| blob | 40767736389d9c1846d315290508b5d524ea9eb2 |
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 /* In GIMPLE empty constructors are allowed in initializers of
3319 aggregates. */
3323 }
3325 /* Don't handle more cases for OEP_BITWISE, since we can't guarantee that
3326 two instances of undefined behavior will give identical results. */
3330 /* Define macros to test an operand from arg0 and arg1 for equality and a
3331 variant that allows null and views null as being different from any
3332 non-null value. In the latter case, if either is null, the both
3333 must be; otherwise, do the normal comparison. */
3334 #define OP_SAME(N) operand_equal_p (TREE_OPERAND (arg0, N), \
3335 TREE_OPERAND (arg1, N), flags)
3337 #define OP_SAME_WITH_NULL(N) \
3338 ((!TREE_OPERAND (arg0, N) || !TREE_OPERAND (arg1, N)) \
3339 ? TREE_OPERAND (arg0, N) == TREE_OPERAND (arg1, N) : OP_SAME (N))
3342 {
3344 /* Two conversions are equal only if signedness and modes match. */
3346 {
3347 CASE_CONVERT:
3355 }
3365 /* For commutative ops, allow the other order. */
3373 /* If either of the pointer (or reference) expressions we are
3374 dereferencing contain a side effect, these cannot be equal,
3375 but their addresses can be. */
3382 {
3385 {
3389 /* Verify that the access types are compatible. */
3393 }
3398 /* Require the same offset. */
3404 /* Fallthru. */
3412 {
3413 /* Require equal access sizes */
3419 flags)))
3421 /* Verify that access happens in similar types. */
3424 /* Verify that accesses are TBAA compatible. */
3433 /* Verify that alignment is compatible. */
3437 }
3440 /* TARGET_MEM_REF require equal extra operands. */
3451 /* Compare the array index by value if it is constant first as we
3452 may have different types but same value here. */
3458 /* Compare low bound and element size as with OEP_ADDRESS_OF
3459 we have to account for the offset of the ref. */
3464 array_ref_low_bound
3468 array_ref_element_size
3470 flags))));
3473 /* Handle operand 2 the same as for ARRAY_REF. Operand 0
3474 may be NULL when we're called to compare MEM_EXPRs. */
3477 {
3480 /* Most of time we only need to compare FIELD_DECLs for equality.
3481 However when determining address look into actual offsets.
3482 These may match for unions and unshared record types. */
3485 {
3488 {
3492 /* Non-FIELD_DECL operands can appear in C++ templates. */
3499 flags))
3501 }
3502 else
3504 }
3505 }
3516 }
3520 {
3522 /* Be sure we pass right ADDRESS_OF flag. */
3539 /* The multiplcation operands are commutative. */
3540 /* FALLTHRU */
3548 /* Otherwise take into account this is a commutative operation. */
3561 /* BIT_INSERT_EXPR has an implict operand as the type precision
3562 of op1. Need to check to make sure they are the same. */
3568 /* FALLTHRU */
3593 /* Virtual table reference. */
3605 {
3610 }
3615 }
3619 {
3623 /* If not both CALL_EXPRs are either internal or normal function
3624 functions, then they are not equal. */
3627 {
3628 /* If the CALL_EXPRs call different internal functions, then they
3629 are not equal. */
3632 }
3633 else
3634 {
3635 /* If the CALL_EXPRs call different functions, then they are not
3636 equal. */
3638 flags))
3640 }
3642 /* FIXME: We could skip this test for OEP_MATCH_SIDE_EFFECTS. */
3643 {
3647 else
3651 }
3653 /* Now see if all the arguments are the same. */
3654 {
3665 /* If we get here and both argument lists are exhausted
3666 then the CALL_EXPRs are equal. */
3668 }
3671 }
3674 /* Consider __builtin_sqrt equal to sqrt. */
3684 {
3685 /* Consider decls with the same name equal. The caller needs
3686 to make sure they refer to the same entity (such as a function
3687 formal parameter). */
3693 }
3698 {
3702 /* In GIMPLE constructors are used only to build vectors from
3703 elements. Individual elements in the constructor must be
3704 indexed in increasing order and form an initial sequence.
3706 We make no effort to compare constructors in generic.
3707 (see sem_variable::equals in ipa-icf which can do so for
3708 constants). */
3713 /* Be sure that vectors constructed have the same representation.
3714 We only tested element precision and modes to match.
3715 Vectors may be BLKmode and thus also check that the number of
3716 parts match. */
3729 {
3734 /* In GIMPLE the indexes can be either NULL or matching i.
3735 Double check this so we won't get false
3736 positives for GENERIC. */
3744 }
3746 }
3749 {
3750 /* Compare the STATEMENT_LISTs. */
3756 {
3757 /* The lists don't have the same number of statements. */
3763 flags & (OEP_LEXICOGRAPHIC
3766 }
3767 }
3772 {
3783 }
3787 }
3789 #undef OP_SAME
3790 #undef OP_SAME_WITH_NULL
3791 }
3793 /* Generate a hash value for an expression. This can be used iteratively
3794 by passing a previous result as the HSTATE argument. */
3796 void
3799 {
3805 {
3808 }
3818 {
3819 /* Alas, constants aren't shared, so we can't rely on pointer
3820 identity. */
3830 {
3834 else
3838 }
3840 {
3844 }
3854 {
3861 }
3863 /* We can just compare by pointer. */
3867 /* The node itself doesn't matter. */
3871 /* Ignore. */
3874 /* A list of expressions, for a CALL_EXPR or as the elements of a
3875 VECTOR_CST. */
3880 {
3886 {
3887 /* In GIMPLE the indexes can be either NULL or matching i. */
3892 }
3894 }
3896 {
3897 tree_stmt_iterator i;
3902 }
3911 /* When referring to a built-in FUNCTION_DECL, use the __builtin__ form.
3912 Otherwise nodes that compare equal according to operand_equal_p might
3913 get different hash codes. However, don't do this for machine specific
3914 or front end builtins, since the function code is overloaded in those
3915 cases. */
3918 {
3921 }
3922 /* FALL THROUGH */
3925 {
3929 }
3933 {
3934 /* DECL's have a unique ID */
3936 }
3938 {
3939 /* For comparisons that can be swapped, use the lower
3940 tree code. */
3947 }
3949 {
3950 /* NOP_EXPR and CONVERT_EXPR are considered equal by
3951 operand_equal_p. */
3955 /* Don't hash the type, that can lead to having nodes which
3956 compare equal according to operand_equal_p, but which
3957 have different hash codes. Make sure to include signedness
3958 in the hash computation. */
3961 }
3962 /* For OEP_ADDRESS_OF, hash MEM_EXPR[&decl, 0] the same as decl. */
3970 /* Don't ICE on FE specific trees, or their arguments etc.
3971 during operand_equal_p hash verification. */
3974 else
3975 {
3981 {
3997 {
4004 }
4018 {
4019 /* The multiplication operands are commutative. */
4026 }
4034 /* For TARGET_EXPR, just hash on the TARGET_EXPR_SLOT.
4035 Usually different TARGET_EXPRs just should use
4036 different temporaries in their slots. */
4041 /* Virtual table reference. */
4049 {
4051 /* We compute mangled names only when free_lang_data is run.
4052 In that case we can hash precisely. */
4055 hstate.add_object
4056 (IDENTIFIER_HASH_VALUE
4058 }
4062 }
4064 /* Don't hash the type, that can lead to having nodes which
4065 compare equal according to operand_equal_p, but which
4066 have different hash codes. */
4068 {
4069 /* Make sure to include signness in the hash computation. */
4072 }
4075 {
4076 /* It's a commutative expression. We want to hash it the same
4077 however it appears. We do this by first hashing both operands
4078 and then rehashing based on the order of their independent
4079 hashes. */
4084 }
4085 else
4089 }
4091 }
4092 }
4094 bool
4097 {
4098 /* When checking and unless comparing DECL names, verify that if
4099 the outermost operand_equal_p call returns non-zero then ARG0
4100 and ARG1 have the same hash value. */
4102 {
4104 {
4106 {
4113 }
4115 }
4116 else
4120 }
4123 }
4128 /* Conveinece wrapper around operand_compare class because usually we do
4129 not need to play with the valueizer. */
4131 bool
4133 {
4135 }
4137 namespace inchash
4138 {
4140 /* Generate a hash value for an expression. This can be used iteratively
4141 by passing a previous result as the HSTATE argument.
4143 This function is intended to produce the same hash for expressions which
4144 would compare equal using operand_equal_p. */
4145 void
4147 {
4149 }
4151 }
4152 \f
4153 /* Similar to operand_equal_p, but see if ARG0 might be a variant of ARG1
4154 with a different signedness or a narrower precision. */
4156 static bool
4158 {
4166 /* Discard any conversions that don't change the modes of ARG0 and ARG1
4167 and see if the inner values are the same. This removes any
4168 signedness comparison, which doesn't matter here. */
4176 /* Discard a single widening conversion from ARG1 and see if the inner
4177 value is the same as ARG0. */
4186 }
4187 \f
4188 /* See if ARG is an expression that is either a comparison or is performing
4189 arithmetic on comparisons. The comparisons must only be comparing
4190 two different values, which will be stored in *CVAL1 and *CVAL2; if
4191 they are nonzero it means that some operands have already been found.
4192 No variables may be used anywhere else in the expression except in the
4193 comparisons.
4195 If this is true, return 1. Otherwise, return zero. */
4197 static bool
4199 {
4203 /* We can handle some of the tcc_expression cases here. */
4212 {
4231 /* First see if we can handle the first operand, then the second. For
4232 the second operand, we know *CVAL1 can't be zero. It must be that
4233 one side of the comparison is each of the values; test for the
4234 case where this isn't true by failing if the two operands
4235 are the same. */
4244 ;
4248 ;
4249 else
4253 ;
4257 ;
4258 else
4265 }
4266 }
4267 \f
4268 /* ARG is a tree that is known to contain just arithmetic operations and
4269 comparisons. Evaluate the operations in the tree substituting NEW0 for
4270 any occurrence of OLD0 as an operand of a comparison and likewise for
4271 NEW1 and OLD1. */
4273 static tree
4276 {
4281 /* We can handle some of the tcc_expression cases here. */
4289 {
4304 {
4323 }
4324 /* Fall through - ??? */
4327 {
4331 /* We need to check both for exact equality and tree equality. The
4332 former will be true if the operand has a side-effect. In that
4333 case, we know the operand occurred exactly once. */
4346 }
4350 }
4351 }
4352 \f
4353 /* Return a tree for the case when the result of an expression is RESULT
4354 converted to TYPE and OMITTED was previously an operand of the expression
4355 but is now not needed (e.g., we folded OMITTED * 0).
4357 If OMITTED has side effects, we must evaluate it. Otherwise, just do
4358 the conversion of RESULT to TYPE. */
4360 tree
4362 {
4365 /* If the resulting operand is an empty statement, just return the omitted
4366 statement casted to void. */
4376 }
4378 /* Return a tree for the case when the result of an expression is RESULT
4379 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
4380 of the expression but are now not needed.
4382 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
4383 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
4384 evaluated before OMITTED2. Otherwise, if neither has side effects,
4385 just do the conversion of RESULT to TYPE. */
4387 tree
4390 {
4399 }
4401 \f
4402 /* Return a simplified tree node for the truth-negation of ARG. This
4403 never alters ARG itself. We assume that ARG is an operation that
4404 returns a truth value (0 or 1).
4406 FIXME: one would think we would fold the result, but it causes
4407 problems with the dominator optimizer. */
4409 static tree
4411 {
4416 /* If this is a comparison, we can simply invert it, except for
4417 floating-point non-equality comparisons, in which case we just
4418 enclose a TRUTH_NOT_EXPR around what we have. */
4421 {
4424 && flag_trapping_math
4437 }
4440 {
4459 /* Here we can invert either operand. We invert the first operand
4460 unless the second operand is a TRUTH_NOT_EXPR in which case our
4461 result is the XOR of the first operand with the inside of the
4462 negation of the second operand. */
4467 else
4490 {
4497 /* A COND_EXPR may have a throw as one operand, which
4498 then has void type. Just leave void operands
4499 as they are. */
4505 }
4517 CASE_CONVERT:
4521 /* fall through */
4543 }
4544 }
4546 /* Fold the truth-negation of ARG. This never alters ARG itself. We
4547 assume that ARG is an operation that returns a truth value (0 or 1
4548 for scalars, 0 or -1 for vectors). Return the folded expression if
4549 folding is successful. Otherwise, return NULL_TREE. */
4551 static tree
4553 {
4556 ? BIT_NOT_EXPR
4559 }
4561 /* Return a simplified tree node for the truth-negation of ARG. This
4562 never alters ARG itself. We assume that ARG is an operation that
4563 returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). */
4565 tree
4567 {
4573 ? BIT_NOT_EXPR
4576 }
4577 \f
4578 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
4579 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero
4580 and uses reverse storage order if REVERSEP is nonzero. ORIG_INNER
4581 is the original memory reference used to preserve the alias set of
4582 the access. */
4584 static tree
4588 {
4591 /* Attempt not to lose the access path if possible. */
4593 {
4595 machine_mode nmode;
4597 tree noffset;
4605 && !reversep
4606 && !nreversep
4608 {
4611 }
4612 }
4621 {
4628 }
4643 }
4645 /* Optimize a bit-field compare.
4647 There are two cases: First is a compare against a constant and the
4648 second is a comparison of two items where the fields are at the same
4649 bit position relative to the start of a chunk (byte, halfword, word)
4650 large enough to contain it. In these cases we can avoid the shift
4651 implicit in bitfield extractions.
4653 For constants, we emit a compare of the shifted constant with the
4654 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
4655 compared. For two fields at the same position, we do the ANDs with the
4656 similar mask and compare the result of the ANDs.
4658 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
4659 COMPARE_TYPE is the type of the comparison, and LHS and RHS
4660 are the left and right operands of the comparison, respectively.
4662 If the optimization described above can be done, we return the resulting
4663 tree. Otherwise we return zero. */
4665 static tree
4668 {
4672 tree unsigned_type;
4675 scalar_int_mode nmode;
4680 tree mask;
4681 tree offset;
4683 /* Get all the information about the extractions being done. If the bit size
4684 is the same as the size of the underlying object, we aren't doing an
4685 extraction at all and so can do nothing. We also don't want to
4686 do anything if the inner expression is a PLACEHOLDER_EXPR since we
4687 then will no longer be able to replace it. */
4702 else
4703 {
4704 /* If this is not a constant, we can only do something if bit positions,
4705 sizes, signedness and storage order are the same. */
4706 rinner
4719 }
4721 /* Honor the C++ memory model and mimic what RTL expansion does. */
4725 {
4729 }
4731 /* See if we can find a mode to refer to this field. We should be able to,
4732 but fail if we can't. */
4740 /* Set signed and unsigned types of the precision of this mode for the
4741 shifts below. */
4744 /* Compute the bit position and size for the new reference and our offset
4745 within it. If the new reference is the same size as the original, we
4746 won't optimize anything, so return zero. */
4756 /* Make the mask to be used against the extracted field. */
4763 {
4767 /* If not comparing with constant, just rework the comparison
4768 and return. */
4776 }
4778 /* Otherwise, we are handling the constant case. See if the constant is too
4779 big for the field. Warn and return a tree for 0 (false) if so. We do
4780 this not only for its own sake, but to avoid having to test for this
4781 error case below. If we didn't, we might generate wrong code.
4783 For unsigned fields, the constant shifted right by the field length should
4784 be all zero. For signed fields, the high-order bits should agree with
4785 the sign bit. */
4788 {
4790 {
4794 }
4795 }
4796 else
4797 {
4800 {
4804 }
4805 }
4810 /* Single-bit compares should always be against zero. */
4812 {
4815 }
4817 /* Make a new bitfield reference, shift the constant over the
4818 appropriate number of bits and mask it with the computed mask
4819 (in case this was a signed field). If we changed it, make a new one. */
4827 mask);
4832 }
4833 \f
4834 /* Subroutine for fold_truth_andor_1: decode a field reference.
4836 If EXP is a comparison reference, we return the innermost reference.
4838 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
4839 set to the starting bit number.
4841 If the innermost field can be completely contained in a mode-sized
4842 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
4844 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
4845 otherwise it is not changed.
4847 *PUNSIGNEDP is set to the signedness of the field.
4849 *PREVERSEP is set to the storage order of the field.
4851 *PMASK is set to the mask used. This is either contained in a
4852 BIT_AND_EXPR or derived from the width of the field.
4854 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
4856 Return 0 if this is not a component reference or is one that we can't
4857 do anything with. */
4859 static tree
4864 {
4869 tree unsigned_type;
4872 /* All the optimizations using this function assume integer fields.
4873 There are problems with FP fields since the type_for_size call
4874 below can fail for, e.g., XFmode. */
4878 /* We are interested in the bare arrangement of bits, so strip everything
4879 that doesn't affect the machine mode. However, record the type of the
4880 outermost expression if it may matter below. */
4887 {
4893 }
4904 /* Reject out-of-bound accesses (PR79731). */
4916 /* If the number of bits in the reference is the same as the bitsize of
4917 the outer type, then the outer type gives the signedness. Otherwise
4918 (in case of a small bitfield) the signedness is unchanged. */
4922 /* Compute the mask to access the bitfield. */
4930 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
4938 }
4940 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
4941 bit positions and MASK is SIGNED. */
4943 static bool
4945 {
4949 /* If this function returns true when the type of the mask is
4950 UNSIGNED, then there will be errors. In particular see
4951 gcc.c-torture/execute/990326-1.c. There does not appear to be
4952 any documentation paper trail as to why this is so. But the pre
4953 wide-int worked with that restriction and it has been preserved
4954 here. */
4959 }
4961 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
4962 represents the sign bit of EXP's type. If EXP represents a sign
4963 or zero extension, also test VAL against the unextended type.
4964 The return value is the (sub)expression whose sign bit is VAL,
4965 or NULL_TREE otherwise. */
4967 tree
4969 {
4971 tree t;
4973 /* Tree EXP must have an integral type. */
4978 /* Tree VAL must be an integer constant. */
4987 /* Handle extension from a narrower type. */
4993 }
4995 /* Subroutine for fold_truth_andor_1 and simple_condition_p: determine if an
4996 operand is simple enough to be evaluated unconditionally. */
4998 static bool
5000 {
5001 /* Strip any conversions that don't change the machine mode. */
5010 /* Don't regard global variables as simple. They may be
5011 allocated in ways unknown to the compiler (shared memory,
5012 #pragma weak, etc). */
5015 /* Weakrefs are not safe to be read, since they can be NULL.
5016 They are !TREE_PUBLIC && !DECL_EXTERNAL but still
5017 have DECL_WEAK flag set. */
5019 /* Loading a static variable is unduly expensive, but global
5020 registers aren't expensive. */
5022 }
5024 /* Determine if an operand is simple enough to be evaluated unconditionally.
5025 In addition to simple_operand_p, we assume that comparisons, conversions,
5026 and logic-not operations are simple, if their operands are simple, too. */
5028 bool
5030 {
5049 }
5051 \f
5052 /* The following functions are subroutines to fold_range_test and allow it to
5053 try to change a logical combination of comparisons into a range test.
5055 For example, both
5056 X == 2 || X == 3 || X == 4 || X == 5
5057 and
5058 X >= 2 && X <= 5
5059 are converted to
5060 (unsigned) (X - 2) <= 3
5062 We describe each set of comparisons as being either inside or outside
5063 a range, using a variable named like IN_P, and then describe the
5064 range with a lower and upper bound. If one of the bounds is omitted,
5065 it represents either the highest or lowest value of the type.
5067 In the comments below, we represent a range by two numbers in brackets
5068 preceded by a "+" to designate being inside that range, or a "-" to
5069 designate being outside that range, so the condition can be inverted by
5070 flipping the prefix. An omitted bound is represented by a "-". For
5071 example, "- [-, 10]" means being outside the range starting at the lowest
5072 possible value and ending at 10, in other words, being greater than 10.
5073 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
5074 always false.
5076 We set up things so that the missing bounds are handled in a consistent
5077 manner so neither a missing bound nor "true" and "false" need to be
5078 handled using a special case. */
5080 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
5081 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
5082 and UPPER1_P are nonzero if the respective argument is an upper bound
5083 and zero for a lower. TYPE, if nonzero, is the type of the result; it
5084 must be specified for a comparison. ARG1 will be converted to ARG0's
5085 type if both are specified. */
5087 static tree
5090 {
5091 tree tem;
5095 /* If neither arg represents infinity, do the normal operation.
5096 Else, if not a comparison, return infinity. Else handle the special
5097 comparison rules. Note that most of the cases below won't occur, but
5098 are handled for consistency. */
5101 {
5106 }
5111 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
5112 for neither. In real maths, we cannot assume open ended ranges are
5113 the same. But, this is computer arithmetic, where numbers are finite.
5114 We can therefore make the transformation of any unbounded range with
5115 the value Z, Z being greater than any representable number. This permits
5116 us to treat unbounded ranges as equal. */
5120 {
5141 }
5144 }
5145 \f
5146 /* Helper routine for make_range. Perform one step for it, return
5147 new expression if the loop should continue or NULL_TREE if it should
5148 stop. */
5150 tree
5154 {
5160 {
5162 /* We can only do something if the range is testing for zero. */
5171 /* We can only do something if the range is testing for zero
5172 and if the second operand is an integer constant. Note that
5173 saying something is "in" the range we make is done by
5174 complementing IN_P since it will set in the initial case of
5175 being not equal to zero; "out" is leaving it alone. */
5182 {
5203 }
5205 /* If this is an unsigned comparison, we also know that EXP is
5206 greater than or equal to zero. We base the range tests we make
5207 on that fact, so we record it here so we can parse existing
5208 range tests. We test arg0_type since often the return type
5209 of, e.g. EQ_EXPR, is boolean. */
5211 {
5215 NULL_TREE))
5220 /* If the high bound is missing, but we have a nonzero low
5221 bound, reverse the range so it goes from zero to the low bound
5222 minus 1. */
5224 {
5229 }
5230 }
5238 /* If flag_wrapv and ARG0_TYPE is signed, make sure
5239 low and high are non-NULL, then normalize will DTRT. */
5242 {
5247 }
5249 /* (-x) IN [a,b] -> x in [-b, -a] */
5261 /* ~ X -> -X - 1 */
5270 /* If flag_wrapv and ARG0_TYPE is signed, then we cannot
5271 move a constant to the other side. */
5276 /* If EXP is signed, any overflow in the computation is undefined,
5277 so we don't worry about it so long as our computations on
5278 the bounds don't overflow. For unsigned, overflow is defined
5279 and this is exactly the right thing. */
5291 normalize:
5292 /* Check for an unsigned range which has wrapped around the maximum
5293 value thus making n_high < n_low, and normalize it. */
5295 {
5301 /* If the range is of the form +/- [ x+1, x ], we won't
5302 be able to normalize it. But then, it represents the
5303 whole range or the empty set, so make it
5304 +/- [ -, - ]. */
5308 else
5310 }
5311 else
5319 CASE_CONVERT:
5337 /* If we're converting arg0 from an unsigned type, to exp,
5338 a signed type, we will be doing the comparison as unsigned.
5339 The tests above have already verified that LOW and HIGH
5340 are both positive.
5342 So we have to ensure that we will handle large unsigned
5343 values the same way that the current signed bounds treat
5344 negative values. */
5347 {
5348 tree high_positive;
5349 tree equiv_type;
5350 /* For fixed-point modes, we need to pass the saturating flag
5351 as the 2nd parameter. */
5353 equiv_type
5358 else
5359 equiv_type
5362 /* A range without an upper bound is, naturally, unbounded.
5363 Since convert would have cropped a very large value, use
5364 the max value for the destination type. */
5365 high_positive
5372 high_positive),
5375 /* If the low bound is specified, "and" the range with the
5376 range for which the original unsigned value will be
5377 positive. */
5379 {
5382 integer_zero_node),
5383 high_positive))
5387 }
5388 else
5389 {
5390 /* Otherwise, "or" the range with the range of the input
5391 that will be interpreted as negative. */
5394 integer_zero_node),
5395 high_positive))
5399 }
5400 }
5402 /* Otherwise, if we are converting arg0 from signed type, to exp,
5403 an unsigned type, we will do the comparison as signed. If
5404 high is non-NULL, we punt above if it doesn't fit in the signed
5405 type, so if we get through here, +[-, high] or +[low, high] are
5406 equivalent to +[-, n_high] or +[n_low, n_high]. Similarly,
5407 +[-, -] or -[-, -] are equivalent too. But if low is specified and
5408 high is not, the +[low, -] range is equivalent to union of
5409 +[n_low, -] and +[-, -1] ranges, so +[low, -] is equivalent to
5410 -[0, n_low-1] and similarly -[low, -] to +[0, n_low-1], except for
5411 low being 0, which should be treated as [-, -]. */
5414 && low
5416 {
5419 else
5420 {
5425 }
5426 }
5435 }
5436 }
5438 /* Given EXP, a logical expression, set the range it is testing into
5439 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
5440 actually being tested. *PLOW and *PHIGH will be made of the same
5441 type as the returned expression. If EXP is not a comparison, we
5442 will most likely not be returning a useful value and range. Set
5443 *STRICT_OVERFLOW_P to true if the return value is only valid
5444 because signed overflow is undefined; otherwise, do not change
5445 *STRICT_OVERFLOW_P. */
5447 tree
5450 {
5458 /* Start with simply saying "EXP != 0" and then look at the code of EXP
5459 and see if we can refine the range. Some of the cases below may not
5460 happen, but it doesn't seem worth worrying about this. We "continue"
5461 the outer loop when we've changed something; otherwise we "break"
5462 the switch, which will "break" the while. */
5468 {
5474 {
5482 }
5491 }
5493 /* If EXP is a constant, we can evaluate whether this is true or false. */
5495 {
5502 }
5506 }
5508 /* Returns TRUE if [LOW, HIGH] range check can be optimized to
5509 a bitwise check i.e. when
5510 LOW == 0xXX...X00...0
5511 HIGH == 0xXX...X11...1
5512 Return corresponding mask in MASK and stem in VALUE. */
5514 static bool
5517 {
5540 }
5541 \f
5542 /* Helper routine for build_range_check and match.pd. Return the type to
5543 perform the check or NULL if it shouldn't be optimized. */
5545 tree
5547 {
5548 /* First make sure that arithmetics in this type is valid, then make sure
5549 that it wraps around. */
5554 {
5557 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
5558 for the type in question, as we rely on this here. */
5568 else
5570 }
5573 /* Right now all BITINT_TYPEs satisfy
5574 (unsigned) max + 1 == (unsigned) min, so no need to verify
5575 that like for INTEGER_TYPEs. */
5579 }
5581 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
5582 type, TYPE, return an expression to test if EXP is in (or out of, depending
5583 on IN_P) the range. Return 0 if the test couldn't be created. */
5585 tree
5588 {
5591 /* Disable this optimization for function pointer expressions
5592 on targets that require function pointer canonicalization. */
5599 {
5605 }
5627 value);
5630 {
5632 {
5636 }
5638 }
5640 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
5642 {
5646 {
5648 {
5651 etype
5653 else
5656 }
5659 }
5660 }
5662 /* Optimize (c>=low) && (c<=high) into (c-low>=0) && (c-low<=high-low).
5663 This requires wrap-around arithmetics for the type of the expression. */
5680 }
5681 \f
5682 /* Return the predecessor of VAL in its type, handling the infinite case. */
5684 static tree
5686 {
5692 else
5695 }
5697 /* Return the successor of VAL in its type, handling the infinite case. */
5699 static tree
5701 {
5707 else
5710 }
5712 /* Given two ranges, see if we can merge them into one. Return 1 if we
5713 can, 0 if we can't. Set the output range into the specified parameters. */
5715 bool
5718 {
5722 tree tem;
5732 /* Make range 0 be the range that starts first, or ends last if they
5733 start at the same value. Swap them if it isn't. */
5736 || (lowequal
5739 {
5743 }
5745 /* If the second range is != high1 where high1 is the type maximum of
5746 the type, try first merging with < high1 range. */
5748 && high1
5755 {
5760 /* Similarly for the second range != low1 where low1 is the type minimum
5761 of the type, try first merging with > low1 range. */
5766 }
5768 /* Now flag two cases, whether the ranges are disjoint or whether the
5769 second range is totally subsumed in the first. Note that the tests
5770 below are simplified by the ones above. */
5776 /* We now have four cases, depending on whether we are including or
5777 excluding the two ranges. */
5779 {
5780 /* If they don't overlap, the result is false. If the second range
5781 is a subset it is the result. Otherwise, the range is from the start
5782 of the second to the end of the first. */
5787 else
5789 }
5792 {
5793 /* If they don't overlap, the result is the first range. If they are
5794 equal, the result is false. If the second range is a subset of the
5795 first, and the ranges begin at the same place, we go from just after
5796 the end of the second range to the end of the first. If the second
5797 range is not a subset of the first, or if it is a subset and both
5798 ranges end at the same place, the range starts at the start of the
5799 first range and ends just before the second range.
5800 Otherwise, we can't describe this as a single range. */
5806 {
5811 {
5812 /* We are in the weird situation where high0 > high1 but
5813 high1 has no successor. Punt. */
5815 }
5816 }
5818 {
5823 {
5824 /* low0 < low1 but low1 has no predecessor. Punt. */
5826 }
5827 }
5828 else
5830 }
5833 {
5834 /* If they don't overlap, the result is the second range. If the second
5835 is a subset of the first, the result is false. Otherwise,
5836 the range starts just after the first range and ends at the
5837 end of the second. */
5842 else
5843 {
5848 {
5849 /* high1 > high0 but high0 has no successor. Punt. */
5851 }
5852 }
5853 }
5855 else
5856 {
5857 /* The case where we are excluding both ranges. Here the complex case
5858 is if they don't overlap. In that case, the only time we have a
5859 range is if they are adjacent. If the second is a subset of the
5860 first, the result is the first. Otherwise, the range to exclude
5861 starts at the beginning of the first range and ends at the end of the
5862 second. */
5864 {
5869 else
5870 {
5871 /* Canonicalize - [min, x] into - [-, x]. */
5874 {
5877 GET_MODE_BITSIZE
5880 /* FALLTHROUGH */
5893 }
5895 /* Canonicalize - [x, max] into - [x, -]. */
5898 {
5901 GET_MODE_BITSIZE
5904 /* FALLTHROUGH */
5920 }
5922 /* The ranges might be also adjacent between the maximum and
5923 minimum values of the given type. For
5924 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
5925 return + [x + 1, y - 1]. */
5927 {
5934 }
5935 else
5937 }
5938 }
5941 else
5943 }
5947 }
5948 \f
5950 /* Subroutine of fold, looking inside expressions of the form
5951 A op B ? A : C, where (ARG00, COMP_CODE, ARG01), ARG1 and ARG2
5952 are the three operands of the COND_EXPR. This function is
5953 being used also to optimize A op B ? C : A, by reversing the
5954 comparison first.
5956 Return a folded expression whose code is not a COND_EXPR
5957 anymore, or NULL_TREE if no folding opportunity is found. */
5959 static tree
5963 {
5965 tree tem;
5970 /* If we have A op 0 ? A : -A, consider applying the following
5971 transformations:
5973 A == 0? A : -A same as -A
5974 A != 0? A : -A same as A
5975 A >= 0? A : -A same as abs (A)
5976 A > 0? A : -A same as abs (A)
5977 A <= 0? A : -A same as -abs (A)
5978 A < 0? A : -A same as -abs (A)
5980 None of these transformations work for modes with signed
5981 zeros. If A is +/-0, the first two transformations will
5982 change the sign of the result (from +0 to -0, or vice
5983 versa). The last four will fix the sign of the result,
5984 even though the original expressions could be positive or
5985 negative, depending on the sign of A.
5987 Note that all these transformations are correct if A is
5988 NaN, since the two alternatives (A and -A) are also NaNs. */
5995 /* In the case that A is of the form X-Y, '-A' (arg2) may
5996 have already been folded to Y-X, check for that. */
6004 {
6016 /* Fall through. */
6027 /* FALLTHRU */
6034 {
6035 /* A <= 0 ? A : -A for A INT_MIN is valid, but -abs(INT_MIN)
6036 is not, invokes UB both in abs and in the negation of it.
6037 So, use ABSU_EXPR instead. */
6042 }
6043 else
6044 {
6047 }
6051 }
6053 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
6054 A == 0 ? A : 0 is always 0 unless A is -0. Note that
6055 both transformations are correct when A is NaN: A != 0
6056 is then true, and A == 0 is false. */
6060 {
6065 }
6067 /* Try some transformations of A op B ? A : B.
6069 A == B? A : B same as B
6070 A != B? A : B same as A
6071 A >= B? A : B same as max (A, B)
6072 A > B? A : B same as max (B, A)
6073 A <= B? A : B same as min (A, B)
6074 A < B? A : B same as min (B, A)
6076 As above, these transformations don't work in the presence
6077 of signed zeros. For example, if A and B are zeros of
6078 opposite sign, the first two transformations will change
6079 the sign of the result. In the last four, the original
6080 expressions give different results for (A=+0, B=-0) and
6081 (A=-0, B=+0), but the transformed expressions do not.
6083 The first two transformations are correct if either A or B
6084 is a NaN. In the first transformation, the condition will
6085 be false, and B will indeed be chosen. In the case of the
6086 second transformation, the condition A != B will be true,
6087 and A will be chosen.
6089 The conversions to max() and min() are not correct if B is
6090 a number and A is not. The conditions in the original
6091 expressions will be false, so all four give B. The min()
6092 and max() versions would give a NaN instead. */
6095 /* Avoid these transformations if the COND_EXPR may be used
6096 as an lvalue in the C++ front-end. PR c++/19199. */
6097 && (in_gimple_form
6103 {
6109 {
6118 /* In C++ a ?: expression can be an lvalue, so put the
6119 operand which will be used if they are equal first
6120 so that we can convert this back to the
6121 corresponding COND_EXPR. */
6123 {
6131 }
6138 {
6146 }
6159 }
6160 }
6163 }
6166 \f
6167 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
6168 #define LOGICAL_OP_NON_SHORT_CIRCUIT \
6169 (BRANCH_COST (optimize_function_for_speed_p (cfun), \
6170 false) >= 2)
6171 #endif
6173 /* EXP is some logical combination of boolean tests. See if we can
6174 merge it into some range test. Return the new tree if so. */
6176 static tree
6179 {
6193 /* If op0 is known true or false and this is a short-circuiting
6194 operation we must not merge with op1 since that makes side-effects
6195 unconditional. So special-case this. */
6202 /* If this is an OR operation, invert both sides; we will invert
6203 again at the end. */
6207 /* If both expressions are the same, if we can merge the ranges, and we
6208 can build the range test, return it or it inverted. If one of the
6209 ranges is always true or always false, consider it to be the same
6210 expression as the other. */
6218 {
6222 }
6224 /* On machines where the branch cost is expensive, if this is a
6225 short-circuited branch and the underlying object on both sides
6226 is the same, make a non-short-circuit operation. */
6229 logical_op_non_short_circuit
6236 {
6237 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
6238 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
6239 which cases we can't do this. */
6247 {
6256 {
6259 WARN_STRICT_OVERFLOW_COMPARISON);
6263 }
6264 }
6265 }
6268 }
6269 \f
6270 /* Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P
6271 bit value. Arrange things so the extra bits will be set to zero if and
6272 only if C is signed-extended to its full width. If MASK is nonzero,
6273 it is an INTEGER_CST that should be AND'ed with the extra bits. */
6275 static tree
6277 {
6280 tree temp;
6285 /* We work by getting just the sign bit into the low-order bit, then
6286 into the high-order bit, then sign-extend. We then XOR that value
6287 with C. */
6291 /* We must use a signed type in order to get an arithmetic right shift.
6292 However, we must also avoid introducing accidental overflows, so that
6293 a subsequent call to integer_zerop will work. Hence we must
6294 do the type conversion here. At this point, the constant is either
6295 zero or one, and the conversion to a signed type can never overflow.
6296 We could get an overflow if this conversion is done anywhere else. */
6305 /* If necessary, convert the type back to match the type of C. */
6310 }
6311 \f
6312 /* For an expression that has the form
6313 (A && B) || ~B
6314 or
6315 (A || B) && ~B,
6316 we can drop one of the inner expressions and simplify to
6317 A || ~B
6318 or
6319 A && ~B
6320 LOC is the location of the resulting expression. OP is the inner
6321 logical operation; the left-hand side in the examples above, while CMPOP
6322 is the right-hand side. RHS_ONLY is used to prevent us from accidentally
6323 removing a condition that guards another, as in
6324 (A != NULL && A->...) || A == NULL
6325 which we must not transform. If RHS_ONLY is true, only eliminate the
6326 right-most operand of the inner logical operation. */
6328 static tree
6331 {
6349 {
6352 {
6355 }
6356 }
6358 {
6361 {
6364 }
6365 }
6380 }
6382 /* Find ways of folding logical expressions of LHS and RHS:
6383 Try to merge two comparisons to the same innermost item.
6384 Look for range tests like "ch >= '0' && ch <= '9'".
6385 Look for combinations of simple terms on machines with expensive branches
6386 and evaluate the RHS unconditionally.
6388 For example, if we have p->a == 2 && p->b == 4 and we can make an
6389 object large enough to span both A and B, we can do this with a comparison
6390 against the object ANDed with the a mask.
6392 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
6393 operations to do this with one comparison.
6395 We check for both normal comparisons and the BIT_AND_EXPRs made this by
6396 function and the one above.
6398 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
6399 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
6401 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
6402 two operands.
6404 We return the simplified tree or 0 if no optimization is possible. */
6406 static tree
6409 {
6410 /* If this is the "or" of two comparisons, we can do something if
6411 the comparisons are NE_EXPR. If this is the "and", we can do something
6412 if the comparisons are EQ_EXPR. I.e.,
6413 (a->b == 2 && a->c == 4) can become (a->new == NEW).
6415 WANTED_CODE is this operation code. For single bit fields, we can
6416 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
6417 comparison for one-bit fields. */
6438 /* Start by getting the comparison codes. Fail if anything is volatile.
6439 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
6440 it were surrounded with a NE_EXPR. */
6449 {
6453 }
6456 {
6460 }
6471 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
6474 {
6477 {
6482 }
6485 {
6491 }
6492 }
6497 /* If the RHS can be evaluated unconditionally and its operands are
6498 simple, it wins to evaluate the RHS unconditionally on machines
6499 with expensive branches. In this case, this isn't a comparison
6500 that can be merged. */
6507 {
6508 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
6519 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
6529 }
6531 /* See if the comparisons can be merged. Then get all the parameters for
6532 each side. */
6557 /* It must be true that the inner operation on the lhs of each
6558 comparison must be the same if we are to be able to do anything.
6559 Then see if we have constants. If not, the same must be true for
6560 the rhs's. */
6569 {
6572 }
6577 else
6580 /* If either comparison code is not correct for our logical operation,
6581 fail. However, we can convert a one-bit comparison against zero into
6582 the opposite comparison against that bit being set in the field. */
6586 {
6588 {
6589 /* Make the left operand unsigned, since we are only interested
6590 in the value of one bit. Otherwise we are doing the wrong
6591 thing below. */
6594 }
6595 else
6597 }
6599 /* This is analogous to the code for l_const above. */
6601 {
6603 {
6606 }
6607 else
6609 }
6611 /* See if we can find a mode that contains both fields being compared on
6612 the left. If we can't, fail. Otherwise, update all constants and masks
6613 to be relative to a field of that size. */
6627 {
6630 }
6640 {
6649 {
6653 }
6654 }
6656 {
6665 {
6669 }
6670 }
6672 /* If the right sides are not constant, do the same for it. Also,
6673 disallow this optimization if a size, signedness or storage order
6674 mismatch occurs between the left and right sides. */
6676 {
6680 /* Make sure the two fields on the right
6681 correspond to the left without being swapped. */
6698 {
6701 }
6712 /* Make a mask that corresponds to both fields being compared.
6713 Do this for both items being compared. If the operands are the
6714 same size and the bits being compared are in the same position
6715 then we can do this by masking both and comparing the masked
6716 results. */
6723 {
6737 }
6739 /* There is still another way we can do something: If both pairs of
6740 fields being compared are adjacent, we may be able to make a wider
6741 field containing them both.
6743 Note that we still must mask the lhs/rhs expressions. Furthermore,
6744 the mask must be shifted to account for the shift done by
6745 make_bit_field_ref. */
6754 {
6755 tree type;
6773 /* Convert to the smaller type before masking out unwanted bits. */
6776 {
6778 {
6782 }
6784 {
6788 }
6789 }
6798 }
6801 }
6803 /* Handle the case of comparisons with constants. If there is something in
6804 common between the masks, those bits of the constants must be the same.
6805 If not, the condition is always false. Test for this to avoid generating
6806 incorrect code below. */
6811 {
6813 {
6816 }
6817 else
6818 {
6821 }
6822 }
6827 /* Construct the expression we will return. First get the component
6828 reference we will make. Unless the mask is all ones the width of
6829 that field, perform the mask operation. Then compare with the
6830 merged constant. */
6841 }
6842 \f
6843 /* T is an integer expression that is being multiplied, divided, or taken a
6844 modulus (CODE says which and what kind of divide or modulus) by a
6845 constant C. See if we can eliminate that operation by folding it with
6846 other operations already in T. WIDE_TYPE, if non-null, is a type that
6847 should be used for the computation if wider than our type.
6849 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
6850 (X * 2) + (Y * 4). We must, however, be assured that either the original
6851 expression would not overflow or that overflow is undefined for the type
6852 in the language in question.
6854 If we return a non-null expression, it is an equivalent form of the
6855 original computation, but need not be in the original type.
6857 We set *STRICT_OVERFLOW_P to true if the return values depends on
6858 signed overflow being undefined. Otherwise we do not change
6859 *STRICT_OVERFLOW_P. */
6861 static tree
6864 {
6865 /* To avoid exponential search depth, refuse to allow recursion past
6866 three levels. Beyond that (1) it's highly unlikely that we'll find
6867 something interesting and (2) we've probably processed it before
6868 when we built the inner expression. */
6871 tree ret;
6876 depth++;
6878 depth--;
6881 }
6883 static tree
6886 {
6891 {
6894 {
6897 }
6901 }
6907 /* Don't deal with constants of zero here; they confuse the code below. */
6917 /* Note that we need not handle conditional operations here since fold
6918 already handles those cases. So just do arithmetic here. */
6920 {
6922 /* For a constant, we can always simplify if we are a multiply
6923 or (for divide and modulus) if it is a multiple of our constant. */
6927 {
6930 /* If the multiplication overflowed, we lost information on it.
6931 See PR68142 and PR69845. */
6935 }
6941 /* If op0 is an expression ... */
6947 /* ... and has wrapping overflow, and its type is smaller
6948 than ctype, then we cannot pass through as widening. */
6952 /* ... or this is a truncation (t is narrower than op0),
6953 then we cannot pass through this narrowing. */
6956 /* ... or signedness changes for division or modulus,
6957 then we cannot pass through this conversion. */
6961 /* ... or has undefined overflow while the converted to
6962 type has not, we cannot do the operation in the inner type
6963 as that would introduce undefined overflow. */
6968 /* Pass the constant down and see if we can make a simplification. If
6969 we can, replace this expression with the inner simplification for
6970 possible later conversion to our or some other type. */
6981 /* If widening the type changes it from signed to unsigned, then we
6982 must avoid building ABS_EXPR itself as unsigned. */
6984 {
6988 {
6991 }
6993 }
6994 /* If the constant is negative, we cannot simplify this. */
6997 /* FALLTHROUGH */
6999 /* For division and modulus, type can't be unsigned, as e.g.
7000 (-(x / 2U)) / 2U isn't equal to -((x / 2U) / 2U) for x >= 2.
7001 For signed types, even with wrapping overflow, this is fine. */
7010 /* If widening the type changes the signedness, then we can't perform
7011 this optimization as that changes the result. */
7015 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
7021 {
7028 }
7032 /* If the second operand is constant, this is a multiplication
7033 or floor division, by a power of two, so we can treat it that
7034 way unless the multiplier or divisor overflows. Signed
7035 left-shift overflow is implementation-defined rather than
7036 undefined in C90, so do not convert signed left shift into
7037 multiplication. */
7040 /* const_binop may not detect overflow correctly,
7041 so check for it explicitly here. */
7050 ctype,
7052 t1),
7057 /* See if we can eliminate the operation on both sides. If we can, we
7058 can return a new PLUS or MINUS. If we can't, the only remaining
7059 cases where we can do anything are if the second operand is a
7060 constant. */
7067 /* If not multiplication, we can only do this if both operands
7068 are divisible by c. */
7071 {
7076 }
7078 /* If this was a subtraction, negate OP1 and set it to be an addition.
7079 This simplifies the logic below. */
7081 {
7083 /* If OP1 was not easily negatable, the constant may be OP0. */
7085 {
7088 }
7089 }
7094 /* If either OP1 or C are negative, this optimization is not safe for
7095 some of the division and remainder types while for others we need
7096 to change the code. */
7098 {
7106 }
7108 /* If it's a multiply or a division/modulus operation of a multiple
7109 of our constant, do the operation and verify it doesn't overflow. */
7113 {
7116 /* We allow the constant to overflow with wrapping semantics. */
7120 }
7121 else
7124 /* If we have an unsigned type, we cannot widen the operation since it
7125 will change the result if the original computation overflowed. */
7129 /* The last case is if we are a multiply. In that case, we can
7130 apply the distributive law to commute the multiply and addition
7131 if the multiplication of the constants doesn't overflow
7132 and overflow is defined. With undefined overflow
7133 op0 * c might overflow, while (op0 + orig_op1) * c doesn't.
7134 But fold_plusminus_mult_expr would factor back any power-of-two
7135 value so do not distribute in the first place in this case. */
7143 op1);
7148 /* We have a special case here if we are doing something like
7149 (C * 8) % 4 since we know that's zero. */
7152 /* If the multiplication can overflow we cannot optimize this. */
7157 {
7160 }
7162 /* ... fall through ... */
7166 /* If we can extract our operation from the LHS, do so and return a
7167 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
7168 do something only if the second operand is a constant. */
7184 /* If these are the same operation types, we can associate them
7185 assuming no overflow. */
7187 {
7202 }
7204 /* If these operations "cancel" each other, we have the main
7205 optimizations of this pass, which occur when either constant is a
7206 multiple of the other, in which case we replace this with either an
7207 operation or CODE or TCODE.
7209 If we have an unsigned type, we cannot do this since it will change
7210 the result if the original computation overflowed. */
7218 {
7221 {
7227 }
7230 {
7236 }
7237 }
7242 }
7245 }
7246 \f
7247 /* Return a node which has the indicated constant VALUE (either 0 or
7248 1 for scalars or {-1,-1,..} or {0,0,...} for vectors),
7249 and is of the indicated TYPE. */
7251 tree
7253 {
7262 else
7264 }
7267 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
7268 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
7269 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
7270 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
7271 COND is the first argument to CODE; otherwise (as in the example
7272 given here), it is the second argument. TYPE is the type of the
7273 original expression. Return NULL_TREE if no simplification is
7274 possible. */
7276 static tree
7281 {
7289 /* Do not move possibly trapping operations into the conditional as this
7290 pessimizes code and causes gimplification issues when applied late. */
7298 {
7302 /* If this operand throws an expression, then it does not make
7303 sense to try to perform a logical or arithmetic operation
7304 involving it. */
7309 }
7312 {
7317 }
7318 else
7319 /* Detect the case of mixing vector and scalar types - bail out. */
7325 /* This transformation is only worthwhile if we don't have to wrap ARG
7326 in a SAVE_EXPR and the operation can be simplified without recursing
7327 on at least one of the branches once its pushed inside the COND_EXPR. */
7336 {
7340 else
7342 }
7344 {
7348 else
7350 }
7352 /* Check that we have simplified at least one of the branches. */
7357 }
7359 \f
7360 /* Subroutine of fold() that checks for the addition of ARG +/- 0.0.
7362 If !NEGATE, return true if ZERO_ARG is +/-0.0 and, for all ARG of
7363 type TYPE, ARG + ZERO_ARG is the same as ARG. If NEGATE, return true
7364 if ARG - ZERO_ARG is the same as X.
7366 If ARG is NULL, check for any value of type TYPE.
7368 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
7369 and finite. The problematic cases are when X is zero, and its mode
7370 has signed zeros. In the case of rounding towards -infinity,
7371 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
7372 modes, X + 0 is not the same as X because -0 + 0 is 0. */
7374 bool
7377 {
7381 /* Don't allow the fold with -fsignaling-nans. */
7385 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
7389 /* There is no case that is safe for all rounding modes. */
7393 /* In a vector or complex, we would need to check the sign of all zeros. */
7399 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
7403 /* The mode has signed zeros, and we have to honor their sign.
7404 In this situation, there are only two cases we can return true for.
7405 (i) X - 0 is the same as X with default rounding.
7406 (ii) X + 0 is X when X can't possibly be -0.0. */
7408 }
7410 /* Subroutine of match.pd that optimizes comparisons of a division by
7411 a nonzero integer constant against an integer constant, i.e.
7412 X/C1 op C2.
7414 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
7415 GE_EXPR or LE_EXPR. ARG01 and ARG1 must be a INTEGER_CST. */
7417 enum tree_code
7420 {
7425 /* We have to do this the hard way to detect unsigned overflow.
7426 prod = int_const_binop (MULT_EXPR, c1, c2); */
7432 {
7436 /* Likewise *hi = int_const_binop (PLUS_EXPR, prod, tmp). */
7439 }
7441 {
7444 {
7463 }
7464 }
7465 else
7466 {
7467 /* A negative divisor reverses the relational operators. */
7472 {
7491 }
7492 }
7505 }
7507 /* Test whether it is preferable to swap two operands, ARG0 and
7508 ARG1, for example because ARG0 is an integer constant and ARG1
7509 isn't. */
7511 bool
7513 {
7527 /* It is preferable to swap two SSA_NAME to ensure a canonical form
7528 for commutative and comparison operators. Ensuring a canonical
7529 form allows the optimizers to find additional redundancies without
7530 having to explicitly check for both orderings. */
7536 /* Put SSA_NAMEs last. */
7542 /* Put variables last. */
7549 }
7552 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
7553 means A >= Y && A != MAX, but in this case we know that
7554 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
7556 static tree
7558 {
7565 else
7574 {
7577 }
7579 {
7582 }
7583 else
7590 {
7591 /* Convert the pointer types into integer before taking the difference. */
7595 }
7596 else
7603 }
7605 /* Fold a sum or difference of at least one multiplication.
7606 Returns the folded tree or NULL if no simplification could be made. */
7608 static tree
7611 {
7615 /* (A * C) +- (B * C) -> (A+-B) * C.
7616 (A * C) +- A -> A * (C+-1).
7617 We are most concerned about the case where C is a constant,
7618 but other combinations show up during loop reduction. Since
7619 it is not difficult, try all four possibilities. */
7622 {
7625 }
7627 {
7630 }
7631 else
7632 {
7633 /* We cannot generate constant 1 for fract. */
7638 }
7640 {
7643 }
7645 {
7647 /* As we canonicalize A - 2 to A + -2 get rid of that sign for
7648 the purpose of this canonicalization. */
7652 {
7655 }
7656 else
7658 }
7659 else
7660 {
7661 /* We cannot generate constant 1 for fract. */
7666 }
7669 /* Prefer factoring a common non-constant. */
7679 /* No identical multiplicands; see if we can find a common
7680 power-of-two factor in non-power-of-two multiplies. This
7681 can help in multi-dimensional array access. */
7683 {
7686 HOST_WIDE_INT tmp;
7688 tree maybe_same;
7690 /* Move min of absolute values to int11. */
7692 {
7697 }
7698 else
7705 /* The remainder should not be a constant, otherwise we
7706 end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
7707 increased the number of multiplications necessary. */
7709 {
7717 }
7718 }
7725 /* We are neither factoring zero nor minus one. */
7733 /* Same may be zero and thus the operation 'code' may overflow. Likewise
7734 same may be minus one and thus the multiplication may overflow. Perform
7735 the sum operation in an unsigned type. */
7740 /* If the sum evaluated to a constant that is not -INF the multiplication
7741 cannot overflow. */
7748 /* Do not resort to unsigned multiplication because
7749 we lose the no-overflow property of the expression. */
7751 }
7753 /* Subroutine of native_encode_expr. Encode the INTEGER_CST
7754 specified by EXPR into the buffer PTR of length LEN bytes.
7755 Return the number of bytes placed in the buffer, or zero
7756 upon failure. */
7758 static int
7760 {
7764 {
7770 {
7772 /* More work is needed when adding _BitInt support to PDP endian
7773 if limb is smaller than word, or if _BitInt limb ordering doesn't
7774 match target endianity here. */
7779 }
7780 else
7782 }
7783 else
7794 /* Dry run. */
7800 {
7802 /* Extend EXPR according to TYPE_SIGN if the precision isn't a whole
7803 number of bytes. */
7807 {
7814 else
7816 }
7817 else
7821 }
7823 }
7826 /* Subroutine of native_encode_expr. Encode the FIXED_CST
7827 specified by EXPR into the buffer PTR of length LEN bytes.
7828 Return the number of bytes placed in the buffer, or zero
7829 upon failure. */
7831 static int
7833 {
7837 FIXED_VALUE_TYPE value;
7852 }
7855 /* Subroutine of native_encode_expr. Encode the REAL_CST
7856 specified by EXPR into the buffer PTR of length LEN bytes.
7857 Return the number of bytes placed in the buffer, or zero
7858 upon failure. */
7860 static int
7862 {
7868 /* There are always 32 bits in each long, no matter the size of
7869 the hosts long. We handle floating point representations with
7870 up to 192 bits. */
7879 /* Dry run. */
7888 {
7893 {
7900 else
7902 }
7903 else
7904 {
7907 {
7908 /* Reverse bytes within each long, or within the entire float
7909 if it's smaller than a long (for HFmode). */
7912 }
7913 }
7918 }
7920 }
7922 /* Subroutine of native_encode_expr. Encode the COMPLEX_CST
7923 specified by EXPR into the buffer PTR of length LEN bytes.
7924 Return the number of bytes placed in the buffer, or zero
7925 upon failure. */
7927 static int
7929 {
7931 tree part;
7945 }
7947 /* Like native_encode_vector, but only encode the first COUNT elements.
7948 The other arguments are as for native_encode_vector. */
7950 static int
7953 {
7957 {
7958 /* This is the only case in which elements can be smaller than a byte.
7959 Element 0 is always in the lsb of the containing byte. */
7968 /* Zero the buffer and then set bits later where necessary. */
7977 {
7983 {
7986 }
7987 }
7989 }
7994 {
7996 {
7999 }
8010 }
8012 }
8014 /* Subroutine of native_encode_expr. Encode the VECTOR_CST
8015 specified by EXPR into the buffer PTR of length LEN bytes.
8016 Return the number of bytes placed in the buffer, or zero
8017 upon failure. */
8019 static int
8021 {
8026 }
8029 /* Subroutine of native_encode_expr. Encode the STRING_CST
8030 specified by EXPR into the buffer PTR of length LEN bytes.
8031 Return the number of bytes placed in the buffer, or zero
8032 upon failure. */
8034 static int
8036 {
8039 /* Wide-char strings are encoded in target byte-order so native
8040 encoding them is trivial. */
8055 else
8056 {
8059 {
8062 }
8064 }
8066 }
8069 /* Subroutine of fold_view_convert_expr. Encode the INTEGER_CST, REAL_CST,
8070 FIXED_CST, COMPLEX_CST, STRING_CST, or VECTOR_CST specified by EXPR into
8071 the buffer PTR of size LEN bytes. If PTR is NULL, don't actually store
8072 anything, just do a dry run. Fail either if OFF is -1 and LEN isn't
8073 sufficient to encode the entire EXPR, or if OFF is out of bounds.
8074 Otherwise, start at byte offset OFF and encode at most LEN bytes.
8075 Return the number of bytes placed in the buffer, or zero upon failure. */
8077 int
8079 {
8080 /* We don't support starting at negative offset and -1 is special. */
8085 {
8106 }
8107 }
8109 /* Try to find a type whose byte size is smaller or equal to LEN bytes larger
8110 or equal to FIELDSIZE bytes, with underlying mode precision/size multiple
8111 of BITS_PER_UNIT. As native_{interpret,encode}_int works in term of
8112 machine modes, we can't just use build_nonstandard_integer_type. */
8114 tree
8116 {
8117 machine_mode mode;
8119 {
8126 {
8130 }
8131 }
8137 {
8145 }
8148 }
8150 /* Similar to native_encode_expr, but also handle CONSTRUCTORs, VCEs,
8151 NON_LVALUE_EXPRs and nops. If MASK is non-NULL (then PTR has
8152 to be non-NULL and OFF zero), then in addition to filling the
8153 bytes pointed by PTR with the value also clear any bits pointed
8154 by MASK that are known to be initialized, keep them as is for
8155 e.g. uninitialized padding bits or uninitialized fields. */
8157 int
8160 {
8163 /* We don't support starting at negative offset and -1 is special. */
8171 {
8175 mask);
8190 {
8191 tree min_index;
8209 {
8214 {
8217 }
8222 else
8226 {
8230 offset_int wpos
8238 offset_int wcount
8245 }
8247 {
8250 offset_int wpos
8258 }
8261 {
8263 {
8274 }
8276 {
8280 }
8281 }
8285 do
8286 {
8291 {
8293 {
8296 fieldsize);
8299 }
8301 ptr
8304 fieldsize,
8307 mask
8311 else
8312 {
8315 }
8316 }
8319 {
8320 /* Partial overlap. */
8326 {
8330 fieldsize);
8331 }
8332 else
8333 {
8337 }
8340 }
8342 }
8344 }
8346 }
8349 {
8359 {
8365 {
8368 }
8371 {
8380 }
8384 else
8388 {
8389 tree fld;
8391 {
8404 }
8406 {
8410 }
8413 {
8414 cnt--;
8420 }
8421 }
8426 {
8439 {
8444 }
8445 }
8446 else
8447 {
8452 }
8456 /* Prepare to deal with integral bit-fields and filter out other
8457 bit-fields that do not start and end on a byte boundary. */
8459 {
8464 {
8470 }
8476 }
8486 {
8487 /* FIXME: Handle PDP endian. */
8498 {
8501 }
8502 else
8503 {
8511 else
8512 {
8515 }
8516 }
8526 else
8539 /* If the bitfield does not start at byte boundary, handle
8540 the partial byte at the start. */
8543 {
8545 {
8550 {
8553 else
8555 }
8556 }
8557 else
8558 {
8563 {
8566 else
8570 }
8571 }
8572 }
8573 /* If the bitfield does not end at byte boundary, handle
8574 the partial byte at the end. */
8578 {
8580 {
8587 }
8588 else
8589 {
8596 }
8597 }
8601 {
8606 }
8607 else
8608 {
8609 /* Partial overlap. */
8613 {
8616 }
8620 }
8622 }
8627 {
8630 {
8633 {
8637 }
8640 }
8643 fldsize,
8653 }
8654 else
8655 {
8656 /* Partial overlap. */
8662 {
8666 fieldsize);
8667 }
8668 else
8669 {
8673 }
8676 }
8677 }
8679 }
8681 }
8682 }
8685 /* Subroutine of native_interpret_expr. Interpret the contents of
8686 the buffer PTR of length LEN as an INTEGER_CST of type TYPE.
8687 If the buffer cannot be interpreted, return NULL_TREE. */
8689 static tree
8691 {
8694 {
8700 {
8702 /* More work is needed when adding _BitInt support to PDP endian
8703 if limb is smaller than word, or if _BitInt limb ordering doesn't
8704 match target endianity here. */
8709 }
8710 else
8712 }
8713 else
8723 }
8726 /* Subroutine of native_interpret_expr. Interpret the contents of
8727 the buffer PTR of length LEN as a FIXED_CST of type TYPE.
8728 If the buffer cannot be interpreted, return NULL_TREE. */
8730 static tree
8732 {
8735 double_int result;
8736 FIXED_VALUE_TYPE fixed_value;
8746 }
8749 /* Subroutine of native_interpret_expr. Interpret the contents of
8750 the buffer PTR of length LEN as a REAL_CST of type TYPE.
8751 If the buffer cannot be interpreted, return NULL_TREE. */
8753 tree
8755 {
8759 /* There are always 32 bits in each long, no matter the size of
8760 the hosts long. We handle floating point representations with
8761 up to 192 bits. */
8762 REAL_VALUE_TYPE r;
8772 {
8773 /* Both OFFSET and BYTE index within a long;
8774 bitpos indexes the whole float. */
8777 {
8784 else
8786 }
8787 else
8788 {
8791 {
8792 /* Reverse bytes within each long, or within the entire float
8793 if it's smaller than a long (for HFmode). */
8796 }
8797 }
8801 }
8805 }
8808 /* Subroutine of native_interpret_expr. Interpret the contents of
8809 the buffer PTR of length LEN as a COMPLEX_CST of type TYPE.
8810 If the buffer cannot be interpreted, return NULL_TREE. */
8812 static tree
8814 {
8829 }
8831 /* Read a vector of type TYPE from the target memory image given by BYTES,
8832 which contains LEN bytes. The vector is known to be encodable using
8833 NPATTERNS interleaved patterns with NELTS_PER_PATTERN elements each.
8835 Return the vector on success, otherwise return null. */
8837 static tree
8841 {
8845 {
8846 /* This is the only case in which elements can be smaller than a byte.
8847 Element 0 is always in the lsb of the containing byte. */
8854 {
8861 }
8863 }
8871 {
8877 }
8879 }
8881 /* Subroutine of native_interpret_expr. Interpret the contents of
8882 the buffer PTR of length LEN as a VECTOR_CST of type TYPE.
8883 If the buffer cannot be interpreted, return NULL_TREE. */
8885 static tree
8887 {
8896 }
8899 /* Subroutine of fold_view_convert_expr. Interpret the contents of
8900 the buffer PTR of length LEN as a constant of type TYPE. For
8901 INTEGRAL_TYPE_P we return an INTEGER_CST, for SCALAR_FLOAT_TYPE_P
8902 we return a REAL_CST, etc... If the buffer cannot be interpreted,
8903 return NULL_TREE. */
8905 tree
8907 {
8909 {
8921 {
8922 /* For floating point values in composite modes, punt if this
8923 folding doesn't preserve bit representation. As the mode doesn't
8924 have fixed precision while GCC pretends it does, there could be
8925 valid values that GCC can't really represent accurately.
8926 See PR95450. Even for other modes, e.g. x86 XFmode can have some
8927 bit combinationations which GCC doesn't preserve. */
8937 }
8951 }
8952 }
8954 /* Returns true if we can interpret the contents of a native encoding
8955 as TYPE. */
8957 bool
8959 {
8961 {
8975 }
8976 }
8978 /* Attempt to interpret aggregate of TYPE from bytes encoded in target
8979 byte order at PTR + OFF with LEN bytes. Does not handle unions. */
8981 tree
8984 {
8987 {
8994 {
8998 }
9003 {
9008 {
9013 }
9017 eltsz);
9021 }
9023 }
9027 {
9036 {
9039 {
9040 poly_int64 bitoffset;
9045 else
9055 }
9056 else
9057 {
9075 HOST_WIDE_INT rpos;
9078 else
9079 {
9082 }
9091 }
9092 }
9095 {
9105 }
9109 {
9114 }
9121 {
9125 /* FIXME: Figure out how to handle PDP endian bitfields. */
9131 else
9135 }
9137 }
9139 }
9141 /* Routines for manipulation of native_encode_expr encoded data if the encoded
9142 or extracted constant positions and/or sizes aren't byte aligned. */
9144 /* Shift left the bytes in PTR of SZ elements by AMNT bits, carrying over the
9145 bits between adjacent elements. AMNT should be within
9146 [0, BITS_PER_UNIT).
9147 Example, AMNT = 2:
9148 00011111|11100000 << 2 = 01111111|10000000
9149 PTR[1] | PTR[0] PTR[1] | PTR[0]. */
9151 void
9154 {
9163 {
9169 {
9172 }
9173 }
9174 }
9176 /* Like shift_bytes_in_array_left but for big-endian.
9177 Shift right the bytes in PTR of SZ elements by AMNT bits, carrying over the
9178 bits between adjacent elements. AMNT should be within
9179 [0, BITS_PER_UNIT).
9180 Example, AMNT = 2:
9181 00011111|11100000 >> 2 = 00000111|11111000
9182 PTR[0] | PTR[1] PTR[0] | PTR[1]. */
9184 void
9187 {
9195 {
9202 }
9203 }
9205 /* Try to view-convert VECTOR_CST EXPR to VECTOR_TYPE TYPE by operating
9206 directly on the VECTOR_CST encoding, in a way that works for variable-
9207 length vectors. Return the resulting VECTOR_CST on success or null
9208 on failure. */
9210 static tree
9212 {
9224 /* We can only preserve the semantics of a stepped pattern if the new
9225 vector element is an integer of the same size. */
9230 /* The number of bits needed to encode one element from every pattern
9231 of the original vector. */
9232 unsigned int expr_sequence_bits
9235 /* The number of bits needed to encode one element from every pattern
9236 of the result. */
9237 unsigned int type_sequence_bits
9240 /* Don't try to read more bytes than are available, which can happen
9241 for constant-sized vectors if TYPE has larger elements than EXPR_TYPE.
9242 The general VIEW_CONVERT handling can cope with that case, so there's
9243 no point complicating things here. */
9246 BITS_PER_UNIT);
9251 /* Get enough bytes of EXPR to form the new encoding. */
9259 /* Reencode the bytes as TYPE. */
9263 }
9265 /* Fold a VIEW_CONVERT_EXPR of a constant expression EXPR to type
9266 TYPE at compile-time. If we're unable to perform the conversion
9267 return NULL_TREE. */
9269 static tree
9271 {
9272 /* We support up to 1024-bit values (for GCN/RISC-V V128QImode). */
9276 /* Check that the host and target are sane. */
9289 }
9291 /* Build an expression for the address of T. Folds away INDIRECT_REF
9292 to avoid confusing the gimplify process. */
9294 tree
9296 {
9297 /* The size of the object is not relevant when talking about its address. */
9302 {
9307 }
9310 {
9315 }
9322 {
9327 }
9328 else
9332 }
9334 /* Build an expression for the address of T. */
9336 tree
9338 {
9342 }
9344 /* Fold a unary expression of code CODE and type TYPE with operand
9345 OP0. Return the folded expression if folding is successful.
9346 Otherwise, return NULL_TREE. */
9348 tree
9350 {
9351 tree tem;
9352 tree arg0;
9360 {
9363 {
9364 /* Don't use STRIP_NOPS, because signedness of argument type
9365 matters. */
9367 }
9368 else
9369 {
9370 /* Strip any conversions that don't change the mode. This
9371 is safe for every expression, except for a comparison
9372 expression because its signedness is derived from its
9373 operands.
9375 Note that this is done as an internal manipulation within
9376 the constant folder, in order to find the simplest
9377 representation of the arguments so that their form can be
9378 studied. In any cases, the appropriate type conversions
9379 should be put back in the tree that will get out of the
9380 constant folder. */
9382 }
9385 {
9388 {
9392 }
9393 }
9394 }
9401 {
9408 {
9422 /* If this was a conversion, and all we did was to move into
9423 inside the COND_EXPR, bring it back out. But leave it if
9424 it is a conversion from integer to integer and the
9425 result precision is no wider than a word since such a
9426 conversion is cheap and may be optimized away by combine,
9427 while it couldn't if it were outside the COND_EXPR. Then return
9428 so we don't get into an infinite recursion loop taking the
9429 conversion out and then back in. */
9441 && (INTEGRAL_TYPE_P
9454 }
9455 }
9458 {
9464 CASE_CONVERT:
9468 {
9469 /* If we have (type) (a CMP b) and type is an integral type, return
9470 new expression involving the new type. Canonicalize
9471 (type) (a CMP b) to (a CMP b) ? (type) true : (type) false for
9472 non-integral type.
9473 Do not fold the result as that would not simplify further, also
9474 folding again results in recursions. */
9484 }
9486 /* Handle (T *)&A.B.C for A being of type T and B and C
9487 living at offset zero. This occurs frequently in
9488 C++ upcasting and then accessing the base. */
9492 {
9494 tree offset;
9495 machine_mode mode;
9497 tree base
9501 /* If the reference was to a (constant) zero offset, we can use
9502 the address of the base if it has the same base type
9503 as the result type and the pointer type is unqualified. */
9511 }
9515 /* Detect assigning a bitfield. */
9517 && DECL_BIT_FIELD
9519 {
9520 /* Don't leave an assignment inside a conversion
9521 unless assigning a bitfield. */
9523 /* First do the assignment, then return converted constant. */
9528 }
9530 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
9531 constants (if x has signed type, the sign bit cannot be set
9532 in c). This folds extension into the BIT_AND_EXPR.
9533 ??? We don't do it for BOOLEAN_TYPE or ENUMERAL_TYPE because they
9534 very likely don't have maximal range for their precision and this
9535 transformation effectively doesn't preserve non-maximal ranges. */
9539 {
9550 <= HOST_BITS_PER_WIDE_INT
9552 {
9556 cst &= HOST_WIDE_INT_M1U
9560 && !flag_syntax_only
9563 {
9567 }
9568 }
9570 {
9580 }
9581 }
9583 /* Convert (T1)(X p+ Y) into ((T1)X p+ Y), for pointer type, when the new
9584 cast (T1)X will fold away. We assume that this happens when X itself
9585 is a cast. */
9589 {
9593 /* If -fsanitize=alignment, avoid this optimization in GENERIC
9594 when the pointed type needs higher alignment than
9595 the p+ first operand's pointed type. */
9602 /* Similarly, avoid this optimization in GENERIC for -fsanitize=null
9603 when type is a reference type and arg00's type is not,
9604 because arg00 could be validly nullptr and if arg01 doesn't return,
9605 we don't want false positive binding of reference to nullptr. */
9607 && !in_gimple_form
9614 }
9616 /* Convert (T1)(~(T2)X) into ~(T1)X if T1 and T2 are integral types
9617 of the same precision, and X is an integer type not narrower than
9618 types T1 or T2, i.e. the cast (T2)X isn't an extension. */
9624 {
9630 }
9632 /* Convert (T1)(X * Y) into (T1)X * (T1)Y if T1 is narrower than the
9633 type of X and Y (integer types only). */
9640 {
9641 /* Be careful not to introduce new overflows. */
9642 tree mult_type;
9645 else
9649 {
9656 }
9657 }
9663 {
9670 }
9681 /* Convert fabs((double)float) into (double)fabsf(float). */
9684 {
9690 targ0));
9691 }
9695 /* Convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
9714 /* Note that the operand of this must be an int
9715 and its values must be 0 or 1.
9716 ("true" is a fixed value perhaps depending on the language,
9717 but we don't handle values other than 1 correctly yet.) */
9724 /* Fold *&X to X if X is an lvalue. */
9726 {
9733 }
9739 }
9742 /* If the operation was a conversion do _not_ mark a resulting constant
9743 with TREE_OVERFLOW if the original constant was not. These conversions
9744 have implementation defined behavior and retaining the TREE_OVERFLOW
9745 flag here would confuse later passes such as VRP. */
9746 tree
9749 {
9758 }
9760 /* Fold a binary bitwise/truth expression of code CODE and type TYPE with
9761 operands OP0 and OP1. LOC is the location of the resulting expression.
9762 ARG0 and ARG1 are the NOP_STRIPed results of OP0 and OP1.
9763 Return the folded expression if folding is successful. Otherwise,
9764 return NULL_TREE. */
9765 static tree
9768 {
9769 tree tem;
9771 /* We only do these simplifications if we are optimizing. */
9775 /* Check for things like (A || B) && (A || C). We can convert this
9776 to A || (B && C). Note that either operator can be any of the four
9777 truth and/or operations and the transformation will still be
9778 valid. Also note that we only care about order for the
9779 ANDIF and ORIF operators. If B contains side effects, this
9780 might change the truth-value of A. */
9787 {
9807 /* This case if tricky because we must either have commutative
9808 operators or else A10 must not have side-effects. */
9814 a01);
9815 }
9817 /* See if we can build a range comparison. */
9823 {
9827 }
9831 {
9835 }
9837 /* Check for the possibility of merging component references. If our
9838 lhs is another similar operation, try to merge its rhs with our
9839 rhs. Then try to merge our lhs and rhs. */
9850 logical_op_non_short_circuit
9858 {
9865 /* Transform ((A AND-IF B) AND[-IF] C) into (A AND-IF (B AND C)),
9866 or ((A OR-IF B) OR[-IF] C) into (A OR-IF (B OR C))
9867 We don't want to pack more than two leafs to a non-IF AND/OR
9868 expression.
9869 If tree-code of left-hand operand isn't an AND/OR-IF code and not
9870 equal to IF-CODE, then we don't want to add right-hand operand.
9871 If the inner right-hand side of left-hand operand has
9872 side-effects, or isn't simple, then we can't add to it,
9873 as otherwise we might destroy if-sequence. */
9876 /* Needed for sequence points to handle trappings, and
9877 side-effects. */
9879 {
9881 arg1);
9883 tem);
9884 }
9885 /* Same as above but for (A AND[-IF] (B AND-IF C)) -> ((A AND B) AND-IF C),
9886 or (A OR[-IF] (B OR-IF C) -> ((A OR B) OR-IF C). */
9889 /* Needed for sequence points to handle trappings, and
9890 side-effects. */
9892 {
9897 }
9898 /* Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
9899 into (A OR B).
9900 For sequence point consistancy, we need to check for trapping,
9901 and side-effects. */
9905 }
9908 }
9910 /* Helper that tries to canonicalize the comparison ARG0 CODE ARG1
9911 by changing CODE to reduce the magnitude of constants involved in
9912 ARG0 of the comparison.
9913 Returns a canonicalized comparison tree if a simplification was
9914 possible, otherwise returns NULL_TREE.
9915 Set *STRICT_OVERFLOW_P to true if the canonicalization is only
9916 valid if signed overflow is undefined. */
9918 static tree
9922 {
9927 /* Match A +- CST code arg1. We can change this only if overflow
9928 is undefined. */
9931 /* In principle pointers also have undefined overflow behavior,
9932 but that causes problems elsewhere. */
9939 /* Identify the constant in arg0 and its sign. */
9943 /* Overflowed constants and zero will cause problems. */
9948 /* See if we can reduce the magnitude of the constant in
9949 arg0 by changing the comparison code. */
9950 /* A - CST < arg1 -> A - CST-1 <= arg1. */
9954 /* A + CST > arg1 -> A + CST-1 >= arg1. */
9958 /* A + CST <= arg1 -> A + CST-1 < arg1. */
9962 /* A - CST >= arg1 -> A - CST-1 > arg1. */
9966 else
9970 /* Now build the constant reduced in magnitude. But not if that
9971 would produce one outside of its types range. */
9987 }
9989 /* Canonicalize the comparison ARG0 CODE ARG1 with type TYPE with undefined
9990 overflow further. Try to decrease the magnitude of constants involved
9991 by changing LE_EXPR and GE_EXPR to LT_EXPR and GT_EXPR or vice versa
9992 and put sole constants at the second argument position.
9993 Returns the canonicalized tree if changed, otherwise NULL_TREE. */
9995 static tree
9998 {
9999 tree t;
10004 /* Try canonicalization by simplifying arg0. */
10009 {
10013 }
10015 /* Try canonicalization by simplifying arg1 using the swapped
10016 comparison. */
10024 }
10026 /* Return whether BASE + OFFSET + BITPOS may wrap around the address
10027 space. This is used to avoid issuing overflow warnings for
10028 expressions like &p->x which cannot wrap. */
10030 static bool
10032 {
10039 poly_wide_int wi_offset;
10045 else
10050 precision);
10065 /* We can do slightly better for SIZE if we have an ADDR_EXPR of an
10066 array. */
10075 }
10077 /* Return a positive integer when the symbol DECL is known to have
10078 a nonzero address, zero when it's known not to (e.g., it's a weak
10079 symbol), and a negative integer when the symbol is not yet in the
10080 symbol table and so whether or not its address is zero is unknown.
10081 For function local objects always return positive integer. */
10082 static int
10084 {
10085 /* Normally, don't do anything for variables and functions before symtab is
10086 built; it is quite possible that DECL will be declared weak later.
10087 But if folding_initializer, we need a constant answer now, so create
10088 the symtab entry and prevent later weak declaration. */
10091 = (folding_initializer
10096 /* Function local objects are never NULL. */
10104 }
10106 /* Subroutine of fold_binary. This routine performs all of the
10107 transformations that are common to the equality/inequality
10108 operators (EQ_EXPR and NE_EXPR) and the ordering operators
10109 (LT_EXPR, LE_EXPR, GE_EXPR and GT_EXPR). Callers other than
10110 fold_binary should call fold_binary. Fold a comparison with
10111 tree code CODE and type TYPE with operands OP0 and OP1. Return
10112 the folded comparison or NULL_TREE. */
10114 static tree
10117 {
10127 /* For comparisons of pointers we can decompose it to a compile time
10128 comparison of the base objects and the offsets into the object.
10129 This requires at least one operand being an ADDR_EXPR or a
10130 POINTER_PLUS_EXPR to do more than the operand_equal_p test below. */
10136 {
10139 machine_mode mode;
10143 /* Get base and offset for the access. Strip ADDR_EXPR for
10144 get_inner_reference, but put it back by stripping INDIRECT_REF
10145 off the base object if possible. indirect_baseN will be true
10146 if baseN is not an address but refers to the object itself. */
10149 {
10150 base0
10156 else
10158 }
10160 {
10164 {
10165 base0
10171 else
10173 }
10176 else
10180 {
10187 }
10188 }
10192 {
10193 base1
10199 else
10201 }
10203 {
10207 {
10208 base1
10214 else
10216 }
10219 else
10223 {
10230 }
10231 }
10233 /* If we have equivalent bases we might be able to simplify. */
10237 {
10238 /* We can fold this expression to a constant if the non-constant
10239 offset parts are equal. */
10243 && (equality_code
10244 || (indirect_base0
10247 {
10253 "occur when comparing P +- C1 with "
10255 WARN_STRICT_OVERFLOW_CONDITIONAL);
10258 {
10296 }
10297 }
10298 /* We can simplify the comparison to a comparison of the variable
10299 offset parts if the constant offset parts are equal.
10300 Be careful to use signed sizetype here because otherwise we
10301 mess with array offsets in the wrong way. This is possible
10302 because pointer arithmetic is restricted to retain within an
10303 object and overflow on pointer differences is undefined as of
10304 6.5.6/8 and /9 with respect to the signed ptrdiff_t. */
10306 && (equality_code
10307 || (indirect_base0
10310 {
10311 /* By converting to signed sizetype we cover middle-end pointer
10312 arithmetic which operates on unsigned pointer types of size
10313 type size and ARRAY_REF offsets which are properly sign or
10314 zero extended from their type in case it is narrower than
10315 sizetype. */
10318 else
10322 else
10329 "occur when comparing P +- C1 with "
10331 WARN_STRICT_OVERFLOW_COMPARISON);
10334 }
10335 }
10336 /* For equal offsets we can simplify to a comparison of the
10337 base addresses. */
10339 && (indirect_base0
10341 && (indirect_base1
10346 {
10352 }
10353 /* Comparison between an ordinary (non-weak) symbol and a null
10354 pointer can be eliminated since such symbols must have a non
10355 null address. In C, relational expressions between pointers
10356 to objects and null pointers are undefined. The results
10357 below follow the C++ rules with the additional property that
10358 every object pointer compares greater than a null pointer.
10359 */
10362 /* Avoid folding references to struct members at offset 0 to
10363 prevent tests like '&ptr->firstmember == 0' from getting
10364 eliminated. When ptr is null, although the -> expression
10365 is strictly speaking invalid, GCC retains it as a matter
10366 of QoI. See PR c/44555. */
10369 && indirect_base0
10370 /* The caller guarantees that when one of the arguments is
10371 constant (i.e., null in this case) it is second. */
10373 {
10375 {
10386 }
10387 }
10388 }
10390 /* Transform comparisons of the form X +- C1 CMP Y +- C2 to
10391 X CMP Y +- C2 +- C1 for signed X, Y. This is valid if
10392 the resulting offset is smaller in absolute value than the
10393 original one and has the same sign. */
10402 {
10407 tree cst;
10409 "occur when combining constants around "
10412 /* Put the constant on the side where it doesn't overflow and is
10413 of lower absolute value and of same sign than before. */
10420 {
10423 variable1,
10427 }
10435 {
10441 variable2);
10442 }
10443 }
10449 /* If we are comparing an expression that just has comparisons
10450 of two integer values, arithmetic expressions of those comparisons,
10451 and constants, we can simplify it. There are only three cases
10452 to check: the two values can either be equal, the first can be
10453 greater, or the second can be greater. Fold the expression for
10454 those three values. Since each value must be 0 or 1, we have
10455 eight possibilities, each of which corresponds to the constant 0
10456 or 1 or one of the six possible comparisons.
10458 This handles common cases like (a > b) == 0 but also handles
10459 expressions like ((x > y) - (y > x)) > 0, which supposedly
10460 occur in macroized code. */
10463 {
10467 /* Don't handle degenerate cases here; they should already
10468 have been handled anyway. */
10477 {
10481 /* We can't just pass T to eval_subst in case cval1 or cval2
10482 was the same as ARG1. */
10484 tree high_result
10488 arg1);
10489 tree equal_result
10493 arg1);
10494 tree low_result
10498 arg1);
10500 /* All three of these results should be 0 or 1. Confirm they are.
10501 Then use those values to select the proper code to use. */
10506 {
10507 /* Make a 3-bit mask with the high-order bit being the
10508 value for `>', the next for '=', and the low for '<'. */
10512 {
10514 /* Always false. */
10535 /* Always true. */
10537 }
10540 }
10541 }
10542 }
10545 }
10548 /* Subroutine of fold_binary. Optimize complex multiplications of the
10549 form z * conj(z), as pow(realpart(z),2) + pow(imagpart(z),2). The
10550 argument EXPR represents the expression "z" of type TYPE. */
10552 static tree
10554 {
10559 {
10562 }
10564 {
10567 }
10568 else
10569 {
10573 }
10582 }
10585 /* Helper function for fold_vec_perm. Store elements of VECTOR_CST or
10586 CONSTRUCTOR ARG into array ELTS, which has NELTS elements, and return
10587 true if successful. */
10589 static bool
10591 {
10596 {
10599 }
10601 {
10607 else
10609 }
10610 else
10616 }
10618 /* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
10619 mask for VLA vec_perm folding.
10620 REASON if specified, will contain the reason why SEL is not suitable.
10621 Used only for debugging and unit-testing. */
10623 static bool
10627 {
10634 {
10638 }
10640 /* We want to avoid cases where sel.length is not a multiple of npatterns.
10641 For eg: sel.length = 2 + 2x, and sel npatterns = 4. */
10642 poly_uint64 esel;
10644 {
10648 }
10654 {
10657 HOST_WIDE_INT step;
10659 {
10663 }
10664 // FIXME: Punt on step < 0 for now, revisit later.
10671 {
10675 }
10677 /* Ensure that stepped sequence of the pattern selects elements
10678 only from the same input vector. */
10687 {
10691 }
10693 /* Ensure that the stepped sequence always selects from the same
10694 input pattern. */
10699 {
10703 }
10705 /* If a1 chooses base element from arg, ensure that it's a natural
10706 stepped sequence, ie, (arg[2] - arg[1]) == (arg[1] - arg[0])
10707 to preserve arg's encoding. */
10710 {
10723 {
10727 }
10728 }
10729 }
10732 }
10734 /* Try to fold permutation of ARG0 and ARG1 with SEL selector when
10735 the input vectors are VECTOR_CST. Return NULL_TREE otherwise.
10736 REASON has same purpose as described in
10737 valid_mask_for_fold_vec_perm_cst_p. */
10739 static tree
10742 {
10746 /* (1) If SEL is a suitable mask as determined by
10747 valid_mask_for_fold_vec_perm_cst_p, then:
10748 res_npatterns = max of npatterns between ARG0, ARG1, and SEL
10749 res_nelts_per_pattern = max of nelts_per_pattern between
10750 ARG0, ARG1 and SEL.
10751 (2) If SEL is not a suitable mask, and TYPE is VLS then:
10752 res_npatterns = nelts in result vector.
10753 res_nelts_per_pattern = 1.
10754 This exception is made so that VLS ARG0, ARG1 and SEL work as before. */
10756 {
10757 res_npatterns
10762 res_nelts_per_pattern
10768 }
10770 {
10773 }
10774 else
10779 {
10782 poly_uint64 r;
10785 /* Punt if sel[i] /trunc_div len cannot be determined,
10786 because the input vector to be chosen will depend on
10787 runtime vector length.
10788 For example if len == 4 + 4x, and sel[i] == 4,
10789 If len at runtime equals 4, we choose arg1[0].
10790 For any other value of len > 4 at runtime, we choose arg0[4].
10791 which makes the element choice dependent on runtime vector length. */
10793 {
10797 }
10799 /* sel[i] % len will give the index of element in the chosen input
10800 vector. For example if sel[i] == 5 + 4x and len == 4 + 4x,
10801 we will choose arg1[1] since (5 + 4x) % (4 + 4x) == 1. */
10803 {
10807 }
10812 }
10815 }
10817 /* Attempt to fold vector permutation of ARG0 and ARG1 vectors using SEL
10818 selector. Return the folded VECTOR_CST or CONSTRUCTOR if successful,
10819 NULL_TREE otherwise. */
10821 tree
10823 {
10839 /* For fall back case, we want to ensure we have VLS vectors
10840 with equal length. */
10854 {
10855 HOST_WIDE_INT index;
10859 }
10861 }
10863 /* Try to fold a pointer difference of type TYPE two address expressions of
10864 array references AREF0 and AREF1 using location LOC. Return a
10865 simplified expression for the difference or NULL_TREE. */
10867 static tree
10871 {
10876 /* If the bases are array references as well, recurse. If the bases
10877 are pointer indirections compute the difference of the pointers.
10878 If the bases are equal, we are set. */
10881 && (base_offset
10883 use_pointer_diff)))
10886 && (base_offset
10887 = use_pointer_diff
10897 {
10903 base_offset,
10906 }
10908 }
10910 /* If the real or vector real constant CST of type TYPE has an exact
10911 inverse, return it, else return NULL. */
10913 tree
10915 {
10916 REAL_VALUE_TYPE r;
10917 tree unit_type;
10918 machine_mode mode;
10921 {
10931 {
10935 tree_vector_builder elts;
10940 {
10945 }
10948 }
10952 }
10953 }
10955 /* Mask out the tz least significant bits of X of type TYPE where
10956 tz is the number of trailing zeroes in Y. */
10957 static wide_int
10959 {
10964 }
10966 /* Return true when T is an address and is known to be nonzero.
10967 For floating point we further ensure that T is not denormal.
10968 Similar logic is present in nonzero_address in rtlanal.h.
10970 If the return value is based on the assumption that signed overflow
10971 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
10972 change *STRICT_OVERFLOW_P. */
10974 static bool
10976 {
10980 /* Doing something useful for floating point would need more work. */
10986 {
10989 strict_overflow_p);
10995 strict_overflow_p);
11003 }
11006 {
11009 strict_overflow_p);
11017 strict_overflow_p);
11031 strict_overflow_p);
11035 strict_overflow_p);
11038 {
11050 }
11054 }
11056 }
11058 /* Return true when T is an address and is known to be nonzero.
11059 Handle warnings about undefined signed overflow. */
11061 bool
11063 {
11070 "determining that expression is always "
11072 WARN_STRICT_OVERFLOW_MISC);
11074 }
11076 /* Return true if T is known not to be equal to an integer W. */
11078 bool
11080 {
11081 int_range_max vr;
11083 {
11094 /* If T has some known zero bits and W has any of those bits set,
11095 then T is known not to be equal to W. */
11103 }
11104 }
11106 /* Fold a binary expression of code CODE and type TYPE with operands
11107 OP0 and OP1. LOC is the location of the resulting expression.
11108 Return the folded expression if folding is successful. Otherwise,
11109 return NULL_TREE. */
11111 tree
11114 {
11129 /* Strip any conversions that don't change the mode. This is
11130 safe for every expression, except for a comparison expression
11131 because its signedness is derived from its operands. So, in
11132 the latter case, only strip conversions that don't change the
11133 signedness. MIN_EXPR/MAX_EXPR also need signedness of arguments
11134 preserved.
11136 Note that this is done as an internal manipulation within the
11137 constant folder, in order to find the simplest representation
11138 of the arguments so that their form can be studied. In any
11139 cases, the appropriate type conversions should be put back in
11140 the tree that will get out of the constant folder. */
11143 {
11146 }
11147 else
11148 {
11151 }
11153 /* Note that TREE_CONSTANT isn't enough: static var addresses are
11154 constant but we can't do arithmetic on them. */
11156 {
11159 {
11163 }
11164 }
11166 /* If this is a commutative operation, and ARG0 is a constant, move it
11167 to ARG1 to reduce the number of tests below. */
11172 /* Likewise if this is a comparison, and ARG0 is a constant, move it
11173 to ARG1 to reduce the number of tests below. */
11182 /* ARG0 is the first operand of EXPR, and ARG1 is the second operand.
11184 First check for cases where an arithmetic operation is applied to a
11185 compound, conditional, or comparison operation. Push the arithmetic
11186 operation inside the compound or conditional to see if any folding
11187 can then be done. Convert comparison to conditional for this purpose.
11188 The also optimizes non-constant cases that used to be done in
11189 expand_expr.
11191 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
11192 one of the operands is a comparison and the other is a comparison, a
11193 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
11194 code below would make the expression more complex. Change it to a
11195 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
11196 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
11209 {
11213 boolean_type_node,
11221 }
11225 {
11227 {
11232 tem);
11233 }
11235 {
11240 tem);
11241 }
11246 {
11252 }
11257 {
11263 }
11264 }
11267 {
11269 /* MEM[&MEM[p, CST1], CST2] -> MEM[p, CST1 + CST2]. */
11272 {
11278 }
11280 /* MEM[&a.b, CST2] -> MEM[&a, offsetof (a, b) + CST2]. */
11283 {
11284 tree base;
11285 poly_int64 coffset;
11294 }
11299 /* INT +p INT -> (PTR)(INT + INT). Stripping types allows for this. */
11305 arg1),
11307 arg0)));
11313 {
11314 /* X + (X / CST) * -CST is X % CST. */
11319 {
11328 cst0));
11329 }
11330 }
11332 /* Handle (A1 * C1) + (A2 * C2) with A1, A2 or C1, C2 being the same or
11333 one. Make sure the type is not saturating and has the signedness of
11334 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11335 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11342 {
11346 }
11349 {
11350 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
11351 (plus (plus (mult) (mult)) (foo)) so that we can
11352 take advantage of the factoring cases below. */
11361 {
11367 else
11380 parg0),
11382 marg)),
11386 return
11392 parg1)));
11393 }
11394 }
11395 else
11396 {
11397 /* Fold __complex__ ( x, 0 ) + __complex__ ( 0, y )
11398 to __complex__ ( x, y ). This is not the same for SNaNs or
11399 if signed zeros are involved. */
11403 {
11410 {
11414 {
11420 }
11422 {
11428 }
11429 }
11430 }
11432 /* Convert a + (b*c + d*e) into (a + b*c) + d*e.
11433 We associate floats only if the user has specified
11434 -fassociative-math. */
11438 {
11443 {
11444 tree tree0;
11447 }
11448 }
11449 /* Convert (b*c + d*e) + a into b*c + (d*e +a).
11450 We associate floats only if the user has specified
11451 -fassociative-math. */
11455 {
11460 {
11461 tree tree0;
11464 }
11465 }
11466 }
11468 bit_rotate:
11469 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
11470 is a rotate of A by C1 bits. */
11471 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
11472 is a rotate of A by B bits.
11473 Similarly for (A << B) | (A >> (-B & C3)) where C3 is Z-1,
11474 though in this case CODE must be | and not + or ^, otherwise
11475 it doesn't return A when B is 0. */
11476 {
11478 tree rtype;
11487 /* Only create rotates in complete modes. Other cases are not
11488 expanded properly. */
11491 {
11504 {
11509 }
11514 {
11517 code0 == LSHIFT_EXPR
11520 }
11522 {
11532 {
11536 orig_tree01);
11538 }
11539 }
11543 {
11554 {
11558 orig_tree01);
11560 }
11561 }
11562 }
11563 }
11565 associate:
11566 /* In most languages, can't associate operations on floats through
11567 parentheses. Rather than remember where the parentheses were, we
11568 don't associate floats at all, unless the user has specified
11569 -fassociative-math.
11570 And, we need to make sure type is not saturating. */
11575 {
11581 /* Split both trees into variables, constants, and literals. Then
11582 associate each group together, the constants with literals,
11583 then the result with variables. This increases the chances of
11584 literals being recombined later and of generating relocatable
11585 expressions for the sum of a constant and literal. */
11593 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
11597 /* With undefined overflow prefer doing association in a type
11598 which wraps on overflow, if that is one of the operand types. */
11601 {
11609 }
11611 /* With undefined overflow we can only associate constants with one
11612 variable, and constants whose association doesn't overflow. */
11615 {
11617 {
11618 /* ??? If split_tree would handle NEGATE_EXPR we could
11619 simply reject these cases and the allowed cases would
11620 be the var0/minus_var1 ones. */
11626 {
11629 }
11636 {
11639 }
11645 /* The only case we can still associate with two variables
11646 is if they cancel out. */
11650 }
11656 }
11658 /* Only do something if we found more than two objects. Otherwise,
11659 nothing has changed and we risk infinite recursion. */
11667 {
11679 {
11683 }
11685 {
11689 }
11691 /* Preserve the MINUS_EXPR if the negative part of the literal is
11692 greater than the positive part. Otherwise, the multiplicative
11693 folding code (i.e extract_muldiv) may be fooled in case
11694 unsigned constants are subtracted, like in the following
11695 example: ((X*2 + 4) - 8U)/2. */
11697 {
11701 /* But avoid ending up with only negated parts. */
11703 {
11707 }
11708 else
11709 {
11713 }
11714 }
11716 /* Don't introduce overflows through reassociation. */
11721 /* Eliminate lit0 and minus_lit0 to con0 and minus_con0. */
11728 /* Eliminate minus_con0. */
11730 {
11737 else
11740 }
11742 /* Eliminate minus_var0. */
11744 {
11748 else
11751 }
11753 return
11756 }
11757 }
11763 /* Fold &a[i] - &a[j] to i-j. */
11768 {
11772 code
11776 }
11778 /* Further transformations are not for pointers. */
11782 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
11785 /* If arg0 is e.g. unsigned int and type is int, then this could
11786 introduce UB, because if A is INT_MIN at runtime, the original
11787 expression can be well defined while the latter is not.
11788 See PR83269. */
11797 /* Fold __complex__ ( x, 0 ) - __complex__ ( 0, y ) to
11798 __complex__ ( x, -y ). This is not the same for SNaNs or if
11799 signed zeros are involved. */
11803 {
11810 {
11814 {
11816 arg1r ? arg1r
11821 }
11823 {
11827 arg1i ? arg1i
11830 }
11831 }
11832 }
11834 /* A - B -> A + (-B) if B is easily negatable. */
11838 /* Avoid this transformation if B is a positive REAL_CST. */
11846 /* Handle (A1 * C1) - (A2 * C2) with A1, A2 or C1, C2 being the same or
11847 one. Make sure the type is not saturating and has the signedness of
11848 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11849 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11856 {
11860 }
11866 {
11867 /* Transform x * -C into -x * C if x is easily negatable. */
11882 {
11885 "occur when simplifying "
11887 WARN_STRICT_OVERFLOW_MISC);
11889 }
11891 /* Optimize z * conj(z) for integer complex numbers. */
11898 }
11899 else
11900 {
11901 /* Fold z * +-I to __complex__ (-+__imag z, +-__real z).
11902 This is not the same for NaNs or if signed zeros are
11903 involved. */
11909 {
11912 return
11918 return
11923 }
11925 /* Optimize z * conj(z) for floating point complex numbers.
11926 Guarded by flag_unsafe_math_optimizations as non-finite
11927 imaginary components don't produce scalar results. */
11936 }
11940 /* Canonicalize (X & C1) | C2. */
11944 {
11949 /* If (C1&C2) == C1, then (X&C1)|C2 becomes (X,C2). */
11957 /* If (C1|C2) == ~0 then (X&C1)|C2 becomes X|C2. */
11959 {
11962 }
11964 /* Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2,
11965 unless (C1 & ~C2) | (C2 & C3) for some C3 is a mask of some
11966 mode which allows further optimizations. */
11971 {
11976 {
11979 }
11980 }
11983 {
11988 }
11989 }
11991 /* See if this can be simplified into a rotate first. If that
11992 is unsuccessful continue in the association code. */
11996 /* Fold (X & 1) ^ 1 as (X & 1) == 0. */
12004 /* See if this can be simplified into a rotate first. If that
12005 is unsuccessful continue in the association code. */
12009 /* Fold (X ^ 1) & 1 as (X & 1) == 0. */
12014 {
12015 tree tem2;
12022 }
12023 /* Fold ~X & 1 as (X & 1) == 0. */
12027 {
12028 tree tem2;
12035 }
12036 /* Fold !X & 1 as X == 0. */
12039 {
12043 }
12045 /* Fold (X * Y) & -(1 << CST) to X * Y if Y is a constant
12046 multiple of 1 << CST. */
12048 {
12055 }
12057 /* Fold (X * CST1) & CST2 to zero if we can, or drop known zero
12058 bits from CST2. */
12062 {
12064 wide_int masked
12071 {
12072 /* Avoid the transform if arg1 is a mask of some
12073 mode which allows further optimizations. */
12080 }
12081 }
12083 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
12086 {
12091 return
12093 }
12098 /* Don't touch a floating-point divide by zero unless the mode
12099 of the constant can represent infinity. */
12105 /* (-A) / (-B) -> A / B */
12117 /* Fall through */
12120 /* Simplify A / (B << N) where A and B are positive and B is
12121 a power of 2, to A >> (N + log2(B)). */
12126 {
12129 {
12137 WARN_STRICT_OVERFLOW_MISC);
12143 }
12144 }
12146 /* Fall through */
12154 /* Convert -A / -B to A / B when the type is signed and overflow is
12155 undefined. */
12159 {
12162 "when distributing negation across "
12164 WARN_STRICT_OVERFLOW_MISC);
12169 }
12173 {
12176 "when distributing negation across "
12178 WARN_STRICT_OVERFLOW_MISC);
12183 }
12185 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
12186 operation, EXACT_DIV_EXPR.
12188 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
12189 At one time others generated faster code, it's not clear if they do
12190 after the last round to changes to the DIV code in expmed.cc. */
12201 {
12205 WARN_STRICT_OVERFLOW_MISC);
12207 }
12219 {
12223 WARN_STRICT_OVERFLOW_MISC);
12225 }
12233 /* Since negative shift count is not well-defined,
12234 don't try to compute it in the compiler. */
12240 /* If we have a rotate of a bit operation with the rotate count and
12241 the second operand of the bit operation both constant,
12242 permute the two operations. */
12248 {
12256 }
12258 /* Two consecutive rotates adding up to the some integer
12259 multiple of the precision of the type can be ignored. */
12275 /* Note that the operands of this must be ints
12276 and their values must be 0 or 1.
12277 ("true" is a fixed value perhaps depending on the language.) */
12278 /* If first arg is constant zero, return it. */
12281 /* FALLTHRU */
12283 /* If either arg is constant true, drop it. */
12287 /* Preserve sequence points. */
12290 /* If second arg is constant zero, result is zero, but first arg
12291 must be evaluated. */
12294 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
12295 case will be handled here. */
12299 /* !X && X is always false. */
12303 /* X && !X is always false. */
12308 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
12309 means A >= Y && A != MAX, but in this case we know that
12310 A < X <= MAX. */
12314 {
12326 }
12335 /* Note that the operands of this must be ints
12336 and their values must be 0 or true.
12337 ("true" is a fixed value perhaps depending on the language.) */
12338 /* If first arg is constant true, return it. */
12341 /* FALLTHRU */
12343 /* If either arg is constant zero, drop it. */
12347 /* Preserve sequence points. */
12350 /* If second arg is constant true, result is true, but we must
12351 evaluate first arg. */
12354 /* Likewise for first arg, but note this only occurs here for
12355 TRUTH_OR_EXPR. */
12359 /* !X || X is always true. */
12363 /* X || !X is always true. */
12368 /* (X && !Y) || (!X && Y) is X ^ Y */
12371 {
12388 }
12397 /* If the second arg is constant zero, drop it. */
12400 /* If the second arg is constant true, this is a logical inversion. */
12402 {
12405 }
12406 /* Identical arguments cancel to zero. */
12410 /* !X ^ X is always true. */
12415 /* X ^ !X is always true. */
12431 /* bool_var != 1 becomes !bool_var. */
12438 /* bool_var == 0 becomes !bool_var. */
12445 /* !exp != 0 becomes !exp */
12450 /* If this is an EQ or NE comparison with zero and ARG0 is
12451 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
12452 two operations, but the latter can be done in one less insn
12453 on machines that have only two-operand insns or on which a
12454 constant cannot be the first operand. */
12457 {
12462 {
12470 }
12473 {
12481 }
12482 }
12484 /* If this is a comparison of a field, we may be able to simplify it. */
12487 /* Handle the constant case even without -O
12488 to make sure the warnings are given. */
12490 {
12494 }
12496 /* Optimize comparisons of strlen vs zero to a compare of the
12497 first character of the string vs zero. To wit,
12498 strlen(ptr) == 0 => *ptr == 0
12499 strlen(ptr) != 0 => *ptr != 0
12500 Other cases should reduce to one of these two (or a constant)
12501 due to the return value of strlen being unsigned. */
12503 {
12511 {
12512 tree ptrtype
12514 TYPE_QUAL_CONST));
12520 }
12521 }
12523 /* Fold (X >> C) != 0 into X < 0 if C is one less than the width
12524 of X. Similarly fold (X >> C) == 0 into X >= 0. */
12528 {
12533 {
12535 {
12538 }
12541 }
12542 }
12544 /* Fold (~X & C) == 0 into (X & C) != 0 and (~X & C) != 0 into
12545 (X & C) == 0 when C is a single bit. */
12550 {
12557 arg1));
12558 }
12560 /* Fold ((X & C) ^ C) eq/ne 0 into (X & C) ne/eq 0, when the
12561 constant C is a power of two, i.e. a single bit. */
12568 {
12572 }
12574 /* Likewise, fold ((X ^ C) & C) eq/ne 0 into (X & C) ne/eq 0,
12575 when is C is a power of two, i.e. a single bit. */
12582 {
12588 }
12592 {
12599 /* Optimize (X ^ Z) op (Y ^ Z) as X op Y, and symmetries.
12600 operand_equal_p guarantees no side-effects so we don't need
12601 to use omit_one_operand on Z. */
12605 arg10));
12609 arg11));
12613 arg10));
12617 arg11));
12619 /* Optimize (X ^ C1) op (Y ^ C2) as (X ^ (C1 ^ C2)) op Y. */
12622 {
12628 }
12629 }
12631 /* Attempt to simplify equality/inequality comparisons of complex
12632 values. Only lower the comparison if the result is known or
12633 can be simplified to a single scalar comparison. */
12638 {
12643 {
12646 }
12647 else
12648 {
12651 }
12654 {
12657 }
12658 else
12659 {
12662 }
12666 {
12668 {
12673 }
12674 else
12675 {
12680 }
12681 }
12685 {
12687 {
12692 }
12693 else
12694 {
12699 }
12700 }
12701 }
12713 /* Transform comparisons of the form X +- C CMP X. */
12718 {
12723 /* (X - c) > X becomes false. */
12729 /* Likewise (X + c) < X becomes false. */
12735 /* Convert (X - c) <= X to true. */
12742 /* Convert (X + c) >= X to true. */
12748 }
12750 /* If we are comparing an ABS_EXPR with a constant, we can
12751 convert all the cases into explicit comparisons, but they may
12752 well not be faster than doing the ABS and one comparison.
12753 But ABS (X) <= C is a range comparison, which becomes a subtraction
12754 and a comparison, and is probably faster. */
12768 /* Convert ABS_EXPR<x> >= 0 to true. */
12775 {
12778 "when simplifying comparison of "
12780 WARN_STRICT_OVERFLOW_CONDITIONAL);
12783 arg0);
12784 }
12786 /* Convert ABS_EXPR<x> < 0 to false. */
12791 {
12794 "when simplifying comparison of "
12796 WARN_STRICT_OVERFLOW_CONDITIONAL);
12799 arg0);
12800 }
12802 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
12803 and similarly for >= into !=. */
12813 /* Similarly for X < (cast) (1 << Y). But cast can't be narrowing,
12814 otherwise Y might be >= # of bits in X's type and thus e.g.
12815 (unsigned char) (1 << Y) for Y 15 might be 0.
12816 If the cast is widening, then 1 << Y should have unsigned type,
12817 otherwise if Y is number of bits in the signed shift type minus 1,
12818 we can't optimize this. E.g. (unsigned long long) (1 << Y) for Y
12819 31 might be 0xffffffff80000000. */
12832 {
12838 }
12850 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
12851 {
12863 }
12868 /* When pedantic, a compound expression can be neither an lvalue
12869 nor an integer constant expression. */
12872 /* Don't let (0, 0) be null pointer constant. */
12880 }
12882 /* For constants M and N, if M == (1LL << cst) - 1 && (N & M) == M,
12883 ((A & N) + B) & M -> (A + B) & M
12884 Similarly if (N & M) == 0,
12885 ((A | N) + B) & M -> (A + B) & M
12886 and for - instead of + (or unary - instead of +)
12887 and/or ^ instead of |.
12888 If B is constant and (B & M) == 0, fold into A & M.
12890 This function is a helper for match.pd patterns. Return non-NULL
12891 type in which the simplified operation should be performed only
12892 if any optimization is possible.
12894 ARG1 is M above, ARG00 is left operand of +/-, if CODE00 is BIT_*_EXPR,
12895 then ARG00{0,1} are operands of that bitop, otherwise CODE00 is ERROR_MARK.
12896 Similarly for ARG01, CODE01 and ARG01{0,1}, just for the right operand of
12897 +/-. */
12898 tree
12903 {
12918 wide_int cst0;
12920 /* Now we know that arg0 is (C + D) or (C - D) or -C and
12921 arg1 (M) is == (1LL << cst) - 1.
12922 Store C into PMOP[0] and D into PMOP[1]. */
12929 {
12936 {
12939 }
12942 /* If C or D is of the form (A & N) where
12943 (N & M) == M, or of the form (A | N) or
12944 (A ^ N) where (N & M) == 0, replace it with A. */
12950 /* If C or D is a N where (N & M) == 0, it can be
12951 omitted (replaced with 0). */
12956 /* Similarly, with C - N where (-N & M) == 0. */
12964 }
12966 /* Only build anything new if we optimized one or both arguments above. */
12972 else
12974 }
12976 /* Used by contains_label_[p1]. */
12978 struct contains_label_data
12979 {
12982 };
12984 /* Callback for walk_tree, looking for LABEL_EXPR. Return *TP if it is
12985 a LABEL_EXPR or CASE_LABEL_EXPR not inside of another SWITCH_EXPR; otherwise
12986 return NULL_TREE. Do not check the subtrees of GOTO_EXPR. */
12988 static tree
12990 {
12993 {
13004 {
13012 }
13021 }
13022 }
13024 /* Return whether the sub-tree ST contains a label which is accessible from
13025 outside the sub-tree. */
13027 static bool
13029 {
13033 }
13035 /* Fold a ternary expression of code CODE and type TYPE with operands
13036 OP0, OP1, and OP2. Return the folded expression if folding is
13037 successful. Otherwise, return NULL_TREE. */
13039 tree
13042 {
13043 tree tem;
13050 /* If this is a commutative operation, and OP0 is a constant, move it
13051 to OP1 to reduce the number of tests below. */
13060 /* Strip any conversions that don't change the mode. This is safe
13061 for every expression, except for a comparison expression because
13062 its signedness is derived from its operands. So, in the latter
13063 case, only strip conversions that don't change the signedness.
13065 Note that this is done as an internal manipulation within the
13066 constant folder, in order to find the simplest representation of
13067 the arguments so that their form can be studied. In any cases,
13068 the appropriate type conversions should be put back in the tree
13069 that will get out of the constant folder. */
13071 {
13074 }
13077 {
13080 }
13083 {
13086 }
13089 {
13093 {
13099 }
13104 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
13105 so all simple results must be passed through pedantic_non_lvalue. */
13107 {
13110 /* Only optimize constant conditions when the selected branch
13111 has the same type as the COND_EXPR. This avoids optimizing
13112 away "c ? x : throw", where the throw has a void type.
13113 Avoid throwing away that operand which contains label. */
13120 }
13122 {
13129 {
13132 {
13140 }
13145 }
13146 }
13148 /* If we have A op B ? A : C, we may be able to convert this to a
13149 simpler expression, depending on the operation and the values
13150 of B and C. Signed zeros prevent all of these transformations,
13151 for reasons given above each one.
13153 Also try swapping the arguments and inverting the conditional. */
13157 {
13164 }
13169 {
13176 arg00,
13177 arg01,
13181 }
13183 /* If the second operand is simpler than the third, swap them
13184 since that produces better jump optimization results. */
13187 {
13189 /* See if this can be inverted. If it can't, possibly because
13190 it was a floating-point inequality comparison, don't do
13191 anything. */
13195 }
13197 /* Convert A ? 1 : 0 to simply A. */
13202 /* If we try to convert OP0 to our type, the
13203 call to fold will try to move the conversion inside
13204 a COND, which will recurse. In that case, the COND_EXPR
13205 is probably the best choice, so leave it alone. */
13209 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
13210 over COND_EXPR in cases such as floating point comparisons. */
13219 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
13224 {
13225 /* sign_bit_p looks through both zero and sign extensions,
13226 but for this optimization only sign extensions are
13227 usable. */
13230 {
13233 {
13236 }
13238 }
13239 /* sign_bit_p only checks ARG1 bits within A's precision.
13240 If <sign bit of A> has wider type than A, bits outside
13241 of A's precision in <sign bit of A> need to be checked.
13242 If they are all 0, this optimization needs to be done
13243 in unsigned A's type, if they are all 1 in signed A's type,
13244 otherwise this can't be done. */
13250 {
13252 tree tem_type;
13265 {
13268 }
13270 {
13273 }
13274 else
13276 }
13279 return
13285 arg1)));
13286 }
13288 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
13289 already handled above. */
13294 {
13304 op1);
13305 }
13307 /* A & N ? N : 0 is simply A & N if N is a power of two. This
13308 is probably obsolete because the first operand should be a
13309 truth value (that's why we have the two cases above), but let's
13310 leave it in until we can confirm this for all front-ends. */
13318 /* operand_equal_p compares just value, not precision, so e.g.
13319 arg1 could be 8-bit -128 and be power of two, but BIT_AND_EXPR
13320 second operand 32-bit -128, which is not a power of two (or vice
13321 versa. */
13325 /* Disable the transformations below for vectors, since
13326 fold_binary_op_with_conditional_arg may undo them immediately,
13327 yielding an infinite loop. */
13331 /* Convert A ? B : 0 into A && B if A and B are truth values. */
13340 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
13345 {
13347 /* Only perform transformation if ARG0 is easily inverted. */
13351 ? BIT_IOR_EXPR
13354 op1);
13355 }
13357 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
13362 {
13364 /* Only perform transformation if ARG0 is easily inverted. */
13370 op2);
13371 }
13373 /* Convert A ? 1 : B into A || B if A and B are truth values. */
13385 /* CALL_EXPRs used to be ternary exprs. Catch any mistaken uses
13386 of fold_ternary on them. */
13396 {
13398 unsigned HOST_WIDE_INT width
13409 {
13414 {
13416 {
13421 }
13427 }
13428 }
13429 }
13431 /* On constants we can use native encode/interpret to constant
13432 fold (nearly) all BIT_FIELD_REFs. */
13438 {
13441 /* Limit us to a reasonable amount of work. To relax the
13442 other limitations we need bit-shifting of the buffer
13443 and rounding up the size. */
13447 {
13449 unsigned HOST_WIDE_INT len
13454 {
13459 }
13460 }
13461 }
13466 /* Perform constant folding of BIT_INSERT_EXPR. */
13470 {
13471 /* Build a vector of integers from the tree mask. */
13472 vec_perm_builder builder;
13476 /* Create a vec_perm_indices for the integer vector. */
13481 }
13485 /* Perform (partial) constant folding of BIT_INSERT_EXPR. */
13488 {
13494 wide_int tem2
13498 }
13503 {
13505 unsigned HOST_WIDE_INT elsize
13508 {
13514 {
13520 }
13521 }
13522 }
13528 }
13530 /* Gets the element ACCESS_INDEX from CTOR, which must be a CONSTRUCTOR
13531 of an array (or vector). *CTOR_IDX if non-NULL is updated with the
13532 constructor element index of the value returned. If the element is
13533 not found NULL_TREE is returned and *CTOR_IDX is updated to
13534 the index of the element after the ACCESS_INDEX position (which
13535 may be outside of the CTOR array). */
13537 tree
13540 {
13546 {
13549 {
13550 /* Static constructors for variably sized objects makes no sense. */
13553 /* ??? When it is obvious that the range is signed, treat it so. */
13558 {
13560 low_bound
13562 SIGNED);
13563 }
13564 else
13565 {
13568 }
13569 }
13570 }
13574 index_sgn);
13586 {
13587 /* Array constructor might explicitly set index, or specify a range,
13588 or leave index NULL meaning that it is next index after previous
13589 one. */
13591 {
13593 max_index = index
13595 else
13596 {
13599 index_sgn);
13600 max_index
13602 index_sgn);
13604 }
13605 }
13607 {
13613 }
13614 else
13617 /* Do we have match? */
13619 {
13621 {
13625 }
13626 }
13628 /* We're past the element we search for. Note during parsing
13629 the elements might not be sorted.
13630 ??? We should use a binary search and a flag on the
13631 CONSTRUCTOR as to whether elements are sorted in declaration
13632 order. */
13634 }
13638 }
13640 /* Perform constant folding and related simplification of EXPR.
13641 The related simplifications include x*1 => x, x*0 => 0, etc.,
13642 and application of the associative law.
13643 NOP_EXPR conversions may be removed freely (as long as we
13644 are careful not to change the type of the overall expression).
13645 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
13646 but we can constant-fold them if they have constant operands. */
13648 #ifdef ENABLE_FOLD_CHECKING
13649 # define fold(x) fold_1 (x)
13651 static
13652 #endif
13653 tree
13655 {
13659 tree tem;
13662 /* Return right away if a constant. */
13666 /* CALL_EXPR-like objects with variable numbers of operands are
13667 treated specially. */
13669 {
13671 {
13674 }
13676 }
13679 {
13684 {
13702 }
13703 }
13706 {
13708 {
13715 {
13720 }
13723 }
13725 /* Return a VECTOR_CST if possible. */
13727 {
13733 tree val;
13739 }
13747 }
13749 #ifdef ENABLE_FOLD_CHECKING
13750 #undef fold
13757 /* When --enable-checking=fold, compute a digest of expr before
13758 and after actual fold call to see if fold did not accidentally
13759 change original expr. */
13761 tree
13763 {
13764 tree ret;
13784 }
13786 void
13788 {
13799 }
13801 static void
13803 {
13805 }
13807 static void
13810 {
13816 recursive_label:
13826 {
13827 /* Allow DECL_ASSEMBLER_NAME and symtab_node to be modified. */
13835 }
13843 {
13844 /* Allow these fields to be modified. */
13845 tree tmp;
13856 {
13859 }
13860 }
13862 {
13863 /* Allow the no-warning bit to be set. Perhaps we shouldn't allow
13864 that and change builtins.cc etc. instead - see PR89543. */
13870 }
13881 {
13884 {
13900 }
13904 {
13917 }
13934 {
13940 }
13943 {
13945 {
13948 }
13950 }
13961 {
13964 }
13972 }
13973 }
13975 /* Helper function for outputting the checksum of a tree T. When
13976 debugging with gdb, you can "define mynext" to be "next" followed
13977 by "call debug_fold_checksum (op0)", then just trace down till the
13978 outputs differ. */
13980 DEBUG_FUNCTION void
13982 {
13997 }
13999 #endif
14001 /* Fold a unary tree expression with code CODE of type TYPE with an
14002 operand OP0. LOC is the location of the resulting expression.
14003 Return a folded expression if successful. Otherwise, return a tree
14004 expression with code CODE of type TYPE with an operand OP0. */
14006 tree
14009 {
14010 tree tem;
14011 #ifdef ENABLE_FOLD_CHECKING
14020 #endif
14026 #ifdef ENABLE_FOLD_CHECKING
14033 #endif
14035 }
14037 /* Fold a binary tree expression with code CODE of type TYPE with
14038 operands OP0 and OP1. LOC is the location of the resulting
14039 expression. Return a folded expression if successful. Otherwise,
14040 return a tree expression with code CODE of type TYPE with operands
14041 OP0 and OP1. */
14043 tree
14046 MEM_STAT_DECL)
14047 {
14048 tree tem;
14049 #ifdef ENABLE_FOLD_CHECKING
14066 #endif
14072 #ifdef ENABLE_FOLD_CHECKING
14087 #endif
14089 }
14091 /* Fold a ternary tree expression with code CODE of type TYPE with
14092 operands OP0, OP1, and OP2. Return a folded expression if
14093 successful. Otherwise, return a tree expression with code CODE of
14094 type TYPE with operands OP0, OP1, and OP2. */
14096 tree
14099 {
14100 tree tem;
14101 #ifdef ENABLE_FOLD_CHECKING
14125 #endif
14132 #ifdef ENABLE_FOLD_CHECKING
14155 #endif
14157 }
14159 /* Fold a CALL_EXPR expression of type TYPE with operands FN and NARGS
14160 arguments in ARGARRAY, and a null static chain.
14161 Return a folded expression if successful. Otherwise, return a CALL_EXPR
14162 of type TYPE from the given operands as constructed by build_call_array. */
14164 tree
14167 {
14168 tree tem;
14169 #ifdef ENABLE_FOLD_CHECKING
14188 #endif
14194 #ifdef ENABLE_FOLD_CHECKING
14210 #endif
14212 }
14214 /* Perform constant folding and related simplification of initializer
14215 expression EXPR. These behave identically to "fold_buildN" but ignore
14216 potential run-time traps and exceptions that fold must preserve. */
14218 #define START_FOLD_INIT \
14219 int saved_signaling_nans = flag_signaling_nans;\
14220 int saved_trapping_math = flag_trapping_math;\
14221 int saved_rounding_math = flag_rounding_math;\
14222 int saved_trapv = flag_trapv;\
14223 int saved_folding_initializer = folding_initializer;\
14224 flag_signaling_nans = 0;\
14225 flag_trapping_math = 0;\
14226 flag_rounding_math = 0;\
14227 flag_trapv = 0;\
14228 folding_initializer = 1;
14230 #define END_FOLD_INIT \
14231 flag_signaling_nans = saved_signaling_nans;\
14232 flag_trapping_math = saved_trapping_math;\
14233 flag_rounding_math = saved_rounding_math;\
14234 flag_trapv = saved_trapv;\
14235 folding_initializer = saved_folding_initializer;
14237 tree
14239 {
14240 tree result;
14241 START_FOLD_INIT;
14245 END_FOLD_INIT;
14247 }
14249 tree
14252 {
14253 tree result;
14254 START_FOLD_INIT;
14258 END_FOLD_INIT;
14260 }
14262 tree
14265 {
14266 tree result;
14267 START_FOLD_INIT;
14271 END_FOLD_INIT;
14273 }
14275 tree
14278 {
14279 tree result;
14280 START_FOLD_INIT;
14284 END_FOLD_INIT;
14286 }
14288 tree
14291 {
14292 tree result;
14293 START_FOLD_INIT;
14297 END_FOLD_INIT;
14299 }
14301 #undef START_FOLD_INIT
14302 #undef END_FOLD_INIT
14304 /* Determine if first argument is a multiple of second argument. Return
14305 false if it is not, or we cannot easily determined it to be.
14307 An example of the sort of thing we care about (at this point; this routine
14308 could surely be made more general, and expanded to do what the *_DIV_EXPR's
14309 fold cases do now) is discovering that
14311 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14313 is a multiple of
14315 SAVE_EXPR (J * 8)
14317 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
14319 This code also handles discovering that
14321 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14323 is a multiple of 8 so we don't have to worry about dealing with a
14324 possible remainder.
14326 Note that we *look* inside a SAVE_EXPR only to determine how it was
14327 calculated; it is not safe for fold to do much of anything else with the
14328 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
14329 at run time. For example, the latter example above *cannot* be implemented
14330 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
14331 evaluation time of the original SAVE_EXPR is not necessarily the same at
14332 the time the new expression is evaluated. The only optimization of this
14333 sort that would be valid is changing
14335 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
14337 divided by 8 to
14339 SAVE_EXPR (I) * SAVE_EXPR (J)
14341 (where the same SAVE_EXPR (J) is used in the original and the
14342 transformed version).
14344 NOWRAP specifies whether all outer operations in TYPE should
14345 be considered not wrapping. Any type conversion within TOP acts
14346 as a barrier and we will fall back to NOWRAP being false.
14347 NOWRAP is mostly used to treat expressions in TYPE_SIZE and friends
14348 as not wrapping even though they are generally using unsigned arithmetic. */
14350 bool
14352 {
14363 {
14365 /* Bitwise and provides a power of two multiple. If the mask is
14366 a multiple of BOTTOM then TOP is a multiple of BOTTOM. */
14373 /* If the multiplication can wrap we cannot recurse further unless
14374 the bottom is a power of two which is where wrapping does not
14375 matter. */
14381 {
14387 {
14390 /* Handle multiple_of_p ((x * 2 + 2) * 4, 8). */
14392 {
14396 {
14399 }
14400 }
14402 }
14403 }
14408 /* Handle X << CST as X * (1 << CST) and only process the constant. */
14410 {
14413 {
14414 wide_int mul_op
14418 nowrap);
14419 }
14420 }
14425 /* If the addition or subtraction can wrap we cannot recurse further
14426 unless bottom is a power of two which is where wrapping does not
14427 matter. */
14433 /* Handle cases like op0 + 0xfffffffd as op0 - 3 if the expression has
14434 unsigned type. For example, (X / 3) + 0xfffffffd is multiple of 3,
14435 but 0xfffffffd is not. */
14438 && nowrap
14443 /* It is impossible to prove if op0 +- op1 is multiple of bottom
14444 precisely, so be conservative here checking if both op0 and op1
14445 are multiple of bottom. Note we check the second operand first
14446 since it's usually simpler. */
14450 CASE_CONVERT:
14451 /* Can't handle conversions from non-integral or wider integral type. */
14456 /* NOWRAP only extends to operations in the outermost type so
14457 make sure to strip it off here. */
14472 SIGNED);
14478 {
14481 /* Check for special cases to see if top is defined as multiple
14482 of bottom:
14484 top = (X & ~(bottom - 1) ; bottom is power of 2
14486 or
14488 Y = X % bottom
14489 top = X - Y. */
14508 }
14510 /* fall through */
14518 }
14519 }
14521 /* Return true if expression X cannot be (or contain) a NaN or infinity.
14522 This function returns true for integer expressions, and returns
14523 false if uncertain. */
14525 bool
14527 {
14532 {
14555 {
14556 CASE_CFN_FABS:
14557 CASE_CFN_FABS_FN:
14559 CASE_CFN_FMAX:
14560 CASE_CFN_FMAX_FN:
14561 CASE_CFN_FMIN:
14562 CASE_CFN_FMIN_FN:
14567 }
14571 }
14572 }
14574 /* Return true if expression X evaluates to an infinity.
14575 This function returns false for integer expressions. */
14577 bool
14579 {
14583 {
14596 }
14597 }
14599 /* Return true if expression X could evaluate to an infinity.
14600 This function returns false for integer expressions, and returns
14601 true if uncertain. */
14603 bool
14605 {
14609 {
14622 }
14623 }
14625 /* Return true if expression X evaluates to a signaling NaN.
14626 This function returns false for integer expressions. */
14628 bool
14630 {
14634 {
14645 }
14646 }
14648 /* Return true if expression X could evaluate to a signaling NaN.
14649 This function returns false for integer expressions, and returns
14650 true if uncertain. */
14652 bool
14654 {
14658 {
14678 {
14679 CASE_CFN_FABS:
14680 CASE_CFN_FABS_FN:
14682 CASE_CFN_FMAX:
14683 CASE_CFN_FMAX_FN:
14684 CASE_CFN_FMIN:
14685 CASE_CFN_FMIN_FN:
14690 }
14693 }
14694 }
14696 /* Return true if expression X evaluates to a NaN.
14697 This function returns false for integer expressions. */
14699 bool
14701 {
14705 {
14716 }
14717 }
14719 /* Return true if expression X could evaluate to a NaN.
14720 This function returns false for integer expressions, and returns
14721 true if uncertain. */
14723 bool
14725 {
14729 {
14754 {
14755 CASE_CFN_FABS:
14756 CASE_CFN_FABS_FN:
14758 CASE_CFN_FMAX:
14759 CASE_CFN_FMAX_FN:
14760 CASE_CFN_FMIN:
14761 CASE_CFN_FMIN_FN:
14766 }
14769 }
14770 }
14772 /* Return true if expression X could evaluate to -0.0.
14773 This function returns true if uncertain. */
14775 bool
14777 {
14781 {
14796 {
14797 CASE_CFN_FABS:
14798 CASE_CFN_FABS_FN:
14802 }
14805 }
14806 /* Ideally !(tree_expr_nonzero_p (X) || tree_expr_nonnegative_p (X))
14807 * but currently those predicates require tree and not const_tree. */
14809 }
14811 #define tree_expr_nonnegative_warnv_p(X, Y) \
14814 #define RECURSE(X) \
14815 ((tree_expr_nonnegative_warnv_p) (X, strict_overflow_p, depth + 1))
14817 /* Return true if CODE or TYPE is known to be non-negative. */
14819 static bool
14821 {
14825 /* Truth values evaluate to 0 or 1, which is nonnegative unless we
14826 have a signed:1 type (where the value is -1 and 0). */
14829 }
14831 /* Return true if (CODE OP0) is known to be non-negative. If the return
14832 value is based on the assumption that signed overflow is undefined,
14833 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14834 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14836 bool
14839 {
14844 {
14846 /* We can't return 1 if flag_wrapv is set because
14847 ABS_EXPR<INT_MIN> = INT_MIN. */
14851 {
14854 }
14862 CASE_CONVERT:
14863 {
14868 {
14872 {
14876 }
14877 }
14879 {
14885 }
14886 }
14891 }
14893 /* We don't know sign of `t', so be conservative and return false. */
14895 }
14897 /* Return true if (CODE OP0 OP1) is known to be non-negative. If the return
14898 value is based on the assumption that signed overflow is undefined,
14899 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14900 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14902 bool
14906 {
14911 {
14917 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
14918 both unsigned and at least 2 bits shorter than the result. */
14922 {
14927 {
14931 }
14932 }
14937 {
14938 /* x * x is always non-negative for floating point x
14939 or without overflow. */
14942 {
14947 }
14948 }
14950 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
14951 both unsigned and their total bits is shorter than the result. */
14955 {
14974 {
14984 }
14985 }
14992 /* Usually RECURSE (op0) || RECURSE (op1) but NaNs complicate
14993 things. */
15018 }
15020 /* We don't know sign of `t', so be conservative and return false. */
15022 }
15024 /* Return true if T is known to be non-negative. If the return
15025 value is based on the assumption that signed overflow is undefined,
15026 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15027 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15029 bool
15031 {
15036 {
15050 /* Limit the depth of recursion to avoid quadratic behavior.
15051 This is expected to catch almost all occurrences in practice.
15052 If this code misses important cases that unbounded recursion
15053 would not, passes that need this information could be revised
15054 to provide it through dataflow propagation. */
15062 }
15063 }
15065 /* Return true if T is known to be non-negative. If the return
15066 value is based on the assumption that signed overflow is undefined,
15067 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15068 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15070 bool
15073 {
15075 {
15076 CASE_CFN_ACOS:
15077 CASE_CFN_ACOS_FN:
15078 CASE_CFN_ACOSH:
15079 CASE_CFN_ACOSH_FN:
15080 CASE_CFN_CABS:
15081 CASE_CFN_CABS_FN:
15082 CASE_CFN_COSH:
15083 CASE_CFN_COSH_FN:
15084 CASE_CFN_ERFC:
15085 CASE_CFN_ERFC_FN:
15086 CASE_CFN_EXP:
15087 CASE_CFN_EXP_FN:
15088 CASE_CFN_EXP10:
15089 CASE_CFN_EXP2:
15090 CASE_CFN_EXP2_FN:
15091 CASE_CFN_FABS:
15092 CASE_CFN_FABS_FN:
15093 CASE_CFN_FDIM:
15094 CASE_CFN_FDIM_FN:
15095 CASE_CFN_HYPOT:
15096 CASE_CFN_HYPOT_FN:
15097 CASE_CFN_POW10:
15098 CASE_CFN_FFS:
15099 CASE_CFN_PARITY:
15100 CASE_CFN_POPCOUNT:
15101 CASE_CFN_CLZ:
15102 CASE_CFN_CLRSB:
15107 /* Always true. */
15110 CASE_CFN_SQRT:
15111 CASE_CFN_SQRT_FN:
15112 /* sqrt(-0.0) is -0.0. */
15117 CASE_CFN_ASINH:
15118 CASE_CFN_ASINH_FN:
15119 CASE_CFN_ATAN:
15120 CASE_CFN_ATAN_FN:
15121 CASE_CFN_ATANH:
15122 CASE_CFN_ATANH_FN:
15123 CASE_CFN_CBRT:
15124 CASE_CFN_CBRT_FN:
15125 CASE_CFN_CEIL:
15126 CASE_CFN_CEIL_FN:
15127 CASE_CFN_ERF:
15128 CASE_CFN_ERF_FN:
15129 CASE_CFN_EXPM1:
15130 CASE_CFN_EXPM1_FN:
15131 CASE_CFN_FLOOR:
15132 CASE_CFN_FLOOR_FN:
15133 CASE_CFN_FMOD:
15134 CASE_CFN_FMOD_FN:
15135 CASE_CFN_FREXP:
15136 CASE_CFN_FREXP_FN:
15137 CASE_CFN_ICEIL:
15138 CASE_CFN_IFLOOR:
15139 CASE_CFN_IRINT:
15140 CASE_CFN_IROUND:
15141 CASE_CFN_LCEIL:
15142 CASE_CFN_LDEXP:
15143 CASE_CFN_LFLOOR:
15144 CASE_CFN_LLCEIL:
15145 CASE_CFN_LLFLOOR:
15146 CASE_CFN_LLRINT:
15147 CASE_CFN_LLRINT_FN:
15148 CASE_CFN_LLROUND:
15149 CASE_CFN_LLROUND_FN:
15150 CASE_CFN_LRINT:
15151 CASE_CFN_LRINT_FN:
15152 CASE_CFN_LROUND:
15153 CASE_CFN_LROUND_FN:
15154 CASE_CFN_MODF:
15155 CASE_CFN_MODF_FN:
15156 CASE_CFN_NEARBYINT:
15157 CASE_CFN_NEARBYINT_FN:
15158 CASE_CFN_RINT:
15159 CASE_CFN_RINT_FN:
15160 CASE_CFN_ROUND:
15161 CASE_CFN_ROUND_FN:
15162 CASE_CFN_ROUNDEVEN:
15163 CASE_CFN_ROUNDEVEN_FN:
15164 CASE_CFN_SCALB:
15165 CASE_CFN_SCALBLN:
15166 CASE_CFN_SCALBLN_FN:
15167 CASE_CFN_SCALBN:
15168 CASE_CFN_SCALBN_FN:
15169 CASE_CFN_SIGNBIT:
15170 CASE_CFN_SIGNIFICAND:
15171 CASE_CFN_SINH:
15172 CASE_CFN_SINH_FN:
15173 CASE_CFN_TANH:
15174 CASE_CFN_TANH_FN:
15175 CASE_CFN_TRUNC:
15176 CASE_CFN_TRUNC_FN:
15177 /* True if the 1st argument is nonnegative. */
15180 CASE_CFN_FMAX:
15181 CASE_CFN_FMAX_FN:
15182 /* Usually RECURSE (arg0) || RECURSE (arg1) but NaNs complicate
15183 things. In the presence of sNaNs, we're only guaranteed to be
15184 non-negative if both operands are non-negative. In the presence
15185 of qNaNs, we're non-negative if either operand is non-negative
15186 and can't be a qNaN, or if both operands are non-negative. */
15195 CASE_CFN_FMIN:
15196 CASE_CFN_FMIN_FN:
15197 /* True if the 1st AND 2nd arguments are nonnegative. */
15200 CASE_CFN_COPYSIGN:
15201 CASE_CFN_COPYSIGN_FN:
15202 /* True if the 2nd argument is nonnegative. */
15205 CASE_CFN_POWI:
15206 /* True if the 1st argument is nonnegative or the second
15207 argument is an even integer. */
15213 CASE_CFN_POW:
15214 CASE_CFN_POW_FN:
15215 /* True if the 1st argument is nonnegative or the second
15216 argument is an even integer valued real. */
15218 {
15219 REAL_VALUE_TYPE c;
15220 HOST_WIDE_INT n;
15225 {
15226 REAL_VALUE_TYPE cint;
15230 }
15231 }
15236 }
15238 }
15240 /* Return true if T is known to be non-negative. If the return
15241 value is based on the assumption that signed overflow is undefined,
15242 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15243 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15245 static bool
15247 {
15253 {
15255 {
15259 /* If the initializer is non-void, then it's a normal expression
15260 that will be assigned to the slot. */
15264 /* Otherwise, the initializer sets the slot in some way. One common
15265 way is an assignment statement at the end of the initializer. */
15267 {
15275 else
15277 }
15283 }
15286 {
15292 arg0,
15293 arg1,
15295 }
15308 }
15309 }
15311 #undef RECURSE
15312 #undef tree_expr_nonnegative_warnv_p
15314 /* Return true if T is known to be non-negative. If the return
15315 value is based on the assumption that signed overflow is undefined,
15316 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15317 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15319 bool
15321 {
15328 {
15350 }
15353 {
15378 }
15379 }
15381 /* Return true if `t' is known to be non-negative. Handle warnings
15382 about undefined signed overflow. */
15384 bool
15386 {
15393 "determining that expression is always "
15395 WARN_STRICT_OVERFLOW_MISC);
15397 }
15400 /* Return true when (CODE OP0) is an address and is known to be nonzero.
15401 For floating point we further ensure that T is not denormal.
15402 Similar logic is present in nonzero_address in rtlanal.h.
15404 If the return value is based on the assumption that signed overflow
15405 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15406 change *STRICT_OVERFLOW_P. */
15408 bool
15411 {
15413 {
15416 strict_overflow_p);
15419 {
15425 strict_overflow_p));
15426 }
15431 strict_overflow_p);
15435 }
15438 }
15440 /* Return true when (CODE OP0 OP1) is an address and is known to be nonzero.
15441 For floating point we further ensure that T is not denormal.
15442 Similar logic is present in nonzero_address in rtlanal.h.
15444 If the return value is based on the assumption that signed overflow
15445 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15446 change *STRICT_OVERFLOW_P. */
15448 bool
15450 tree type,
15451 tree op0,
15453 {
15456 {
15460 {
15461 /* With the presence of negative values it is hard
15462 to say something. */
15469 /* One of operands must be positive and the other non-negative. */
15470 /* We don't set *STRICT_OVERFLOW_P here: even if this value
15471 overflows, on a twos-complement machine the sum of two
15472 nonnegative numbers can never be zero. */
15474 strict_overflow_p)
15476 strict_overflow_p));
15477 }
15482 {
15484 strict_overflow_p)
15486 strict_overflow_p))
15487 {
15490 }
15491 }
15500 {
15503 }
15510 {
15514 /* When both operands are nonzero, then MAX must be too. */
15516 strict_overflow_p))
15519 /* MAX where operand 0 is positive is positive. */
15521 strict_overflow_p);
15522 }
15523 /* MAX where operand 1 is positive is positive. */
15528 {
15532 }
15537 strict_overflow_p)
15539 strict_overflow_p));
15543 }
15546 }
15548 /* Return true when T is an address and is known to be nonzero.
15549 For floating point we further ensure that T is not denormal.
15550 Similar logic is present in nonzero_address in rtlanal.h.
15552 If the return value is based on the assumption that signed overflow
15553 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15554 change *STRICT_OVERFLOW_P. */
15556 bool
15558 {
15561 {
15566 {
15578 /* For objects in symbol table check if we know they are non-zero.
15579 Don't do anything for variables and functions before symtab is built;
15580 it is quite possible that they will be declared weak later. */
15585 /* Constants are never weak. */
15590 }
15598 {
15602 }
15612 }
15614 }
15616 #define integer_valued_real_p(X) \
15619 #define RECURSE(X) \
15620 ((integer_valued_real_p) (X, depth + 1))
15622 /* Return true if the floating point result of (CODE OP0) has an
15623 integer value. We also allow +Inf, -Inf and NaN to be considered
15624 integer values. Return false for signaling NaN.
15626 DEPTH is the current nesting depth of the query. */
15628 bool
15630 {
15632 {
15639 CASE_CONVERT:
15640 {
15647 }
15651 }
15653 }
15655 /* Return true if the floating point result of (CODE OP0 OP1) has an
15656 integer value. We also allow +Inf, -Inf and NaN to be considered
15657 integer values. Return false for signaling NaN.
15659 DEPTH is the current nesting depth of the query. */
15661 bool
15663 {
15665 {
15675 }
15677 }
15679 /* Return true if the floating point result of calling FNDECL with arguments
15680 ARG0 and ARG1 has an integer value. We also allow +Inf, -Inf and NaN to be
15681 considered integer values. Return false for signaling NaN. If FNDECL
15682 takes fewer than 2 arguments, the remaining ARGn are null.
15684 DEPTH is the current nesting depth of the query. */
15686 bool
15688 {
15690 {
15691 CASE_CFN_CEIL:
15692 CASE_CFN_CEIL_FN:
15693 CASE_CFN_FLOOR:
15694 CASE_CFN_FLOOR_FN:
15695 CASE_CFN_NEARBYINT:
15696 CASE_CFN_NEARBYINT_FN:
15697 CASE_CFN_RINT:
15698 CASE_CFN_RINT_FN:
15699 CASE_CFN_ROUND:
15700 CASE_CFN_ROUND_FN:
15701 CASE_CFN_ROUNDEVEN:
15702 CASE_CFN_ROUNDEVEN_FN:
15703 CASE_CFN_TRUNC:
15704 CASE_CFN_TRUNC_FN:
15707 CASE_CFN_FMIN:
15708 CASE_CFN_FMIN_FN:
15709 CASE_CFN_FMAX:
15710 CASE_CFN_FMAX_FN:
15715 }
15717 }
15719 /* Return true if the floating point expression T (a GIMPLE_SINGLE_RHS)
15720 has an integer value. We also allow +Inf, -Inf and NaN to be
15721 considered integer values. Return false for signaling NaN.
15723 DEPTH is the current nesting depth of the query. */
15725 bool
15727 {
15729 {
15737 /* Limit the depth of recursion to avoid quadratic behavior.
15738 This is expected to catch almost all occurrences in practice.
15739 If this code misses important cases that unbounded recursion
15740 would not, passes that need this information could be revised
15741 to provide it through dataflow propagation. */
15745 depth));
15749 }
15751 }
15753 /* Return true if the floating point expression T (a GIMPLE_INVALID_RHS)
15754 has an integer value. We also allow +Inf, -Inf and NaN to be
15755 considered integer values. Return false for signaling NaN.
15757 DEPTH is the current nesting depth of the query. */
15759 static bool
15761 {
15763 {
15774 }
15776 }
15778 #undef RECURSE
15779 #undef integer_valued_real_p
15781 /* Return true if the floating point expression T has an integer value.
15782 We also allow +Inf, -Inf and NaN to be considered integer values.
15783 Return false for signaling NaN.
15785 DEPTH is the current nesting depth of the query. */
15787 bool
15789 {
15797 {
15813 }
15816 {
15822 {
15831 }
15835 }
15836 }
15838 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
15839 attempt to fold the expression to a constant without modifying TYPE,
15840 OP0 or OP1.
15842 If the expression could be simplified to a constant, then return
15843 the constant. If the expression would not be simplified to a
15844 constant, then return NULL_TREE. */
15846 tree
15848 {
15851 }
15853 /* Given the components of a unary expression CODE, TYPE and OP0,
15854 attempt to fold the expression to a constant without modifying
15855 TYPE or OP0.
15857 If the expression could be simplified to a constant, then return
15858 the constant. If the expression would not be simplified to a
15859 constant, then return NULL_TREE. */
15861 tree
15863 {
15866 }
15868 /* If EXP represents referencing an element in a constant string
15869 (either via pointer arithmetic or array indexing), return the
15870 tree representing the value accessed, otherwise return NULL. */
15872 tree
15874 {
15878 {
15880 tree index;
15881 tree string;
15886 else
15887 {
15891 /* Optimize the special-case of a zero lower bound.
15893 We convert the low_bound to sizetype to avoid some problems
15894 with constant folding. (E.g. suppose the lower bound is 1,
15895 and its mode is QI. Without the conversion,l (ARRAY
15896 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
15897 +INDEX), which becomes (ARRAY+255+INDEX). Oops!) */
15903 }
15905 scalar_int_mode char_mode;
15917 }
15919 }
15921 /* Folds a read from vector element at IDX of vector ARG. */
15923 tree
15925 {
15930 {
15934 {
15941 }
15942 }
15944 }
15946 /* Return the tree for neg (ARG0) when ARG0 is known to be either
15947 an integer constant, real, or fixed-point constant.
15949 TYPE is the type of the result. */
15951 static tree
15953 {
15957 {
15963 {
15964 FIXED_VALUE_TYPE f;
15969 /* Propagate overflow flags. */
15973 }
15977 {
15984 }
15987 }
15990 }
15992 /* Return the tree for abs (ARG0) when ARG0 is known to be either
15993 an integer constant or real constant.
15995 TYPE is the type of the result. */
15997 tree
15999 {
16003 {
16005 {
16006 /* If the value is unsigned or non-negative, then the absolute value
16007 is the same as the ordinary value. */
16011 ;
16013 /* If the value is negative, then the absolute value is
16014 its negation. */
16015 else
16018 /* Force to the destination type, set TREE_OVERFLOW for signed
16019 TYPE only. */
16021 }
16027 else
16033 }
16036 }
16038 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
16039 constant. TYPE is the type of the result. */
16041 static tree
16043 {
16047 }
16049 /* Given CODE, a relational operator, the target type, TYPE and two
16050 constant operands OP0 and OP1, return the result of the
16051 relational operation. If the result is not a compile time
16052 constant, then return NULL_TREE. */
16054 static tree
16056 {
16059 /* From here on, the only cases we handle are when the result is
16060 known to be a constant. */
16063 {
16067 /* Handle the cases where either operand is a NaN. */
16069 {
16071 {
16099 }
16102 }
16105 }
16108 {
16112 }
16114 /* Handle equality/inequality of complex constants. */
16116 {
16127 else
16129 }
16132 {
16134 {
16135 /* Have vector comparison with scalar boolean result. */
16143 {
16151 }
16153 }
16154 tree_vector_builder elts;
16159 {
16172 }
16175 }
16177 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
16179 To compute GT, swap the arguments and do LT.
16180 To compute GE, do LT and invert the result.
16181 To compute LE, swap the arguments, do LT and invert the result.
16182 To compute NE, do EQ and invert the result.
16184 Therefore, the code below must handle only EQ and LT. */
16187 {
16190 }
16192 /* Note that it is safe to invert for real values here because we
16193 have already handled the one case that it matters. */
16197 {
16200 }
16202 /* Compute a result for LT or EQ if args permit;
16203 Otherwise return T. */
16205 {
16208 else
16210 }
16211 else
16217 }
16219 /* If necessary, return a CLEANUP_POINT_EXPR for EXPR with the
16220 indicated TYPE. If no CLEANUP_POINT_EXPR is necessary, return EXPR
16221 itself. */
16223 tree
16225 {
16226 /* If the expression does not have side effects then we don't have to wrap
16227 it with a cleanup point expression. */
16231 /* If the expression is a return, check to see if the expression inside the
16232 return has no side effects or the right hand side of the modify expression
16233 inside the return. If either don't have side effects set we don't need to
16234 wrap the expression in a cleanup point expression. Note we don't check the
16235 left hand side of the modify because it should always be a return decl. */
16237 {
16244 }
16247 }
16249 /* Given a pointer value OP0 and a type TYPE, return a simplified version
16250 of an indirection through OP0, or NULL_TREE if no simplification is
16251 possible. */
16253 tree
16255 {
16257 tree subtype;
16258 poly_uint64 const_op01;
16267 {
16271 /* *&CONST_DECL -> to the value of the const decl. */
16274 /* *&p => p; make sure to handle *&"str"[cst] here. */
16276 {
16280 else
16282 }
16283 /* *(foo *)&fooarray => fooarray[0] */
16286 && (!in_gimple_form
16288 {
16298 }
16299 /* *(foo *)&complexfoo => __real__ complexfoo */
16303 /* *(foo *)&vectorfoo => BIT_FIELD_REF<vectorfoo,...> */
16306 {
16310 index);
16311 }
16312 }
16316 {
16322 {
16323 tree op00type;
16327 /* ((foo*)&vectorfoo)[1] => BIT_FIELD_REF<vectorfoo,...> */
16330 /* POINTER_PLUS_EXPR second operand is sizetype, unsigned,
16331 but we want to treat offsets with MSB set as negative.
16332 For the code below negative offsets are invalid and
16333 TYPE_SIZE of the element is something unsigned, so
16334 check whether op01 fits into poly_int64, which implies
16335 it is from 0 to INTTYPE_MAXIMUM (HOST_WIDE_INT), and
16336 then just use poly_uint64 because we want to treat the
16337 value as unsigned. */
16339 {
16341 poly_uint64 max_offset
16345 {
16350 }
16351 }
16352 /* ((foo*)&complexfoo)[1] => __imag__ complexfoo */
16355 {
16357 const_op01))
16359 }
16360 /* ((foo *)&fooarray)[1] => fooarray[1] */
16363 {
16372 {
16377 }
16378 }
16379 }
16380 }
16382 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
16385 && (!in_gimple_form
16387 {
16388 tree type_domain;
16398 NULL_TREE);
16399 }
16402 }
16404 /* Builds an expression for an indirection through T, simplifying some
16405 cases. */
16407 tree
16409 {
16417 }
16419 /* Given an INDIRECT_REF T, return either T or a simplified version. */
16421 tree
16423 {
16428 else
16430 }
16432 /* Strip non-trapping, non-side-effecting tree nodes from an expression
16433 whose result is ignored. The type of the returned tree need not be
16434 the same as the original expression. */
16436 tree
16438 {
16444 {
16455 else
16461 {
16477 }
16482 }
16483 }
16485 /* Return the value of VALUE, rounded up to a multiple of DIVISOR. */
16487 tree
16489 {
16495 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16496 have to do anything. Only do this when we are not given a const,
16497 because in that case, this check is more expensive than just
16498 doing it. */
16500 {
16505 }
16507 /* If divisor is a power of two, simplify this to bit manipulation. */
16509 {
16511 {
16525 }
16526 else
16527 {
16528 tree t;
16534 }
16535 }
16536 else
16537 {
16542 }
16545 }
16547 /* Likewise, but round down. */
16549 tree
16551 {
16558 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16559 have to do anything. Only do this when we are not given a const,
16560 because in that case, this check is more expensive than just
16561 doing it. */
16563 {
16568 }
16570 /* If divisor is a power of two, simplify this to bit manipulation. */
16572 {
16573 tree t;
16577 }
16578 else
16579 {
16584 }
16587 }
16589 /* Returns the pointer to the base of the object addressed by EXP and
16590 extracts the information about the offset of the access, storing it
16591 to PBITPOS and POFFSET. */
16593 static tree
16596 {
16597 tree core;
16598 machine_mode mode;
16600 poly_int64 bitsize;
16609 {
16614 }
16616 {
16622 {
16623 poly_offset_int tem
16629 }
16630 }
16631 else
16632 {
16636 }
16639 }
16641 /* Returns true if addresses of E1 and E2 differ by a constant, false
16642 otherwise. If they do, E1 - E2 is stored in *DIFF. */
16644 bool
16646 {
16661 {
16671 }
16673 {
16674 /* If only one of the offsets is non-constant, the difference cannot
16675 be a constant. */
16677 }
16678 else
16683 }
16685 /* Return OFF converted to a pointer offset type suitable as offset for
16686 POINTER_PLUS_EXPR. Use location LOC for this conversion. */
16687 tree
16689 {
16693 }
16695 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16696 tree
16698 {
16701 }
16703 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16704 tree
16706 {
16709 }
16711 /* Return a pointer to a NUL-terminated string containing the sequence
16712 of bytes corresponding to the representation of the object referred to
16713 by SRC (or a subsequence of such bytes within it if SRC is a reference
16714 to an initialized constant array plus some constant offset).
16715 Set *STRSIZE the number of bytes in the constant sequence including
16716 the terminating NUL byte. *STRSIZE is equal to sizeof(A) - OFFSET
16717 where A is the array that stores the constant sequence that SRC points
16718 to and OFFSET is the byte offset of SRC from the beginning of A. SRC
16719 need not point to a string or even an array of characters but may point
16720 to an object of any type. */
16724 {
16725 /* The offset into the array A storing the string, and A's byte size. */
16726 tree offset_node;
16727 tree mem_size;
16734 else
16741 {
16744 else
16746 }
16751 /* ARRAY_SIZE is the byte size of the array the constant sequence
16752 is stored in and equal to sizeof A. INIT_BYTES is the number
16753 of bytes in the constant sequence used to initialize the array,
16754 including any embedded NULs as well as the terminating NUL (for
16755 strings), but not including any trailing zeros/NULs past
16756 the terminating one appended implicitly to a string literal to
16757 zero out the remainder of the array it's stored in. For example,
16758 given:
16759 const char a[7] = "abc\0d";
16760 n = strlen (a + 1);
16761 ARRAY_SIZE is 7, INIT_BYTES is 6, and OFFSET is 1. For a valid
16762 (i.e., nul-terminated) string with no embedded nuls, INIT_BYTES
16763 is equal to strlen (A) + 1. */
16768 /* Ideally this would turn into a gcc_checking_assert over time. */
16776 {
16777 /* Compute and store the number of characters from the beginning
16778 of the substring at OFFSET to the end, including the terminating
16779 nul. Offsets past the initial length refer to null strings. */
16782 else
16784 }
16785 else
16786 {
16788 /* Support only properly NUL-terminated single byte strings. */
16793 }
16796 }
16798 /* Return a pointer to a NUL-terminated string corresponding to
16799 the expression STR referencing a constant string, possibly
16800 involving a constant offset. Return null if STR either doesn't
16801 reference a constant string or if it involves a nonconstant
16802 offset. */
16806 {
16808 }
16810 /* Given a tree T, compute which bits in T may be nonzero. */
16812 wide_int
16814 {
16816 {
16834 CASE_CONVERT:
16840 {
16845 }
16849 {
16857 }
16861 {
16869 }
16873 }
16876 }
16878 /* Helper function for address compare simplifications in match.pd.
16879 OP0 and OP1 are ADDR_EXPR operands being compared by CODE.
16880 TYPE is the type of comparison operands.
16881 BASE0, BASE1, OFF0 and OFF1 are set by the function.
16882 GENERIC is true if GENERIC folding and false for GIMPLE folding.
16883 Returns 0 if OP0 is known to be unequal to OP1 regardless of OFF{0,1},
16884 1 if bases are known to be equal and OP0 cmp OP1 depends on OFF0 cmp OFF1,
16885 and 2 if unknown. */
16887 int
16891 {
16901 {
16904 }
16906 {
16909 }
16914 /* Punt in GENERIC on variables with value expressions;
16915 the value expressions might point to fields/elements
16916 of other vars etc. */
16922 {
16926 }
16934 /* Assume different STRING_CSTs with the same content will be
16935 merged. */
16944 {
16947 /* If the offsets are equal we can ignore overflow. */
16950 /* Or if we compare using pointers to decls or strings. */
16955 }
16961 /* At this point we know (or assume) the two pointers point at
16962 different objects. */
16966 /* Punt on non-zero offsets from functions. */
16970 /* Or if the bases are neither decls nor string literals. */
16975 /* For initializers, assume addresses of different functions are
16976 different. */
16982 /* Compute whether one address points to the start of one
16983 object and another one to the end of another one. */
16986 {
16989 else
16991 }
16994 else
16995 {
16999 else
17001 }
17003 {
17006 else
17008 }
17011 else
17012 {
17016 else
17018 }
17020 {
17021 /* If one offset is pointing (or could be) to the beginning of one
17022 object and the other is pointing to one past the last byte of the
17023 other object, punt. */
17028 /* If both offsets are the same, there are some cases we know that are
17029 ok. Either if we know they aren't zero, or if we know both sizes
17030 are no zero. */
17036 }
17038 /* At this point, equal is 2 if either one or both pointers are out of
17039 bounds of their object, or one points to start of its object and the
17040 other points to end of its object. This is unspecified behavior
17041 e.g. in C++. Otherwise equal is 0. */
17045 /* When both pointers point to string literals, even when equal is 0,
17046 due to tail merging of string literals the pointers might be the same. */
17048 {
17055 /* If the bytes in the string literals starting at the pointers
17056 differ, the pointers need to be different. */
17061 {
17066 /* If even the bytes in the string literal before the
17067 pointers are the same, the string literals could be
17068 tail merged. */
17070 }
17072 }
17077 /* If this is a pointer comparison, ignore for now even
17078 valid equalities where one pointer is the offset zero
17079 of one object and the other to one past end of another one. */
17083 /* Assume that string literals can't be adjacent to variables
17084 (automatic or global). */
17088 /* Assume that automatic variables can't be adjacent to global
17089 variables. */
17094 }
17096 /* Return the single non-zero element of a CONSTRUCTOR or NULL_TREE. */
17097 tree
17099 {
17108 {
17112 }
17114 }
17116 #if CHECKING_P
17120 /* Helper functions for writing tests of folding trees. */
17122 /* Verify that the binary op (LHS CODE RHS) folds to CONSTANT. */
17124 static void
17126 tree constant)
17127 {
17129 }
17131 /* Verify that the binary op (LHS CODE RHS) folds to an NON_LVALUE_EXPR
17132 wrapping WRAPPED_EXPR. */
17134 static void
17136 tree wrapped_expr)
17137 {
17142 }
17144 /* Verify that various arithmetic binary operations are folded
17145 correctly. */
17147 static void
17149 {
17155 /* Addition. */
17156 /* 1 <-- (0 + 1) */
17158 one);
17160 one);
17162 /* (nonlvalue)x <-- (x + 0) */
17164 x);
17166 /* Subtraction. */
17167 /* 0 <-- (x - x) */
17169 zero);
17171 x);
17173 /* Multiplication. */
17174 /* 0 <-- (x * 0) */
17176 zero);
17178 /* (nonlvalue)x <-- (x * 1) */
17180 x);
17181 }
17185 /* Build a VECTOR_CST corresponding to VMODE, and has
17186 encoding given by NPATTERNS, NELTS_PER_PATTERN and STEP.
17187 Fill it with randomized elements, using rand() % THRESHOLD. */
17189 static tree
17194 {
17199 // Fill a0 for each pattern
17206 // Fill a1 for each pattern
17208 {
17209 tree a1;
17211 {
17216 }
17217 else
17220 }
17225 {
17230 }
17233 }
17235 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17236 when result is VLA. */
17238 static void
17241 {
17242 /* Actual npatterns and encoded_elts in res may be less than expected due
17243 to canonicalization. */
17250 }
17252 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17253 when the result is VLS. */
17255 static void
17257 {
17261 }
17263 /* Helper routine to push multiple elements into BUILDER. */
17267 {
17270 }
17272 #define ARG0(index) vector_cst_elt (arg0, index)
17273 #define ARG1(index) vector_cst_elt (arg1, index)
17275 /* Test cases where result is VNx4SI and input vectors are V4SI. */
17277 static void
17279 {
17281 {
17282 /* Case 1:
17283 sel = { 0, 4, 1, 5, ... }
17284 res = { arg[0], arg1[0], arg0[1], arg1[1], ...} // (4, 1) */
17285 {
17289 tree inner_type
17303 }
17305 /* Case 2: Same as case 1, but contains an out of bounds access which
17306 should wrap around.
17307 sel = {0, 8, 4, 12, ...} (4, 1)
17308 res = { arg0[0], arg0[0], arg1[0], arg1[0], ... } (4, 1). */
17309 {
17313 tree inner_type
17327 }
17328 }
17329 }
17331 /* Test cases where result is V4SI and input vectors are VNx4SI. */
17333 static void
17335 {
17337 {
17338 /* Case 1:
17339 sel = { 0, 1, 2, 3}
17340 res = { arg0[0], arg0[1], arg0[2], arg0[3] }. */
17341 {
17345 tree inner_type
17359 }
17361 /* Case 2: Same as Case 1, but crossing input vector.
17362 sel = {0, 2, 4, 6}
17363 In this case,the index 4 is ambiguous since len = 4 + 4x.
17364 Since we cannot determine, which vector to choose from during
17365 compile time, should return NULL_TREE. */
17366 {
17370 tree inner_type
17385 }
17386 }
17387 }
17389 /* Test all input vectors. */
17391 static void
17393 {
17394 /* Test with 10 different inputs. */
17396 {
17401 /* Case 1: mask = {0, ...} // (1, 1)
17402 res = { arg0[0], ... } // (1, 1) */
17403 {
17410 }
17412 /* Case 2: mask = {len, ...} // (1, 1)
17413 res = { arg1[0], ... } // (1, 1) */
17414 {
17422 }
17423 }
17424 }
17426 /* Test all vectors which contain at-least 2 elements. */
17428 static void
17430 {
17432 {
17433 /* Case 1: mask = { 0, len, ... } // (2, 1)
17434 res = { arg0[0], arg1[0], ... } // (2, 1) */
17435 {
17449 }
17451 /* Case 2: mask = { 0, len, 1, len+1, ... } // (2, 2)
17452 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2) */
17453 {
17467 }
17469 /* Case 4: mask = {0, 0, 1, ...} // (1, 3)
17470 Test that the stepped sequence of the pattern selects from
17471 same input pattern. Since input vectors have npatterns = 2,
17472 and step (a2 - a1) = 1, step is not a multiple of npatterns
17473 in input vector. So return NULL_TREE. */
17474 {
17489 }
17491 /* Case 5: mask = {len, 0, 1, ...} // (1, 3)
17492 Test that stepped sequence of the pattern selects from arg0.
17493 res = { arg1[0], arg0[0], arg0[1], ... } // (1, 3) */
17494 {
17508 }
17510 /* Case 6: PR111648 - a1 chooses base element from input vector arg.
17511 In this case ensure that arg has a natural stepped sequence
17512 to preserve arg's encoding.
17514 As a concrete example, consider:
17515 arg0: { -16, -9, -10, ... } // (1, 3)
17516 arg1: { -12, -5, -6, ... } // (1, 3)
17517 sel = { 0, len, len + 1, ... } // (1, 3)
17519 This will create res with following encoding:
17520 res = { arg0[0], arg1[0], arg1[1], ... } // (1, 3)
17521 = { -16, -12, -5, ... }
17523 The step in above encoding would be: (-5) - (-12) = 7
17524 And hence res[3] would be computed as -5 + 7 = 2.
17525 instead of arg1[2], ie, -6.
17526 Ensure that valid_mask_for_fold_vec_perm_cst returns false
17527 for this case. */
17528 {
17539 /* FIXME: It may happen that build_vec_cst_rand may build a natural
17540 stepped pattern, even if we didn't explicitly tell it to. So folding
17541 may not always fail, but if it does, ensure that's because arg1 does
17542 not have a natural stepped sequence (and not due to other reason) */
17546 }
17548 /* Case 7: Same as Case 6, except that arg1 contains natural stepped
17549 sequence and thus folding should be valid for this case. */
17550 {
17564 }
17565 }
17566 }
17568 /* Test all vectors which contain at-least 4 elements. */
17570 static void
17572 {
17574 {
17575 /* Case 1: mask = { 0, len, 1, len+1, ... } // (4, 1)
17576 res: { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (4, 1) */
17577 {
17591 }
17593 /* Case 2: sel = {0, 1, 2, ...} // (1, 3)
17594 res: { arg0[0], arg0[1], arg0[2], ... } // (1, 3) */
17595 {
17608 }
17610 /* Case 3: sel = {len, len+1, len+2, ...} // (1, 3)
17611 res: { arg1[0], arg1[1], arg1[2], ... } // (1, 3) */
17612 {
17625 }
17627 /* Case 4:
17628 sel = { len, 0, 2, ... } // (1, 3)
17629 This should return NULL because we cross the input vectors.
17630 Because,
17631 Let's assume len = C + Cx
17632 a1 = 0
17633 S = 2
17634 esel = arg0_len / sel_npatterns = C + Cx
17635 ae = 0 + (esel - 2) * S
17636 = 0 + (C + Cx - 2) * 2
17637 = 2(C-2) + 2Cx
17639 For C >= 4:
17640 Let q1 = a1 / arg0_len = 0 / (C + Cx) = 0
17641 Let qe = ae / arg0_len = (2(C-2) + 2Cx) / (C + Cx) = 1
17642 Since q1 != qe, we cross input vectors.
17643 So return NULL_TREE. */
17644 {
17658 }
17660 /* Case 5: npatterns(arg0) = 4 > npatterns(sel) = 2
17661 mask = { 0, len, 1, len + 1, ...} // (2, 2)
17662 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2)
17664 Note that fold_vec_perm_cst will set
17665 res_npatterns = max(4, max(4, 2)) = 4
17666 However after canonicalizing, we will end up with shape (2, 2). */
17667 {
17680 }
17682 /* Case 6: Test combination in sel, where one pattern is dup and other
17683 is stepped sequence.
17684 sel = { 0, 0, 0, 1, 0, 2, ... } // (2, 3)
17685 res = { arg0[0], arg0[0], arg0[0],
17686 arg0[1], arg0[0], arg0[2], ... } // (2, 3) */
17687 {
17702 }
17704 /* Case 7: PR111048: Check that we set arg_npatterns correctly,
17705 when arg0, arg1 and sel have different number of patterns.
17706 arg0 is of shape (1, 1)
17707 arg1 is of shape (4, 1)
17708 sel is of shape (2, 3) = {1, len, 2, len+1, 3, len+2, ...}
17710 In this case the pattern: {len, len+1, len+2, ...} chooses arg1.
17711 However,
17712 step = (len+2) - (len+1) = 1
17713 arg_npatterns = VECTOR_CST_NPATTERNS (arg1) = 4
17714 Since step is not a multiple of arg_npatterns,
17715 valid_mask_for_fold_vec_perm_cst should return false,
17716 and thus fold_vec_perm_cst should return NULL_TREE. */
17717 {
17732 }
17733 }
17734 }
17736 /* Test all vectors which contain at-least 8 elements. */
17738 static void
17740 {
17742 {
17743 /* Case 1: sel_npatterns (4) > input npatterns (2)
17744 sel: { 0, 0, 1, len, 2, 0, 3, len, 4, 0, 5, len, ...} // (4, 3)
17745 res: { arg0[0], arg0[0], arg0[0], arg1[0],
17746 arg0[2], arg0[0], arg0[3], arg1[0],
17747 arg0[4], arg0[0], arg0[5], arg1[0], ... } // (4, 3) */
17748 {
17765 }
17766 }
17767 }
17769 /* Test vectors for which nunits[0] <= 4. */
17771 static void
17773 {
17774 /* Case 1: mask = {0, 4, ...} // (1, 2)
17775 This should return NULL_TREE because the index 4 may choose
17776 from either arg0 or arg1 depending on vector length. */
17777 {
17792 }
17793 }
17795 #undef ARG0
17796 #undef ARG1
17798 /* Return true if SIZE is of the form C + Cx and C is power of 2. */
17800 static bool
17802 {
17810 }
17812 /* Execute fold_vec_perm_cst unit tests. */
17814 static void
17816 {
17820 machine_mode vmode;
17822 {
17823 /* Obtain modes corresponding to VNx4SI and V4SI,
17824 to call mixed mode tests below.
17825 FIXME: Is there a better way to do this ? */
17827 {
17834 }
17851 }
17856 {
17859 }
17860 }
17863 /* Verify that various binary operations on vectors are folded
17864 correctly. */
17866 static void
17868 {
17875 /* Verify equality tests that return a scalar boolean result. */
17888 }
17890 /* Verify folding of VEC_DUPLICATE_EXPRs. */
17892 static void
17894 {
17897 /* This will be 1 if VEC_MODE isn't a vector mode. */
17904 }
17906 /* Run all of the selftests within this file. */
17908 void
17910 {
17915 }