 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
7449 /* There is no case that is safe for all rounding modes. */
7453 /* In a vector or complex, we would need to check the sign of all zeros. */
7459 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
7463 /* The mode has signed zeros, and we have to honor their sign.
7464 In this situation, there are only two cases we can return true for.
7465 (i) X - 0 is the same as X with default rounding.
7466 (ii) X + 0 is X when X can't possibly be -0.0. */
7468 }
7470 /* Subroutine of match.pd that optimizes comparisons of a division by
7471 a nonzero integer constant against an integer constant, i.e.
7472 X/C1 op C2.
7474 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
7475 GE_EXPR or LE_EXPR. ARG01 and ARG1 must be a INTEGER_CST. */
7477 enum tree_code
7480 {
7485 /* We have to do this the hard way to detect unsigned overflow.
7486 prod = int_const_binop (MULT_EXPR, c1, c2); */
7492 {
7496 /* Likewise *hi = int_const_binop (PLUS_EXPR, prod, tmp). */
7499 }
7501 {
7504 {
7523 }
7524 }
7525 else
7526 {
7527 /* A negative divisor reverses the relational operators. */
7532 {
7551 }
7552 }
7565 }
7567 /* Test whether it is preferable to swap two operands, ARG0 and
7568 ARG1, for example because ARG0 is an integer constant and ARG1
7569 isn't. */
7571 bool
7573 {
7587 /* It is preferable to swap two SSA_NAME to ensure a canonical form
7588 for commutative and comparison operators. Ensuring a canonical
7589 form allows the optimizers to find additional redundancies without
7590 having to explicitly check for both orderings. */
7596 /* Put SSA_NAMEs last. */
7602 /* Put variables last. */
7609 }
7612 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
7613 means A >= Y && A != MAX, but in this case we know that
7614 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
7616 static tree
7618 {
7625 else
7634 {
7637 }
7639 {
7642 }
7643 else
7650 {
7651 /* Convert the pointer types into integer before taking the difference. */
7655 }
7656 else
7663 }
7665 /* Fold a sum or difference of at least one multiplication.
7666 Returns the folded tree or NULL if no simplification could be made. */
7668 static tree
7671 {
7675 /* (A * C) +- (B * C) -> (A+-B) * C.
7676 (A * C) +- A -> A * (C+-1).
7677 We are most concerned about the case where C is a constant,
7678 but other combinations show up during loop reduction. Since
7679 it is not difficult, try all four possibilities. */
7682 {
7685 }
7687 {
7690 }
7691 else
7692 {
7693 /* We cannot generate constant 1 for fract. */
7698 }
7700 {
7703 }
7705 {
7707 /* As we canonicalize A - 2 to A + -2 get rid of that sign for
7708 the purpose of this canonicalization. */
7712 {
7715 }
7716 else
7718 }
7719 else
7720 {
7721 /* We cannot generate constant 1 for fract. */
7726 }
7729 /* Prefer factoring a common non-constant. */
7739 /* No identical multiplicands; see if we can find a common
7740 power-of-two factor in non-power-of-two multiplies. This
7741 can help in multi-dimensional array access. */
7743 {
7746 HOST_WIDE_INT tmp;
7748 tree maybe_same;
7750 /* Move min of absolute values to int11. */
7752 {
7757 }
7758 else
7765 /* The remainder should not be a constant, otherwise we
7766 end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
7767 increased the number of multiplications necessary. */
7769 {
7777 }
7778 }
7785 /* We are neither factoring zero nor minus one. */
7793 /* Same may be zero and thus the operation 'code' may overflow. Likewise
7794 same may be minus one and thus the multiplication may overflow. Perform
7795 the sum operation in an unsigned type. */
7800 /* If the sum evaluated to a constant that is not -INF the multiplication
7801 cannot overflow. */
7808 /* Do not resort to unsigned multiplication because
7809 we lose the no-overflow property of the expression. */
7811 }
7813 /* Subroutine of native_encode_expr. Encode the INTEGER_CST
7814 specified by EXPR into the buffer PTR of length LEN bytes.
7815 Return the number of bytes placed in the buffer, or zero
7816 upon failure. */
7818 static int
7820 {
7824 {
7830 {
7832 /* More work is needed when adding _BitInt support to PDP endian
7833 if limb is smaller than word, or if _BitInt limb ordering doesn't
7834 match target endianity here. */
7839 }
7840 else
7842 }
7843 else
7854 /* Dry run. */
7860 {
7862 /* Extend EXPR according to TYPE_SIGN if the precision isn't a whole
7863 number of bytes. */
7867 {
7874 else
7876 }
7877 else
7881 }
7883 }
7886 /* Subroutine of native_encode_expr. Encode the FIXED_CST
7887 specified by EXPR into the buffer PTR of length LEN bytes.
7888 Return the number of bytes placed in the buffer, or zero
7889 upon failure. */
7891 static int
7893 {
7897 FIXED_VALUE_TYPE value;
7912 }
7915 /* Subroutine of native_encode_expr. Encode the REAL_CST
7916 specified by EXPR into the buffer PTR of length LEN bytes.
7917 Return the number of bytes placed in the buffer, or zero
7918 upon failure. */
7920 static int
7922 {
7928 /* There are always 32 bits in each long, no matter the size of
7929 the hosts long. We handle floating point representations with
7930 up to 192 bits. */
7939 /* Dry run. */
7948 {
7953 {
7960 else
7962 }
7963 else
7964 {
7967 {
7968 /* Reverse bytes within each long, or within the entire float
7969 if it's smaller than a long (for HFmode). */
7972 }
7973 }
7978 }
7980 }
7982 /* Subroutine of native_encode_expr. Encode the COMPLEX_CST
7983 specified by EXPR into the buffer PTR of length LEN bytes.
7984 Return the number of bytes placed in the buffer, or zero
7985 upon failure. */
7987 static int
7989 {
7991 tree part;
8005 }
8007 /* Like native_encode_vector, but only encode the first COUNT elements.
8008 The other arguments are as for native_encode_vector. */
8010 static int
8013 {
8017 {
8018 /* This is the only case in which elements can be smaller than a byte.
8019 Element 0 is always in the lsb of the containing byte. */
8028 /* Zero the buffer and then set bits later where necessary. */
8037 {
8043 {
8046 }
8047 }
8049 }
8054 {
8056 {
8059 }
8070 }
8072 }
8074 /* Subroutine of native_encode_expr. Encode the VECTOR_CST
8075 specified by EXPR into the buffer PTR of length LEN bytes.
8076 Return the number of bytes placed in the buffer, or zero
8077 upon failure. */
8079 static int
8081 {
8086 }
8089 /* Subroutine of native_encode_expr. Encode the STRING_CST
8090 specified by EXPR into the buffer PTR of length LEN bytes.
8091 Return the number of bytes placed in the buffer, or zero
8092 upon failure. */
8094 static int
8096 {
8099 /* Wide-char strings are encoded in target byte-order so native
8100 encoding them is trivial. */
8115 else
8116 {
8119 {
8122 }
8124 }
8126 }
8129 /* Subroutine of fold_view_convert_expr. Encode the INTEGER_CST, REAL_CST,
8130 FIXED_CST, COMPLEX_CST, STRING_CST, or VECTOR_CST specified by EXPR into
8131 the buffer PTR of size LEN bytes. If PTR is NULL, don't actually store
8132 anything, just do a dry run. Fail either if OFF is -1 and LEN isn't
8133 sufficient to encode the entire EXPR, or if OFF is out of bounds.
8134 Otherwise, start at byte offset OFF and encode at most LEN bytes.
8135 Return the number of bytes placed in the buffer, or zero upon failure. */
8137 int
8139 {
8140 /* We don't support starting at negative offset and -1 is special. */
8145 {
8166 }
8167 }
8169 /* Try to find a type whose byte size is smaller or equal to LEN bytes larger
8170 or equal to FIELDSIZE bytes, with underlying mode precision/size multiple
8171 of BITS_PER_UNIT. As native_{interpret,encode}_int works in term of
8172 machine modes, we can't just use build_nonstandard_integer_type. */
8174 tree
8176 {
8177 machine_mode mode;
8179 {
8186 {
8190 }
8191 }
8197 {
8205 }
8208 }
8210 /* Similar to native_encode_expr, but also handle CONSTRUCTORs, VCEs,
8211 NON_LVALUE_EXPRs and nops. If MASK is non-NULL (then PTR has
8212 to be non-NULL and OFF zero), then in addition to filling the
8213 bytes pointed by PTR with the value also clear any bits pointed
8214 by MASK that are known to be initialized, keep them as is for
8215 e.g. uninitialized padding bits or uninitialized fields. */
8217 int
8220 {
8223 /* We don't support starting at negative offset and -1 is special. */
8231 {
8235 mask);
8250 {
8251 tree min_index;
8269 {
8274 {
8277 }
8282 else
8286 {
8290 offset_int wpos
8298 offset_int wcount
8305 }
8307 {
8310 offset_int wpos
8318 }
8321 {
8323 {
8334 }
8336 {
8340 }
8341 }
8345 do
8346 {
8351 {
8353 {
8356 fieldsize);
8359 }
8361 ptr
8364 fieldsize,
8367 mask
8371 else
8372 {
8375 }
8376 }
8379 {
8380 /* Partial overlap. */
8386 {
8390 fieldsize);
8391 }
8392 else
8393 {
8397 }
8400 }
8402 }
8404 }
8406 }
8409 {
8419 {
8425 {
8428 }
8431 {
8440 }
8444 else
8448 {
8449 tree fld;
8451 {
8464 }
8466 {
8470 }
8473 {
8474 cnt--;
8480 }
8481 }
8486 {
8499 {
8504 }
8505 }
8506 else
8507 {
8512 }
8516 /* Prepare to deal with integral bit-fields and filter out other
8517 bit-fields that do not start and end on a byte boundary. */
8519 {
8524 {
8530 }
8536 }
8546 {
8547 /* FIXME: Handle PDP endian. */
8558 {
8561 }
8562 else
8563 {
8571 else
8572 {
8575 }
8576 }
8586 else
8599 /* If the bitfield does not start at byte boundary, handle
8600 the partial byte at the start. */
8603 {
8605 {
8610 {
8613 else
8615 }
8616 }
8617 else
8618 {
8623 {
8626 else
8630 }
8631 }
8632 }
8633 /* If the bitfield does not end at byte boundary, handle
8634 the partial byte at the end. */
8638 {
8640 {
8647 }
8648 else
8649 {
8656 }
8657 }
8661 {
8666 }
8667 else
8668 {
8669 /* Partial overlap. */
8673 {
8676 }
8680 }
8682 }
8687 {
8690 {
8693 {
8697 }
8700 }
8703 fldsize,
8713 }
8714 else
8715 {
8716 /* Partial overlap. */
8722 {
8726 fieldsize);
8727 }
8728 else
8729 {
8733 }
8736 }
8737 }
8739 }
8741 }
8742 }
8745 /* Subroutine of native_interpret_expr. Interpret the contents of
8746 the buffer PTR of length LEN as an INTEGER_CST of type TYPE.
8747 If the buffer cannot be interpreted, return NULL_TREE. */
8749 static tree
8751 {
8754 {
8760 {
8762 /* More work is needed when adding _BitInt support to PDP endian
8763 if limb is smaller than word, or if _BitInt limb ordering doesn't
8764 match target endianity here. */
8769 }
8770 else
8772 }
8773 else
8783 }
8786 /* Subroutine of native_interpret_expr. Interpret the contents of
8787 the buffer PTR of length LEN as a FIXED_CST of type TYPE.
8788 If the buffer cannot be interpreted, return NULL_TREE. */
8790 static tree
8792 {
8795 double_int result;
8796 FIXED_VALUE_TYPE fixed_value;
8806 }
8809 /* Subroutine of native_interpret_expr. Interpret the contents of
8810 the buffer PTR of length LEN as a REAL_CST of type TYPE.
8811 If the buffer cannot be interpreted, return NULL_TREE. */
8813 tree
8815 {
8819 /* There are always 32 bits in each long, no matter the size of
8820 the hosts long. We handle floating point representations with
8821 up to 192 bits. */
8822 REAL_VALUE_TYPE r;
8832 {
8833 /* Both OFFSET and BYTE index within a long;
8834 bitpos indexes the whole float. */
8837 {
8844 else
8846 }
8847 else
8848 {
8851 {
8852 /* Reverse bytes within each long, or within the entire float
8853 if it's smaller than a long (for HFmode). */
8856 }
8857 }
8861 }
8865 }
8868 /* Subroutine of native_interpret_expr. Interpret the contents of
8869 the buffer PTR of length LEN as a COMPLEX_CST of type TYPE.
8870 If the buffer cannot be interpreted, return NULL_TREE. */
8872 static tree
8874 {
8889 }
8891 /* Read a vector of type TYPE from the target memory image given by BYTES,
8892 which contains LEN bytes. The vector is known to be encodable using
8893 NPATTERNS interleaved patterns with NELTS_PER_PATTERN elements each.
8895 Return the vector on success, otherwise return null. */
8897 static tree
8901 {
8905 {
8906 /* This is the only case in which elements can be smaller than a byte.
8907 Element 0 is always in the lsb of the containing byte. */
8914 {
8921 }
8923 }
8931 {
8937 }
8939 }
8941 /* Subroutine of native_interpret_expr. Interpret the contents of
8942 the buffer PTR of length LEN as a VECTOR_CST of type TYPE.
8943 If the buffer cannot be interpreted, return NULL_TREE. */
8945 static tree
8947 {
8956 }
8959 /* Subroutine of fold_view_convert_expr. Interpret the contents of
8960 the buffer PTR of length LEN as a constant of type TYPE. For
8961 INTEGRAL_TYPE_P we return an INTEGER_CST, for SCALAR_FLOAT_TYPE_P
8962 we return a REAL_CST, etc... If the buffer cannot be interpreted,
8963 return NULL_TREE. */
8965 tree
8967 {
8969 {
8981 {
8982 /* For floating point values in composite modes, punt if this
8983 folding doesn't preserve bit representation. As the mode doesn't
8984 have fixed precision while GCC pretends it does, there could be
8985 valid values that GCC can't really represent accurately.
8986 See PR95450. Even for other modes, e.g. x86 XFmode can have some
8987 bit combinationations which GCC doesn't preserve. */
8997 }
9011 }
9012 }
9014 /* Returns true if we can interpret the contents of a native encoding
9015 as TYPE. */
9017 bool
9019 {
9021 {
9035 }
9036 }
9038 /* Attempt to interpret aggregate of TYPE from bytes encoded in target
9039 byte order at PTR + OFF with LEN bytes. Does not handle unions. */
9041 tree
9044 {
9047 {
9054 {
9058 }
9063 {
9068 {
9073 }
9077 eltsz);
9081 }
9083 }
9087 {
9096 {
9099 {
9100 poly_int64 bitoffset;
9105 else
9115 }
9116 else
9117 {
9135 HOST_WIDE_INT rpos;
9138 else
9139 {
9142 }
9151 }
9152 }
9155 {
9165 }
9169 {
9174 }
9181 {
9185 /* FIXME: Figure out how to handle PDP endian bitfields. */
9191 else
9195 }
9197 }
9199 }
9201 /* Routines for manipulation of native_encode_expr encoded data if the encoded
9202 or extracted constant positions and/or sizes aren't byte aligned. */
9204 /* Shift left the bytes in PTR of SZ elements by AMNT bits, carrying over the
9205 bits between adjacent elements. AMNT should be within
9206 [0, BITS_PER_UNIT).
9207 Example, AMNT = 2:
9208 00011111|11100000 << 2 = 01111111|10000000
9209 PTR[1] | PTR[0] PTR[1] | PTR[0]. */
9211 void
9214 {
9223 {
9229 {
9232 }
9233 }
9234 }
9236 /* Like shift_bytes_in_array_left but for big-endian.
9237 Shift right the bytes in PTR of SZ elements by AMNT bits, carrying over the
9238 bits between adjacent elements. AMNT should be within
9239 [0, BITS_PER_UNIT).
9240 Example, AMNT = 2:
9241 00011111|11100000 >> 2 = 00000111|11111000
9242 PTR[0] | PTR[1] PTR[0] | PTR[1]. */
9244 void
9247 {
9255 {
9262 }
9263 }
9265 /* Try to view-convert VECTOR_CST EXPR to VECTOR_TYPE TYPE by operating
9266 directly on the VECTOR_CST encoding, in a way that works for variable-
9267 length vectors. Return the resulting VECTOR_CST on success or null
9268 on failure. */
9270 static tree
9272 {
9284 /* We can only preserve the semantics of a stepped pattern if the new
9285 vector element is an integer of the same size. */
9290 /* The number of bits needed to encode one element from every pattern
9291 of the original vector. */
9292 unsigned int expr_sequence_bits
9295 /* The number of bits needed to encode one element from every pattern
9296 of the result. */
9297 unsigned int type_sequence_bits
9300 /* Don't try to read more bytes than are available, which can happen
9301 for constant-sized vectors if TYPE has larger elements than EXPR_TYPE.
9302 The general VIEW_CONVERT handling can cope with that case, so there's
9303 no point complicating things here. */
9306 BITS_PER_UNIT);
9311 /* Get enough bytes of EXPR to form the new encoding. */
9319 /* Reencode the bytes as TYPE. */
9323 }
9325 /* Fold a VIEW_CONVERT_EXPR of a constant expression EXPR to type
9326 TYPE at compile-time. If we're unable to perform the conversion
9327 return NULL_TREE. */
9329 static tree
9331 {
9332 /* We support up to 1024-bit values (for GCN/RISC-V V128QImode). */
9336 /* Check that the host and target are sane. */
9349 }
9351 /* Build an expression for the address of T. Folds away INDIRECT_REF
9352 to avoid confusing the gimplify process. */
9354 tree
9356 {
9357 /* The size of the object is not relevant when talking about its address. */
9362 {
9367 }
9370 {
9375 }
9382 {
9387 }
9388 else
9392 }
9394 /* Build an expression for the address of T. */
9396 tree
9398 {
9402 }
9404 /* Fold a unary expression of code CODE and type TYPE with operand
9405 OP0. Return the folded expression if folding is successful.
9406 Otherwise, return NULL_TREE. */
9408 tree
9410 {
9411 tree tem;
9412 tree arg0;
9420 {
9423 {
9424 /* Don't use STRIP_NOPS, because signedness of argument type
9425 matters. */
9427 }
9428 else
9429 {
9430 /* Strip any conversions that don't change the mode. This
9431 is safe for every expression, except for a comparison
9432 expression because its signedness is derived from its
9433 operands.
9435 Note that this is done as an internal manipulation within
9436 the constant folder, in order to find the simplest
9437 representation of the arguments so that their form can be
9438 studied. In any cases, the appropriate type conversions
9439 should be put back in the tree that will get out of the
9440 constant folder. */
9442 }
9445 {
9448 {
9452 }
9453 }
9454 }
9461 {
9468 {
9482 /* If this was a conversion, and all we did was to move into
9483 inside the COND_EXPR, bring it back out. But leave it if
9484 it is a conversion from integer to integer and the
9485 result precision is no wider than a word since such a
9486 conversion is cheap and may be optimized away by combine,
9487 while it couldn't if it were outside the COND_EXPR. Then return
9488 so we don't get into an infinite recursion loop taking the
9489 conversion out and then back in. */
9501 && (INTEGRAL_TYPE_P
9514 }
9515 }
9518 {
9524 CASE_CONVERT:
9528 {
9529 /* If we have (type) (a CMP b) and type is an integral type, return
9530 new expression involving the new type. Canonicalize
9531 (type) (a CMP b) to (a CMP b) ? (type) true : (type) false for
9532 non-integral type.
9533 Do not fold the result as that would not simplify further, also
9534 folding again results in recursions. */
9544 }
9546 /* Handle (T *)&A.B.C for A being of type T and B and C
9547 living at offset zero. This occurs frequently in
9548 C++ upcasting and then accessing the base. */
9552 {
9554 tree offset;
9555 machine_mode mode;
9557 tree base
9561 /* If the reference was to a (constant) zero offset, we can use
9562 the address of the base if it has the same base type
9563 as the result type and the pointer type is unqualified. */
9571 }
9575 /* Detect assigning a bitfield. */
9577 && DECL_BIT_FIELD
9579 {
9580 /* Don't leave an assignment inside a conversion
9581 unless assigning a bitfield. */
9583 /* First do the assignment, then return converted constant. */
9588 }
9590 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
9591 constants (if x has signed type, the sign bit cannot be set
9592 in c). This folds extension into the BIT_AND_EXPR.
9593 ??? We don't do it for BOOLEAN_TYPE or ENUMERAL_TYPE because they
9594 very likely don't have maximal range for their precision and this
9595 transformation effectively doesn't preserve non-maximal ranges. */
9599 {
9610 <= HOST_BITS_PER_WIDE_INT
9612 {
9616 cst &= HOST_WIDE_INT_M1U
9620 && !flag_syntax_only
9623 {
9627 }
9628 }
9630 {
9640 }
9641 }
9643 /* Convert (T1)(X p+ Y) into ((T1)X p+ Y), for pointer type, when the new
9644 cast (T1)X will fold away. We assume that this happens when X itself
9645 is a cast. */
9649 {
9653 /* If -fsanitize=alignment, avoid this optimization in GENERIC
9654 when the pointed type needs higher alignment than
9655 the p+ first operand's pointed type. */
9662 /* Similarly, avoid this optimization in GENERIC for -fsanitize=null
9663 when type is a reference type and arg00's type is not,
9664 because arg00 could be validly nullptr and if arg01 doesn't return,
9665 we don't want false positive binding of reference to nullptr. */
9667 && !in_gimple_form
9674 }
9676 /* Convert (T1)(~(T2)X) into ~(T1)X if T1 and T2 are integral types
9677 of the same precision, and X is an integer type not narrower than
9678 types T1 or T2, i.e. the cast (T2)X isn't an extension. */
9684 {
9690 }
9692 /* Convert (T1)(X * Y) into (T1)X * (T1)Y if T1 is narrower than the
9693 type of X and Y (integer types only). */
9700 {
9701 /* Be careful not to introduce new overflows. */
9702 tree mult_type;
9705 else
9709 {
9716 }
9717 }
9723 {
9730 }
9741 /* Convert fabs((double)float) into (double)fabsf(float). */
9744 {
9750 targ0));
9751 }
9755 /* Convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
9774 /* Note that the operand of this must be an int
9775 and its values must be 0 or 1.
9776 ("true" is a fixed value perhaps depending on the language,
9777 but we don't handle values other than 1 correctly yet.) */
9784 /* Fold *&X to X if X is an lvalue. */
9786 {
9793 }
9799 }
9802 /* If the operation was a conversion do _not_ mark a resulting constant
9803 with TREE_OVERFLOW if the original constant was not. These conversions
9804 have implementation defined behavior and retaining the TREE_OVERFLOW
9805 flag here would confuse later passes such as VRP. */
9806 tree
9809 {
9818 }
9820 /* Fold a binary bitwise/truth expression of code CODE and type TYPE with
9821 operands OP0 and OP1. LOC is the location of the resulting expression.
9822 ARG0 and ARG1 are the NOP_STRIPed results of OP0 and OP1.
9823 Return the folded expression if folding is successful. Otherwise,
9824 return NULL_TREE. */
9825 static tree
9828 {
9829 tree tem;
9831 /* We only do these simplifications if we are optimizing. */
9835 /* Check for things like (A || B) && (A || C). We can convert this
9836 to A || (B && C). Note that either operator can be any of the four
9837 truth and/or operations and the transformation will still be
9838 valid. Also note that we only care about order for the
9839 ANDIF and ORIF operators. If B contains side effects, this
9840 might change the truth-value of A. */
9847 {
9867 /* This case if tricky because we must either have commutative
9868 operators or else A10 must not have side-effects. */
9874 a01);
9875 }
9877 /* See if we can build a range comparison. */
9883 {
9887 }
9891 {
9895 }
9897 /* Check for the possibility of merging component references. If our
9898 lhs is another similar operation, try to merge its rhs with our
9899 rhs. Then try to merge our lhs and rhs. */
9910 logical_op_non_short_circuit
9918 {
9925 /* Transform ((A AND-IF B) AND[-IF] C) into (A AND-IF (B AND C)),
9926 or ((A OR-IF B) OR[-IF] C) into (A OR-IF (B OR C))
9927 We don't want to pack more than two leafs to a non-IF AND/OR
9928 expression.
9929 If tree-code of left-hand operand isn't an AND/OR-IF code and not
9930 equal to IF-CODE, then we don't want to add right-hand operand.
9931 If the inner right-hand side of left-hand operand has
9932 side-effects, or isn't simple, then we can't add to it,
9933 as otherwise we might destroy if-sequence. */
9936 /* Needed for sequence points to handle trappings, and
9937 side-effects. */
9939 {
9941 arg1);
9943 tem);
9944 }
9945 /* Same as above but for (A AND[-IF] (B AND-IF C)) -> ((A AND B) AND-IF C),
9946 or (A OR[-IF] (B OR-IF C) -> ((A OR B) OR-IF C). */
9949 /* Needed for sequence points to handle trappings, and
9950 side-effects. */
9952 {
9957 }
9958 /* Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
9959 into (A OR B).
9960 For sequence point consistancy, we need to check for trapping,
9961 and side-effects. */
9965 }
9968 }
9970 /* Helper that tries to canonicalize the comparison ARG0 CODE ARG1
9971 by changing CODE to reduce the magnitude of constants involved in
9972 ARG0 of the comparison.
9973 Returns a canonicalized comparison tree if a simplification was
9974 possible, otherwise returns NULL_TREE.
9975 Set *STRICT_OVERFLOW_P to true if the canonicalization is only
9976 valid if signed overflow is undefined. */
9978 static tree
9982 {
9987 /* Match A +- CST code arg1. We can change this only if overflow
9988 is undefined. */
9991 /* In principle pointers also have undefined overflow behavior,
9992 but that causes problems elsewhere. */
9999 /* Identify the constant in arg0 and its sign. */
10003 /* Overflowed constants and zero will cause problems. */
10008 /* See if we can reduce the magnitude of the constant in
10009 arg0 by changing the comparison code. */
10010 /* A - CST < arg1 -> A - CST-1 <= arg1. */
10014 /* A + CST > arg1 -> A + CST-1 >= arg1. */
10018 /* A + CST <= arg1 -> A + CST-1 < arg1. */
10022 /* A - CST >= arg1 -> A - CST-1 > arg1. */
10026 else
10030 /* Now build the constant reduced in magnitude. But not if that
10031 would produce one outside of its types range. */
10047 }
10049 /* Canonicalize the comparison ARG0 CODE ARG1 with type TYPE with undefined
10050 overflow further. Try to decrease the magnitude of constants involved
10051 by changing LE_EXPR and GE_EXPR to LT_EXPR and GT_EXPR or vice versa
10052 and put sole constants at the second argument position.
10053 Returns the canonicalized tree if changed, otherwise NULL_TREE. */
10055 static tree
10058 {
10059 tree t;
10064 /* Try canonicalization by simplifying arg0. */
10069 {
10073 }
10075 /* Try canonicalization by simplifying arg1 using the swapped
10076 comparison. */
10084 }
10086 /* Return whether BASE + OFFSET + BITPOS may wrap around the address
10087 space. This is used to avoid issuing overflow warnings for
10088 expressions like &p->x which cannot wrap. */
10090 static bool
10092 {
10099 poly_wide_int wi_offset;
10105 else
10110 precision);
10125 /* We can do slightly better for SIZE if we have an ADDR_EXPR of an
10126 array. */
10135 }
10137 /* Return a positive integer when the symbol DECL is known to have
10138 a nonzero address, zero when it's known not to (e.g., it's a weak
10139 symbol), and a negative integer when the symbol is not yet in the
10140 symbol table and so whether or not its address is zero is unknown.
10141 For function local objects always return positive integer. */
10142 static int
10144 {
10145 /* Normally, don't do anything for variables and functions before symtab is
10146 built; it is quite possible that DECL will be declared weak later.
10147 But if folding_initializer, we need a constant answer now, so create
10148 the symtab entry and prevent later weak declaration. */
10151 = (folding_initializer
10156 /* Function local objects are never NULL. */
10164 }
10166 /* Subroutine of fold_binary. This routine performs all of the
10167 transformations that are common to the equality/inequality
10168 operators (EQ_EXPR and NE_EXPR) and the ordering operators
10169 (LT_EXPR, LE_EXPR, GE_EXPR and GT_EXPR). Callers other than
10170 fold_binary should call fold_binary. Fold a comparison with
10171 tree code CODE and type TYPE with operands OP0 and OP1. Return
10172 the folded comparison or NULL_TREE. */
10174 static tree
10177 {
10187 /* For comparisons of pointers we can decompose it to a compile time
10188 comparison of the base objects and the offsets into the object.
10189 This requires at least one operand being an ADDR_EXPR or a
10190 POINTER_PLUS_EXPR to do more than the operand_equal_p test below. */
10196 {
10199 machine_mode mode;
10203 /* Get base and offset for the access. Strip ADDR_EXPR for
10204 get_inner_reference, but put it back by stripping INDIRECT_REF
10205 off the base object if possible. indirect_baseN will be true
10206 if baseN is not an address but refers to the object itself. */
10209 {
10210 base0
10216 else
10218 }
10220 {
10224 {
10225 base0
10231 else
10233 }
10236 else
10240 {
10247 }
10248 }
10252 {
10253 base1
10259 else
10261 }
10263 {
10267 {
10268 base1
10274 else
10276 }
10279 else
10283 {
10290 }
10291 }
10293 /* If we have equivalent bases we might be able to simplify. */
10297 {
10298 /* We can fold this expression to a constant if the non-constant
10299 offset parts are equal. */
10303 && (equality_code
10304 || (indirect_base0
10307 {
10313 "occur when comparing P +- C1 with "
10315 WARN_STRICT_OVERFLOW_CONDITIONAL);
10318 {
10356 }
10357 }
10358 /* We can simplify the comparison to a comparison of the variable
10359 offset parts if the constant offset parts are equal.
10360 Be careful to use signed sizetype here because otherwise we
10361 mess with array offsets in the wrong way. This is possible
10362 because pointer arithmetic is restricted to retain within an
10363 object and overflow on pointer differences is undefined as of
10364 6.5.6/8 and /9 with respect to the signed ptrdiff_t. */
10366 && (equality_code
10367 || (indirect_base0
10370 {
10371 /* By converting to signed sizetype we cover middle-end pointer
10372 arithmetic which operates on unsigned pointer types of size
10373 type size and ARRAY_REF offsets which are properly sign or
10374 zero extended from their type in case it is narrower than
10375 sizetype. */
10378 else
10382 else
10389 "occur when comparing P +- C1 with "
10391 WARN_STRICT_OVERFLOW_COMPARISON);
10394 }
10395 }
10396 /* For equal offsets we can simplify to a comparison of the
10397 base addresses. */
10399 && (indirect_base0
10401 && (indirect_base1
10406 {
10412 }
10413 /* Comparison between an ordinary (non-weak) symbol and a null
10414 pointer can be eliminated since such symbols must have a non
10415 null address. In C, relational expressions between pointers
10416 to objects and null pointers are undefined. The results
10417 below follow the C++ rules with the additional property that
10418 every object pointer compares greater than a null pointer.
10419 */
10422 /* Avoid folding references to struct members at offset 0 to
10423 prevent tests like '&ptr->firstmember == 0' from getting
10424 eliminated. When ptr is null, although the -> expression
10425 is strictly speaking invalid, GCC retains it as a matter
10426 of QoI. See PR c/44555. */
10429 && indirect_base0
10430 /* The caller guarantees that when one of the arguments is
10431 constant (i.e., null in this case) it is second. */
10433 {
10435 {
10446 }
10447 }
10448 }
10450 /* Transform comparisons of the form X +- C1 CMP Y +- C2 to
10451 X CMP Y +- C2 +- C1 for signed X, Y. This is valid if
10452 the resulting offset is smaller in absolute value than the
10453 original one and has the same sign. */
10462 {
10467 tree cst;
10469 "occur when combining constants around "
10472 /* Put the constant on the side where it doesn't overflow and is
10473 of lower absolute value and of same sign than before. */
10480 {
10483 variable1,
10487 }
10495 {
10501 variable2);
10502 }
10503 }
10509 /* If we are comparing an expression that just has comparisons
10510 of two integer values, arithmetic expressions of those comparisons,
10511 and constants, we can simplify it. There are only three cases
10512 to check: the two values can either be equal, the first can be
10513 greater, or the second can be greater. Fold the expression for
10514 those three values. Since each value must be 0 or 1, we have
10515 eight possibilities, each of which corresponds to the constant 0
10516 or 1 or one of the six possible comparisons.
10518 This handles common cases like (a > b) == 0 but also handles
10519 expressions like ((x > y) - (y > x)) > 0, which supposedly
10520 occur in macroized code. */
10523 {
10527 /* Don't handle degenerate cases here; they should already
10528 have been handled anyway. */
10537 {
10541 /* We can't just pass T to eval_subst in case cval1 or cval2
10542 was the same as ARG1. */
10544 tree high_result
10548 arg1);
10549 tree equal_result
10553 arg1);
10554 tree low_result
10558 arg1);
10560 /* All three of these results should be 0 or 1. Confirm they are.
10561 Then use those values to select the proper code to use. */
10566 {
10567 /* Make a 3-bit mask with the high-order bit being the
10568 value for `>', the next for '=', and the low for '<'. */
10572 {
10574 /* Always false. */
10595 /* Always true. */
10597 }
10600 }
10601 }
10602 }
10605 }
10608 /* Subroutine of fold_binary. Optimize complex multiplications of the
10609 form z * conj(z), as pow(realpart(z),2) + pow(imagpart(z),2). The
10610 argument EXPR represents the expression "z" of type TYPE. */
10612 static tree
10614 {
10619 {
10622 }
10624 {
10627 }
10628 else
10629 {
10633 }
10642 }
10645 /* Helper function for fold_vec_perm. Store elements of VECTOR_CST or
10646 CONSTRUCTOR ARG into array ELTS, which has NELTS elements, and return
10647 true if successful. */
10649 static bool
10651 {
10656 {
10659 }
10661 {
10667 else
10669 }
10670 else
10676 }
10678 /* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
10679 mask for VLA vec_perm folding.
10680 REASON if specified, will contain the reason why SEL is not suitable.
10681 Used only for debugging and unit-testing. */
10683 static bool
10687 {
10694 {
10698 }
10700 /* We want to avoid cases where sel.length is not a multiple of npatterns.
10701 For eg: sel.length = 2 + 2x, and sel npatterns = 4. */
10702 poly_uint64 esel;
10704 {
10708 }
10714 {
10717 HOST_WIDE_INT step;
10719 {
10723 }
10724 // FIXME: Punt on step < 0 for now, revisit later.
10731 {
10735 }
10737 /* Ensure that stepped sequence of the pattern selects elements
10738 only from the same input vector. */
10747 {
10751 }
10753 /* Ensure that the stepped sequence always selects from the same
10754 input pattern. */
10759 {
10763 }
10765 /* If a1 chooses base element from arg, ensure that it's a natural
10766 stepped sequence, ie, (arg[2] - arg[1]) == (arg[1] - arg[0])
10767 to preserve arg's encoding. */
10770 {
10783 {
10787 }
10788 }
10789 }
10792 }
10794 /* Try to fold permutation of ARG0 and ARG1 with SEL selector when
10795 the input vectors are VECTOR_CST. Return NULL_TREE otherwise.
10796 REASON has same purpose as described in
10797 valid_mask_for_fold_vec_perm_cst_p. */
10799 static tree
10802 {
10806 /* First try to implement the fold in a VLA-friendly way.
10808 (1) If the selector is simply a duplication of N elements, the
10809 result is likewise a duplication of N elements.
10811 (2) If the selector is N elements followed by a duplication
10812 of N elements, the result is too.
10814 (3) If the selector is N elements followed by an interleaving
10815 of N linear series, the situation is more complex.
10817 valid_mask_for_fold_vec_perm_cst_p detects whether we
10818 can handle this case. If we can, then each of the N linear
10819 series either (a) selects the same element each time or
10820 (b) selects a linear series from one of the input patterns.
10822 If (b) holds for one of the linear series, the result
10823 will contain a linear series, and so the result will have
10824 the same shape as the selector. If (a) holds for all of
10825 the linear series, the result will be the same as (2) above.
10827 (b) can only hold if one of the input patterns has a
10828 stepped encoding. */
10831 {
10839 }
10841 {
10844 }
10845 else
10850 {
10853 poly_uint64 r;
10856 /* Punt if sel[i] /trunc_div len cannot be determined,
10857 because the input vector to be chosen will depend on
10858 runtime vector length.
10859 For example if len == 4 + 4x, and sel[i] == 4,
10860 If len at runtime equals 4, we choose arg1[0].
10861 For any other value of len > 4 at runtime, we choose arg0[4].
10862 which makes the element choice dependent on runtime vector length. */
10864 {
10868 }
10870 /* sel[i] % len will give the index of element in the chosen input
10871 vector. For example if sel[i] == 5 + 4x and len == 4 + 4x,
10872 we will choose arg1[1] since (5 + 4x) % (4 + 4x) == 1. */
10874 {
10878 }
10883 }
10886 }
10888 /* Attempt to fold vector permutation of ARG0 and ARG1 vectors using SEL
10889 selector. Return the folded VECTOR_CST or CONSTRUCTOR if successful,
10890 NULL_TREE otherwise. */
10892 tree
10894 {
10910 /* For fall back case, we want to ensure we have VLS vectors
10911 with equal length. */
10925 {
10926 HOST_WIDE_INT index;
10930 }
10932 }
10934 /* Try to fold a pointer difference of type TYPE two address expressions of
10935 array references AREF0 and AREF1 using location LOC. Return a
10936 simplified expression for the difference or NULL_TREE. */
10938 static tree
10942 {
10947 /* If the bases are array references as well, recurse. If the bases
10948 are pointer indirections compute the difference of the pointers.
10949 If the bases are equal, we are set. */
10952 && (base_offset
10954 use_pointer_diff)))
10957 && (base_offset
10958 = use_pointer_diff
10968 {
10974 base_offset,
10977 }
10979 }
10981 /* If the real or vector real constant CST of type TYPE has an exact
10982 inverse, return it, else return NULL. */
10984 tree
10986 {
10987 REAL_VALUE_TYPE r;
10988 tree unit_type;
10989 machine_mode mode;
10992 {
11002 {
11006 tree_vector_builder elts;
11011 {
11016 }
11019 }
11023 }
11024 }
11026 /* Mask out the tz least significant bits of X of type TYPE where
11027 tz is the number of trailing zeroes in Y. */
11028 static wide_int
11030 {
11035 }
11037 /* Return true when T is an address and is known to be nonzero.
11038 For floating point we further ensure that T is not denormal.
11039 Similar logic is present in nonzero_address in rtlanal.h.
11041 If the return value is based on the assumption that signed overflow
11042 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
11043 change *STRICT_OVERFLOW_P. */
11045 static bool
11047 {
11051 /* Doing something useful for floating point would need more work. */
11057 {
11060 strict_overflow_p);
11066 strict_overflow_p);
11074 }
11077 {
11080 strict_overflow_p);
11088 strict_overflow_p);
11102 strict_overflow_p);
11106 strict_overflow_p);
11109 {
11121 }
11125 }
11127 }
11129 /* Return true when T is an address and is known to be nonzero.
11130 Handle warnings about undefined signed overflow. */
11132 bool
11134 {
11141 "determining that expression is always "
11143 WARN_STRICT_OVERFLOW_MISC);
11145 }
11147 /* Return true if T is known not to be equal to an integer W. */
11149 bool
11151 {
11152 int_range_max vr;
11154 {
11165 /* If T has some known zero bits and W has any of those bits set,
11166 then T is known not to be equal to W. */
11174 }
11175 }
11177 /* Fold a binary expression of code CODE and type TYPE with operands
11178 OP0 and OP1. LOC is the location of the resulting expression.
11179 Return the folded expression if folding is successful. Otherwise,
11180 return NULL_TREE. */
11182 tree
11185 {
11200 /* Strip any conversions that don't change the mode. This is
11201 safe for every expression, except for a comparison expression
11202 because its signedness is derived from its operands. So, in
11203 the latter case, only strip conversions that don't change the
11204 signedness. MIN_EXPR/MAX_EXPR also need signedness of arguments
11205 preserved.
11207 Note that this is done as an internal manipulation within the
11208 constant folder, in order to find the simplest representation
11209 of the arguments so that their form can be studied. In any
11210 cases, the appropriate type conversions should be put back in
11211 the tree that will get out of the constant folder. */
11214 {
11217 }
11218 else
11219 {
11222 }
11224 /* Note that TREE_CONSTANT isn't enough: static var addresses are
11225 constant but we can't do arithmetic on them. */
11227 {
11230 {
11234 }
11235 }
11237 /* If this is a commutative operation, and ARG0 is a constant, move it
11238 to ARG1 to reduce the number of tests below. */
11243 /* Likewise if this is a comparison, and ARG0 is a constant, move it
11244 to ARG1 to reduce the number of tests below. */
11253 /* ARG0 is the first operand of EXPR, and ARG1 is the second operand.
11255 First check for cases where an arithmetic operation is applied to a
11256 compound, conditional, or comparison operation. Push the arithmetic
11257 operation inside the compound or conditional to see if any folding
11258 can then be done. Convert comparison to conditional for this purpose.
11259 The also optimizes non-constant cases that used to be done in
11260 expand_expr.
11262 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
11263 one of the operands is a comparison and the other is a comparison, a
11264 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
11265 code below would make the expression more complex. Change it to a
11266 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
11267 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
11280 {
11284 boolean_type_node,
11292 }
11296 {
11298 {
11303 tem);
11304 }
11306 {
11311 tem);
11312 }
11317 {
11323 }
11328 {
11334 }
11335 }
11338 {
11340 /* MEM[&MEM[p, CST1], CST2] -> MEM[p, CST1 + CST2]. */
11343 {
11349 }
11351 /* MEM[&a.b, CST2] -> MEM[&a, offsetof (a, b) + CST2]. */
11354 {
11355 tree base;
11356 poly_int64 coffset;
11365 }
11370 /* INT +p INT -> (PTR)(INT + INT). Stripping types allows for this. */
11376 arg1),
11378 arg0)));
11384 {
11385 /* X + (X / CST) * -CST is X % CST. */
11390 {
11399 cst0));
11400 }
11401 }
11403 /* Handle (A1 * C1) + (A2 * C2) with A1, A2 or C1, C2 being the same or
11404 one. Make sure the type is not saturating and has the signedness of
11405 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11406 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11413 {
11417 }
11420 {
11421 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
11422 (plus (plus (mult) (mult)) (foo)) so that we can
11423 take advantage of the factoring cases below. */
11432 {
11438 else
11451 parg0),
11453 marg)),
11457 return
11463 parg1)));
11464 }
11465 }
11466 else
11467 {
11468 /* Fold __complex__ ( x, 0 ) + __complex__ ( 0, y )
11469 to __complex__ ( x, y ). This is not the same for SNaNs or
11470 if signed zeros are involved. */
11474 {
11481 {
11485 {
11491 }
11493 {
11499 }
11500 }
11501 }
11503 /* Convert a + (b*c + d*e) into (a + b*c) + d*e.
11504 We associate floats only if the user has specified
11505 -fassociative-math. */
11509 {
11514 {
11515 tree tree0;
11518 }
11519 }
11520 /* Convert (b*c + d*e) + a into b*c + (d*e +a).
11521 We associate floats only if the user has specified
11522 -fassociative-math. */
11526 {
11531 {
11532 tree tree0;
11535 }
11536 }
11537 }
11539 bit_rotate:
11540 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
11541 is a rotate of A by C1 bits. */
11542 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
11543 is a rotate of A by B bits.
11544 Similarly for (A << B) | (A >> (-B & C3)) where C3 is Z-1,
11545 though in this case CODE must be | and not + or ^, otherwise
11546 it doesn't return A when B is 0. */
11547 {
11549 tree rtype;
11558 /* Only create rotates in complete modes. Other cases are not
11559 expanded properly. */
11562 {
11575 {
11580 }
11585 {
11588 code0 == LSHIFT_EXPR
11591 }
11593 {
11603 {
11607 orig_tree01);
11609 }
11610 }
11614 {
11625 {
11629 orig_tree01);
11631 }
11632 }
11633 }
11634 }
11636 associate:
11637 /* In most languages, can't associate operations on floats through
11638 parentheses. Rather than remember where the parentheses were, we
11639 don't associate floats at all, unless the user has specified
11640 -fassociative-math.
11641 And, we need to make sure type is not saturating. */
11646 {
11652 /* Split both trees into variables, constants, and literals. Then
11653 associate each group together, the constants with literals,
11654 then the result with variables. This increases the chances of
11655 literals being recombined later and of generating relocatable
11656 expressions for the sum of a constant and literal. */
11664 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
11668 /* With undefined overflow prefer doing association in a type
11669 which wraps on overflow, if that is one of the operand types. */
11672 {
11680 }
11682 /* With undefined overflow we can only associate constants with one
11683 variable, and constants whose association doesn't overflow. */
11686 {
11688 {
11689 /* ??? If split_tree would handle NEGATE_EXPR we could
11690 simply reject these cases and the allowed cases would
11691 be the var0/minus_var1 ones. */
11697 {
11700 }
11707 {
11710 }
11716 /* The only case we can still associate with two variables
11717 is if they cancel out. */
11721 }
11727 }
11729 /* Only do something if we found more than two objects. Otherwise,
11730 nothing has changed and we risk infinite recursion. */
11738 {
11750 {
11754 }
11756 {
11760 }
11762 /* Preserve the MINUS_EXPR if the negative part of the literal is
11763 greater than the positive part. Otherwise, the multiplicative
11764 folding code (i.e extract_muldiv) may be fooled in case
11765 unsigned constants are subtracted, like in the following
11766 example: ((X*2 + 4) - 8U)/2. */
11768 {
11772 /* But avoid ending up with only negated parts. */
11774 {
11778 }
11779 else
11780 {
11784 }
11785 }
11787 /* Don't introduce overflows through reassociation. */
11792 /* Eliminate lit0 and minus_lit0 to con0 and minus_con0. */
11799 /* Eliminate minus_con0. */
11801 {
11808 else
11811 }
11813 /* Eliminate minus_var0. */
11815 {
11819 else
11822 }
11824 return
11827 }
11828 }
11834 /* Fold &a[i] - &a[j] to i-j. */
11839 {
11843 code
11847 }
11849 /* Further transformations are not for pointers. */
11853 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
11856 /* If arg0 is e.g. unsigned int and type is int, then this could
11857 introduce UB, because if A is INT_MIN at runtime, the original
11858 expression can be well defined while the latter is not.
11859 See PR83269. */
11868 /* Fold __complex__ ( x, 0 ) - __complex__ ( 0, y ) to
11869 __complex__ ( x, -y ). This is not the same for SNaNs or if
11870 signed zeros are involved. */
11874 {
11881 {
11885 {
11887 arg1r ? arg1r
11892 }
11894 {
11898 arg1i ? arg1i
11901 }
11902 }
11903 }
11905 /* A - B -> A + (-B) if B is easily negatable. */
11909 /* Avoid this transformation if B is a positive REAL_CST. */
11917 /* Handle (A1 * C1) - (A2 * C2) with A1, A2 or C1, C2 being the same or
11918 one. Make sure the type is not saturating and has the signedness of
11919 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11920 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11927 {
11931 }
11937 {
11938 /* Transform x * -C into -x * C if x is easily negatable. */
11953 {
11956 "occur when simplifying "
11958 WARN_STRICT_OVERFLOW_MISC);
11960 }
11962 /* Optimize z * conj(z) for integer complex numbers. */
11969 }
11970 else
11971 {
11972 /* Fold z * +-I to __complex__ (-+__imag z, +-__real z).
11973 This is not the same for NaNs or if signed zeros are
11974 involved. */
11980 {
11983 return
11989 return
11994 }
11996 /* Optimize z * conj(z) for floating point complex numbers.
11997 Guarded by flag_unsafe_math_optimizations as non-finite
11998 imaginary components don't produce scalar results. */
12007 }
12011 /* Canonicalize (X & C1) | C2. */
12015 {
12020 /* If (C1&C2) == C1, then (X&C1)|C2 becomes (X,C2). */
12028 /* If (C1|C2) == ~0 then (X&C1)|C2 becomes X|C2. */
12030 {
12033 }
12035 /* Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2,
12036 unless (C1 & ~C2) | (C2 & C3) for some C3 is a mask of some
12037 mode which allows further optimizations. */
12042 {
12047 {
12050 }
12051 }
12054 {
12059 }
12060 }
12062 /* See if this can be simplified into a rotate first. If that
12063 is unsuccessful continue in the association code. */
12067 /* Fold (X & 1) ^ 1 as (X & 1) == 0. */
12075 /* See if this can be simplified into a rotate first. If that
12076 is unsuccessful continue in the association code. */
12080 /* Fold !X & 1 as X == 0. */
12083 {
12087 }
12089 /* Fold (X * Y) & -(1 << CST) to X * Y if Y is a constant
12090 multiple of 1 << CST. */
12092 {
12099 }
12101 /* Fold (X * CST1) & CST2 to zero if we can, or drop known zero
12102 bits from CST2. */
12106 {
12108 wide_int masked
12115 {
12116 /* Avoid the transform if arg1 is a mask of some
12117 mode which allows further optimizations. */
12124 }
12125 }
12127 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
12130 {
12135 return
12137 }
12142 /* Don't touch a floating-point divide by zero unless the mode
12143 of the constant can represent infinity. */
12149 /* (-A) / (-B) -> A / B */
12161 /* Fall through */
12164 /* Simplify A / (B << N) where A and B are positive and B is
12165 a power of 2, to A >> (N + log2(B)). */
12170 {
12173 {
12181 WARN_STRICT_OVERFLOW_MISC);
12187 }
12188 }
12190 /* Fall through */
12198 /* Convert -A / -B to A / B when the type is signed and overflow is
12199 undefined. */
12203 {
12206 "when distributing negation across "
12208 WARN_STRICT_OVERFLOW_MISC);
12213 }
12217 {
12220 "when distributing negation across "
12222 WARN_STRICT_OVERFLOW_MISC);
12227 }
12229 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
12230 operation, EXACT_DIV_EXPR.
12232 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
12233 At one time others generated faster code, it's not clear if they do
12234 after the last round to changes to the DIV code in expmed.cc. */
12245 {
12249 WARN_STRICT_OVERFLOW_MISC);
12251 }
12263 {
12267 WARN_STRICT_OVERFLOW_MISC);
12269 }
12277 /* Since negative shift count is not well-defined,
12278 don't try to compute it in the compiler. */
12284 /* If we have a rotate of a bit operation with the rotate count and
12285 the second operand of the bit operation both constant,
12286 permute the two operations. */
12292 {
12300 }
12302 /* Two consecutive rotates adding up to the some integer
12303 multiple of the precision of the type can be ignored. */
12319 /* Note that the operands of this must be ints
12320 and their values must be 0 or 1.
12321 ("true" is a fixed value perhaps depending on the language.) */
12322 /* If first arg is constant zero, return it. */
12325 /* FALLTHRU */
12327 /* If either arg is constant true, drop it. */
12331 /* Preserve sequence points. */
12334 /* If second arg is constant zero, result is zero, but first arg
12335 must be evaluated. */
12338 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
12339 case will be handled here. */
12343 /* !X && X is always false. */
12347 /* X && !X is always false. */
12352 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
12353 means A >= Y && A != MAX, but in this case we know that
12354 A < X <= MAX. */
12358 {
12370 }
12379 /* Note that the operands of this must be ints
12380 and their values must be 0 or true.
12381 ("true" is a fixed value perhaps depending on the language.) */
12382 /* If first arg is constant true, return it. */
12385 /* FALLTHRU */
12387 /* If either arg is constant zero, drop it. */
12391 /* Preserve sequence points. */
12394 /* If second arg is constant true, result is true, but we must
12395 evaluate first arg. */
12398 /* Likewise for first arg, but note this only occurs here for
12399 TRUTH_OR_EXPR. */
12403 /* !X || X is always true. */
12407 /* X || !X is always true. */
12412 /* (X && !Y) || (!X && Y) is X ^ Y */
12415 {
12432 }
12441 /* If the second arg is constant zero, drop it. */
12444 /* If the second arg is constant true, this is a logical inversion. */
12446 {
12449 }
12450 /* Identical arguments cancel to zero. */
12454 /* !X ^ X is always true. */
12459 /* X ^ !X is always true. */
12475 /* bool_var != 1 becomes !bool_var. */
12482 /* bool_var == 0 becomes !bool_var. */
12489 /* !exp != 0 becomes !exp */
12494 /* If this is an EQ or NE comparison with zero and ARG0 is
12495 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
12496 two operations, but the latter can be done in one less insn
12497 on machines that have only two-operand insns or on which a
12498 constant cannot be the first operand. */
12501 {
12506 {
12514 }
12517 {
12525 }
12526 }
12528 /* If this is a comparison of a field, we may be able to simplify it. */
12531 /* Handle the constant case even without -O
12532 to make sure the warnings are given. */
12534 {
12538 }
12540 /* Optimize comparisons of strlen vs zero to a compare of the
12541 first character of the string vs zero. To wit,
12542 strlen(ptr) == 0 => *ptr == 0
12543 strlen(ptr) != 0 => *ptr != 0
12544 Other cases should reduce to one of these two (or a constant)
12545 due to the return value of strlen being unsigned. */
12547 {
12555 {
12556 tree ptrtype
12558 TYPE_QUAL_CONST));
12564 }
12565 }
12567 /* Fold (X >> C) != 0 into X < 0 if C is one less than the width
12568 of X. Similarly fold (X >> C) == 0 into X >= 0. */
12572 {
12577 {
12579 {
12582 }
12585 }
12586 }
12588 /* Fold (~X & C) == 0 into (X & C) != 0 and (~X & C) != 0 into
12589 (X & C) == 0 when C is a single bit. */
12594 {
12601 arg1));
12602 }
12604 /* Fold ((X & C) ^ C) eq/ne 0 into (X & C) ne/eq 0, when the
12605 constant C is a power of two, i.e. a single bit. */
12612 {
12616 }
12618 /* Likewise, fold ((X ^ C) & C) eq/ne 0 into (X & C) ne/eq 0,
12619 when is C is a power of two, i.e. a single bit. */
12626 {
12632 }
12636 {
12643 /* Optimize (X ^ Z) op (Y ^ Z) as X op Y, and symmetries.
12644 operand_equal_p guarantees no side-effects so we don't need
12645 to use omit_one_operand on Z. */
12649 arg10));
12653 arg11));
12657 arg10));
12661 arg11));
12663 /* Optimize (X ^ C1) op (Y ^ C2) as (X ^ (C1 ^ C2)) op Y. */
12666 {
12672 }
12673 }
12675 /* Attempt to simplify equality/inequality comparisons of complex
12676 values. Only lower the comparison if the result is known or
12677 can be simplified to a single scalar comparison. */
12682 {
12687 {
12690 }
12691 else
12692 {
12695 }
12698 {
12701 }
12702 else
12703 {
12706 }
12710 {
12712 {
12717 }
12718 else
12719 {
12724 }
12725 }
12729 {
12731 {
12736 }
12737 else
12738 {
12743 }
12744 }
12745 }
12757 /* Transform comparisons of the form X +- C CMP X. */
12762 {
12767 /* (X - c) > X becomes false. */
12773 /* Likewise (X + c) < X becomes false. */
12779 /* Convert (X - c) <= X to true. */
12786 /* Convert (X + c) >= X to true. */
12792 }
12794 /* If we are comparing an ABS_EXPR with a constant, we can
12795 convert all the cases into explicit comparisons, but they may
12796 well not be faster than doing the ABS and one comparison.
12797 But ABS (X) <= C is a range comparison, which becomes a subtraction
12798 and a comparison, and is probably faster. */
12812 /* Convert ABS_EXPR<x> >= 0 to true. */
12819 {
12822 "when simplifying comparison of "
12824 WARN_STRICT_OVERFLOW_CONDITIONAL);
12827 arg0);
12828 }
12830 /* Convert ABS_EXPR<x> < 0 to false. */
12835 {
12838 "when simplifying comparison of "
12840 WARN_STRICT_OVERFLOW_CONDITIONAL);
12843 arg0);
12844 }
12846 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
12847 and similarly for >= into !=. */
12857 /* Similarly for X < (cast) (1 << Y). But cast can't be narrowing,
12858 otherwise Y might be >= # of bits in X's type and thus e.g.
12859 (unsigned char) (1 << Y) for Y 15 might be 0.
12860 If the cast is widening, then 1 << Y should have unsigned type,
12861 otherwise if Y is number of bits in the signed shift type minus 1,
12862 we can't optimize this. E.g. (unsigned long long) (1 << Y) for Y
12863 31 might be 0xffffffff80000000. */
12876 {
12882 }
12894 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
12895 {
12907 }
12912 /* When pedantic, a compound expression can be neither an lvalue
12913 nor an integer constant expression. */
12916 /* Don't let (0, 0) be null pointer constant. */
12924 }
12926 /* For constants M and N, if M == (1LL << cst) - 1 && (N & M) == M,
12927 ((A & N) + B) & M -> (A + B) & M
12928 Similarly if (N & M) == 0,
12929 ((A | N) + B) & M -> (A + B) & M
12930 and for - instead of + (or unary - instead of +)
12931 and/or ^ instead of |.
12932 If B is constant and (B & M) == 0, fold into A & M.
12934 This function is a helper for match.pd patterns. Return non-NULL
12935 type in which the simplified operation should be performed only
12936 if any optimization is possible.
12938 ARG1 is M above, ARG00 is left operand of +/-, if CODE00 is BIT_*_EXPR,
12939 then ARG00{0,1} are operands of that bitop, otherwise CODE00 is ERROR_MARK.
12940 Similarly for ARG01, CODE01 and ARG01{0,1}, just for the right operand of
12941 +/-. */
12942 tree
12947 {
12962 wide_int cst0;
12964 /* Now we know that arg0 is (C + D) or (C - D) or -C and
12965 arg1 (M) is == (1LL << cst) - 1.
12966 Store C into PMOP[0] and D into PMOP[1]. */
12973 {
12980 {
12983 }
12986 /* If C or D is of the form (A & N) where
12987 (N & M) == M, or of the form (A | N) or
12988 (A ^ N) where (N & M) == 0, replace it with A. */
12994 /* If C or D is a N where (N & M) == 0, it can be
12995 omitted (replaced with 0). */
13000 /* Similarly, with C - N where (-N & M) == 0. */
13008 }
13010 /* Only build anything new if we optimized one or both arguments above. */
13016 else
13018 }
13020 /* Used by contains_label_[p1]. */
13022 struct contains_label_data
13023 {
13026 };
13028 /* Callback for walk_tree, looking for LABEL_EXPR. Return *TP if it is
13029 a LABEL_EXPR or CASE_LABEL_EXPR not inside of another SWITCH_EXPR; otherwise
13030 return NULL_TREE. Do not check the subtrees of GOTO_EXPR. */
13032 static tree
13034 {
13037 {
13048 {
13056 }
13065 }
13066 }
13068 /* Return whether the sub-tree ST contains a label which is accessible from
13069 outside the sub-tree. */
13071 static bool
13073 {
13077 }
13079 /* Fold a ternary expression of code CODE and type TYPE with operands
13080 OP0, OP1, and OP2. Return the folded expression if folding is
13081 successful. Otherwise, return NULL_TREE. */
13083 tree
13086 {
13087 tree tem;
13094 /* If this is a commutative operation, and OP0 is a constant, move it
13095 to OP1 to reduce the number of tests below. */
13104 /* Strip any conversions that don't change the mode. This is safe
13105 for every expression, except for a comparison expression because
13106 its signedness is derived from its operands. So, in the latter
13107 case, only strip conversions that don't change the signedness.
13109 Note that this is done as an internal manipulation within the
13110 constant folder, in order to find the simplest representation of
13111 the arguments so that their form can be studied. In any cases,
13112 the appropriate type conversions should be put back in the tree
13113 that will get out of the constant folder. */
13115 {
13118 }
13121 {
13124 }
13127 {
13130 }
13133 {
13137 {
13143 }
13148 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
13149 so all simple results must be passed through pedantic_non_lvalue. */
13151 {
13154 /* Only optimize constant conditions when the selected branch
13155 has the same type as the COND_EXPR. This avoids optimizing
13156 away "c ? x : throw", where the throw has a void type.
13157 Avoid throwing away that operand which contains label. */
13164 }
13166 {
13173 {
13176 {
13184 }
13189 }
13190 }
13192 /* If we have A op B ? A : C, we may be able to convert this to a
13193 simpler expression, depending on the operation and the values
13194 of B and C. Signed zeros prevent all of these transformations,
13195 for reasons given above each one.
13197 Also try swapping the arguments and inverting the conditional. */
13201 {
13208 }
13213 {
13220 arg00,
13221 arg01,
13225 }
13227 /* If the second operand is simpler than the third, swap them
13228 since that produces better jump optimization results. */
13231 {
13233 /* See if this can be inverted. If it can't, possibly because
13234 it was a floating-point inequality comparison, don't do
13235 anything. */
13239 }
13241 /* Convert A ? 1 : 0 to simply A. */
13246 /* If we try to convert OP0 to our type, the
13247 call to fold will try to move the conversion inside
13248 a COND, which will recurse. In that case, the COND_EXPR
13249 is probably the best choice, so leave it alone. */
13253 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
13254 over COND_EXPR in cases such as floating point comparisons. */
13263 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
13268 {
13269 /* sign_bit_p looks through both zero and sign extensions,
13270 but for this optimization only sign extensions are
13271 usable. */
13274 {
13277 {
13280 }
13282 }
13283 /* sign_bit_p only checks ARG1 bits within A's precision.
13284 If <sign bit of A> has wider type than A, bits outside
13285 of A's precision in <sign bit of A> need to be checked.
13286 If they are all 0, this optimization needs to be done
13287 in unsigned A's type, if they are all 1 in signed A's type,
13288 otherwise this can't be done. */
13294 {
13296 tree tem_type;
13309 {
13312 }
13314 {
13317 }
13318 else
13320 }
13323 return
13329 arg1)));
13330 }
13332 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
13333 already handled above. */
13338 {
13348 op1);
13349 }
13351 /* A & N ? N : 0 is simply A & N if N is a power of two. This
13352 is probably obsolete because the first operand should be a
13353 truth value (that's why we have the two cases above), but let's
13354 leave it in until we can confirm this for all front-ends. */
13362 /* operand_equal_p compares just value, not precision, so e.g.
13363 arg1 could be 8-bit -128 and be power of two, but BIT_AND_EXPR
13364 second operand 32-bit -128, which is not a power of two (or vice
13365 versa. */
13369 /* Disable the transformations below for vectors, since
13370 fold_binary_op_with_conditional_arg may undo them immediately,
13371 yielding an infinite loop. */
13375 /* Convert A ? B : 0 into A && B if A and B are truth values. */
13384 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
13389 {
13391 /* Only perform transformation if ARG0 is easily inverted. */
13395 ? BIT_IOR_EXPR
13398 op1);
13399 }
13401 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
13406 {
13408 /* Only perform transformation if ARG0 is easily inverted. */
13414 op2);
13415 }
13417 /* Convert A ? 1 : B into A || B if A and B are truth values. */
13429 /* CALL_EXPRs used to be ternary exprs. Catch any mistaken uses
13430 of fold_ternary on them. */
13440 {
13442 unsigned HOST_WIDE_INT width
13453 {
13458 {
13460 {
13465 }
13471 }
13472 }
13473 }
13475 /* On constants we can use native encode/interpret to constant
13476 fold (nearly) all BIT_FIELD_REFs. */
13482 {
13485 /* Limit us to a reasonable amount of work. To relax the
13486 other limitations we need bit-shifting of the buffer
13487 and rounding up the size. */
13491 {
13493 unsigned HOST_WIDE_INT len
13498 {
13503 }
13504 }
13505 }
13510 /* Perform constant folding of BIT_INSERT_EXPR. */
13514 {
13515 /* Build a vector of integers from the tree mask. */
13516 vec_perm_builder builder;
13520 /* Create a vec_perm_indices for the integer vector. */
13525 }
13529 /* Perform (partial) constant folding of BIT_INSERT_EXPR. */
13532 {
13538 wide_int tem2
13542 }
13547 {
13549 unsigned HOST_WIDE_INT elsize
13552 {
13558 {
13564 }
13565 }
13566 }
13572 }
13574 /* Gets the element ACCESS_INDEX from CTOR, which must be a CONSTRUCTOR
13575 of an array (or vector). *CTOR_IDX if non-NULL is updated with the
13576 constructor element index of the value returned. If the element is
13577 not found NULL_TREE is returned and *CTOR_IDX is updated to
13578 the index of the element after the ACCESS_INDEX position (which
13579 may be outside of the CTOR array). */
13581 tree
13584 {
13590 {
13593 {
13594 /* Static constructors for variably sized objects makes no sense. */
13597 /* ??? When it is obvious that the range is signed, treat it so. */
13602 {
13604 low_bound
13606 SIGNED);
13607 }
13608 else
13609 {
13612 }
13613 }
13614 }
13618 index_sgn);
13630 {
13631 /* Array constructor might explicitly set index, or specify a range,
13632 or leave index NULL meaning that it is next index after previous
13633 one. */
13635 {
13637 max_index = index
13639 else
13640 {
13643 index_sgn);
13644 max_index
13646 index_sgn);
13648 }
13649 }
13651 {
13657 }
13658 else
13661 /* Do we have match? */
13663 {
13665 {
13669 }
13670 }
13672 /* We're past the element we search for. Note during parsing
13673 the elements might not be sorted.
13674 ??? We should use a binary search and a flag on the
13675 CONSTRUCTOR as to whether elements are sorted in declaration
13676 order. */
13678 }
13682 }
13684 /* Perform constant folding and related simplification of EXPR.
13685 The related simplifications include x*1 => x, x*0 => 0, etc.,
13686 and application of the associative law.
13687 NOP_EXPR conversions may be removed freely (as long as we
13688 are careful not to change the type of the overall expression).
13689 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
13690 but we can constant-fold them if they have constant operands. */
13692 #ifdef ENABLE_FOLD_CHECKING
13693 # define fold(x) fold_1 (x)
13695 static
13696 #endif
13697 tree
13699 {
13703 tree tem;
13706 /* Return right away if a constant. */
13710 /* CALL_EXPR-like objects with variable numbers of operands are
13711 treated specially. */
13713 {
13715 {
13718 }
13720 }
13723 {
13728 {
13746 }
13747 }
13750 {
13752 {
13759 {
13764 }
13767 }
13769 /* Return a VECTOR_CST if possible. */
13771 {
13777 tree val;
13783 }
13791 }
13793 #ifdef ENABLE_FOLD_CHECKING
13794 #undef fold
13801 /* When --enable-checking=fold, compute a digest of expr before
13802 and after actual fold call to see if fold did not accidentally
13803 change original expr. */
13805 tree
13807 {
13808 tree ret;
13828 }
13830 void
13832 {
13843 }
13845 static void
13847 {
13849 }
13851 static void
13854 {
13860 recursive_label:
13870 {
13871 /* Allow DECL_ASSEMBLER_NAME and symtab_node to be modified. */
13879 }
13887 {
13888 /* Allow these fields to be modified. */
13889 tree tmp;
13900 {
13903 }
13904 }
13906 {
13907 /* Allow the no-warning bit to be set. Perhaps we shouldn't allow
13908 that and change builtins.cc etc. instead - see PR89543. */
13914 }
13925 {
13928 {
13944 }
13948 {
13961 }
13978 {
13984 }
13987 {
13989 {
13992 }
13994 }
14005 {
14008 }
14016 }
14017 }
14019 /* Helper function for outputting the checksum of a tree T. When
14020 debugging with gdb, you can "define mynext" to be "next" followed
14021 by "call debug_fold_checksum (op0)", then just trace down till the
14022 outputs differ. */
14024 DEBUG_FUNCTION void
14026 {
14041 }
14043 #endif
14045 /* Fold a unary tree expression with code CODE of type TYPE with an
14046 operand OP0. LOC is the location of the resulting expression.
14047 Return a folded expression if successful. Otherwise, return a tree
14048 expression with code CODE of type TYPE with an operand OP0. */
14050 tree
14053 {
14054 tree tem;
14055 #ifdef ENABLE_FOLD_CHECKING
14064 #endif
14070 #ifdef ENABLE_FOLD_CHECKING
14077 #endif
14079 }
14081 /* Fold a binary tree expression with code CODE of type TYPE with
14082 operands OP0 and OP1. LOC is the location of the resulting
14083 expression. Return a folded expression if successful. Otherwise,
14084 return a tree expression with code CODE of type TYPE with operands
14085 OP0 and OP1. */
14087 tree
14090 MEM_STAT_DECL)
14091 {
14092 tree tem;
14093 #ifdef ENABLE_FOLD_CHECKING
14110 #endif
14116 #ifdef ENABLE_FOLD_CHECKING
14131 #endif
14133 }
14135 /* Fold a ternary tree expression with code CODE of type TYPE with
14136 operands OP0, OP1, and OP2. Return a folded expression if
14137 successful. Otherwise, return a tree expression with code CODE of
14138 type TYPE with operands OP0, OP1, and OP2. */
14140 tree
14143 {
14144 tree tem;
14145 #ifdef ENABLE_FOLD_CHECKING
14169 #endif
14176 #ifdef ENABLE_FOLD_CHECKING
14199 #endif
14201 }
14203 /* Fold a CALL_EXPR expression of type TYPE with operands FN and NARGS
14204 arguments in ARGARRAY, and a null static chain.
14205 Return a folded expression if successful. Otherwise, return a CALL_EXPR
14206 of type TYPE from the given operands as constructed by build_call_array. */
14208 tree
14211 {
14212 tree tem;
14213 #ifdef ENABLE_FOLD_CHECKING
14232 #endif
14238 #ifdef ENABLE_FOLD_CHECKING
14254 #endif
14256 }
14258 /* Perform constant folding and related simplification of initializer
14259 expression EXPR. These behave identically to "fold_buildN" but ignore
14260 potential run-time traps and exceptions that fold must preserve. */
14262 #define START_FOLD_INIT \
14263 int saved_signaling_nans = flag_signaling_nans;\
14264 int saved_trapping_math = flag_trapping_math;\
14265 int saved_rounding_math = flag_rounding_math;\
14266 int saved_trapv = flag_trapv;\
14267 int saved_folding_initializer = folding_initializer;\
14268 flag_signaling_nans = 0;\
14269 flag_trapping_math = 0;\
14270 flag_rounding_math = 0;\
14271 flag_trapv = 0;\
14272 folding_initializer = 1;
14274 #define END_FOLD_INIT \
14275 flag_signaling_nans = saved_signaling_nans;\
14276 flag_trapping_math = saved_trapping_math;\
14277 flag_rounding_math = saved_rounding_math;\
14278 flag_trapv = saved_trapv;\
14279 folding_initializer = saved_folding_initializer;
14281 tree
14283 {
14284 tree result;
14285 START_FOLD_INIT;
14289 END_FOLD_INIT;
14291 }
14293 tree
14296 {
14297 tree result;
14298 START_FOLD_INIT;
14302 END_FOLD_INIT;
14304 }
14306 tree
14309 {
14310 tree result;
14311 START_FOLD_INIT;
14315 END_FOLD_INIT;
14317 }
14319 tree
14322 {
14323 tree result;
14324 START_FOLD_INIT;
14328 END_FOLD_INIT;
14330 }
14332 tree
14335 {
14336 tree result;
14337 START_FOLD_INIT;
14341 END_FOLD_INIT;
14343 }
14345 #undef START_FOLD_INIT
14346 #undef END_FOLD_INIT
14348 /* Determine if first argument is a multiple of second argument. Return
14349 false if it is not, or we cannot easily determined it to be.
14351 An example of the sort of thing we care about (at this point; this routine
14352 could surely be made more general, and expanded to do what the *_DIV_EXPR's
14353 fold cases do now) is discovering that
14355 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14357 is a multiple of
14359 SAVE_EXPR (J * 8)
14361 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
14363 This code also handles discovering that
14365 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14367 is a multiple of 8 so we don't have to worry about dealing with a
14368 possible remainder.
14370 Note that we *look* inside a SAVE_EXPR only to determine how it was
14371 calculated; it is not safe for fold to do much of anything else with the
14372 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
14373 at run time. For example, the latter example above *cannot* be implemented
14374 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
14375 evaluation time of the original SAVE_EXPR is not necessarily the same at
14376 the time the new expression is evaluated. The only optimization of this
14377 sort that would be valid is changing
14379 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
14381 divided by 8 to
14383 SAVE_EXPR (I) * SAVE_EXPR (J)
14385 (where the same SAVE_EXPR (J) is used in the original and the
14386 transformed version).
14388 NOWRAP specifies whether all outer operations in TYPE should
14389 be considered not wrapping. Any type conversion within TOP acts
14390 as a barrier and we will fall back to NOWRAP being false.
14391 NOWRAP is mostly used to treat expressions in TYPE_SIZE and friends
14392 as not wrapping even though they are generally using unsigned arithmetic. */
14394 bool
14396 {
14407 {
14409 /* Bitwise and provides a power of two multiple. If the mask is
14410 a multiple of BOTTOM then TOP is a multiple of BOTTOM. */
14417 /* If the multiplication can wrap we cannot recurse further unless
14418 the bottom is a power of two which is where wrapping does not
14419 matter. */
14425 {
14431 {
14434 /* Handle multiple_of_p ((x * 2 + 2) * 4, 8). */
14436 {
14440 {
14443 }
14444 }
14446 }
14447 }
14452 /* Handle X << CST as X * (1 << CST) and only process the constant. */
14454 {
14457 {
14458 wide_int mul_op
14462 nowrap);
14463 }
14464 }
14469 /* If the addition or subtraction can wrap we cannot recurse further
14470 unless bottom is a power of two which is where wrapping does not
14471 matter. */
14477 /* Handle cases like op0 + 0xfffffffd as op0 - 3 if the expression has
14478 unsigned type. For example, (X / 3) + 0xfffffffd is multiple of 3,
14479 but 0xfffffffd is not. */
14482 && nowrap
14487 /* It is impossible to prove if op0 +- op1 is multiple of bottom
14488 precisely, so be conservative here checking if both op0 and op1
14489 are multiple of bottom. Note we check the second operand first
14490 since it's usually simpler. */
14494 CASE_CONVERT:
14495 /* Can't handle conversions from non-integral or wider integral type. */
14500 /* NOWRAP only extends to operations in the outermost type so
14501 make sure to strip it off here. */
14516 SIGNED);
14522 {
14525 /* Check for special cases to see if top is defined as multiple
14526 of bottom:
14528 top = (X & ~(bottom - 1) ; bottom is power of 2
14530 or
14532 Y = X % bottom
14533 top = X - Y. */
14552 }
14554 /* fall through */
14562 }
14563 }
14565 /* Return true if expression X cannot be (or contain) a NaN or infinity.
14566 This function returns true for integer expressions, and returns
14567 false if uncertain. */
14569 bool
14571 {
14576 {
14599 {
14600 CASE_CFN_FABS:
14601 CASE_CFN_FABS_FN:
14603 CASE_CFN_FMAX:
14604 CASE_CFN_FMAX_FN:
14605 CASE_CFN_FMIN:
14606 CASE_CFN_FMIN_FN:
14611 }
14615 }
14616 }
14618 /* Return true if expression X evaluates to an infinity.
14619 This function returns false for integer expressions. */
14621 bool
14623 {
14627 {
14640 }
14641 }
14643 /* Return true if expression X could evaluate to an infinity.
14644 This function returns false for integer expressions, and returns
14645 true if uncertain. */
14647 bool
14649 {
14653 {
14666 }
14667 }
14669 /* Return true if expression X evaluates to a signaling NaN.
14670 This function returns false for integer expressions. */
14672 bool
14674 {
14678 {
14689 }
14690 }
14692 /* Return true if expression X could evaluate to a signaling NaN.
14693 This function returns false for integer expressions, and returns
14694 true if uncertain. */
14696 bool
14698 {
14702 {
14722 {
14723 CASE_CFN_FABS:
14724 CASE_CFN_FABS_FN:
14726 CASE_CFN_FMAX:
14727 CASE_CFN_FMAX_FN:
14728 CASE_CFN_FMIN:
14729 CASE_CFN_FMIN_FN:
14734 }
14737 }
14738 }
14740 /* Return true if expression X evaluates to a NaN.
14741 This function returns false for integer expressions. */
14743 bool
14745 {
14749 {
14760 }
14761 }
14763 /* Return true if expression X could evaluate to a NaN.
14764 This function returns false for integer expressions, and returns
14765 true if uncertain. */
14767 bool
14769 {
14773 {
14798 {
14799 CASE_CFN_FABS:
14800 CASE_CFN_FABS_FN:
14802 CASE_CFN_FMAX:
14803 CASE_CFN_FMAX_FN:
14804 CASE_CFN_FMIN:
14805 CASE_CFN_FMIN_FN:
14810 }
14813 }
14814 }
14816 /* Return true if expression X could evaluate to -0.0.
14817 This function returns true if uncertain. */
14819 bool
14821 {
14825 {
14840 {
14841 CASE_CFN_FABS:
14842 CASE_CFN_FABS_FN:
14846 }
14849 }
14850 /* Ideally !(tree_expr_nonzero_p (X) || tree_expr_nonnegative_p (X))
14851 * but currently those predicates require tree and not const_tree. */
14853 }
14855 #define tree_expr_nonnegative_warnv_p(X, Y) \
14858 #define RECURSE(X) \
14859 ((tree_expr_nonnegative_warnv_p) (X, strict_overflow_p, depth + 1))
14861 /* Return true if CODE or TYPE is known to be non-negative. */
14863 static bool
14865 {
14869 /* Truth values evaluate to 0 or 1, which is nonnegative unless we
14870 have a signed:1 type (where the value is -1 and 0). */
14873 }
14875 /* Return true if (CODE OP0) is known to be non-negative. If the return
14876 value is based on the assumption that signed overflow is undefined,
14877 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14878 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14880 bool
14883 {
14888 {
14890 /* We can't return 1 if flag_wrapv is set because
14891 ABS_EXPR<INT_MIN> = INT_MIN. */
14895 {
14898 }
14906 CASE_CONVERT:
14907 {
14912 {
14916 {
14920 }
14921 }
14923 {
14929 }
14930 }
14935 }
14937 /* We don't know sign of `t', so be conservative and return false. */
14939 }
14941 /* Return true if (CODE OP0 OP1) is known to be non-negative. If the return
14942 value is based on the assumption that signed overflow is undefined,
14943 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14944 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14946 bool
14950 {
14955 {
14961 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
14962 both unsigned and at least 2 bits shorter than the result. */
14966 {
14971 {
14975 }
14976 }
14981 {
14982 /* x * x is always non-negative for floating point x
14983 or without overflow. */
14986 {
14991 }
14992 }
14994 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
14995 both unsigned and their total bits is shorter than the result. */
14999 {
15018 {
15028 }
15029 }
15036 /* Usually RECURSE (op0) || RECURSE (op1) but NaNs complicate
15037 things. */
15062 }
15064 /* We don't know sign of `t', so be conservative and return false. */
15066 }
15068 /* Return true if T is known to be non-negative. If the return
15069 value is based on the assumption that signed overflow is undefined,
15070 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15071 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15073 bool
15075 {
15080 {
15094 /* Limit the depth of recursion to avoid quadratic behavior.
15095 This is expected to catch almost all occurrences in practice.
15096 If this code misses important cases that unbounded recursion
15097 would not, passes that need this information could be revised
15098 to provide it through dataflow propagation. */
15106 }
15107 }
15109 /* Return true if T is known to be non-negative. If the return
15110 value is based on the assumption that signed overflow is undefined,
15111 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15112 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15114 bool
15117 {
15119 {
15120 CASE_CFN_ACOS:
15121 CASE_CFN_ACOS_FN:
15122 CASE_CFN_ACOSH:
15123 CASE_CFN_ACOSH_FN:
15124 CASE_CFN_CABS:
15125 CASE_CFN_CABS_FN:
15126 CASE_CFN_COSH:
15127 CASE_CFN_COSH_FN:
15128 CASE_CFN_ERFC:
15129 CASE_CFN_ERFC_FN:
15130 CASE_CFN_EXP:
15131 CASE_CFN_EXP_FN:
15132 CASE_CFN_EXP10:
15133 CASE_CFN_EXP2:
15134 CASE_CFN_EXP2_FN:
15135 CASE_CFN_FABS:
15136 CASE_CFN_FABS_FN:
15137 CASE_CFN_FDIM:
15138 CASE_CFN_FDIM_FN:
15139 CASE_CFN_HYPOT:
15140 CASE_CFN_HYPOT_FN:
15141 CASE_CFN_POW10:
15142 CASE_CFN_FFS:
15143 CASE_CFN_PARITY:
15144 CASE_CFN_POPCOUNT:
15145 CASE_CFN_CLZ:
15146 CASE_CFN_CLRSB:
15151 /* Always true. */
15154 CASE_CFN_SQRT:
15155 CASE_CFN_SQRT_FN:
15156 /* sqrt(-0.0) is -0.0. */
15161 CASE_CFN_ASINH:
15162 CASE_CFN_ASINH_FN:
15163 CASE_CFN_ATAN:
15164 CASE_CFN_ATAN_FN:
15165 CASE_CFN_ATANH:
15166 CASE_CFN_ATANH_FN:
15167 CASE_CFN_CBRT:
15168 CASE_CFN_CBRT_FN:
15169 CASE_CFN_CEIL:
15170 CASE_CFN_CEIL_FN:
15171 CASE_CFN_ERF:
15172 CASE_CFN_ERF_FN:
15173 CASE_CFN_EXPM1:
15174 CASE_CFN_EXPM1_FN:
15175 CASE_CFN_FLOOR:
15176 CASE_CFN_FLOOR_FN:
15177 CASE_CFN_FMOD:
15178 CASE_CFN_FMOD_FN:
15179 CASE_CFN_FREXP:
15180 CASE_CFN_FREXP_FN:
15181 CASE_CFN_ICEIL:
15182 CASE_CFN_IFLOOR:
15183 CASE_CFN_IRINT:
15184 CASE_CFN_IROUND:
15185 CASE_CFN_LCEIL:
15186 CASE_CFN_LDEXP:
15187 CASE_CFN_LFLOOR:
15188 CASE_CFN_LLCEIL:
15189 CASE_CFN_LLFLOOR:
15190 CASE_CFN_LLRINT:
15191 CASE_CFN_LLRINT_FN:
15192 CASE_CFN_LLROUND:
15193 CASE_CFN_LLROUND_FN:
15194 CASE_CFN_LRINT:
15195 CASE_CFN_LRINT_FN:
15196 CASE_CFN_LROUND:
15197 CASE_CFN_LROUND_FN:
15198 CASE_CFN_MODF:
15199 CASE_CFN_MODF_FN:
15200 CASE_CFN_NEARBYINT:
15201 CASE_CFN_NEARBYINT_FN:
15202 CASE_CFN_RINT:
15203 CASE_CFN_RINT_FN:
15204 CASE_CFN_ROUND:
15205 CASE_CFN_ROUND_FN:
15206 CASE_CFN_ROUNDEVEN:
15207 CASE_CFN_ROUNDEVEN_FN:
15208 CASE_CFN_SCALB:
15209 CASE_CFN_SCALBLN:
15210 CASE_CFN_SCALBLN_FN:
15211 CASE_CFN_SCALBN:
15212 CASE_CFN_SCALBN_FN:
15213 CASE_CFN_SIGNBIT:
15214 CASE_CFN_SIGNIFICAND:
15215 CASE_CFN_SINH:
15216 CASE_CFN_SINH_FN:
15217 CASE_CFN_TANH:
15218 CASE_CFN_TANH_FN:
15219 CASE_CFN_TRUNC:
15220 CASE_CFN_TRUNC_FN:
15221 /* True if the 1st argument is nonnegative. */
15224 CASE_CFN_FMAX:
15225 CASE_CFN_FMAX_FN:
15226 /* Usually RECURSE (arg0) || RECURSE (arg1) but NaNs complicate
15227 things. In the presence of sNaNs, we're only guaranteed to be
15228 non-negative if both operands are non-negative. In the presence
15229 of qNaNs, we're non-negative if either operand is non-negative
15230 and can't be a qNaN, or if both operands are non-negative. */
15239 CASE_CFN_FMIN:
15240 CASE_CFN_FMIN_FN:
15241 /* True if the 1st AND 2nd arguments are nonnegative. */
15244 CASE_CFN_COPYSIGN:
15245 CASE_CFN_COPYSIGN_FN:
15246 /* True if the 2nd argument is nonnegative. */
15249 CASE_CFN_POWI:
15250 /* True if the 1st argument is nonnegative or the second
15251 argument is an even integer. */
15257 CASE_CFN_POW:
15258 CASE_CFN_POW_FN:
15259 /* True if the 1st argument is nonnegative or the second
15260 argument is an even integer valued real. */
15262 {
15263 REAL_VALUE_TYPE c;
15264 HOST_WIDE_INT n;
15269 {
15270 REAL_VALUE_TYPE cint;
15274 }
15275 }
15280 }
15282 }
15284 /* Return true if T is known to be non-negative. If the return
15285 value is based on the assumption that signed overflow is undefined,
15286 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15287 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15289 static bool
15291 {
15297 {
15299 {
15303 /* If the initializer is non-void, then it's a normal expression
15304 that will be assigned to the slot. */
15308 /* Otherwise, the initializer sets the slot in some way. One common
15309 way is an assignment statement at the end of the initializer. */
15311 {
15319 else
15321 }
15327 }
15330 {
15336 arg0,
15337 arg1,
15339 }
15352 }
15353 }
15355 #undef RECURSE
15356 #undef tree_expr_nonnegative_warnv_p
15358 /* Return true if T is known to be non-negative. If the return
15359 value is based on the assumption that signed overflow is undefined,
15360 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15361 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15363 bool
15365 {
15372 {
15394 }
15397 {
15422 }
15423 }
15425 /* Return true if `t' is known to be non-negative. Handle warnings
15426 about undefined signed overflow. */
15428 bool
15430 {
15437 "determining that expression is always "
15439 WARN_STRICT_OVERFLOW_MISC);
15441 }
15444 /* Return true when (CODE OP0) is an address and is known to be nonzero.
15445 For floating point we further ensure that T is not denormal.
15446 Similar logic is present in nonzero_address in rtlanal.h.
15448 If the return value is based on the assumption that signed overflow
15449 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15450 change *STRICT_OVERFLOW_P. */
15452 bool
15455 {
15457 {
15460 strict_overflow_p);
15463 {
15469 strict_overflow_p));
15470 }
15475 strict_overflow_p);
15479 }
15482 }
15484 /* Return true when (CODE OP0 OP1) is an address and is known to be nonzero.
15485 For floating point we further ensure that T is not denormal.
15486 Similar logic is present in nonzero_address in rtlanal.h.
15488 If the return value is based on the assumption that signed overflow
15489 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15490 change *STRICT_OVERFLOW_P. */
15492 bool
15494 tree type,
15495 tree op0,
15497 {
15500 {
15504 {
15505 /* With the presence of negative values it is hard
15506 to say something. */
15513 /* One of operands must be positive and the other non-negative. */
15514 /* We don't set *STRICT_OVERFLOW_P here: even if this value
15515 overflows, on a twos-complement machine the sum of two
15516 nonnegative numbers can never be zero. */
15518 strict_overflow_p)
15520 strict_overflow_p));
15521 }
15526 {
15528 strict_overflow_p)
15530 strict_overflow_p))
15531 {
15534 }
15535 }
15544 {
15547 }
15554 {
15558 /* When both operands are nonzero, then MAX must be too. */
15560 strict_overflow_p))
15563 /* MAX where operand 0 is positive is positive. */
15565 strict_overflow_p);
15566 }
15567 /* MAX where operand 1 is positive is positive. */
15572 {
15576 }
15581 strict_overflow_p)
15583 strict_overflow_p));
15587 }
15590 }
15592 /* Return true when T is an address and is known to be nonzero.
15593 For floating point we further ensure that T is not denormal.
15594 Similar logic is present in nonzero_address in rtlanal.h.
15596 If the return value is based on the assumption that signed overflow
15597 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15598 change *STRICT_OVERFLOW_P. */
15600 bool
15602 {
15605 {
15610 {
15622 /* For objects in symbol table check if we know they are non-zero.
15623 Don't do anything for variables and functions before symtab is built;
15624 it is quite possible that they will be declared weak later. */
15629 /* Constants are never weak. */
15634 }
15642 {
15646 }
15656 }
15658 }
15660 #define integer_valued_real_p(X) \
15663 #define RECURSE(X) \
15664 ((integer_valued_real_p) (X, depth + 1))
15666 /* Return true if the floating point result of (CODE OP0) has an
15667 integer value. We also allow +Inf, -Inf and NaN to be considered
15668 integer values. Return false for signaling NaN.
15670 DEPTH is the current nesting depth of the query. */
15672 bool
15674 {
15676 {
15683 CASE_CONVERT:
15684 {
15691 }
15695 }
15697 }
15699 /* Return true if the floating point result of (CODE OP0 OP1) has an
15700 integer value. We also allow +Inf, -Inf and NaN to be considered
15701 integer values. Return false for signaling NaN.
15703 DEPTH is the current nesting depth of the query. */
15705 bool
15707 {
15709 {
15719 }
15721 }
15723 /* Return true if the floating point result of calling FNDECL with arguments
15724 ARG0 and ARG1 has an integer value. We also allow +Inf, -Inf and NaN to be
15725 considered integer values. Return false for signaling NaN. If FNDECL
15726 takes fewer than 2 arguments, the remaining ARGn are null.
15728 DEPTH is the current nesting depth of the query. */
15730 bool
15732 {
15734 {
15735 CASE_CFN_CEIL:
15736 CASE_CFN_CEIL_FN:
15737 CASE_CFN_FLOOR:
15738 CASE_CFN_FLOOR_FN:
15739 CASE_CFN_NEARBYINT:
15740 CASE_CFN_NEARBYINT_FN:
15741 CASE_CFN_RINT:
15742 CASE_CFN_RINT_FN:
15743 CASE_CFN_ROUND:
15744 CASE_CFN_ROUND_FN:
15745 CASE_CFN_ROUNDEVEN:
15746 CASE_CFN_ROUNDEVEN_FN:
15747 CASE_CFN_TRUNC:
15748 CASE_CFN_TRUNC_FN:
15751 CASE_CFN_FMIN:
15752 CASE_CFN_FMIN_FN:
15753 CASE_CFN_FMAX:
15754 CASE_CFN_FMAX_FN:
15759 }
15761 }
15763 /* Return true if the floating point expression T (a GIMPLE_SINGLE_RHS)
15764 has an integer value. We also allow +Inf, -Inf and NaN to be
15765 considered integer values. Return false for signaling NaN.
15767 DEPTH is the current nesting depth of the query. */
15769 bool
15771 {
15773 {
15781 /* Limit the depth of recursion to avoid quadratic behavior.
15782 This is expected to catch almost all occurrences in practice.
15783 If this code misses important cases that unbounded recursion
15784 would not, passes that need this information could be revised
15785 to provide it through dataflow propagation. */
15789 depth));
15793 }
15795 }
15797 /* Return true if the floating point expression T (a GIMPLE_INVALID_RHS)
15798 has an integer value. We also allow +Inf, -Inf and NaN to be
15799 considered integer values. Return false for signaling NaN.
15801 DEPTH is the current nesting depth of the query. */
15803 static bool
15805 {
15807 {
15818 }
15820 }
15822 #undef RECURSE
15823 #undef integer_valued_real_p
15825 /* Return true if the floating point expression T has an integer value.
15826 We also allow +Inf, -Inf and NaN to be considered integer values.
15827 Return false for signaling NaN.
15829 DEPTH is the current nesting depth of the query. */
15831 bool
15833 {
15841 {
15857 }
15860 {
15866 {
15875 }
15879 }
15880 }
15882 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
15883 attempt to fold the expression to a constant without modifying TYPE,
15884 OP0 or OP1.
15886 If the expression could be simplified to a constant, then return
15887 the constant. If the expression would not be simplified to a
15888 constant, then return NULL_TREE. */
15890 tree
15892 {
15895 }
15897 /* Given the components of a unary expression CODE, TYPE and OP0,
15898 attempt to fold the expression to a constant without modifying
15899 TYPE or OP0.
15901 If the expression could be simplified to a constant, then return
15902 the constant. If the expression would not be simplified to a
15903 constant, then return NULL_TREE. */
15905 tree
15907 {
15910 }
15912 /* If EXP represents referencing an element in a constant string
15913 (either via pointer arithmetic or array indexing), return the
15914 tree representing the value accessed, otherwise return NULL. */
15916 tree
15918 {
15922 {
15924 tree index;
15925 tree string;
15930 else
15931 {
15935 /* Optimize the special-case of a zero lower bound.
15937 We convert the low_bound to sizetype to avoid some problems
15938 with constant folding. (E.g. suppose the lower bound is 1,
15939 and its mode is QI. Without the conversion,l (ARRAY
15940 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
15941 +INDEX), which becomes (ARRAY+255+INDEX). Oops!) */
15947 }
15949 scalar_int_mode char_mode;
15961 }
15963 }
15965 /* Folds a read from vector element at IDX of vector ARG. */
15967 tree
15969 {
15974 {
15978 {
15985 }
15986 }
15988 }
15990 /* Return the tree for neg (ARG0) when ARG0 is known to be either
15991 an integer constant, real, or fixed-point constant.
15993 TYPE is the type of the result. */
15995 static tree
15997 {
16001 {
16007 {
16008 FIXED_VALUE_TYPE f;
16013 /* Propagate overflow flags. */
16017 }
16021 {
16028 }
16031 }
16034 }
16036 /* Return the tree for abs (ARG0) when ARG0 is known to be either
16037 an integer constant or real constant.
16039 TYPE is the type of the result. */
16041 tree
16043 {
16047 {
16049 {
16050 /* If the value is unsigned or non-negative, then the absolute value
16051 is the same as the ordinary value. */
16055 ;
16057 /* If the value is negative, then the absolute value is
16058 its negation. */
16059 else
16062 /* Force to the destination type, set TREE_OVERFLOW for signed
16063 TYPE only. */
16065 }
16071 else
16077 }
16080 }
16082 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
16083 constant. TYPE is the type of the result. */
16085 static tree
16087 {
16091 }
16093 /* Given CODE, a relational operator, the target type, TYPE and two
16094 constant operands OP0 and OP1, return the result of the
16095 relational operation. If the result is not a compile time
16096 constant, then return NULL_TREE. */
16098 static tree
16100 {
16103 /* From here on, the only cases we handle are when the result is
16104 known to be a constant. */
16107 {
16111 /* Handle the cases where either operand is a NaN. */
16113 {
16115 {
16143 }
16146 }
16149 }
16152 {
16156 }
16158 /* Handle equality/inequality of complex constants. */
16160 {
16171 else
16173 }
16176 {
16178 {
16179 /* Have vector comparison with scalar boolean result. */
16187 {
16195 }
16197 }
16198 tree_vector_builder elts;
16203 {
16216 }
16219 }
16221 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
16223 To compute GT, swap the arguments and do LT.
16224 To compute GE, do LT and invert the result.
16225 To compute LE, swap the arguments, do LT and invert the result.
16226 To compute NE, do EQ and invert the result.
16228 Therefore, the code below must handle only EQ and LT. */
16231 {
16234 }
16236 /* Note that it is safe to invert for real values here because we
16237 have already handled the one case that it matters. */
16241 {
16244 }
16246 /* Compute a result for LT or EQ if args permit;
16247 Otherwise return T. */
16249 {
16252 else
16254 }
16255 else
16261 }
16263 /* If necessary, return a CLEANUP_POINT_EXPR for EXPR with the
16264 indicated TYPE. If no CLEANUP_POINT_EXPR is necessary, return EXPR
16265 itself. */
16267 tree
16269 {
16270 /* If the expression does not have side effects then we don't have to wrap
16271 it with a cleanup point expression. */
16275 /* If the expression is a return, check to see if the expression inside the
16276 return has no side effects or the right hand side of the modify expression
16277 inside the return. If either don't have side effects set we don't need to
16278 wrap the expression in a cleanup point expression. Note we don't check the
16279 left hand side of the modify because it should always be a return decl. */
16281 {
16288 }
16291 }
16293 /* Given a pointer value OP0 and a type TYPE, return a simplified version
16294 of an indirection through OP0, or NULL_TREE if no simplification is
16295 possible. */
16297 tree
16299 {
16301 tree subtype;
16302 poly_uint64 const_op01;
16311 {
16315 /* *&CONST_DECL -> to the value of the const decl. */
16318 /* *&p => p; make sure to handle *&"str"[cst] here. */
16320 {
16324 else
16326 }
16327 /* *(foo *)&fooarray => fooarray[0] */
16330 && (!in_gimple_form
16332 {
16342 }
16343 /* *(foo *)&complexfoo => __real__ complexfoo */
16347 /* *(foo *)&vectorfoo => BIT_FIELD_REF<vectorfoo,...> */
16350 {
16354 index);
16355 }
16356 }
16360 {
16366 {
16367 tree op00type;
16371 /* ((foo*)&vectorfoo)[1] => BIT_FIELD_REF<vectorfoo,...> */
16374 /* POINTER_PLUS_EXPR second operand is sizetype, unsigned,
16375 but we want to treat offsets with MSB set as negative.
16376 For the code below negative offsets are invalid and
16377 TYPE_SIZE of the element is something unsigned, so
16378 check whether op01 fits into poly_int64, which implies
16379 it is from 0 to INTTYPE_MAXIMUM (HOST_WIDE_INT), and
16380 then just use poly_uint64 because we want to treat the
16381 value as unsigned. */
16383 {
16385 poly_uint64 max_offset
16389 {
16394 }
16395 }
16396 /* ((foo*)&complexfoo)[1] => __imag__ complexfoo */
16399 {
16401 const_op01))
16403 }
16404 /* ((foo *)&fooarray)[1] => fooarray[1] */
16407 {
16416 {
16421 }
16422 }
16423 }
16424 }
16426 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
16429 && (!in_gimple_form
16431 {
16432 tree type_domain;
16442 NULL_TREE);
16443 }
16446 }
16448 /* Builds an expression for an indirection through T, simplifying some
16449 cases. */
16451 tree
16453 {
16461 }
16463 /* Given an INDIRECT_REF T, return either T or a simplified version. */
16465 tree
16467 {
16472 else
16474 }
16476 /* Strip non-trapping, non-side-effecting tree nodes from an expression
16477 whose result is ignored. The type of the returned tree need not be
16478 the same as the original expression. */
16480 tree
16482 {
16488 {
16499 else
16505 {
16521 }
16526 }
16527 }
16529 /* Return the value of VALUE, rounded up to a multiple of DIVISOR. */
16531 tree
16533 {
16539 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16540 have to do anything. Only do this when we are not given a const,
16541 because in that case, this check is more expensive than just
16542 doing it. */
16544 {
16549 }
16551 /* If divisor is a power of two, simplify this to bit manipulation. */
16553 {
16555 {
16569 }
16570 else
16571 {
16572 tree t;
16578 }
16579 }
16580 else
16581 {
16586 }
16589 }
16591 /* Likewise, but round down. */
16593 tree
16595 {
16602 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16603 have to do anything. Only do this when we are not given a const,
16604 because in that case, this check is more expensive than just
16605 doing it. */
16607 {
16612 }
16614 /* If divisor is a power of two, simplify this to bit manipulation. */
16616 {
16617 tree t;
16621 }
16622 else
16623 {
16628 }
16631 }
16633 /* Returns the pointer to the base of the object addressed by EXP and
16634 extracts the information about the offset of the access, storing it
16635 to PBITPOS and POFFSET. */
16637 static tree
16640 {
16641 tree core;
16642 machine_mode mode;
16644 poly_int64 bitsize;
16653 {
16658 }
16660 {
16666 {
16667 poly_offset_int tem
16673 }
16674 }
16675 else
16676 {
16680 }
16683 }
16685 /* Returns true if addresses of E1 and E2 differ by a constant, false
16686 otherwise. If they do, E1 - E2 is stored in *DIFF. */
16688 bool
16690 {
16705 {
16715 }
16717 {
16718 /* If only one of the offsets is non-constant, the difference cannot
16719 be a constant. */
16721 }
16722 else
16727 }
16729 /* Return OFF converted to a pointer offset type suitable as offset for
16730 POINTER_PLUS_EXPR. Use location LOC for this conversion. */
16731 tree
16733 {
16737 }
16739 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16740 tree
16742 {
16745 }
16747 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16748 tree
16750 {
16753 }
16755 /* Return a pointer to a NUL-terminated string containing the sequence
16756 of bytes corresponding to the representation of the object referred to
16757 by SRC (or a subsequence of such bytes within it if SRC is a reference
16758 to an initialized constant array plus some constant offset).
16759 Set *STRSIZE the number of bytes in the constant sequence including
16760 the terminating NUL byte. *STRSIZE is equal to sizeof(A) - OFFSET
16761 where A is the array that stores the constant sequence that SRC points
16762 to and OFFSET is the byte offset of SRC from the beginning of A. SRC
16763 need not point to a string or even an array of characters but may point
16764 to an object of any type. */
16768 {
16769 /* The offset into the array A storing the string, and A's byte size. */
16770 tree offset_node;
16771 tree mem_size;
16778 else
16785 {
16788 else
16790 }
16795 /* ARRAY_SIZE is the byte size of the array the constant sequence
16796 is stored in and equal to sizeof A. INIT_BYTES is the number
16797 of bytes in the constant sequence used to initialize the array,
16798 including any embedded NULs as well as the terminating NUL (for
16799 strings), but not including any trailing zeros/NULs past
16800 the terminating one appended implicitly to a string literal to
16801 zero out the remainder of the array it's stored in. For example,
16802 given:
16803 const char a[7] = "abc\0d";
16804 n = strlen (a + 1);
16805 ARRAY_SIZE is 7, INIT_BYTES is 6, and OFFSET is 1. For a valid
16806 (i.e., nul-terminated) string with no embedded nuls, INIT_BYTES
16807 is equal to strlen (A) + 1. */
16812 /* Ideally this would turn into a gcc_checking_assert over time. */
16820 {
16821 /* Compute and store the number of characters from the beginning
16822 of the substring at OFFSET to the end, including the terminating
16823 nul. Offsets past the initial length refer to null strings. */
16826 else
16828 }
16829 else
16830 {
16832 /* Support only properly NUL-terminated single byte strings. */
16837 }
16840 }
16842 /* Return a pointer to a NUL-terminated string corresponding to
16843 the expression STR referencing a constant string, possibly
16844 involving a constant offset. Return null if STR either doesn't
16845 reference a constant string or if it involves a nonconstant
16846 offset. */
16850 {
16852 }
16854 /* Given a tree T, compute which bits in T may be nonzero. */
16856 wide_int
16858 {
16860 {
16878 CASE_CONVERT:
16884 {
16889 }
16893 {
16901 }
16905 {
16913 }
16917 }
16920 }
16922 /* Helper function for address compare simplifications in match.pd.
16923 OP0 and OP1 are ADDR_EXPR operands being compared by CODE.
16924 TYPE is the type of comparison operands.
16925 BASE0, BASE1, OFF0 and OFF1 are set by the function.
16926 GENERIC is true if GENERIC folding and false for GIMPLE folding.
16927 Returns 0 if OP0 is known to be unequal to OP1 regardless of OFF{0,1},
16928 1 if bases are known to be equal and OP0 cmp OP1 depends on OFF0 cmp OFF1,
16929 and 2 if unknown. */
16931 int
16935 {
16945 {
16948 }
16950 {
16953 }
16958 /* Punt in GENERIC on variables with value expressions;
16959 the value expressions might point to fields/elements
16960 of other vars etc. */
16966 {
16970 }
16978 /* Assume different STRING_CSTs with the same content will be
16979 merged. */
16988 {
16991 /* If the offsets are equal we can ignore overflow. */
16994 /* Or if we compare using pointers to decls or strings. */
16999 }
17005 /* At this point we know (or assume) the two pointers point at
17006 different objects. */
17010 /* Punt on non-zero offsets from functions. */
17014 /* Or if the bases are neither decls nor string literals. */
17019 /* For initializers, assume addresses of different functions are
17020 different. */
17026 /* Compute whether one address points to the start of one
17027 object and another one to the end of another one. */
17030 {
17033 else
17035 }
17038 else
17039 {
17043 else
17045 }
17047 {
17050 else
17052 }
17055 else
17056 {
17060 else
17062 }
17064 {
17065 /* If one offset is pointing (or could be) to the beginning of one
17066 object and the other is pointing to one past the last byte of the
17067 other object, punt. */
17072 /* If both offsets are the same, there are some cases we know that are
17073 ok. Either if we know they aren't zero, or if we know both sizes
17074 are no zero. */
17080 }
17082 /* At this point, equal is 2 if either one or both pointers are out of
17083 bounds of their object, or one points to start of its object and the
17084 other points to end of its object. This is unspecified behavior
17085 e.g. in C++. Otherwise equal is 0. */
17089 /* When both pointers point to string literals, even when equal is 0,
17090 due to tail merging of string literals the pointers might be the same. */
17092 {
17099 /* If the bytes in the string literals starting at the pointers
17100 differ, the pointers need to be different. */
17105 {
17110 /* If even the bytes in the string literal before the
17111 pointers are the same, the string literals could be
17112 tail merged. */
17114 }
17116 }
17121 /* If this is a pointer comparison, ignore for now even
17122 valid equalities where one pointer is the offset zero
17123 of one object and the other to one past end of another one. */
17127 /* Assume that string literals can't be adjacent to variables
17128 (automatic or global). */
17132 /* Assume that automatic variables can't be adjacent to global
17133 variables. */
17138 }
17140 /* Return the single non-zero element of a CONSTRUCTOR or NULL_TREE. */
17141 tree
17143 {
17152 {
17156 }
17158 }
17160 #if CHECKING_P
17164 /* Helper functions for writing tests of folding trees. */
17166 /* Verify that the binary op (LHS CODE RHS) folds to CONSTANT. */
17168 static void
17170 tree constant)
17171 {
17173 }
17175 /* Verify that the binary op (LHS CODE RHS) folds to an NON_LVALUE_EXPR
17176 wrapping WRAPPED_EXPR. */
17178 static void
17180 tree wrapped_expr)
17181 {
17186 }
17188 /* Verify that various arithmetic binary operations are folded
17189 correctly. */
17191 static void
17193 {
17199 /* Addition. */
17200 /* 1 <-- (0 + 1) */
17202 one);
17204 one);
17206 /* (nonlvalue)x <-- (x + 0) */
17208 x);
17210 /* Subtraction. */
17211 /* 0 <-- (x - x) */
17213 zero);
17215 x);
17217 /* Multiplication. */
17218 /* 0 <-- (x * 0) */
17220 zero);
17222 /* (nonlvalue)x <-- (x * 1) */
17224 x);
17225 }
17229 /* Build a VECTOR_CST corresponding to VMODE, and has
17230 encoding given by NPATTERNS, NELTS_PER_PATTERN and STEP.
17231 Fill it with randomized elements, using rand() % THRESHOLD. */
17233 static tree
17238 {
17243 // Fill a0 for each pattern
17250 // Fill a1 for each pattern
17252 {
17253 tree a1;
17255 {
17260 }
17261 else
17264 }
17269 {
17274 }
17277 }
17279 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17280 when result is VLA. */
17282 static void
17285 {
17286 /* Actual npatterns and encoded_elts in res may be less than expected due
17287 to canonicalization. */
17294 }
17296 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17297 when the result is VLS. */
17299 static void
17301 {
17305 }
17307 /* Helper routine to push multiple elements into BUILDER. */
17311 {
17314 }
17316 #define ARG0(index) vector_cst_elt (arg0, index)
17317 #define ARG1(index) vector_cst_elt (arg1, index)
17319 /* Test cases where result is VNx4SI and input vectors are V4SI. */
17321 static void
17323 {
17325 {
17326 /* Case 1:
17327 sel = { 0, 4, 1, 5, ... }
17328 res = { arg[0], arg1[0], arg0[1], arg1[1], ...} // (4, 1) */
17329 {
17333 tree inner_type
17347 }
17349 /* Case 2: Same as case 1, but contains an out of bounds access which
17350 should wrap around.
17351 sel = {0, 8, 4, 12, ...} (4, 1)
17352 res = { arg0[0], arg0[0], arg1[0], arg1[0], ... } (4, 1). */
17353 {
17357 tree inner_type
17371 }
17372 }
17373 }
17375 /* Test cases where result is V4SI and input vectors are VNx4SI. */
17377 static void
17379 {
17381 {
17382 /* Case 1:
17383 sel = { 0, 1, 2, 3}
17384 res = { arg0[0], arg0[1], arg0[2], arg0[3] }. */
17385 {
17389 tree inner_type
17403 }
17405 /* Case 2: Same as Case 1, but crossing input vector.
17406 sel = {0, 2, 4, 6}
17407 In this case,the index 4 is ambiguous since len = 4 + 4x.
17408 Since we cannot determine, which vector to choose from during
17409 compile time, should return NULL_TREE. */
17410 {
17414 tree inner_type
17429 }
17430 }
17431 }
17433 /* Test all input vectors. */
17435 static void
17437 {
17438 /* Test with 10 different inputs. */
17440 {
17445 /* Case 1: mask = {0, ...} // (1, 1)
17446 res = { arg0[0], ... } // (1, 1) */
17447 {
17454 }
17456 /* Case 2: mask = {len, ...} // (1, 1)
17457 res = { arg1[0], ... } // (1, 1) */
17458 {
17466 }
17467 }
17468 }
17470 /* Test all vectors which contain at-least 2 elements. */
17472 static void
17474 {
17476 {
17477 /* Case 1: mask = { 0, len, ... } // (2, 1)
17478 res = { arg0[0], arg1[0], ... } // (2, 1) */
17479 {
17493 }
17495 /* Case 2: mask = { 0, len, 1, len+1, ... } // (2, 2)
17496 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2) */
17497 {
17511 }
17513 /* Case 4: mask = {0, 0, 1, ...} // (1, 3)
17514 Test that the stepped sequence of the pattern selects from
17515 same input pattern. Since input vectors have npatterns = 2,
17516 and step (a2 - a1) = 1, step is not a multiple of npatterns
17517 in input vector. So return NULL_TREE. */
17518 {
17533 }
17535 /* Case 5: mask = {len, 0, 1, ...} // (1, 3)
17536 Test that stepped sequence of the pattern selects from arg0.
17537 res = { arg1[0], arg0[0], arg0[1], ... } // (1, 3) */
17538 {
17552 }
17554 /* Case 6: PR111648 - a1 chooses base element from input vector arg.
17555 In this case ensure that arg has a natural stepped sequence
17556 to preserve arg's encoding.
17558 As a concrete example, consider:
17559 arg0: { -16, -9, -10, ... } // (1, 3)
17560 arg1: { -12, -5, -6, ... } // (1, 3)
17561 sel = { 0, len, len + 1, ... } // (1, 3)
17563 This will create res with following encoding:
17564 res = { arg0[0], arg1[0], arg1[1], ... } // (1, 3)
17565 = { -16, -12, -5, ... }
17567 The step in above encoding would be: (-5) - (-12) = 7
17568 And hence res[3] would be computed as -5 + 7 = 2.
17569 instead of arg1[2], ie, -6.
17570 Ensure that valid_mask_for_fold_vec_perm_cst returns false
17571 for this case. */
17572 {
17583 /* FIXME: It may happen that build_vec_cst_rand may build a natural
17584 stepped pattern, even if we didn't explicitly tell it to. So folding
17585 may not always fail, but if it does, ensure that's because arg1 does
17586 not have a natural stepped sequence (and not due to other reason) */
17590 }
17592 /* Case 7: Same as Case 6, except that arg1 contains natural stepped
17593 sequence and thus folding should be valid for this case. */
17594 {
17608 }
17610 /* Case 8: Same as aarch64/sve/slp_3.c:
17611 arg0, arg1 are dup vectors.
17612 sel = { 0, len, 1, len+1, 2, len+2, ... } // (2, 3)
17613 So res = { arg0[0], arg1[0], ... } // (2, 1)
17615 In this case, since the input vectors are dup, only the first two
17616 elements per pattern in sel are considered significant. */
17617 {
17631 }
17632 }
17633 }
17635 /* Test all vectors which contain at-least 4 elements. */
17637 static void
17639 {
17641 {
17642 /* Case 1: mask = { 0, len, 1, len+1, ... } // (4, 1)
17643 res: { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (4, 1) */
17644 {
17658 }
17660 /* Case 2: sel = {0, 1, 2, ...} // (1, 3)
17661 res: { arg0[0], arg0[1], arg0[2], ... } // (1, 3) */
17662 {
17675 }
17677 /* Case 3: sel = {len, len+1, len+2, ...} // (1, 3)
17678 res: { arg1[0], arg1[1], arg1[2], ... } // (1, 3) */
17679 {
17692 }
17694 /* Case 4:
17695 sel = { len, 0, 2, ... } // (1, 3)
17696 This should return NULL because we cross the input vectors.
17697 Because,
17698 Let's assume len = C + Cx
17699 a1 = 0
17700 S = 2
17701 esel = arg0_len / sel_npatterns = C + Cx
17702 ae = 0 + (esel - 2) * S
17703 = 0 + (C + Cx - 2) * 2
17704 = 2(C-2) + 2Cx
17706 For C >= 4:
17707 Let q1 = a1 / arg0_len = 0 / (C + Cx) = 0
17708 Let qe = ae / arg0_len = (2(C-2) + 2Cx) / (C + Cx) = 1
17709 Since q1 != qe, we cross input vectors.
17710 So return NULL_TREE. */
17711 {
17725 }
17727 /* Case 5: npatterns(arg0) = 4 > npatterns(sel) = 2
17728 mask = { 0, len, 1, len + 1, ...} // (2, 2)
17729 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2)
17731 Note that fold_vec_perm_cst will set
17732 res_npatterns = max(4, max(4, 2)) = 4
17733 However after canonicalizing, we will end up with shape (2, 2). */
17734 {
17747 }
17749 /* Case 6: Test combination in sel, where one pattern is dup and other
17750 is stepped sequence.
17751 sel = { 0, 0, 0, 1, 0, 2, ... } // (2, 3)
17752 res = { arg0[0], arg0[0], arg0[0],
17753 arg0[1], arg0[0], arg0[2], ... } // (2, 3) */
17754 {
17769 }
17771 /* Case 7: PR111048: Check that we set arg_npatterns correctly,
17772 when arg0, arg1 and sel have different number of patterns.
17773 arg0 is of shape (1, 1)
17774 arg1 is of shape (4, 1)
17775 sel is of shape (2, 3) = {1, len, 2, len+1, 3, len+2, ...}
17777 In this case the pattern: {len, len+1, len+2, ...} chooses arg1.
17778 However,
17779 step = (len+2) - (len+1) = 1
17780 arg_npatterns = VECTOR_CST_NPATTERNS (arg1) = 4
17781 Since step is not a multiple of arg_npatterns,
17782 valid_mask_for_fold_vec_perm_cst should return false,
17783 and thus fold_vec_perm_cst should return NULL_TREE. */
17784 {
17799 }
17801 /* Case 8: PR111754: When input vector is not a stepped sequence,
17802 check that the result is not a stepped sequence either, even
17803 if sel has a stepped sequence. */
17804 {
17817 }
17819 /* Case 9: If sel doesn't contain a stepped sequence,
17820 check that the result has same encoding as sel, irrespective
17821 of shape of input vectors. */
17822 {
17837 }
17838 }
17839 }
17841 /* Test all vectors which contain at-least 8 elements. */
17843 static void
17845 {
17847 {
17848 /* Case 1: sel_npatterns (4) > input npatterns (2)
17849 sel: { 0, 0, 1, len, 2, 0, 3, len, 4, 0, 5, len, ...} // (4, 3)
17850 res: { arg0[0], arg0[0], arg0[0], arg1[0],
17851 arg0[2], arg0[0], arg0[3], arg1[0],
17852 arg0[4], arg0[0], arg0[5], arg1[0], ... } // (4, 3) */
17853 {
17870 }
17871 }
17872 }
17874 /* Test vectors for which nunits[0] <= 4. */
17876 static void
17878 {
17879 /* Case 1: mask = {0, 4, ...} // (1, 2)
17880 This should return NULL_TREE because the index 4 may choose
17881 from either arg0 or arg1 depending on vector length. */
17882 {
17897 }
17898 }
17900 #undef ARG0
17901 #undef ARG1
17903 /* Return true if SIZE is of the form C + Cx and C is power of 2. */
17905 static bool
17907 {
17915 }
17917 /* Execute fold_vec_perm_cst unit tests. */
17919 static void
17921 {
17925 machine_mode vmode;
17927 {
17928 /* Obtain modes corresponding to VNx4SI and V4SI,
17929 to call mixed mode tests below.
17930 FIXME: Is there a better way to do this ? */
17932 {
17939 }
17956 }
17961 {
17964 }
17965 }
17968 /* Verify that various binary operations on vectors are folded
17969 correctly. */
17971 static void
17973 {
17980 /* Verify equality tests that return a scalar boolean result. */
17993 }
17995 /* Verify folding of VEC_DUPLICATE_EXPRs. */
17997 static void
17999 {
18002 /* This will be 1 if VEC_MODE isn't a vector mode. */
18009 }
18011 /* Run all of the selftests within this file. */
18013 void
18015 {
18020 }