| blob | c6fb083027db7193f16dd29af21cf8b3d0dbd5c0 |
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 /* Combine two poly int's ARG1 and ARG2 under operation CODE to
1217 produce a new constant in RES. Return FALSE if we don't know how
1218 to evaluate CODE at compile-time. */
1220 static bool
1224 {
1228 {
1246 else
1253 else
1266 }
1268 }
1270 /* Combine two integer constants ARG1 and ARG2 under operation CODE to
1271 produce a new constant. Return NULL_TREE if we don't know how to
1272 evaluate CODE at compile-time. */
1274 tree
1277 {
1278 poly_wide_int poly_res;
1284 {
1290 }
1299 }
1301 /* Return true if binary operation OP distributes over addition in operand
1302 OPNO, with the other operand being held constant. OPNO counts from 1. */
1304 static bool
1306 {
1308 {
1319 }
1320 }
1322 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1323 constant. We assume ARG1 and ARG2 have the same data type, or at least
1324 are the same kind of constant and the same machine mode. Return zero if
1325 combining the constants is not allowed in the current operating mode. */
1327 static tree
1329 {
1330 /* Sanity check for the recursive cases. */
1338 {
1344 }
1347 {
1348 machine_mode mode;
1349 REAL_VALUE_TYPE d1;
1350 REAL_VALUE_TYPE d2;
1351 REAL_VALUE_TYPE value;
1352 REAL_VALUE_TYPE result;
1356 /* The following codes are handled by real_arithmetic. */
1358 {
1369 }
1377 /* Don't perform operation if we honor signaling NaNs and
1378 either operand is a signaling NaN. */
1384 /* Don't perform operation if it would raise a division
1385 by zero exception. */
1391 /* If either operand is a NaN, just return it. Otherwise, set up
1392 for floating-point trap; we return an overflow. */
1394 {
1395 /* Make resulting NaN value to be qNaN when flag_signaling_nans
1396 is off. */
1400 }
1402 {
1403 /* Make resulting NaN value to be qNaN when flag_signaling_nans
1404 is off. */
1408 }
1413 /* Don't constant fold this floating point operation if
1414 both operands are not NaN but the result is NaN, and
1415 flag_trapping_math. Such operations should raise an
1416 invalid operation exception. */
1424 /* Don't constant fold this floating point operation if
1425 the result has overflowed and flag_trapping_math. */
1433 /* Don't constant fold this floating point operation if the
1434 result may dependent upon the run-time rounding mode and
1435 flag_rounding_math is set, or if GCC's software emulation
1436 is unable to accurately represent the result. */
1446 }
1449 {
1450 FIXED_VALUE_TYPE f1;
1451 FIXED_VALUE_TYPE f2;
1452 FIXED_VALUE_TYPE result;
1457 /* The following codes are handled by fixed_arithmetic. */
1459 {
1471 {
1478 }
1483 }
1490 /* Propagate overflow flags. */
1494 }
1497 {
1506 {
1517 mpc_mul);
1531 mpc_div);
1532 /* Fallthru. */
1538 {
1539 /* Keep this algorithm in sync with
1540 tree-complex.cc:expand_complex_div_straight().
1542 Expand complex division to scalars, straightforward algorithm.
1543 a / b = ((ar*br + ai*bi)/t) + i((ai*br - ar*bi)/t)
1544 t = br*br + bi*bi
1545 */
1546 tree magsquared
1550 tree t1
1554 tree t2
1561 }
1562 else
1563 {
1564 /* Keep this algorithm in sync with
1565 tree-complex.cc:expand_complex_div_wide().
1567 Expand complex division to scalars, modified algorithm to minimize
1568 overflow with wide input ranges. */
1574 {
1575 /* In the TRUE branch, we compute
1576 ratio = br/bi;
1577 div = (br * ratio) + bi;
1578 tr = (ar * ratio) + ai;
1579 ti = (ai * ratio) - ar;
1580 tr = tr / div;
1581 ti = ti / div; */
1592 }
1593 else
1594 {
1595 /* In the FALSE branch, we compute
1596 ratio = d/c;
1597 divisor = (d * ratio) + c;
1598 tr = (b * ratio) + a;
1599 ti = b - (a * ratio);
1600 tr = tr / div;
1601 ti = ti / div; */
1613 }
1614 }
1619 }
1623 }
1629 {
1634 /* We can operate directly on the encoding if:
1636 a3 - a2 == a2 - a1 && b3 - b2 == b2 - b1
1637 implies
1638 (a3 op b3) - (a2 op b2) == (a2 op b2) - (a1 op b1)
1640 Addition and subtraction are the supported operators
1641 for which this is true. */
1644 /* We can operate directly on stepped encodings if:
1646 a3 - a2 == a2 - a1
1647 implies:
1648 (a3 op c) - (a2 op c) == (a2 op c) - (a1 op c)
1650 which is true if (x -> x op c) distributes over addition. */
1652 else
1653 /* Similarly in reverse. */
1655 tree_vector_builder elts;
1660 {
1666 /* It is possible that const_binop cannot handle the given
1667 code and return NULL_TREE */
1671 }
1674 }
1676 /* Shifts allow a scalar offset for a vector. */
1679 {
1682 tree_vector_builder elts;
1687 {
1692 /* It is possible that const_binop cannot handle the given
1693 code and return NULL_TREE. */
1697 }
1700 }
1702 }
1704 /* Overload that adds a TYPE parameter to be able to dispatch
1705 to fold_relational_const. */
1707 tree
1709 {
1713 /* ??? Until we make the const_binop worker take the type of the
1714 result as argument put those cases that need it here. */
1716 {
1733 {
1738 }
1744 {
1760 {
1765 ? NOP_EXPR
1772 }
1775 }
1781 {
1804 {
1817 }
1820 }
1823 }
1828 /* Make sure type and arg0 have the same saturating flag. */
1833 }
1835 /* Compute CODE ARG1 with resulting type TYPE with ARG1 being constant.
1836 Return zero if computing the constants is not possible. */
1838 tree
1840 {
1841 /* Don't perform the operation, other than NEGATE and ABS, if
1842 flag_signaling_nans is on and the operand is a signaling NaN. */
1852 {
1853 CASE_CONVERT:
1860 /* If the source address is 0, and the source address space
1861 cannot have a valid object at 0, fold to dest type null. */
1872 {
1873 /* Can't call fold_negate_const directly here as that doesn't
1874 handle all cases and we might not be able to negate some
1875 constants. */
1880 }
1890 {
1894 }
1902 /* Perform BIT_NOT_EXPR on each element individually. */
1904 {
1905 tree elem;
1907 /* This can cope with stepped encodings because ~x == -1 - x. */
1908 tree_vector_builder elements;
1912 {
1918 }
1921 }
1945 {
1968 else
1973 {
1979 }
1982 }
1991 }
1994 }
1996 /* Create a sizetype INT_CST node with NUMBER sign extended. KIND
1997 indicates which particular sizetype to create. */
1999 tree
2001 {
2003 }
2004 \f
2005 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
2006 is a tree code. The type of the result is taken from the operands.
2007 Both must be equivalent integer types, ala int_binop_types_match_p.
2008 If the operands are constant, so is the result. */
2010 tree
2012 {
2021 /* Handle the special case of two poly_int constants faster. */
2023 {
2024 /* And some specific cases even faster than that. */
2026 {
2033 }
2035 {
2039 }
2041 {
2045 }
2047 /* Handle general case of two integer constants. For sizetype
2048 constant calculations we always want to know about overflow,
2049 even in the unsigned case. */
2053 }
2056 }
2058 /* Given two values, either both of sizetype or both of bitsizetype,
2059 compute the difference between the two values. Return the value
2060 in signed type corresponding to the type of the operands. */
2062 tree
2064 {
2066 tree ctype;
2071 /* If the type is already signed, just do the simple thing. */
2079 else
2082 /* If either operand is not a constant, do the conversions to the signed
2083 type and subtract. The hardware will do the right thing with any
2084 overflow in the subtraction. */
2090 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
2091 Otherwise, subtract the other way, convert to CTYPE (we know that can't
2092 overflow) and negate (which can't either). Special-case a result
2093 of zero while we're here. */
2099 else
2103 MINUS_EXPR,
2105 }
2106 \f
2107 /* A subroutine of fold_convert_const handling conversions of an
2108 INTEGER_CST to another integer type. */
2110 static tree
2112 {
2113 /* Given an integer constant, make new constant with new type,
2114 appropriately sign-extended or truncated. Use widest_int
2115 so that any extension is done according ARG1's type. */
2119 }
2121 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2122 to an integer type. */
2124 static tree
2126 {
2128 tree t;
2130 /* The following code implements the floating point to integer
2131 conversion rules required by the Java Language Specification,
2132 that IEEE NaNs are mapped to zero and values that overflow
2133 the target precision saturate, i.e. values greater than
2134 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
2135 are mapped to INT_MIN. These semantics are allowed by the
2136 C and C++ standards that simply state that the behavior of
2137 FP-to-integer conversion is unspecified upon overflow. */
2139 wide_int val;
2140 REAL_VALUE_TYPE r;
2144 {
2151 }
2153 /* If R is NaN, return zero and show we have an overflow. */
2155 {
2158 }
2160 /* See if R is less than the lower bound or greater than the
2161 upper bound. */
2164 {
2168 {
2171 }
2172 }
2175 {
2178 {
2181 {
2184 }
2185 }
2186 }
2193 }
2195 /* A subroutine of fold_convert_const handling conversions of a
2196 FIXED_CST to an integer type. */
2198 static tree
2200 {
2201 tree t;
2203 scalar_mode mode;
2205 /* Right shift FIXED_CST to temp by fbit. */
2209 {
2211 HOST_BITS_PER_DOUBLE_INT,
2214 /* Left shift temp to temp_trunc by fbit. */
2216 HOST_BITS_PER_DOUBLE_INT,
2218 }
2219 else
2220 {
2223 }
2225 /* If FIXED_CST is negative, we need to round the value toward 0.
2226 By checking if the fractional bits are not zero to add 1 to temp. */
2232 /* Given a fixed-point constant, make new constant with new type,
2233 appropriately sign-extended or truncated. */
2241 }
2243 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2244 to another floating point type. */
2246 static tree
2248 {
2249 REAL_VALUE_TYPE value;
2250 tree t;
2252 /* If the underlying modes are the same, simply treat it as
2253 copy and rebuild with TREE_REAL_CST information and the
2254 given type. */
2256 {
2259 }
2261 /* Don't perform the operation if flag_signaling_nans is on
2262 and the operand is a signaling NaN. */
2267 /* With flag_rounding_math we should respect the current rounding mode
2268 unless the conversion is exact. */
2276 /* If converting an infinity or NAN to a representation that doesn't
2277 have one, set the overflow bit so that we can produce some kind of
2278 error message at the appropriate point if necessary. It's not the
2279 most user-friendly message, but it's better than nothing. */
2286 /* Regular overflow, conversion produced an infinity in a mode that
2287 can't represent them. */
2292 else
2295 }
2297 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2298 to a floating point type. */
2300 static tree
2302 {
2303 REAL_VALUE_TYPE value;
2304 tree t;
2312 }
2314 /* A subroutine of fold_convert_const handling conversions a FIXED_CST
2315 to another fixed-point type. */
2317 static tree
2319 {
2320 FIXED_VALUE_TYPE value;
2321 tree t;
2328 /* Propagate overflow flags. */
2332 }
2334 /* A subroutine of fold_convert_const handling conversions an INTEGER_CST
2335 to a fixed-point type. */
2337 static tree
2339 {
2340 FIXED_VALUE_TYPE value;
2341 tree t;
2343 double_int di;
2350 else
2358 /* Propagate overflow flags. */
2362 }
2364 /* A subroutine of fold_convert_const handling conversions a REAL_CST
2365 to a fixed-point type. */
2367 static tree
2369 {
2370 FIXED_VALUE_TYPE value;
2371 tree t;
2379 /* Propagate overflow flags. */
2383 }
2385 /* Attempt to fold type conversion operation CODE of expression ARG1 to
2386 type TYPE. If no simplification can be done return NULL_TREE. */
2388 static tree
2390 {
2395 /* We can't widen types, since the runtime value could overflow the
2396 original type before being extended to the new type. */
2407 {
2414 }
2416 {
2418 {
2420 /* Avoid the folding if flag_rounding_math is on and the
2421 conversion is not exact. */
2423 {
2429 }
2431 }
2436 }
2438 {
2445 }
2447 {
2450 {
2453 /* We can't handle steps directly when extending, since the
2454 values need to wrap at the original precision first. */
2455 bool step_ok_p
2459 tree_vector_builder v;
2464 {
2470 }
2472 }
2473 }
2475 }
2477 /* Construct a vector of zero elements of vector type TYPE. */
2479 static tree
2481 {
2482 tree t;
2486 }
2488 /* Returns true, if ARG is convertible to TYPE using a NOP_EXPR. */
2490 bool
2492 {
2507 {
2529 }
2530 }
2532 /* Convert expression ARG to type TYPE. Used by the middle-end for
2533 simple conversions in preference to calling the front-end's convert. */
2535 tree
2537 {
2539 tree tem;
2550 {
2553 /* Handle conversions between pointers to different address spaces. */
2558 /* fall through */
2563 {
2567 }
2581 {
2585 }
2587 {
2591 }
2593 {
2597 }
2600 {
2618 }
2623 {
2627 }
2630 {
2644 }
2648 {
2657 integer_zero_node));
2659 {
2663 {
2669 }
2677 }
2681 }
2699 }
2700 fold_convert_exit:
2703 }
2704 \f
2705 /* Return false if expr can be assumed not to be an lvalue, true
2706 otherwise. */
2708 static bool
2710 {
2711 /* We only need to wrap lvalue tree codes. */
2713 {
2746 /* Assume the worst for front-end tree codes. */
2750 }
2753 }
2755 /* Return an expr equal to X but certainly not valid as an lvalue. */
2757 tree
2759 {
2760 /* While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to
2761 us. */
2768 }
2769 \f
2770 /* Given a tree comparison code, return the code that is the logical inverse.
2771 It is generally not safe to do this for floating-point comparisons, except
2772 for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return
2773 ERROR_MARK in this case. */
2775 enum tree_code
2777 {
2783 {
2814 }
2815 }
2817 /* Similar, but return the comparison that results if the operands are
2818 swapped. This is safe for floating-point. */
2820 enum tree_code
2822 {
2824 {
2850 }
2851 }
2854 /* Convert a comparison tree code from an enum tree_code representation
2855 into a compcode bit-based encoding. This function is the inverse of
2856 compcode_to_comparison. */
2858 static enum comparison_code
2860 {
2862 {
2893 }
2894 }
2896 /* Convert a compcode bit-based encoding of a comparison operator back
2897 to GCC's enum tree_code representation. This function is the
2898 inverse of comparison_to_compcode. */
2900 static enum tree_code
2902 {
2904 {
2935 }
2936 }
2938 /* Return true if COND1 tests the opposite condition of COND2. */
2940 bool
2942 {
2945 && (invert_tree_comparison
2952 }
2954 /* Return a tree for the comparison which is the combination of
2955 doing the AND or OR (depending on CODE) of the two operations LCODE
2956 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2957 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2958 if this makes the transformation invalid. */
2960 tree
2965 {
2972 {
2983 }
2986 {
2987 /* Eliminate unordered comparisons, as well as LTGT and ORD
2988 which are not used unless the mode has NaNs. */
2994 }
2996 {
2997 /* Check that the original operation and the optimized ones will trap
2998 under the same condition. */
3009 /* In a short-circuited boolean expression the LHS might be
3010 such that the RHS, if evaluated, will never trap. For
3011 example, in ORD (x, y) && (x < y), we evaluate the RHS only
3012 if neither x nor y is NaN. (This is a mixed blessing: for
3013 example, the expression above will never trap, hence
3014 optimizing it to x < y would be invalid). */
3019 /* If the comparison was short-circuited, and only the RHS
3020 trapped, we may now generate a spurious trap. */
3025 /* If we changed the conditions that cause a trap, we lose. */
3028 }
3034 else
3035 {
3040 }
3041 }
3042 \f
3043 /* Return nonzero if two operands (typically of the same tree node)
3044 are necessarily equal. FLAGS modifies behavior as follows:
3046 If OEP_ONLY_CONST is set, only return nonzero for constants.
3047 This function tests whether the operands are indistinguishable;
3048 it does not test whether they are equal using C's == operation.
3049 The distinction is important for IEEE floating point, because
3050 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
3051 (2) two NaNs may be indistinguishable, but NaN!=NaN.
3053 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
3054 even though it may hold multiple values during a function.
3055 This is because a GCC tree node guarantees that nothing else is
3056 executed between the evaluation of its "operands" (which may often
3057 be evaluated in arbitrary order). Hence if the operands themselves
3058 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
3059 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
3060 unset means assuming isochronic (or instantaneous) tree equivalence.
3061 Unless comparing arbitrary expression trees, such as from different
3062 statements, this flag can usually be left unset.
3064 If OEP_PURE_SAME is set, then pure functions with identical arguments
3065 are considered the same. It is used when the caller has other ways
3066 to ensure that global memory is unchanged in between.
3068 If OEP_ADDRESS_OF is set, we are actually comparing addresses of objects,
3069 not values of expressions.
3071 If OEP_LEXICOGRAPHIC is set, then also handle expressions with side-effects
3072 such as MODIFY_EXPR, RETURN_EXPR, as well as STATEMENT_LISTs.
3074 If OEP_BITWISE is set, then require the values to be bitwise identical
3075 rather than simply numerically equal. Do not take advantage of things
3076 like math-related flags or undefined behavior; only return true for
3077 values that are provably bitwise identical in all circumstances.
3079 Unless OEP_MATCH_SIDE_EFFECTS is set, the function returns false on
3080 any operand with side effect. This is unnecesarily conservative in the
3081 case we know that arg0 and arg1 are in disjoint code paths (such as in
3082 ?: operator). In addition OEP_MATCH_SIDE_EFFECTS is used when comparing
3083 addresses with TREE_CONSTANT flag set so we know that &var == &var
3084 even if var is volatile. */
3086 bool
3089 {
3097 /* If either is ERROR_MARK, they aren't equal. */
3103 /* Similar, if either does not have a type (like a template id),
3104 they aren't equal. */
3108 /* Bitwise identity makes no sense if the values have different layouts. */
3113 /* We cannot consider pointers to different address space equal. */
3120 /* Check equality of integer constants before bailing out due to
3121 precision differences. */
3123 {
3124 /* Address of INTEGER_CST is not defined; check that we did not forget
3125 to drop the OEP_ADDRESS_OF flags. */
3128 }
3131 {
3132 /* If both types don't have the same signedness, then we can't consider
3133 them equal. We must check this before the STRIP_NOPS calls
3134 because they may change the signedness of the arguments. As pointers
3135 strictly don't have a signedness, require either two pointers or
3136 two non-pointers as well. */
3142 /* If both types don't have the same precision, then it is not safe
3143 to strip NOPs. */
3150 }
3151 #if 0
3152 /* FIXME: Fortran FE currently produce ADDR_EXPR of NOP_EXPR. Enable the
3153 sanity check once the issue is solved. */
3154 else
3155 /* Addresses of conversions and SSA_NAMEs (and many other things)
3156 are not defined. Check that we did not forget to drop the
3157 OEP_ADDRESS_OF/OEP_CONSTANT_ADDRESS_OF flags. */
3160 #endif
3162 /* In case both args are comparisons but with different comparison
3163 code, try to swap the comparison operands of one arg to produce
3164 a match and compare that variant. */
3168 {
3176 }
3179 {
3180 /* NOP_EXPR and CONVERT_EXPR are considered equal. */
3182 ;
3184 {
3185 /* If we are interested in comparing addresses ignore
3186 MEM_REF wrappings of the base that can appear just for
3187 TBAA reasons. */
3201 }
3202 else
3204 }
3206 /* When not checking adddresses, this is needed for conversions and for
3207 COMPONENT_REF. Might as well play it safe and always test this. */
3214 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
3215 We don't care about side effects in that case because the SAVE_EXPR
3216 takes care of that for us. In all other cases, two expressions are
3217 equal if they have no side effects. If we have two identical
3218 expressions with side effects that should be treated the same due
3219 to the only side effects being identical SAVE_EXPR's, that will
3220 be detected in the recursive calls below.
3221 If we are taking an invariant address of two identical objects
3222 they are necessarily equal as well. */
3229 /* Next handle constant cases, those for which we can return 1 even
3230 if ONLY_CONST is set. */
3233 {
3246 {
3247 /* If we do not distinguish between signed and unsigned zero,
3248 consider them equal. */
3251 }
3255 {
3270 }
3274 flags)
3276 flags));
3287 flags | OEP_ADDRESS_OF
3290 /* In GIMPLE empty constructors are allowed in initializers of
3291 aggregates. */
3295 }
3297 /* Don't handle more cases for OEP_BITWISE, since we can't guarantee that
3298 two instances of undefined behavior will give identical results. */
3302 /* Define macros to test an operand from arg0 and arg1 for equality and a
3303 variant that allows null and views null as being different from any
3304 non-null value. In the latter case, if either is null, the both
3305 must be; otherwise, do the normal comparison. */
3306 #define OP_SAME(N) operand_equal_p (TREE_OPERAND (arg0, N), \
3307 TREE_OPERAND (arg1, N), flags)
3309 #define OP_SAME_WITH_NULL(N) \
3310 ((!TREE_OPERAND (arg0, N) || !TREE_OPERAND (arg1, N)) \
3311 ? TREE_OPERAND (arg0, N) == TREE_OPERAND (arg1, N) : OP_SAME (N))
3314 {
3316 /* Two conversions are equal only if signedness and modes match. */
3318 {
3319 CASE_CONVERT:
3327 }
3337 /* For commutative ops, allow the other order. */
3345 /* If either of the pointer (or reference) expressions we are
3346 dereferencing contain a side effect, these cannot be equal,
3347 but their addresses can be. */
3354 {
3357 {
3361 /* Verify that the access types are compatible. */
3365 }
3370 /* Require the same offset. */
3376 /* Fallthru. */
3384 {
3385 /* Require equal access sizes */
3391 flags)))
3393 /* Verify that access happens in similar types. */
3396 /* Verify that accesses are TBAA compatible. */
3405 /* Verify that alignment is compatible. */
3409 }
3412 /* TARGET_MEM_REF require equal extra operands. */
3423 /* Compare the array index by value if it is constant first as we
3424 may have different types but same value here. */
3430 /* Compare low bound and element size as with OEP_ADDRESS_OF
3431 we have to account for the offset of the ref. */
3436 array_ref_low_bound
3440 array_ref_element_size
3442 flags))));
3445 /* Handle operand 2 the same as for ARRAY_REF. Operand 0
3446 may be NULL when we're called to compare MEM_EXPRs. */
3449 {
3452 /* Most of time we only need to compare FIELD_DECLs for equality.
3453 However when determining address look into actual offsets.
3454 These may match for unions and unshared record types. */
3457 {
3460 {
3464 /* Non-FIELD_DECL operands can appear in C++ templates. */
3471 flags))
3473 }
3474 else
3476 }
3477 }
3488 }
3492 {
3494 /* Be sure we pass right ADDRESS_OF flag. */
3511 /* The multiplcation operands are commutative. */
3512 /* FALLTHRU */
3520 /* Otherwise take into account this is a commutative operation. */
3533 /* BIT_INSERT_EXPR has an implict operand as the type precision
3534 of op1. Need to check to make sure they are the same. */
3540 /* FALLTHRU */
3565 /* Virtual table reference. */
3577 {
3582 }
3587 }
3591 {
3595 /* If not both CALL_EXPRs are either internal or normal function
3596 functions, then they are not equal. */
3599 {
3600 /* If the CALL_EXPRs call different internal functions, then they
3601 are not equal. */
3604 }
3605 else
3606 {
3607 /* If the CALL_EXPRs call different functions, then they are not
3608 equal. */
3610 flags))
3612 }
3614 /* FIXME: We could skip this test for OEP_MATCH_SIDE_EFFECTS. */
3615 {
3619 else
3623 }
3625 /* Now see if all the arguments are the same. */
3626 {
3637 /* If we get here and both argument lists are exhausted
3638 then the CALL_EXPRs are equal. */
3640 }
3643 }
3646 /* Consider __builtin_sqrt equal to sqrt. */
3656 {
3657 /* Consider decls with the same name equal. The caller needs
3658 to make sure they refer to the same entity (such as a function
3659 formal parameter). */
3665 }
3670 {
3674 /* In GIMPLE constructors are used only to build vectors from
3675 elements. Individual elements in the constructor must be
3676 indexed in increasing order and form an initial sequence.
3678 We make no effort to compare constructors in generic.
3679 (see sem_variable::equals in ipa-icf which can do so for
3680 constants). */
3685 /* Be sure that vectors constructed have the same representation.
3686 We only tested element precision and modes to match.
3687 Vectors may be BLKmode and thus also check that the number of
3688 parts match. */
3701 {
3706 /* In GIMPLE the indexes can be either NULL or matching i.
3707 Double check this so we won't get false
3708 positives for GENERIC. */
3716 }
3718 }
3721 {
3722 /* Compare the STATEMENT_LISTs. */
3728 {
3729 /* The lists don't have the same number of statements. */
3735 flags & (OEP_LEXICOGRAPHIC
3738 }
3739 }
3744 {
3755 }
3759 }
3761 #undef OP_SAME
3762 #undef OP_SAME_WITH_NULL
3763 }
3765 /* Generate a hash value for an expression. This can be used iteratively
3766 by passing a previous result as the HSTATE argument. */
3768 void
3771 {
3777 {
3780 }
3790 {
3791 /* Alas, constants aren't shared, so we can't rely on pointer
3792 identity. */
3802 {
3806 else
3810 }
3812 {
3816 }
3826 {
3833 }
3835 /* We can just compare by pointer. */
3839 /* The node itself doesn't matter. */
3843 /* Ignore. */
3846 /* A list of expressions, for a CALL_EXPR or as the elements of a
3847 VECTOR_CST. */
3852 {
3858 {
3859 /* In GIMPLE the indexes can be either NULL or matching i. */
3864 }
3866 }
3868 {
3869 tree_stmt_iterator i;
3874 }
3883 /* When referring to a built-in FUNCTION_DECL, use the __builtin__ form.
3884 Otherwise nodes that compare equal according to operand_equal_p might
3885 get different hash codes. However, don't do this for machine specific
3886 or front end builtins, since the function code is overloaded in those
3887 cases. */
3890 {
3893 }
3894 /* FALL THROUGH */
3897 {
3901 }
3905 {
3906 /* DECL's have a unique ID */
3908 }
3910 {
3911 /* For comparisons that can be swapped, use the lower
3912 tree code. */
3919 }
3921 {
3922 /* NOP_EXPR and CONVERT_EXPR are considered equal by
3923 operand_equal_p. */
3927 /* Don't hash the type, that can lead to having nodes which
3928 compare equal according to operand_equal_p, but which
3929 have different hash codes. Make sure to include signedness
3930 in the hash computation. */
3933 }
3934 /* For OEP_ADDRESS_OF, hash MEM_EXPR[&decl, 0] the same as decl. */
3942 /* Don't ICE on FE specific trees, or their arguments etc.
3943 during operand_equal_p hash verification. */
3946 else
3947 {
3953 {
3969 {
3976 }
3990 {
3991 /* The multiplication operands are commutative. */
3998 }
4006 /* For TARGET_EXPR, just hash on the TARGET_EXPR_SLOT.
4007 Usually different TARGET_EXPRs just should use
4008 different temporaries in their slots. */
4013 /* Virtual table reference. */
4021 {
4023 /* We compute mangled names only when free_lang_data is run.
4024 In that case we can hash precisely. */
4027 hstate.add_object
4028 (IDENTIFIER_HASH_VALUE
4030 }
4034 }
4036 /* Don't hash the type, that can lead to having nodes which
4037 compare equal according to operand_equal_p, but which
4038 have different hash codes. */
4040 {
4041 /* Make sure to include signness in the hash computation. */
4044 }
4047 {
4048 /* It's a commutative expression. We want to hash it the same
4049 however it appears. We do this by first hashing both operands
4050 and then rehashing based on the order of their independent
4051 hashes. */
4056 }
4057 else
4061 }
4063 }
4064 }
4066 bool
4069 {
4070 /* When checking and unless comparing DECL names, verify that if
4071 the outermost operand_equal_p call returns non-zero then ARG0
4072 and ARG1 have the same hash value. */
4074 {
4076 {
4078 {
4085 }
4087 }
4088 else
4092 }
4095 }
4100 /* Conveinece wrapper around operand_compare class because usually we do
4101 not need to play with the valueizer. */
4103 bool
4105 {
4107 }
4109 namespace inchash
4110 {
4112 /* Generate a hash value for an expression. This can be used iteratively
4113 by passing a previous result as the HSTATE argument.
4115 This function is intended to produce the same hash for expressions which
4116 would compare equal using operand_equal_p. */
4117 void
4119 {
4121 }
4123 }
4124 \f
4125 /* Similar to operand_equal_p, but see if ARG0 might be a variant of ARG1
4126 with a different signedness or a narrower precision. */
4128 static bool
4130 {
4138 /* Discard any conversions that don't change the modes of ARG0 and ARG1
4139 and see if the inner values are the same. This removes any
4140 signedness comparison, which doesn't matter here. */
4148 /* Discard a single widening conversion from ARG1 and see if the inner
4149 value is the same as ARG0. */
4158 }
4159 \f
4160 /* See if ARG is an expression that is either a comparison or is performing
4161 arithmetic on comparisons. The comparisons must only be comparing
4162 two different values, which will be stored in *CVAL1 and *CVAL2; if
4163 they are nonzero it means that some operands have already been found.
4164 No variables may be used anywhere else in the expression except in the
4165 comparisons.
4167 If this is true, return 1. Otherwise, return zero. */
4169 static bool
4171 {
4175 /* We can handle some of the tcc_expression cases here. */
4184 {
4203 /* First see if we can handle the first operand, then the second. For
4204 the second operand, we know *CVAL1 can't be zero. It must be that
4205 one side of the comparison is each of the values; test for the
4206 case where this isn't true by failing if the two operands
4207 are the same. */
4216 ;
4220 ;
4221 else
4225 ;
4229 ;
4230 else
4237 }
4238 }
4239 \f
4240 /* ARG is a tree that is known to contain just arithmetic operations and
4241 comparisons. Evaluate the operations in the tree substituting NEW0 for
4242 any occurrence of OLD0 as an operand of a comparison and likewise for
4243 NEW1 and OLD1. */
4245 static tree
4248 {
4253 /* We can handle some of the tcc_expression cases here. */
4261 {
4276 {
4295 }
4296 /* Fall through - ??? */
4299 {
4303 /* We need to check both for exact equality and tree equality. The
4304 former will be true if the operand has a side-effect. In that
4305 case, we know the operand occurred exactly once. */
4318 }
4322 }
4323 }
4324 \f
4325 /* Return a tree for the case when the result of an expression is RESULT
4326 converted to TYPE and OMITTED was previously an operand of the expression
4327 but is now not needed (e.g., we folded OMITTED * 0).
4329 If OMITTED has side effects, we must evaluate it. Otherwise, just do
4330 the conversion of RESULT to TYPE. */
4332 tree
4334 {
4337 /* If the resulting operand is an empty statement, just return the omitted
4338 statement casted to void. */
4348 }
4350 /* Return a tree for the case when the result of an expression is RESULT
4351 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
4352 of the expression but are now not needed.
4354 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
4355 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
4356 evaluated before OMITTED2. Otherwise, if neither has side effects,
4357 just do the conversion of RESULT to TYPE. */
4359 tree
4362 {
4371 }
4373 \f
4374 /* Return a simplified tree node for the truth-negation of ARG. This
4375 never alters ARG itself. We assume that ARG is an operation that
4376 returns a truth value (0 or 1).
4378 FIXME: one would think we would fold the result, but it causes
4379 problems with the dominator optimizer. */
4381 static tree
4383 {
4388 /* If this is a comparison, we can simply invert it, except for
4389 floating-point non-equality comparisons, in which case we just
4390 enclose a TRUTH_NOT_EXPR around what we have. */
4393 {
4396 && flag_trapping_math
4409 }
4412 {
4431 /* Here we can invert either operand. We invert the first operand
4432 unless the second operand is a TRUTH_NOT_EXPR in which case our
4433 result is the XOR of the first operand with the inside of the
4434 negation of the second operand. */
4439 else
4462 {
4469 /* A COND_EXPR may have a throw as one operand, which
4470 then has void type. Just leave void operands
4471 as they are. */
4477 }
4489 CASE_CONVERT:
4493 /* fall through */
4515 }
4516 }
4518 /* Fold the truth-negation of ARG. This never alters ARG itself. We
4519 assume that ARG is an operation that returns a truth value (0 or 1
4520 for scalars, 0 or -1 for vectors). Return the folded expression if
4521 folding is successful. Otherwise, return NULL_TREE. */
4523 static tree
4525 {
4528 ? BIT_NOT_EXPR
4531 }
4533 /* Return a simplified tree node for the truth-negation of ARG. This
4534 never alters ARG itself. We assume that ARG is an operation that
4535 returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). */
4537 tree
4539 {
4545 ? BIT_NOT_EXPR
4548 }
4549 \f
4550 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
4551 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero
4552 and uses reverse storage order if REVERSEP is nonzero. ORIG_INNER
4553 is the original memory reference used to preserve the alias set of
4554 the access. */
4556 static tree
4560 {
4563 /* Attempt not to lose the access path if possible. */
4565 {
4567 machine_mode nmode;
4569 tree noffset;
4577 && !reversep
4578 && !nreversep
4580 {
4583 }
4584 }
4593 {
4600 }
4615 }
4617 /* Optimize a bit-field compare.
4619 There are two cases: First is a compare against a constant and the
4620 second is a comparison of two items where the fields are at the same
4621 bit position relative to the start of a chunk (byte, halfword, word)
4622 large enough to contain it. In these cases we can avoid the shift
4623 implicit in bitfield extractions.
4625 For constants, we emit a compare of the shifted constant with the
4626 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
4627 compared. For two fields at the same position, we do the ANDs with the
4628 similar mask and compare the result of the ANDs.
4630 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
4631 COMPARE_TYPE is the type of the comparison, and LHS and RHS
4632 are the left and right operands of the comparison, respectively.
4634 If the optimization described above can be done, we return the resulting
4635 tree. Otherwise we return zero. */
4637 static tree
4640 {
4644 tree unsigned_type;
4647 scalar_int_mode nmode;
4652 tree mask;
4653 tree offset;
4655 /* Get all the information about the extractions being done. If the bit size
4656 is the same as the size of the underlying object, we aren't doing an
4657 extraction at all and so can do nothing. We also don't want to
4658 do anything if the inner expression is a PLACEHOLDER_EXPR since we
4659 then will no longer be able to replace it. */
4674 else
4675 {
4676 /* If this is not a constant, we can only do something if bit positions,
4677 sizes, signedness and storage order are the same. */
4678 rinner
4691 }
4693 /* Honor the C++ memory model and mimic what RTL expansion does. */
4697 {
4701 }
4703 /* See if we can find a mode to refer to this field. We should be able to,
4704 but fail if we can't. */
4712 /* Set signed and unsigned types of the precision of this mode for the
4713 shifts below. */
4716 /* Compute the bit position and size for the new reference and our offset
4717 within it. If the new reference is the same size as the original, we
4718 won't optimize anything, so return zero. */
4728 /* Make the mask to be used against the extracted field. */
4735 {
4739 /* If not comparing with constant, just rework the comparison
4740 and return. */
4748 }
4750 /* Otherwise, we are handling the constant case. See if the constant is too
4751 big for the field. Warn and return a tree for 0 (false) if so. We do
4752 this not only for its own sake, but to avoid having to test for this
4753 error case below. If we didn't, we might generate wrong code.
4755 For unsigned fields, the constant shifted right by the field length should
4756 be all zero. For signed fields, the high-order bits should agree with
4757 the sign bit. */
4760 {
4762 {
4766 }
4767 }
4768 else
4769 {
4772 {
4776 }
4777 }
4782 /* Single-bit compares should always be against zero. */
4784 {
4787 }
4789 /* Make a new bitfield reference, shift the constant over the
4790 appropriate number of bits and mask it with the computed mask
4791 (in case this was a signed field). If we changed it, make a new one. */
4799 mask);
4804 }
4805 \f
4806 /* Subroutine for fold_truth_andor_1: decode a field reference.
4808 If EXP is a comparison reference, we return the innermost reference.
4810 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
4811 set to the starting bit number.
4813 If the innermost field can be completely contained in a mode-sized
4814 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
4816 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
4817 otherwise it is not changed.
4819 *PUNSIGNEDP is set to the signedness of the field.
4821 *PREVERSEP is set to the storage order of the field.
4823 *PMASK is set to the mask used. This is either contained in a
4824 BIT_AND_EXPR or derived from the width of the field.
4826 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
4828 Return 0 if this is not a component reference or is one that we can't
4829 do anything with. */
4831 static tree
4836 {
4841 tree unsigned_type;
4844 /* All the optimizations using this function assume integer fields.
4845 There are problems with FP fields since the type_for_size call
4846 below can fail for, e.g., XFmode. */
4850 /* We are interested in the bare arrangement of bits, so strip everything
4851 that doesn't affect the machine mode. However, record the type of the
4852 outermost expression if it may matter below. */
4859 {
4865 }
4876 /* Reject out-of-bound accesses (PR79731). */
4888 /* If the number of bits in the reference is the same as the bitsize of
4889 the outer type, then the outer type gives the signedness. Otherwise
4890 (in case of a small bitfield) the signedness is unchanged. */
4894 /* Compute the mask to access the bitfield. */
4902 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
4910 }
4912 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
4913 bit positions and MASK is SIGNED. */
4915 static bool
4917 {
4921 /* If this function returns true when the type of the mask is
4922 UNSIGNED, then there will be errors. In particular see
4923 gcc.c-torture/execute/990326-1.c. There does not appear to be
4924 any documentation paper trail as to why this is so. But the pre
4925 wide-int worked with that restriction and it has been preserved
4926 here. */
4931 }
4933 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
4934 represents the sign bit of EXP's type. If EXP represents a sign
4935 or zero extension, also test VAL against the unextended type.
4936 The return value is the (sub)expression whose sign bit is VAL,
4937 or NULL_TREE otherwise. */
4939 tree
4941 {
4943 tree t;
4945 /* Tree EXP must have an integral type. */
4950 /* Tree VAL must be an integer constant. */
4959 /* Handle extension from a narrower type. */
4965 }
4967 /* Subroutine for fold_truth_andor_1 and simple_condition_p: determine if an
4968 operand is simple enough to be evaluated unconditionally. */
4970 static bool
4972 {
4973 /* Strip any conversions that don't change the machine mode. */
4982 /* Don't regard global variables as simple. They may be
4983 allocated in ways unknown to the compiler (shared memory,
4984 #pragma weak, etc). */
4987 /* Weakrefs are not safe to be read, since they can be NULL.
4988 They are !TREE_PUBLIC && !DECL_EXTERNAL but still
4989 have DECL_WEAK flag set. */
4991 /* Loading a static variable is unduly expensive, but global
4992 registers aren't expensive. */
4994 }
4996 /* Determine if an operand is simple enough to be evaluated unconditionally.
4997 In addition to simple_operand_p, we assume that comparisons, conversions,
4998 and logic-not operations are simple, if their operands are simple, too. */
5000 bool
5002 {
5021 }
5023 \f
5024 /* The following functions are subroutines to fold_range_test and allow it to
5025 try to change a logical combination of comparisons into a range test.
5027 For example, both
5028 X == 2 || X == 3 || X == 4 || X == 5
5029 and
5030 X >= 2 && X <= 5
5031 are converted to
5032 (unsigned) (X - 2) <= 3
5034 We describe each set of comparisons as being either inside or outside
5035 a range, using a variable named like IN_P, and then describe the
5036 range with a lower and upper bound. If one of the bounds is omitted,
5037 it represents either the highest or lowest value of the type.
5039 In the comments below, we represent a range by two numbers in brackets
5040 preceded by a "+" to designate being inside that range, or a "-" to
5041 designate being outside that range, so the condition can be inverted by
5042 flipping the prefix. An omitted bound is represented by a "-". For
5043 example, "- [-, 10]" means being outside the range starting at the lowest
5044 possible value and ending at 10, in other words, being greater than 10.
5045 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
5046 always false.
5048 We set up things so that the missing bounds are handled in a consistent
5049 manner so neither a missing bound nor "true" and "false" need to be
5050 handled using a special case. */
5052 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
5053 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
5054 and UPPER1_P are nonzero if the respective argument is an upper bound
5055 and zero for a lower. TYPE, if nonzero, is the type of the result; it
5056 must be specified for a comparison. ARG1 will be converted to ARG0's
5057 type if both are specified. */
5059 static tree
5062 {
5063 tree tem;
5067 /* If neither arg represents infinity, do the normal operation.
5068 Else, if not a comparison, return infinity. Else handle the special
5069 comparison rules. Note that most of the cases below won't occur, but
5070 are handled for consistency. */
5073 {
5078 }
5083 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
5084 for neither. In real maths, we cannot assume open ended ranges are
5085 the same. But, this is computer arithmetic, where numbers are finite.
5086 We can therefore make the transformation of any unbounded range with
5087 the value Z, Z being greater than any representable number. This permits
5088 us to treat unbounded ranges as equal. */
5092 {
5113 }
5116 }
5117 \f
5118 /* Helper routine for make_range. Perform one step for it, return
5119 new expression if the loop should continue or NULL_TREE if it should
5120 stop. */
5122 tree
5126 {
5132 {
5134 /* We can only do something if the range is testing for zero. */
5143 /* We can only do something if the range is testing for zero
5144 and if the second operand is an integer constant. Note that
5145 saying something is "in" the range we make is done by
5146 complementing IN_P since it will set in the initial case of
5147 being not equal to zero; "out" is leaving it alone. */
5154 {
5175 }
5177 /* If this is an unsigned comparison, we also know that EXP is
5178 greater than or equal to zero. We base the range tests we make
5179 on that fact, so we record it here so we can parse existing
5180 range tests. We test arg0_type since often the return type
5181 of, e.g. EQ_EXPR, is boolean. */
5183 {
5187 NULL_TREE))
5192 /* If the high bound is missing, but we have a nonzero low
5193 bound, reverse the range so it goes from zero to the low bound
5194 minus 1. */
5196 {
5201 }
5202 }
5210 /* If flag_wrapv and ARG0_TYPE is signed, make sure
5211 low and high are non-NULL, then normalize will DTRT. */
5214 {
5219 }
5221 /* (-x) IN [a,b] -> x in [-b, -a] */
5233 /* ~ X -> -X - 1 */
5242 /* If flag_wrapv and ARG0_TYPE is signed, then we cannot
5243 move a constant to the other side. */
5248 /* If EXP is signed, any overflow in the computation is undefined,
5249 so we don't worry about it so long as our computations on
5250 the bounds don't overflow. For unsigned, overflow is defined
5251 and this is exactly the right thing. */
5263 normalize:
5264 /* Check for an unsigned range which has wrapped around the maximum
5265 value thus making n_high < n_low, and normalize it. */
5267 {
5273 /* If the range is of the form +/- [ x+1, x ], we won't
5274 be able to normalize it. But then, it represents the
5275 whole range or the empty set, so make it
5276 +/- [ -, - ]. */
5280 else
5282 }
5283 else
5291 CASE_CONVERT:
5309 /* If we're converting arg0 from an unsigned type, to exp,
5310 a signed type, we will be doing the comparison as unsigned.
5311 The tests above have already verified that LOW and HIGH
5312 are both positive.
5314 So we have to ensure that we will handle large unsigned
5315 values the same way that the current signed bounds treat
5316 negative values. */
5319 {
5320 tree high_positive;
5321 tree equiv_type;
5322 /* For fixed-point modes, we need to pass the saturating flag
5323 as the 2nd parameter. */
5325 equiv_type
5328 else
5329 equiv_type
5332 /* A range without an upper bound is, naturally, unbounded.
5333 Since convert would have cropped a very large value, use
5334 the max value for the destination type. */
5335 high_positive
5342 high_positive),
5345 /* If the low bound is specified, "and" the range with the
5346 range for which the original unsigned value will be
5347 positive. */
5349 {
5352 integer_zero_node),
5353 high_positive))
5357 }
5358 else
5359 {
5360 /* Otherwise, "or" the range with the range of the input
5361 that will be interpreted as negative. */
5364 integer_zero_node),
5365 high_positive))
5369 }
5370 }
5372 /* Otherwise, if we are converting arg0 from signed type, to exp,
5373 an unsigned type, we will do the comparison as signed. If
5374 high is non-NULL, we punt above if it doesn't fit in the signed
5375 type, so if we get through here, +[-, high] or +[low, high] are
5376 equivalent to +[-, n_high] or +[n_low, n_high]. Similarly,
5377 +[-, -] or -[-, -] are equivalent too. But if low is specified and
5378 high is not, the +[low, -] range is equivalent to union of
5379 +[n_low, -] and +[-, -1] ranges, so +[low, -] is equivalent to
5380 -[0, n_low-1] and similarly -[low, -] to +[0, n_low-1], except for
5381 low being 0, which should be treated as [-, -]. */
5384 && low
5386 {
5389 else
5390 {
5395 }
5396 }
5405 }
5406 }
5408 /* Given EXP, a logical expression, set the range it is testing into
5409 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
5410 actually being tested. *PLOW and *PHIGH will be made of the same
5411 type as the returned expression. If EXP is not a comparison, we
5412 will most likely not be returning a useful value and range. Set
5413 *STRICT_OVERFLOW_P to true if the return value is only valid
5414 because signed overflow is undefined; otherwise, do not change
5415 *STRICT_OVERFLOW_P. */
5417 tree
5420 {
5428 /* Start with simply saying "EXP != 0" and then look at the code of EXP
5429 and see if we can refine the range. Some of the cases below may not
5430 happen, but it doesn't seem worth worrying about this. We "continue"
5431 the outer loop when we've changed something; otherwise we "break"
5432 the switch, which will "break" the while. */
5438 {
5444 {
5452 }
5461 }
5463 /* If EXP is a constant, we can evaluate whether this is true or false. */
5465 {
5472 }
5476 }
5478 /* Returns TRUE if [LOW, HIGH] range check can be optimized to
5479 a bitwise check i.e. when
5480 LOW == 0xXX...X00...0
5481 HIGH == 0xXX...X11...1
5482 Return corresponding mask in MASK and stem in VALUE. */
5484 static bool
5487 {
5510 }
5511 \f
5512 /* Helper routine for build_range_check and match.pd. Return the type to
5513 perform the check or NULL if it shouldn't be optimized. */
5515 tree
5517 {
5518 /* First make sure that arithmetics in this type is valid, then make sure
5519 that it wraps around. */
5524 {
5527 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
5528 for the type in question, as we rely on this here. */
5538 else
5540 }
5544 }
5546 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
5547 type, TYPE, return an expression to test if EXP is in (or out of, depending
5548 on IN_P) the range. Return 0 if the test couldn't be created. */
5550 tree
5553 {
5556 /* Disable this optimization for function pointer expressions
5557 on targets that require function pointer canonicalization. */
5564 {
5570 }
5592 value);
5595 {
5597 {
5601 }
5603 }
5605 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
5607 {
5611 {
5613 {
5616 etype
5618 else
5621 }
5624 }
5625 }
5627 /* Optimize (c>=low) && (c<=high) into (c-low>=0) && (c-low<=high-low).
5628 This requires wrap-around arithmetics for the type of the expression. */
5645 }
5646 \f
5647 /* Return the predecessor of VAL in its type, handling the infinite case. */
5649 static tree
5651 {
5657 else
5660 }
5662 /* Return the successor of VAL in its type, handling the infinite case. */
5664 static tree
5666 {
5672 else
5675 }
5677 /* Given two ranges, see if we can merge them into one. Return 1 if we
5678 can, 0 if we can't. Set the output range into the specified parameters. */
5680 bool
5683 {
5687 tree tem;
5697 /* Make range 0 be the range that starts first, or ends last if they
5698 start at the same value. Swap them if it isn't. */
5701 || (lowequal
5704 {
5708 }
5710 /* If the second range is != high1 where high1 is the type maximum of
5711 the type, try first merging with < high1 range. */
5713 && high1
5720 {
5725 /* Similarly for the second range != low1 where low1 is the type minimum
5726 of the type, try first merging with > low1 range. */
5731 }
5733 /* Now flag two cases, whether the ranges are disjoint or whether the
5734 second range is totally subsumed in the first. Note that the tests
5735 below are simplified by the ones above. */
5741 /* We now have four cases, depending on whether we are including or
5742 excluding the two ranges. */
5744 {
5745 /* If they don't overlap, the result is false. If the second range
5746 is a subset it is the result. Otherwise, the range is from the start
5747 of the second to the end of the first. */
5752 else
5754 }
5757 {
5758 /* If they don't overlap, the result is the first range. If they are
5759 equal, the result is false. If the second range is a subset of the
5760 first, and the ranges begin at the same place, we go from just after
5761 the end of the second range to the end of the first. If the second
5762 range is not a subset of the first, or if it is a subset and both
5763 ranges end at the same place, the range starts at the start of the
5764 first range and ends just before the second range.
5765 Otherwise, we can't describe this as a single range. */
5771 {
5776 {
5777 /* We are in the weird situation where high0 > high1 but
5778 high1 has no successor. Punt. */
5780 }
5781 }
5783 {
5788 {
5789 /* low0 < low1 but low1 has no predecessor. Punt. */
5791 }
5792 }
5793 else
5795 }
5798 {
5799 /* If they don't overlap, the result is the second range. If the second
5800 is a subset of the first, the result is false. Otherwise,
5801 the range starts just after the first range and ends at the
5802 end of the second. */
5807 else
5808 {
5813 {
5814 /* high1 > high0 but high0 has no successor. Punt. */
5816 }
5817 }
5818 }
5820 else
5821 {
5822 /* The case where we are excluding both ranges. Here the complex case
5823 is if they don't overlap. In that case, the only time we have a
5824 range is if they are adjacent. If the second is a subset of the
5825 first, the result is the first. Otherwise, the range to exclude
5826 starts at the beginning of the first range and ends at the end of the
5827 second. */
5829 {
5834 else
5835 {
5836 /* Canonicalize - [min, x] into - [-, x]. */
5839 {
5842 GET_MODE_BITSIZE
5845 /* FALLTHROUGH */
5858 }
5860 /* Canonicalize - [x, max] into - [x, -]. */
5863 {
5866 GET_MODE_BITSIZE
5869 /* FALLTHROUGH */
5885 }
5887 /* The ranges might be also adjacent between the maximum and
5888 minimum values of the given type. For
5889 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
5890 return + [x + 1, y - 1]. */
5892 {
5899 }
5900 else
5902 }
5903 }
5906 else
5908 }
5912 }
5913 \f
5915 /* Subroutine of fold, looking inside expressions of the form
5916 A op B ? A : C, where (ARG00, COMP_CODE, ARG01), ARG1 and ARG2
5917 are the three operands of the COND_EXPR. This function is
5918 being used also to optimize A op B ? C : A, by reversing the
5919 comparison first.
5921 Return a folded expression whose code is not a COND_EXPR
5922 anymore, or NULL_TREE if no folding opportunity is found. */
5924 static tree
5928 {
5930 tree tem;
5935 /* If we have A op 0 ? A : -A, consider applying the following
5936 transformations:
5938 A == 0? A : -A same as -A
5939 A != 0? A : -A same as A
5940 A >= 0? A : -A same as abs (A)
5941 A > 0? A : -A same as abs (A)
5942 A <= 0? A : -A same as -abs (A)
5943 A < 0? A : -A same as -abs (A)
5945 None of these transformations work for modes with signed
5946 zeros. If A is +/-0, the first two transformations will
5947 change the sign of the result (from +0 to -0, or vice
5948 versa). The last four will fix the sign of the result,
5949 even though the original expressions could be positive or
5950 negative, depending on the sign of A.
5952 Note that all these transformations are correct if A is
5953 NaN, since the two alternatives (A and -A) are also NaNs. */
5960 /* In the case that A is of the form X-Y, '-A' (arg2) may
5961 have already been folded to Y-X, check for that. */
5969 {
5981 /* Fall through. */
5992 /* FALLTHRU */
5999 {
6000 /* A <= 0 ? A : -A for A INT_MIN is valid, but -abs(INT_MIN)
6001 is not, invokes UB both in abs and in the negation of it.
6002 So, use ABSU_EXPR instead. */
6007 }
6008 else
6009 {
6012 }
6016 }
6018 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
6019 A == 0 ? A : 0 is always 0 unless A is -0. Note that
6020 both transformations are correct when A is NaN: A != 0
6021 is then true, and A == 0 is false. */
6025 {
6030 }
6032 /* Try some transformations of A op B ? A : B.
6034 A == B? A : B same as B
6035 A != B? A : B same as A
6036 A >= B? A : B same as max (A, B)
6037 A > B? A : B same as max (B, A)
6038 A <= B? A : B same as min (A, B)
6039 A < B? A : B same as min (B, A)
6041 As above, these transformations don't work in the presence
6042 of signed zeros. For example, if A and B are zeros of
6043 opposite sign, the first two transformations will change
6044 the sign of the result. In the last four, the original
6045 expressions give different results for (A=+0, B=-0) and
6046 (A=-0, B=+0), but the transformed expressions do not.
6048 The first two transformations are correct if either A or B
6049 is a NaN. In the first transformation, the condition will
6050 be false, and B will indeed be chosen. In the case of the
6051 second transformation, the condition A != B will be true,
6052 and A will be chosen.
6054 The conversions to max() and min() are not correct if B is
6055 a number and A is not. The conditions in the original
6056 expressions will be false, so all four give B. The min()
6057 and max() versions would give a NaN instead. */
6060 /* Avoid these transformations if the COND_EXPR may be used
6061 as an lvalue in the C++ front-end. PR c++/19199. */
6062 && (in_gimple_form
6068 {
6074 {
6083 /* In C++ a ?: expression can be an lvalue, so put the
6084 operand which will be used if they are equal first
6085 so that we can convert this back to the
6086 corresponding COND_EXPR. */
6088 {
6096 }
6103 {
6111 }
6124 }
6125 }
6128 }
6131 \f
6132 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
6133 #define LOGICAL_OP_NON_SHORT_CIRCUIT \
6134 (BRANCH_COST (optimize_function_for_speed_p (cfun), \
6135 false) >= 2)
6136 #endif
6138 /* EXP is some logical combination of boolean tests. See if we can
6139 merge it into some range test. Return the new tree if so. */
6141 static tree
6144 {
6158 /* If op0 is known true or false and this is a short-circuiting
6159 operation we must not merge with op1 since that makes side-effects
6160 unconditional. So special-case this. */
6167 /* If this is an OR operation, invert both sides; we will invert
6168 again at the end. */
6172 /* If both expressions are the same, if we can merge the ranges, and we
6173 can build the range test, return it or it inverted. If one of the
6174 ranges is always true or always false, consider it to be the same
6175 expression as the other. */
6183 {
6187 }
6189 /* On machines where the branch cost is expensive, if this is a
6190 short-circuited branch and the underlying object on both sides
6191 is the same, make a non-short-circuit operation. */
6194 logical_op_non_short_circuit
6201 {
6202 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
6203 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
6204 which cases we can't do this. */
6212 {
6221 {
6224 WARN_STRICT_OVERFLOW_COMPARISON);
6228 }
6229 }
6230 }
6233 }
6234 \f
6235 /* Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P
6236 bit value. Arrange things so the extra bits will be set to zero if and
6237 only if C is signed-extended to its full width. If MASK is nonzero,
6238 it is an INTEGER_CST that should be AND'ed with the extra bits. */
6240 static tree
6242 {
6245 tree temp;
6250 /* We work by getting just the sign bit into the low-order bit, then
6251 into the high-order bit, then sign-extend. We then XOR that value
6252 with C. */
6256 /* We must use a signed type in order to get an arithmetic right shift.
6257 However, we must also avoid introducing accidental overflows, so that
6258 a subsequent call to integer_zerop will work. Hence we must
6259 do the type conversion here. At this point, the constant is either
6260 zero or one, and the conversion to a signed type can never overflow.
6261 We could get an overflow if this conversion is done anywhere else. */
6270 /* If necessary, convert the type back to match the type of C. */
6275 }
6276 \f
6277 /* For an expression that has the form
6278 (A && B) || ~B
6279 or
6280 (A || B) && ~B,
6281 we can drop one of the inner expressions and simplify to
6282 A || ~B
6283 or
6284 A && ~B
6285 LOC is the location of the resulting expression. OP is the inner
6286 logical operation; the left-hand side in the examples above, while CMPOP
6287 is the right-hand side. RHS_ONLY is used to prevent us from accidentally
6288 removing a condition that guards another, as in
6289 (A != NULL && A->...) || A == NULL
6290 which we must not transform. If RHS_ONLY is true, only eliminate the
6291 right-most operand of the inner logical operation. */
6293 static tree
6296 {
6314 {
6317 {
6320 }
6321 }
6323 {
6326 {
6329 }
6330 }
6345 }
6347 /* Find ways of folding logical expressions of LHS and RHS:
6348 Try to merge two comparisons to the same innermost item.
6349 Look for range tests like "ch >= '0' && ch <= '9'".
6350 Look for combinations of simple terms on machines with expensive branches
6351 and evaluate the RHS unconditionally.
6353 For example, if we have p->a == 2 && p->b == 4 and we can make an
6354 object large enough to span both A and B, we can do this with a comparison
6355 against the object ANDed with the a mask.
6357 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
6358 operations to do this with one comparison.
6360 We check for both normal comparisons and the BIT_AND_EXPRs made this by
6361 function and the one above.
6363 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
6364 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
6366 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
6367 two operands.
6369 We return the simplified tree or 0 if no optimization is possible. */
6371 static tree
6374 {
6375 /* If this is the "or" of two comparisons, we can do something if
6376 the comparisons are NE_EXPR. If this is the "and", we can do something
6377 if the comparisons are EQ_EXPR. I.e.,
6378 (a->b == 2 && a->c == 4) can become (a->new == NEW).
6380 WANTED_CODE is this operation code. For single bit fields, we can
6381 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
6382 comparison for one-bit fields. */
6403 /* Start by getting the comparison codes. Fail if anything is volatile.
6404 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
6405 it were surrounded with a NE_EXPR. */
6414 {
6418 }
6421 {
6425 }
6436 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
6439 {
6442 {
6447 }
6450 {
6456 }
6457 }
6462 /* If the RHS can be evaluated unconditionally and its operands are
6463 simple, it wins to evaluate the RHS unconditionally on machines
6464 with expensive branches. In this case, this isn't a comparison
6465 that can be merged. */
6472 {
6473 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
6484 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
6494 }
6496 /* See if the comparisons can be merged. Then get all the parameters for
6497 each side. */
6522 /* It must be true that the inner operation on the lhs of each
6523 comparison must be the same if we are to be able to do anything.
6524 Then see if we have constants. If not, the same must be true for
6525 the rhs's. */
6534 {
6537 }
6542 else
6545 /* If either comparison code is not correct for our logical operation,
6546 fail. However, we can convert a one-bit comparison against zero into
6547 the opposite comparison against that bit being set in the field. */
6551 {
6553 {
6554 /* Make the left operand unsigned, since we are only interested
6555 in the value of one bit. Otherwise we are doing the wrong
6556 thing below. */
6559 }
6560 else
6562 }
6564 /* This is analogous to the code for l_const above. */
6566 {
6568 {
6571 }
6572 else
6574 }
6576 /* See if we can find a mode that contains both fields being compared on
6577 the left. If we can't, fail. Otherwise, update all constants and masks
6578 to be relative to a field of that size. */
6592 {
6595 }
6605 {
6614 {
6618 }
6619 }
6621 {
6630 {
6634 }
6635 }
6637 /* If the right sides are not constant, do the same for it. Also,
6638 disallow this optimization if a size, signedness or storage order
6639 mismatch occurs between the left and right sides. */
6641 {
6645 /* Make sure the two fields on the right
6646 correspond to the left without being swapped. */
6663 {
6666 }
6677 /* Make a mask that corresponds to both fields being compared.
6678 Do this for both items being compared. If the operands are the
6679 same size and the bits being compared are in the same position
6680 then we can do this by masking both and comparing the masked
6681 results. */
6688 {
6702 }
6704 /* There is still another way we can do something: If both pairs of
6705 fields being compared are adjacent, we may be able to make a wider
6706 field containing them both.
6708 Note that we still must mask the lhs/rhs expressions. Furthermore,
6709 the mask must be shifted to account for the shift done by
6710 make_bit_field_ref. */
6719 {
6720 tree type;
6738 /* Convert to the smaller type before masking out unwanted bits. */
6741 {
6743 {
6747 }
6749 {
6753 }
6754 }
6763 }
6766 }
6768 /* Handle the case of comparisons with constants. If there is something in
6769 common between the masks, those bits of the constants must be the same.
6770 If not, the condition is always false. Test for this to avoid generating
6771 incorrect code below. */
6776 {
6778 {
6781 }
6782 else
6783 {
6786 }
6787 }
6792 /* Construct the expression we will return. First get the component
6793 reference we will make. Unless the mask is all ones the width of
6794 that field, perform the mask operation. Then compare with the
6795 merged constant. */
6806 }
6807 \f
6808 /* T is an integer expression that is being multiplied, divided, or taken a
6809 modulus (CODE says which and what kind of divide or modulus) by a
6810 constant C. See if we can eliminate that operation by folding it with
6811 other operations already in T. WIDE_TYPE, if non-null, is a type that
6812 should be used for the computation if wider than our type.
6814 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
6815 (X * 2) + (Y * 4). We must, however, be assured that either the original
6816 expression would not overflow or that overflow is undefined for the type
6817 in the language in question.
6819 If we return a non-null expression, it is an equivalent form of the
6820 original computation, but need not be in the original type.
6822 We set *STRICT_OVERFLOW_P to true if the return values depends on
6823 signed overflow being undefined. Otherwise we do not change
6824 *STRICT_OVERFLOW_P. */
6826 static tree
6829 {
6830 /* To avoid exponential search depth, refuse to allow recursion past
6831 three levels. Beyond that (1) it's highly unlikely that we'll find
6832 something interesting and (2) we've probably processed it before
6833 when we built the inner expression. */
6836 tree ret;
6841 depth++;
6843 depth--;
6846 }
6848 static tree
6851 {
6863 /* Don't deal with constants of zero here; they confuse the code below. */
6873 /* Note that we need not handle conditional operations here since fold
6874 already handles those cases. So just do arithmetic here. */
6876 {
6878 /* For a constant, we can always simplify if we are a multiply
6879 or (for divide and modulus) if it is a multiple of our constant. */
6883 {
6886 /* If the multiplication overflowed, we lost information on it.
6887 See PR68142 and PR69845. */
6891 }
6897 /* If op0 is an expression ... */
6903 /* ... and has wrapping overflow, and its type is smaller
6904 than ctype, then we cannot pass through as widening. */
6908 /* ... or this is a truncation (t is narrower than op0),
6909 then we cannot pass through this narrowing. */
6912 /* ... or signedness changes for division or modulus,
6913 then we cannot pass through this conversion. */
6917 /* ... or has undefined overflow while the converted to
6918 type has not, we cannot do the operation in the inner type
6919 as that would introduce undefined overflow. */
6924 /* Pass the constant down and see if we can make a simplification. If
6925 we can, replace this expression with the inner simplification for
6926 possible later conversion to our or some other type. */
6937 /* If widening the type changes it from signed to unsigned, then we
6938 must avoid building ABS_EXPR itself as unsigned. */
6940 {
6944 {
6947 }
6949 }
6950 /* If the constant is negative, we cannot simplify this. */
6953 /* FALLTHROUGH */
6955 /* For division and modulus, type can't be unsigned, as e.g.
6956 (-(x / 2U)) / 2U isn't equal to -((x / 2U) / 2U) for x >= 2.
6957 For signed types, even with wrapping overflow, this is fine. */
6966 /* If widening the type changes the signedness, then we can't perform
6967 this optimization as that changes the result. */
6971 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
6977 {
6984 }
6988 /* If the second operand is constant, this is a multiplication
6989 or floor division, by a power of two, so we can treat it that
6990 way unless the multiplier or divisor overflows. Signed
6991 left-shift overflow is implementation-defined rather than
6992 undefined in C90, so do not convert signed left shift into
6993 multiplication. */
6996 /* const_binop may not detect overflow correctly,
6997 so check for it explicitly here. */
7006 ctype,
7008 t1),
7013 /* See if we can eliminate the operation on both sides. If we can, we
7014 can return a new PLUS or MINUS. If we can't, the only remaining
7015 cases where we can do anything are if the second operand is a
7016 constant. */
7023 /* If not multiplication, we can only do this if both operands
7024 are divisible by c. */
7027 {
7032 }
7034 /* If this was a subtraction, negate OP1 and set it to be an addition.
7035 This simplifies the logic below. */
7037 {
7039 /* If OP1 was not easily negatable, the constant may be OP0. */
7041 {
7044 }
7045 }
7050 /* If either OP1 or C are negative, this optimization is not safe for
7051 some of the division and remainder types while for others we need
7052 to change the code. */
7054 {
7062 }
7064 /* If it's a multiply or a division/modulus operation of a multiple
7065 of our constant, do the operation and verify it doesn't overflow. */
7069 {
7072 /* We allow the constant to overflow with wrapping semantics. */
7076 }
7077 else
7080 /* If we have an unsigned type, we cannot widen the operation since it
7081 will change the result if the original computation overflowed. */
7085 /* The last case is if we are a multiply. In that case, we can
7086 apply the distributive law to commute the multiply and addition
7087 if the multiplication of the constants doesn't overflow
7088 and overflow is defined. With undefined overflow
7089 op0 * c might overflow, while (op0 + orig_op1) * c doesn't.
7090 But fold_plusminus_mult_expr would factor back any power-of-two
7091 value so do not distribute in the first place in this case. */
7099 op1);
7104 /* We have a special case here if we are doing something like
7105 (C * 8) % 4 since we know that's zero. */
7108 /* If the multiplication can overflow we cannot optimize this. */
7113 {
7116 }
7118 /* ... fall through ... */
7122 /* If we can extract our operation from the LHS, do so and return a
7123 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
7124 do something only if the second operand is a constant. */
7140 /* If these are the same operation types, we can associate them
7141 assuming no overflow. */
7143 {
7158 }
7160 /* If these operations "cancel" each other, we have the main
7161 optimizations of this pass, which occur when either constant is a
7162 multiple of the other, in which case we replace this with either an
7163 operation or CODE or TCODE.
7165 If we have an unsigned type, we cannot do this since it will change
7166 the result if the original computation overflowed. */
7174 {
7177 {
7183 }
7186 {
7192 }
7193 }
7198 }
7201 }
7202 \f
7203 /* Return a node which has the indicated constant VALUE (either 0 or
7204 1 for scalars or {-1,-1,..} or {0,0,...} for vectors),
7205 and is of the indicated TYPE. */
7207 tree
7209 {
7218 else
7220 }
7223 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
7224 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
7225 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
7226 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
7227 COND is the first argument to CODE; otherwise (as in the example
7228 given here), it is the second argument. TYPE is the type of the
7229 original expression. Return NULL_TREE if no simplification is
7230 possible. */
7232 static tree
7237 {
7245 /* Do not move possibly trapping operations into the conditional as this
7246 pessimizes code and causes gimplification issues when applied late. */
7254 {
7258 /* If this operand throws an expression, then it does not make
7259 sense to try to perform a logical or arithmetic operation
7260 involving it. */
7265 }
7268 {
7273 }
7274 else
7275 /* Detect the case of mixing vector and scalar types - bail out. */
7281 /* This transformation is only worthwhile if we don't have to wrap ARG
7282 in a SAVE_EXPR and the operation can be simplified without recursing
7283 on at least one of the branches once its pushed inside the COND_EXPR. */
7292 {
7296 else
7298 }
7300 {
7304 else
7306 }
7308 /* Check that we have simplified at least one of the branches. */
7313 }
7315 \f
7316 /* Subroutine of fold() that checks for the addition of ARG +/- 0.0.
7318 If !NEGATE, return true if ZERO_ARG is +/-0.0 and, for all ARG of
7319 type TYPE, ARG + ZERO_ARG is the same as ARG. If NEGATE, return true
7320 if ARG - ZERO_ARG is the same as X.
7322 If ARG is NULL, check for any value of type TYPE.
7324 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
7325 and finite. The problematic cases are when X is zero, and its mode
7326 has signed zeros. In the case of rounding towards -infinity,
7327 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
7328 modes, X + 0 is not the same as X because -0 + 0 is 0. */
7330 bool
7333 {
7337 /* Don't allow the fold with -fsignaling-nans. */
7341 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
7345 /* There is no case that is safe for all rounding modes. */
7349 /* In a vector or complex, we would need to check the sign of all zeros. */
7355 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
7359 /* The mode has signed zeros, and we have to honor their sign.
7360 In this situation, there are only two cases we can return true for.
7361 (i) X - 0 is the same as X with default rounding.
7362 (ii) X + 0 is X when X can't possibly be -0.0. */
7364 }
7366 /* Subroutine of match.pd that optimizes comparisons of a division by
7367 a nonzero integer constant against an integer constant, i.e.
7368 X/C1 op C2.
7370 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
7371 GE_EXPR or LE_EXPR. ARG01 and ARG1 must be a INTEGER_CST. */
7373 enum tree_code
7376 {
7381 /* We have to do this the hard way to detect unsigned overflow.
7382 prod = int_const_binop (MULT_EXPR, c1, c2); */
7388 {
7392 /* Likewise *hi = int_const_binop (PLUS_EXPR, prod, tmp). */
7395 }
7397 {
7400 {
7419 }
7420 }
7421 else
7422 {
7423 /* A negative divisor reverses the relational operators. */
7428 {
7447 }
7448 }
7461 }
7463 /* Test whether it is preferable to swap two operands, ARG0 and
7464 ARG1, for example because ARG0 is an integer constant and ARG1
7465 isn't. */
7467 bool
7469 {
7483 /* It is preferable to swap two SSA_NAME to ensure a canonical form
7484 for commutative and comparison operators. Ensuring a canonical
7485 form allows the optimizers to find additional redundancies without
7486 having to explicitly check for both orderings. */
7492 /* Put SSA_NAMEs last. */
7498 /* Put variables last. */
7505 }
7508 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
7509 means A >= Y && A != MAX, but in this case we know that
7510 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
7512 static tree
7514 {
7521 else
7530 {
7533 }
7535 {
7538 }
7539 else
7546 {
7547 /* Convert the pointer types into integer before taking the difference. */
7551 }
7552 else
7559 }
7561 /* Fold a sum or difference of at least one multiplication.
7562 Returns the folded tree or NULL if no simplification could be made. */
7564 static tree
7567 {
7571 /* (A * C) +- (B * C) -> (A+-B) * C.
7572 (A * C) +- A -> A * (C+-1).
7573 We are most concerned about the case where C is a constant,
7574 but other combinations show up during loop reduction. Since
7575 it is not difficult, try all four possibilities. */
7578 {
7581 }
7583 {
7586 }
7587 else
7588 {
7589 /* We cannot generate constant 1 for fract. */
7594 }
7596 {
7599 }
7601 {
7603 /* As we canonicalize A - 2 to A + -2 get rid of that sign for
7604 the purpose of this canonicalization. */
7608 {
7611 }
7612 else
7614 }
7615 else
7616 {
7617 /* We cannot generate constant 1 for fract. */
7622 }
7625 /* Prefer factoring a common non-constant. */
7635 /* No identical multiplicands; see if we can find a common
7636 power-of-two factor in non-power-of-two multiplies. This
7637 can help in multi-dimensional array access. */
7639 {
7642 HOST_WIDE_INT tmp;
7644 tree maybe_same;
7646 /* Move min of absolute values to int11. */
7648 {
7653 }
7654 else
7661 /* The remainder should not be a constant, otherwise we
7662 end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
7663 increased the number of multiplications necessary. */
7665 {
7673 }
7674 }
7681 /* We are neither factoring zero nor minus one. */
7689 /* Same may be zero and thus the operation 'code' may overflow. Likewise
7690 same may be minus one and thus the multiplication may overflow. Perform
7691 the sum operation in an unsigned type. */
7696 /* If the sum evaluated to a constant that is not -INF the multiplication
7697 cannot overflow. */
7704 /* Do not resort to unsigned multiplication because
7705 we lose the no-overflow property of the expression. */
7707 }
7709 /* Subroutine of native_encode_expr. Encode the INTEGER_CST
7710 specified by EXPR into the buffer PTR of length LEN bytes.
7711 Return the number of bytes placed in the buffer, or zero
7712 upon failure. */
7714 static int
7716 {
7728 /* Dry run. */
7734 {
7736 /* Extend EXPR according to TYPE_SIGN if the precision isn't a whole
7737 number of bytes. */
7741 {
7748 else
7750 }
7751 else
7755 }
7757 }
7760 /* Subroutine of native_encode_expr. Encode the FIXED_CST
7761 specified by EXPR into the buffer PTR of length LEN bytes.
7762 Return the number of bytes placed in the buffer, or zero
7763 upon failure. */
7765 static int
7767 {
7771 FIXED_VALUE_TYPE value;
7786 }
7789 /* Subroutine of native_encode_expr. Encode the REAL_CST
7790 specified by EXPR into the buffer PTR of length LEN bytes.
7791 Return the number of bytes placed in the buffer, or zero
7792 upon failure. */
7794 static int
7796 {
7802 /* There are always 32 bits in each long, no matter the size of
7803 the hosts long. We handle floating point representations with
7804 up to 192 bits. */
7813 /* Dry run. */
7822 {
7827 {
7834 else
7836 }
7837 else
7838 {
7841 {
7842 /* Reverse bytes within each long, or within the entire float
7843 if it's smaller than a long (for HFmode). */
7846 }
7847 }
7852 }
7854 }
7856 /* Subroutine of native_encode_expr. Encode the COMPLEX_CST
7857 specified by EXPR into the buffer PTR of length LEN bytes.
7858 Return the number of bytes placed in the buffer, or zero
7859 upon failure. */
7861 static int
7863 {
7865 tree part;
7879 }
7881 /* Like native_encode_vector, but only encode the first COUNT elements.
7882 The other arguments are as for native_encode_vector. */
7884 static int
7887 {
7891 {
7892 /* This is the only case in which elements can be smaller than a byte.
7893 Element 0 is always in the lsb of the containing byte. */
7902 /* Zero the buffer and then set bits later where necessary. */
7911 {
7917 {
7920 }
7921 }
7923 }
7928 {
7930 {
7933 }
7944 }
7946 }
7948 /* Subroutine of native_encode_expr. Encode the VECTOR_CST
7949 specified by EXPR into the buffer PTR of length LEN bytes.
7950 Return the number of bytes placed in the buffer, or zero
7951 upon failure. */
7953 static int
7955 {
7960 }
7963 /* Subroutine of native_encode_expr. Encode the STRING_CST
7964 specified by EXPR into the buffer PTR of length LEN bytes.
7965 Return the number of bytes placed in the buffer, or zero
7966 upon failure. */
7968 static int
7970 {
7973 /* Wide-char strings are encoded in target byte-order so native
7974 encoding them is trivial. */
7989 else
7990 {
7993 {
7996 }
7998 }
8000 }
8003 /* Subroutine of fold_view_convert_expr. Encode the INTEGER_CST, REAL_CST,
8004 FIXED_CST, COMPLEX_CST, STRING_CST, or VECTOR_CST specified by EXPR into
8005 the buffer PTR of size LEN bytes. If PTR is NULL, don't actually store
8006 anything, just do a dry run. Fail either if OFF is -1 and LEN isn't
8007 sufficient to encode the entire EXPR, or if OFF is out of bounds.
8008 Otherwise, start at byte offset OFF and encode at most LEN bytes.
8009 Return the number of bytes placed in the buffer, or zero upon failure. */
8011 int
8013 {
8014 /* We don't support starting at negative offset and -1 is special. */
8019 {
8040 }
8041 }
8043 /* Try to find a type whose byte size is smaller or equal to LEN bytes larger
8044 or equal to FIELDSIZE bytes, with underlying mode precision/size multiple
8045 of BITS_PER_UNIT. As native_{interpret,encode}_int works in term of
8046 machine modes, we can't just use build_nonstandard_integer_type. */
8048 tree
8050 {
8051 machine_mode mode;
8053 {
8060 {
8064 }
8065 }
8071 {
8079 }
8082 }
8084 /* Similar to native_encode_expr, but also handle CONSTRUCTORs, VCEs,
8085 NON_LVALUE_EXPRs and nops. If MASK is non-NULL (then PTR has
8086 to be non-NULL and OFF zero), then in addition to filling the
8087 bytes pointed by PTR with the value also clear any bits pointed
8088 by MASK that are known to be initialized, keep them as is for
8089 e.g. uninitialized padding bits or uninitialized fields. */
8091 int
8094 {
8097 /* We don't support starting at negative offset and -1 is special. */
8105 {
8109 mask);
8124 {
8125 tree min_index;
8143 {
8148 {
8151 }
8156 else
8160 {
8164 offset_int wpos
8172 offset_int wcount
8179 }
8181 {
8184 offset_int wpos
8192 }
8195 {
8197 {
8208 }
8210 {
8214 }
8215 }
8219 do
8220 {
8225 {
8227 {
8230 fieldsize);
8233 }
8235 ptr
8238 fieldsize,
8241 mask
8245 else
8246 {
8249 }
8250 }
8253 {
8254 /* Partial overlap. */
8260 {
8264 fieldsize);
8265 }
8266 else
8267 {
8271 }
8274 }
8276 }
8278 }
8280 }
8283 {
8293 {
8299 {
8302 }
8305 {
8314 }
8318 else
8322 {
8323 tree fld;
8325 {
8338 }
8340 {
8344 }
8347 {
8348 cnt--;
8354 }
8355 }
8360 {
8373 {
8378 }
8379 }
8380 else
8381 {
8386 }
8390 /* Prepare to deal with integral bit-fields and filter out other
8391 bit-fields that do not start and end on a byte boundary. */
8393 {
8398 {
8404 }
8410 }
8420 {
8421 /* FIXME: Handle PDP endian. */
8432 {
8435 }
8436 else
8437 {
8445 else
8446 {
8449 }
8450 }
8460 else
8473 /* If the bitfield does not start at byte boundary, handle
8474 the partial byte at the start. */
8477 {
8479 {
8484 {
8487 else
8489 }
8490 }
8491 else
8492 {
8497 {
8500 else
8504 }
8505 }
8506 }
8507 /* If the bitfield does not end at byte boundary, handle
8508 the partial byte at the end. */
8512 {
8514 {
8521 }
8522 else
8523 {
8530 }
8531 }
8535 {
8540 }
8541 else
8542 {
8543 /* Partial overlap. */
8547 {
8550 }
8554 }
8556 }
8561 {
8564 {
8567 {
8571 }
8574 }
8577 fldsize,
8587 }
8588 else
8589 {
8590 /* Partial overlap. */
8596 {
8600 fieldsize);
8601 }
8602 else
8603 {
8607 }
8610 }
8611 }
8613 }
8615 }
8616 }
8619 /* Subroutine of native_interpret_expr. Interpret the contents of
8620 the buffer PTR of length LEN as an INTEGER_CST of type TYPE.
8621 If the buffer cannot be interpreted, return NULL_TREE. */
8623 static tree
8625 {
8635 }
8638 /* Subroutine of native_interpret_expr. Interpret the contents of
8639 the buffer PTR of length LEN as a FIXED_CST of type TYPE.
8640 If the buffer cannot be interpreted, return NULL_TREE. */
8642 static tree
8644 {
8647 double_int result;
8648 FIXED_VALUE_TYPE fixed_value;
8658 }
8661 /* Subroutine of native_interpret_expr. Interpret the contents of
8662 the buffer PTR of length LEN as a REAL_CST of type TYPE.
8663 If the buffer cannot be interpreted, return NULL_TREE. */
8665 tree
8667 {
8671 /* There are always 32 bits in each long, no matter the size of
8672 the hosts long. We handle floating point representations with
8673 up to 192 bits. */
8674 REAL_VALUE_TYPE r;
8684 {
8685 /* Both OFFSET and BYTE index within a long;
8686 bitpos indexes the whole float. */
8689 {
8696 else
8698 }
8699 else
8700 {
8703 {
8704 /* Reverse bytes within each long, or within the entire float
8705 if it's smaller than a long (for HFmode). */
8708 }
8709 }
8713 }
8717 }
8720 /* Subroutine of native_interpret_expr. Interpret the contents of
8721 the buffer PTR of length LEN as a COMPLEX_CST of type TYPE.
8722 If the buffer cannot be interpreted, return NULL_TREE. */
8724 static tree
8726 {
8741 }
8743 /* Read a vector of type TYPE from the target memory image given by BYTES,
8744 which contains LEN bytes. The vector is known to be encodable using
8745 NPATTERNS interleaved patterns with NELTS_PER_PATTERN elements each.
8747 Return the vector on success, otherwise return null. */
8749 static tree
8753 {
8757 {
8758 /* This is the only case in which elements can be smaller than a byte.
8759 Element 0 is always in the lsb of the containing byte. */
8766 {
8773 }
8775 }
8783 {
8789 }
8791 }
8793 /* Subroutine of native_interpret_expr. Interpret the contents of
8794 the buffer PTR of length LEN as a VECTOR_CST of type TYPE.
8795 If the buffer cannot be interpreted, return NULL_TREE. */
8797 static tree
8799 {
8808 }
8811 /* Subroutine of fold_view_convert_expr. Interpret the contents of
8812 the buffer PTR of length LEN as a constant of type TYPE. For
8813 INTEGRAL_TYPE_P we return an INTEGER_CST, for SCALAR_FLOAT_TYPE_P
8814 we return a REAL_CST, etc... If the buffer cannot be interpreted,
8815 return NULL_TREE. */
8817 tree
8819 {
8821 {
8832 {
8833 /* For floating point values in composite modes, punt if this
8834 folding doesn't preserve bit representation. As the mode doesn't
8835 have fixed precision while GCC pretends it does, there could be
8836 valid values that GCC can't really represent accurately.
8837 See PR95450. Even for other modes, e.g. x86 XFmode can have some
8838 bit combinationations which GCC doesn't preserve. */
8848 }
8862 }
8863 }
8865 /* Returns true if we can interpret the contents of a native encoding
8866 as TYPE. */
8868 bool
8870 {
8872 {
8886 }
8887 }
8889 /* Attempt to interpret aggregate of TYPE from bytes encoded in target
8890 byte order at PTR + OFF with LEN bytes. Does not handle unions. */
8892 tree
8895 {
8898 {
8905 {
8909 }
8914 {
8919 {
8924 }
8928 eltsz);
8932 }
8934 }
8938 {
8947 {
8950 {
8951 poly_int64 bitoffset;
8956 else
8966 }
8967 else
8968 {
8986 HOST_WIDE_INT rpos;
8989 else
8990 {
8993 }
9002 }
9003 }
9006 {
9016 }
9020 {
9025 }
9032 {
9036 /* FIXME: Figure out how to handle PDP endian bitfields. */
9042 else
9046 }
9048 }
9050 }
9052 /* Routines for manipulation of native_encode_expr encoded data if the encoded
9053 or extracted constant positions and/or sizes aren't byte aligned. */
9055 /* Shift left the bytes in PTR of SZ elements by AMNT bits, carrying over the
9056 bits between adjacent elements. AMNT should be within
9057 [0, BITS_PER_UNIT).
9058 Example, AMNT = 2:
9059 00011111|11100000 << 2 = 01111111|10000000
9060 PTR[1] | PTR[0] PTR[1] | PTR[0]. */
9062 void
9065 {
9074 {
9080 {
9083 }
9084 }
9085 }
9087 /* Like shift_bytes_in_array_left but for big-endian.
9088 Shift right the bytes in PTR of SZ elements by AMNT bits, carrying over the
9089 bits between adjacent elements. AMNT should be within
9090 [0, BITS_PER_UNIT).
9091 Example, AMNT = 2:
9092 00011111|11100000 >> 2 = 00000111|11111000
9093 PTR[0] | PTR[1] PTR[0] | PTR[1]. */
9095 void
9098 {
9106 {
9113 }
9114 }
9116 /* Try to view-convert VECTOR_CST EXPR to VECTOR_TYPE TYPE by operating
9117 directly on the VECTOR_CST encoding, in a way that works for variable-
9118 length vectors. Return the resulting VECTOR_CST on success or null
9119 on failure. */
9121 static tree
9123 {
9135 /* We can only preserve the semantics of a stepped pattern if the new
9136 vector element is an integer of the same size. */
9141 /* The number of bits needed to encode one element from every pattern
9142 of the original vector. */
9143 unsigned int expr_sequence_bits
9146 /* The number of bits needed to encode one element from every pattern
9147 of the result. */
9148 unsigned int type_sequence_bits
9151 /* Don't try to read more bytes than are available, which can happen
9152 for constant-sized vectors if TYPE has larger elements than EXPR_TYPE.
9153 The general VIEW_CONVERT handling can cope with that case, so there's
9154 no point complicating things here. */
9157 BITS_PER_UNIT);
9162 /* Get enough bytes of EXPR to form the new encoding. */
9170 /* Reencode the bytes as TYPE. */
9174 }
9176 /* Fold a VIEW_CONVERT_EXPR of a constant expression EXPR to type
9177 TYPE at compile-time. If we're unable to perform the conversion
9178 return NULL_TREE. */
9180 static tree
9182 {
9183 /* We support up to 512-bit values (for V8DFmode). */
9187 /* Check that the host and target are sane. */
9200 }
9202 /* Build an expression for the address of T. Folds away INDIRECT_REF
9203 to avoid confusing the gimplify process. */
9205 tree
9207 {
9208 /* The size of the object is not relevant when talking about its address. */
9213 {
9218 }
9221 {
9226 }
9233 {
9238 }
9239 else
9243 }
9245 /* Build an expression for the address of T. */
9247 tree
9249 {
9253 }
9255 /* Fold a unary expression of code CODE and type TYPE with operand
9256 OP0. Return the folded expression if folding is successful.
9257 Otherwise, return NULL_TREE. */
9259 tree
9261 {
9262 tree tem;
9263 tree arg0;
9271 {
9274 {
9275 /* Don't use STRIP_NOPS, because signedness of argument type
9276 matters. */
9278 }
9279 else
9280 {
9281 /* Strip any conversions that don't change the mode. This
9282 is safe for every expression, except for a comparison
9283 expression because its signedness is derived from its
9284 operands.
9286 Note that this is done as an internal manipulation within
9287 the constant folder, in order to find the simplest
9288 representation of the arguments so that their form can be
9289 studied. In any cases, the appropriate type conversions
9290 should be put back in the tree that will get out of the
9291 constant folder. */
9293 }
9296 {
9299 {
9303 }
9304 }
9305 }
9312 {
9319 {
9333 /* If this was a conversion, and all we did was to move into
9334 inside the COND_EXPR, bring it back out. But leave it if
9335 it is a conversion from integer to integer and the
9336 result precision is no wider than a word since such a
9337 conversion is cheap and may be optimized away by combine,
9338 while it couldn't if it were outside the COND_EXPR. Then return
9339 so we don't get into an infinite recursion loop taking the
9340 conversion out and then back in. */
9352 && (INTEGRAL_TYPE_P
9365 }
9366 }
9369 {
9375 CASE_CONVERT:
9379 {
9380 /* If we have (type) (a CMP b) and type is an integral type, return
9381 new expression involving the new type. Canonicalize
9382 (type) (a CMP b) to (a CMP b) ? (type) true : (type) false for
9383 non-integral type.
9384 Do not fold the result as that would not simplify further, also
9385 folding again results in recursions. */
9395 }
9397 /* Handle (T *)&A.B.C for A being of type T and B and C
9398 living at offset zero. This occurs frequently in
9399 C++ upcasting and then accessing the base. */
9403 {
9405 tree offset;
9406 machine_mode mode;
9408 tree base
9412 /* If the reference was to a (constant) zero offset, we can use
9413 the address of the base if it has the same base type
9414 as the result type and the pointer type is unqualified. */
9422 }
9426 /* Detect assigning a bitfield. */
9428 && DECL_BIT_FIELD
9430 {
9431 /* Don't leave an assignment inside a conversion
9432 unless assigning a bitfield. */
9434 /* First do the assignment, then return converted constant. */
9439 }
9441 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
9442 constants (if x has signed type, the sign bit cannot be set
9443 in c). This folds extension into the BIT_AND_EXPR.
9444 ??? We don't do it for BOOLEAN_TYPE or ENUMERAL_TYPE because they
9445 very likely don't have maximal range for their precision and this
9446 transformation effectively doesn't preserve non-maximal ranges. */
9450 {
9461 <= HOST_BITS_PER_WIDE_INT
9463 {
9467 cst &= HOST_WIDE_INT_M1U
9471 && !flag_syntax_only
9474 {
9478 }
9479 }
9481 {
9486 }
9487 }
9489 /* Convert (T1)(X p+ Y) into ((T1)X p+ Y), for pointer type, when the new
9490 cast (T1)X will fold away. We assume that this happens when X itself
9491 is a cast. */
9495 {
9499 /* If -fsanitize=alignment, avoid this optimization in GENERIC
9500 when the pointed type needs higher alignment than
9501 the p+ first operand's pointed type. */
9508 /* Similarly, avoid this optimization in GENERIC for -fsanitize=null
9509 when type is a reference type and arg00's type is not,
9510 because arg00 could be validly nullptr and if arg01 doesn't return,
9511 we don't want false positive binding of reference to nullptr. */
9513 && !in_gimple_form
9520 }
9522 /* Convert (T1)(~(T2)X) into ~(T1)X if T1 and T2 are integral types
9523 of the same precision, and X is an integer type not narrower than
9524 types T1 or T2, i.e. the cast (T2)X isn't an extension. */
9530 {
9536 }
9538 /* Convert (T1)(X * Y) into (T1)X * (T1)Y if T1 is narrower than the
9539 type of X and Y (integer types only). */
9546 {
9547 /* Be careful not to introduce new overflows. */
9548 tree mult_type;
9551 else
9555 {
9562 }
9563 }
9569 {
9576 }
9587 /* Convert fabs((double)float) into (double)fabsf(float). */
9590 {
9596 targ0));
9597 }
9601 /* Convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
9620 /* Note that the operand of this must be an int
9621 and its values must be 0 or 1.
9622 ("true" is a fixed value perhaps depending on the language,
9623 but we don't handle values other than 1 correctly yet.) */
9630 /* Fold *&X to X if X is an lvalue. */
9632 {
9639 }
9645 }
9648 /* If the operation was a conversion do _not_ mark a resulting constant
9649 with TREE_OVERFLOW if the original constant was not. These conversions
9650 have implementation defined behavior and retaining the TREE_OVERFLOW
9651 flag here would confuse later passes such as VRP. */
9652 tree
9655 {
9664 }
9666 /* Fold a binary bitwise/truth expression of code CODE and type TYPE with
9667 operands OP0 and OP1. LOC is the location of the resulting expression.
9668 ARG0 and ARG1 are the NOP_STRIPed results of OP0 and OP1.
9669 Return the folded expression if folding is successful. Otherwise,
9670 return NULL_TREE. */
9671 static tree
9674 {
9675 tree tem;
9677 /* We only do these simplifications if we are optimizing. */
9681 /* Check for things like (A || B) && (A || C). We can convert this
9682 to A || (B && C). Note that either operator can be any of the four
9683 truth and/or operations and the transformation will still be
9684 valid. Also note that we only care about order for the
9685 ANDIF and ORIF operators. If B contains side effects, this
9686 might change the truth-value of A. */
9693 {
9713 /* This case if tricky because we must either have commutative
9714 operators or else A10 must not have side-effects. */
9720 a01);
9721 }
9723 /* See if we can build a range comparison. */
9729 {
9733 }
9737 {
9741 }
9743 /* Check for the possibility of merging component references. If our
9744 lhs is another similar operation, try to merge its rhs with our
9745 rhs. Then try to merge our lhs and rhs. */
9756 logical_op_non_short_circuit
9764 {
9771 /* Transform ((A AND-IF B) AND[-IF] C) into (A AND-IF (B AND C)),
9772 or ((A OR-IF B) OR[-IF] C) into (A OR-IF (B OR C))
9773 We don't want to pack more than two leafs to a non-IF AND/OR
9774 expression.
9775 If tree-code of left-hand operand isn't an AND/OR-IF code and not
9776 equal to IF-CODE, then we don't want to add right-hand operand.
9777 If the inner right-hand side of left-hand operand has
9778 side-effects, or isn't simple, then we can't add to it,
9779 as otherwise we might destroy if-sequence. */
9782 /* Needed for sequence points to handle trappings, and
9783 side-effects. */
9785 {
9787 arg1);
9789 tem);
9790 }
9791 /* Same as above but for (A AND[-IF] (B AND-IF C)) -> ((A AND B) AND-IF C),
9792 or (A OR[-IF] (B OR-IF C) -> ((A OR B) OR-IF C). */
9795 /* Needed for sequence points to handle trappings, and
9796 side-effects. */
9798 {
9803 }
9804 /* Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
9805 into (A OR B).
9806 For sequence point consistancy, we need to check for trapping,
9807 and side-effects. */
9811 }
9814 }
9816 /* Helper that tries to canonicalize the comparison ARG0 CODE ARG1
9817 by changing CODE to reduce the magnitude of constants involved in
9818 ARG0 of the comparison.
9819 Returns a canonicalized comparison tree if a simplification was
9820 possible, otherwise returns NULL_TREE.
9821 Set *STRICT_OVERFLOW_P to true if the canonicalization is only
9822 valid if signed overflow is undefined. */
9824 static tree
9828 {
9833 /* Match A +- CST code arg1. We can change this only if overflow
9834 is undefined. */
9837 /* In principle pointers also have undefined overflow behavior,
9838 but that causes problems elsewhere. */
9845 /* Identify the constant in arg0 and its sign. */
9849 /* Overflowed constants and zero will cause problems. */
9854 /* See if we can reduce the magnitude of the constant in
9855 arg0 by changing the comparison code. */
9856 /* A - CST < arg1 -> A - CST-1 <= arg1. */
9860 /* A + CST > arg1 -> A + CST-1 >= arg1. */
9864 /* A + CST <= arg1 -> A + CST-1 < arg1. */
9868 /* A - CST >= arg1 -> A - CST-1 > arg1. */
9872 else
9876 /* Now build the constant reduced in magnitude. But not if that
9877 would produce one outside of its types range. */
9893 }
9895 /* Canonicalize the comparison ARG0 CODE ARG1 with type TYPE with undefined
9896 overflow further. Try to decrease the magnitude of constants involved
9897 by changing LE_EXPR and GE_EXPR to LT_EXPR and GT_EXPR or vice versa
9898 and put sole constants at the second argument position.
9899 Returns the canonicalized tree if changed, otherwise NULL_TREE. */
9901 static tree
9904 {
9905 tree t;
9910 /* Try canonicalization by simplifying arg0. */
9915 {
9919 }
9921 /* Try canonicalization by simplifying arg1 using the swapped
9922 comparison. */
9930 }
9932 /* Return whether BASE + OFFSET + BITPOS may wrap around the address
9933 space. This is used to avoid issuing overflow warnings for
9934 expressions like &p->x which cannot wrap. */
9936 static bool
9938 {
9945 poly_wide_int wi_offset;
9951 else
9956 precision);
9971 /* We can do slightly better for SIZE if we have an ADDR_EXPR of an
9972 array. */
9981 }
9983 /* Return a positive integer when the symbol DECL is known to have
9984 a nonzero address, zero when it's known not to (e.g., it's a weak
9985 symbol), and a negative integer when the symbol is not yet in the
9986 symbol table and so whether or not its address is zero is unknown.
9987 For function local objects always return positive integer. */
9988 static int
9990 {
9991 /* Normally, don't do anything for variables and functions before symtab is
9992 built; it is quite possible that DECL will be declared weak later.
9993 But if folding_initializer, we need a constant answer now, so create
9994 the symtab entry and prevent later weak declaration. */
9997 = (folding_initializer
10002 /* Function local objects are never NULL. */
10010 }
10012 /* Subroutine of fold_binary. This routine performs all of the
10013 transformations that are common to the equality/inequality
10014 operators (EQ_EXPR and NE_EXPR) and the ordering operators
10015 (LT_EXPR, LE_EXPR, GE_EXPR and GT_EXPR). Callers other than
10016 fold_binary should call fold_binary. Fold a comparison with
10017 tree code CODE and type TYPE with operands OP0 and OP1. Return
10018 the folded comparison or NULL_TREE. */
10020 static tree
10023 {
10033 /* For comparisons of pointers we can decompose it to a compile time
10034 comparison of the base objects and the offsets into the object.
10035 This requires at least one operand being an ADDR_EXPR or a
10036 POINTER_PLUS_EXPR to do more than the operand_equal_p test below. */
10042 {
10045 machine_mode mode;
10049 /* Get base and offset for the access. Strip ADDR_EXPR for
10050 get_inner_reference, but put it back by stripping INDIRECT_REF
10051 off the base object if possible. indirect_baseN will be true
10052 if baseN is not an address but refers to the object itself. */
10055 {
10056 base0
10062 else
10064 }
10066 {
10070 {
10071 base0
10077 else
10079 }
10082 else
10086 {
10093 }
10094 }
10098 {
10099 base1
10105 else
10107 }
10109 {
10113 {
10114 base1
10120 else
10122 }
10125 else
10129 {
10136 }
10137 }
10139 /* If we have equivalent bases we might be able to simplify. */
10143 {
10144 /* We can fold this expression to a constant if the non-constant
10145 offset parts are equal. */
10149 && (equality_code
10150 || (indirect_base0
10153 {
10159 "occur when comparing P +- C1 with "
10161 WARN_STRICT_OVERFLOW_CONDITIONAL);
10164 {
10202 }
10203 }
10204 /* We can simplify the comparison to a comparison of the variable
10205 offset parts if the constant offset parts are equal.
10206 Be careful to use signed sizetype here because otherwise we
10207 mess with array offsets in the wrong way. This is possible
10208 because pointer arithmetic is restricted to retain within an
10209 object and overflow on pointer differences is undefined as of
10210 6.5.6/8 and /9 with respect to the signed ptrdiff_t. */
10212 && (equality_code
10213 || (indirect_base0
10216 {
10217 /* By converting to signed sizetype we cover middle-end pointer
10218 arithmetic which operates on unsigned pointer types of size
10219 type size and ARRAY_REF offsets which are properly sign or
10220 zero extended from their type in case it is narrower than
10221 sizetype. */
10224 else
10228 else
10235 "occur when comparing P +- C1 with "
10237 WARN_STRICT_OVERFLOW_COMPARISON);
10240 }
10241 }
10242 /* For equal offsets we can simplify to a comparison of the
10243 base addresses. */
10245 && (indirect_base0
10247 && (indirect_base1
10252 {
10258 }
10259 /* Comparison between an ordinary (non-weak) symbol and a null
10260 pointer can be eliminated since such symbols must have a non
10261 null address. In C, relational expressions between pointers
10262 to objects and null pointers are undefined. The results
10263 below follow the C++ rules with the additional property that
10264 every object pointer compares greater than a null pointer.
10265 */
10268 /* Avoid folding references to struct members at offset 0 to
10269 prevent tests like '&ptr->firstmember == 0' from getting
10270 eliminated. When ptr is null, although the -> expression
10271 is strictly speaking invalid, GCC retains it as a matter
10272 of QoI. See PR c/44555. */
10275 && indirect_base0
10276 /* The caller guarantees that when one of the arguments is
10277 constant (i.e., null in this case) it is second. */
10279 {
10281 {
10292 }
10293 }
10294 }
10296 /* Transform comparisons of the form X +- C1 CMP Y +- C2 to
10297 X CMP Y +- C2 +- C1 for signed X, Y. This is valid if
10298 the resulting offset is smaller in absolute value than the
10299 original one and has the same sign. */
10308 {
10313 tree cst;
10315 "occur when combining constants around "
10318 /* Put the constant on the side where it doesn't overflow and is
10319 of lower absolute value and of same sign than before. */
10326 {
10329 variable1,
10333 }
10341 {
10347 variable2);
10348 }
10349 }
10355 /* If we are comparing an expression that just has comparisons
10356 of two integer values, arithmetic expressions of those comparisons,
10357 and constants, we can simplify it. There are only three cases
10358 to check: the two values can either be equal, the first can be
10359 greater, or the second can be greater. Fold the expression for
10360 those three values. Since each value must be 0 or 1, we have
10361 eight possibilities, each of which corresponds to the constant 0
10362 or 1 or one of the six possible comparisons.
10364 This handles common cases like (a > b) == 0 but also handles
10365 expressions like ((x > y) - (y > x)) > 0, which supposedly
10366 occur in macroized code. */
10369 {
10373 /* Don't handle degenerate cases here; they should already
10374 have been handled anyway. */
10383 {
10387 /* We can't just pass T to eval_subst in case cval1 or cval2
10388 was the same as ARG1. */
10390 tree high_result
10394 arg1);
10395 tree equal_result
10399 arg1);
10400 tree low_result
10404 arg1);
10406 /* All three of these results should be 0 or 1. Confirm they are.
10407 Then use those values to select the proper code to use. */
10412 {
10413 /* Make a 3-bit mask with the high-order bit being the
10414 value for `>', the next for '=', and the low for '<'. */
10418 {
10420 /* Always false. */
10441 /* Always true. */
10443 }
10446 }
10447 }
10448 }
10451 }
10454 /* Subroutine of fold_binary. Optimize complex multiplications of the
10455 form z * conj(z), as pow(realpart(z),2) + pow(imagpart(z),2). The
10456 argument EXPR represents the expression "z" of type TYPE. */
10458 static tree
10460 {
10465 {
10468 }
10470 {
10473 }
10474 else
10475 {
10479 }
10488 }
10491 /* Helper function for fold_vec_perm. Store elements of VECTOR_CST or
10492 CONSTRUCTOR ARG into array ELTS, which has NELTS elements, and return
10493 true if successful. */
10495 static bool
10497 {
10502 {
10505 }
10507 {
10513 else
10515 }
10516 else
10522 }
10524 /* Helper routine for fold_vec_perm_cst to check if SEL is a suitable
10525 mask for VLA vec_perm folding.
10526 REASON if specified, will contain the reason why SEL is not suitable.
10527 Used only for debugging and unit-testing. */
10529 static bool
10533 {
10540 {
10544 }
10546 /* We want to avoid cases where sel.length is not a multiple of npatterns.
10547 For eg: sel.length = 2 + 2x, and sel npatterns = 4. */
10548 poly_uint64 esel;
10550 {
10554 }
10560 {
10563 HOST_WIDE_INT step;
10565 {
10569 }
10570 // FIXME: Punt on step < 0 for now, revisit later.
10577 {
10581 }
10583 /* Ensure that stepped sequence of the pattern selects elements
10584 only from the same input vector. */
10593 {
10597 }
10599 /* Ensure that the stepped sequence always selects from the same
10600 input pattern. */
10601 unsigned arg_npatterns
10606 {
10610 }
10611 }
10614 }
10616 /* Try to fold permutation of ARG0 and ARG1 with SEL selector when
10617 the input vectors are VECTOR_CST. Return NULL_TREE otherwise.
10618 REASON has same purpose as described in
10619 valid_mask_for_fold_vec_perm_cst_p. */
10621 static tree
10624 {
10628 /* (1) If SEL is a suitable mask as determined by
10629 valid_mask_for_fold_vec_perm_cst_p, then:
10630 res_npatterns = max of npatterns between ARG0, ARG1, and SEL
10631 res_nelts_per_pattern = max of nelts_per_pattern between
10632 ARG0, ARG1 and SEL.
10633 (2) If SEL is not a suitable mask, and TYPE is VLS then:
10634 res_npatterns = nelts in result vector.
10635 res_nelts_per_pattern = 1.
10636 This exception is made so that VLS ARG0, ARG1 and SEL work as before. */
10638 {
10639 res_npatterns
10644 res_nelts_per_pattern
10650 }
10652 {
10655 }
10656 else
10661 {
10664 poly_uint64 r;
10667 /* Punt if sel[i] /trunc_div len cannot be determined,
10668 because the input vector to be chosen will depend on
10669 runtime vector length.
10670 For example if len == 4 + 4x, and sel[i] == 4,
10671 If len at runtime equals 4, we choose arg1[0].
10672 For any other value of len > 4 at runtime, we choose arg0[4].
10673 which makes the element choice dependent on runtime vector length. */
10675 {
10679 }
10681 /* sel[i] % len will give the index of element in the chosen input
10682 vector. For example if sel[i] == 5 + 4x and len == 4 + 4x,
10683 we will choose arg1[1] since (5 + 4x) % (4 + 4x) == 1. */
10685 {
10689 }
10694 }
10697 }
10699 /* Attempt to fold vector permutation of ARG0 and ARG1 vectors using SEL
10700 selector. Return the folded VECTOR_CST or CONSTRUCTOR if successful,
10701 NULL_TREE otherwise. */
10703 tree
10705 {
10721 /* For fall back case, we want to ensure we have VLS vectors
10722 with equal length. */
10736 {
10737 HOST_WIDE_INT index;
10741 }
10743 }
10745 /* Try to fold a pointer difference of type TYPE two address expressions of
10746 array references AREF0 and AREF1 using location LOC. Return a
10747 simplified expression for the difference or NULL_TREE. */
10749 static tree
10753 {
10758 /* If the bases are array references as well, recurse. If the bases
10759 are pointer indirections compute the difference of the pointers.
10760 If the bases are equal, we are set. */
10763 && (base_offset
10765 use_pointer_diff)))
10768 && (base_offset
10769 = use_pointer_diff
10779 {
10785 base_offset,
10788 }
10790 }
10792 /* If the real or vector real constant CST of type TYPE has an exact
10793 inverse, return it, else return NULL. */
10795 tree
10797 {
10798 REAL_VALUE_TYPE r;
10799 tree unit_type;
10800 machine_mode mode;
10803 {
10813 {
10817 tree_vector_builder elts;
10822 {
10827 }
10830 }
10834 }
10835 }
10837 /* Mask out the tz least significant bits of X of type TYPE where
10838 tz is the number of trailing zeroes in Y. */
10839 static wide_int
10841 {
10846 }
10848 /* Return true when T is an address and is known to be nonzero.
10849 For floating point we further ensure that T is not denormal.
10850 Similar logic is present in nonzero_address in rtlanal.h.
10852 If the return value is based on the assumption that signed overflow
10853 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
10854 change *STRICT_OVERFLOW_P. */
10856 static bool
10858 {
10862 /* Doing something useful for floating point would need more work. */
10868 {
10871 strict_overflow_p);
10877 strict_overflow_p);
10885 }
10888 {
10891 strict_overflow_p);
10899 strict_overflow_p);
10913 strict_overflow_p);
10917 strict_overflow_p);
10920 {
10932 }
10936 }
10938 }
10940 /* Return true when T is an address and is known to be nonzero.
10941 Handle warnings about undefined signed overflow. */
10943 bool
10945 {
10952 "determining that expression is always "
10954 WARN_STRICT_OVERFLOW_MISC);
10956 }
10958 /* Return true if T is known not to be equal to an integer W. */
10960 bool
10962 {
10963 int_range_max vr;
10965 {
10975 else
10980 /* If T has some known zero bits and W has any of those bits set,
10981 then T is known not to be equal to W. */
10989 }
10990 }
10992 /* Fold a binary expression of code CODE and type TYPE with operands
10993 OP0 and OP1. LOC is the location of the resulting expression.
10994 Return the folded expression if folding is successful. Otherwise,
10995 return NULL_TREE. */
10997 tree
11000 {
11015 /* Strip any conversions that don't change the mode. This is
11016 safe for every expression, except for a comparison expression
11017 because its signedness is derived from its operands. So, in
11018 the latter case, only strip conversions that don't change the
11019 signedness. MIN_EXPR/MAX_EXPR also need signedness of arguments
11020 preserved.
11022 Note that this is done as an internal manipulation within the
11023 constant folder, in order to find the simplest representation
11024 of the arguments so that their form can be studied. In any
11025 cases, the appropriate type conversions should be put back in
11026 the tree that will get out of the constant folder. */
11029 {
11032 }
11033 else
11034 {
11037 }
11039 /* Note that TREE_CONSTANT isn't enough: static var addresses are
11040 constant but we can't do arithmetic on them. */
11042 {
11045 {
11049 }
11050 }
11052 /* If this is a commutative operation, and ARG0 is a constant, move it
11053 to ARG1 to reduce the number of tests below. */
11058 /* Likewise if this is a comparison, and ARG0 is a constant, move it
11059 to ARG1 to reduce the number of tests below. */
11068 /* ARG0 is the first operand of EXPR, and ARG1 is the second operand.
11070 First check for cases where an arithmetic operation is applied to a
11071 compound, conditional, or comparison operation. Push the arithmetic
11072 operation inside the compound or conditional to see if any folding
11073 can then be done. Convert comparison to conditional for this purpose.
11074 The also optimizes non-constant cases that used to be done in
11075 expand_expr.
11077 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
11078 one of the operands is a comparison and the other is a comparison, a
11079 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
11080 code below would make the expression more complex. Change it to a
11081 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
11082 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
11095 {
11099 boolean_type_node,
11107 }
11111 {
11113 {
11118 tem);
11119 }
11121 {
11126 tem);
11127 }
11132 {
11138 }
11143 {
11149 }
11150 }
11153 {
11155 /* MEM[&MEM[p, CST1], CST2] -> MEM[p, CST1 + CST2]. */
11158 {
11164 }
11166 /* MEM[&a.b, CST2] -> MEM[&a, offsetof (a, b) + CST2]. */
11169 {
11170 tree base;
11171 poly_int64 coffset;
11180 }
11185 /* INT +p INT -> (PTR)(INT + INT). Stripping types allows for this. */
11191 arg1),
11193 arg0)));
11199 {
11200 /* X + (X / CST) * -CST is X % CST. */
11205 {
11214 cst0));
11215 }
11216 }
11218 /* Handle (A1 * C1) + (A2 * C2) with A1, A2 or C1, C2 being the same or
11219 one. Make sure the type is not saturating and has the signedness of
11220 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11221 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11228 {
11232 }
11235 {
11236 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
11237 (plus (plus (mult) (mult)) (foo)) so that we can
11238 take advantage of the factoring cases below. */
11247 {
11253 else
11266 parg0),
11268 marg)),
11272 return
11278 parg1)));
11279 }
11280 }
11281 else
11282 {
11283 /* Fold __complex__ ( x, 0 ) + __complex__ ( 0, y )
11284 to __complex__ ( x, y ). This is not the same for SNaNs or
11285 if signed zeros are involved. */
11289 {
11296 {
11300 {
11306 }
11308 {
11314 }
11315 }
11316 }
11318 /* Convert a + (b*c + d*e) into (a + b*c) + d*e.
11319 We associate floats only if the user has specified
11320 -fassociative-math. */
11324 {
11329 {
11330 tree tree0;
11333 }
11334 }
11335 /* Convert (b*c + d*e) + a into b*c + (d*e +a).
11336 We associate floats only if the user has specified
11337 -fassociative-math. */
11341 {
11346 {
11347 tree tree0;
11350 }
11351 }
11352 }
11354 bit_rotate:
11355 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
11356 is a rotate of A by C1 bits. */
11357 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
11358 is a rotate of A by B bits.
11359 Similarly for (A << B) | (A >> (-B & C3)) where C3 is Z-1,
11360 though in this case CODE must be | and not + or ^, otherwise
11361 it doesn't return A when B is 0. */
11362 {
11364 tree rtype;
11373 /* Only create rotates in complete modes. Other cases are not
11374 expanded properly. */
11377 {
11390 {
11395 }
11400 {
11403 code0 == LSHIFT_EXPR
11406 }
11408 {
11418 {
11422 orig_tree01);
11424 }
11425 }
11429 {
11440 {
11444 orig_tree01);
11446 }
11447 }
11448 }
11449 }
11451 associate:
11452 /* In most languages, can't associate operations on floats through
11453 parentheses. Rather than remember where the parentheses were, we
11454 don't associate floats at all, unless the user has specified
11455 -fassociative-math.
11456 And, we need to make sure type is not saturating. */
11461 {
11467 /* Split both trees into variables, constants, and literals. Then
11468 associate each group together, the constants with literals,
11469 then the result with variables. This increases the chances of
11470 literals being recombined later and of generating relocatable
11471 expressions for the sum of a constant and literal. */
11479 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
11483 /* With undefined overflow prefer doing association in a type
11484 which wraps on overflow, if that is one of the operand types. */
11487 {
11495 }
11497 /* With undefined overflow we can only associate constants with one
11498 variable, and constants whose association doesn't overflow. */
11501 {
11503 {
11504 /* ??? If split_tree would handle NEGATE_EXPR we could
11505 simply reject these cases and the allowed cases would
11506 be the var0/minus_var1 ones. */
11512 {
11515 }
11522 {
11525 }
11531 /* The only case we can still associate with two variables
11532 is if they cancel out. */
11536 }
11542 }
11544 /* Only do something if we found more than two objects. Otherwise,
11545 nothing has changed and we risk infinite recursion. */
11553 {
11565 {
11569 }
11571 {
11575 }
11577 /* Preserve the MINUS_EXPR if the negative part of the literal is
11578 greater than the positive part. Otherwise, the multiplicative
11579 folding code (i.e extract_muldiv) may be fooled in case
11580 unsigned constants are subtracted, like in the following
11581 example: ((X*2 + 4) - 8U)/2. */
11583 {
11587 /* But avoid ending up with only negated parts. */
11589 {
11593 }
11594 else
11595 {
11599 }
11600 }
11602 /* Don't introduce overflows through reassociation. */
11607 /* Eliminate lit0 and minus_lit0 to con0 and minus_con0. */
11614 /* Eliminate minus_con0. */
11616 {
11623 else
11626 }
11628 /* Eliminate minus_var0. */
11630 {
11634 else
11637 }
11639 return
11642 }
11643 }
11649 /* Fold &a[i] - &a[j] to i-j. */
11654 {
11658 code
11662 }
11664 /* Further transformations are not for pointers. */
11668 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
11671 /* If arg0 is e.g. unsigned int and type is int, then this could
11672 introduce UB, because if A is INT_MIN at runtime, the original
11673 expression can be well defined while the latter is not.
11674 See PR83269. */
11683 /* Fold __complex__ ( x, 0 ) - __complex__ ( 0, y ) to
11684 __complex__ ( x, -y ). This is not the same for SNaNs or if
11685 signed zeros are involved. */
11689 {
11696 {
11700 {
11702 arg1r ? arg1r
11707 }
11709 {
11713 arg1i ? arg1i
11716 }
11717 }
11718 }
11720 /* A - B -> A + (-B) if B is easily negatable. */
11724 /* Avoid this transformation if B is a positive REAL_CST. */
11732 /* Handle (A1 * C1) - (A2 * C2) with A1, A2 or C1, C2 being the same or
11733 one. Make sure the type is not saturating and has the signedness of
11734 the stripped operands, as fold_plusminus_mult_expr will re-associate.
11735 ??? The latter condition should use TYPE_OVERFLOW_* flags instead. */
11742 {
11746 }
11752 {
11753 /* Transform x * -C into -x * C if x is easily negatable. */
11768 {
11771 "occur when simplifying "
11773 WARN_STRICT_OVERFLOW_MISC);
11775 }
11777 /* Optimize z * conj(z) for integer complex numbers. */
11784 }
11785 else
11786 {
11787 /* Fold z * +-I to __complex__ (-+__imag z, +-__real z).
11788 This is not the same for NaNs or if signed zeros are
11789 involved. */
11795 {
11798 return
11804 return
11809 }
11811 /* Optimize z * conj(z) for floating point complex numbers.
11812 Guarded by flag_unsafe_math_optimizations as non-finite
11813 imaginary components don't produce scalar results. */
11822 }
11826 /* Canonicalize (X & C1) | C2. */
11830 {
11835 /* If (C1&C2) == C1, then (X&C1)|C2 becomes (X,C2). */
11843 /* If (C1|C2) == ~0 then (X&C1)|C2 becomes X|C2. */
11845 {
11848 }
11850 /* Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2,
11851 unless (C1 & ~C2) | (C2 & C3) for some C3 is a mask of some
11852 mode which allows further optimizations. */
11857 {
11862 {
11865 }
11866 }
11869 {
11874 }
11875 }
11877 /* See if this can be simplified into a rotate first. If that
11878 is unsuccessful continue in the association code. */
11882 /* Fold (X & 1) ^ 1 as (X & 1) == 0. */
11890 /* See if this can be simplified into a rotate first. If that
11891 is unsuccessful continue in the association code. */
11895 /* Fold (X ^ 1) & 1 as (X & 1) == 0. */
11900 {
11901 tree tem2;
11908 }
11909 /* Fold ~X & 1 as (X & 1) == 0. */
11913 {
11914 tree tem2;
11921 }
11922 /* Fold !X & 1 as X == 0. */
11925 {
11929 }
11931 /* Fold (X * Y) & -(1 << CST) to X * Y if Y is a constant
11932 multiple of 1 << CST. */
11934 {
11941 }
11943 /* Fold (X * CST1) & CST2 to zero if we can, or drop known zero
11944 bits from CST2. */
11948 {
11950 wide_int masked
11957 {
11958 /* Avoid the transform if arg1 is a mask of some
11959 mode which allows further optimizations. */
11966 }
11967 }
11969 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
11972 {
11977 return
11979 }
11984 /* Don't touch a floating-point divide by zero unless the mode
11985 of the constant can represent infinity. */
11991 /* (-A) / (-B) -> A / B */
12003 /* Fall through */
12006 /* Simplify A / (B << N) where A and B are positive and B is
12007 a power of 2, to A >> (N + log2(B)). */
12012 {
12015 {
12023 WARN_STRICT_OVERFLOW_MISC);
12029 }
12030 }
12032 /* Fall through */
12040 /* Convert -A / -B to A / B when the type is signed and overflow is
12041 undefined. */
12045 {
12048 "when distributing negation across "
12050 WARN_STRICT_OVERFLOW_MISC);
12055 }
12059 {
12062 "when distributing negation across "
12064 WARN_STRICT_OVERFLOW_MISC);
12069 }
12071 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
12072 operation, EXACT_DIV_EXPR.
12074 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
12075 At one time others generated faster code, it's not clear if they do
12076 after the last round to changes to the DIV code in expmed.cc. */
12087 {
12091 WARN_STRICT_OVERFLOW_MISC);
12093 }
12105 {
12109 WARN_STRICT_OVERFLOW_MISC);
12111 }
12119 /* Since negative shift count is not well-defined,
12120 don't try to compute it in the compiler. */
12126 /* If we have a rotate of a bit operation with the rotate count and
12127 the second operand of the bit operation both constant,
12128 permute the two operations. */
12134 {
12142 }
12144 /* Two consecutive rotates adding up to the some integer
12145 multiple of the precision of the type can be ignored. */
12161 /* Note that the operands of this must be ints
12162 and their values must be 0 or 1.
12163 ("true" is a fixed value perhaps depending on the language.) */
12164 /* If first arg is constant zero, return it. */
12167 /* FALLTHRU */
12169 /* If either arg is constant true, drop it. */
12173 /* Preserve sequence points. */
12176 /* If second arg is constant zero, result is zero, but first arg
12177 must be evaluated. */
12180 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
12181 case will be handled here. */
12185 /* !X && X is always false. */
12189 /* X && !X is always false. */
12194 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
12195 means A >= Y && A != MAX, but in this case we know that
12196 A < X <= MAX. */
12200 {
12212 }
12221 /* Note that the operands of this must be ints
12222 and their values must be 0 or true.
12223 ("true" is a fixed value perhaps depending on the language.) */
12224 /* If first arg is constant true, return it. */
12227 /* FALLTHRU */
12229 /* If either arg is constant zero, drop it. */
12233 /* Preserve sequence points. */
12236 /* If second arg is constant true, result is true, but we must
12237 evaluate first arg. */
12240 /* Likewise for first arg, but note this only occurs here for
12241 TRUTH_OR_EXPR. */
12245 /* !X || X is always true. */
12249 /* X || !X is always true. */
12254 /* (X && !Y) || (!X && Y) is X ^ Y */
12257 {
12274 }
12283 /* If the second arg is constant zero, drop it. */
12286 /* If the second arg is constant true, this is a logical inversion. */
12288 {
12291 }
12292 /* Identical arguments cancel to zero. */
12296 /* !X ^ X is always true. */
12301 /* X ^ !X is always true. */
12317 /* bool_var != 1 becomes !bool_var. */
12324 /* bool_var == 0 becomes !bool_var. */
12331 /* !exp != 0 becomes !exp */
12336 /* If this is an EQ or NE comparison with zero and ARG0 is
12337 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
12338 two operations, but the latter can be done in one less insn
12339 on machines that have only two-operand insns or on which a
12340 constant cannot be the first operand. */
12343 {
12348 {
12356 }
12359 {
12367 }
12368 }
12370 /* If this is a comparison of a field, we may be able to simplify it. */
12373 /* Handle the constant case even without -O
12374 to make sure the warnings are given. */
12376 {
12380 }
12382 /* Optimize comparisons of strlen vs zero to a compare of the
12383 first character of the string vs zero. To wit,
12384 strlen(ptr) == 0 => *ptr == 0
12385 strlen(ptr) != 0 => *ptr != 0
12386 Other cases should reduce to one of these two (or a constant)
12387 due to the return value of strlen being unsigned. */
12389 {
12397 {
12398 tree ptrtype
12400 TYPE_QUAL_CONST));
12406 }
12407 }
12409 /* Fold (X >> C) != 0 into X < 0 if C is one less than the width
12410 of X. Similarly fold (X >> C) == 0 into X >= 0. */
12414 {
12419 {
12421 {
12424 }
12427 }
12428 }
12430 /* Fold (~X & C) == 0 into (X & C) != 0 and (~X & C) != 0 into
12431 (X & C) == 0 when C is a single bit. */
12436 {
12443 arg1));
12444 }
12446 /* Fold ((X & C) ^ C) eq/ne 0 into (X & C) ne/eq 0, when the
12447 constant C is a power of two, i.e. a single bit. */
12454 {
12458 }
12460 /* Likewise, fold ((X ^ C) & C) eq/ne 0 into (X & C) ne/eq 0,
12461 when is C is a power of two, i.e. a single bit. */
12468 {
12474 }
12478 {
12485 /* Optimize (X ^ Z) op (Y ^ Z) as X op Y, and symmetries.
12486 operand_equal_p guarantees no side-effects so we don't need
12487 to use omit_one_operand on Z. */
12491 arg10));
12495 arg11));
12499 arg10));
12503 arg11));
12505 /* Optimize (X ^ C1) op (Y ^ C2) as (X ^ (C1 ^ C2)) op Y. */
12508 {
12514 }
12515 }
12517 /* Attempt to simplify equality/inequality comparisons of complex
12518 values. Only lower the comparison if the result is known or
12519 can be simplified to a single scalar comparison. */
12524 {
12529 {
12532 }
12533 else
12534 {
12537 }
12540 {
12543 }
12544 else
12545 {
12548 }
12552 {
12554 {
12559 }
12560 else
12561 {
12566 }
12567 }
12571 {
12573 {
12578 }
12579 else
12580 {
12585 }
12586 }
12587 }
12599 /* Transform comparisons of the form X +- C CMP X. */
12604 {
12609 /* (X - c) > X becomes false. */
12615 /* Likewise (X + c) < X becomes false. */
12621 /* Convert (X - c) <= X to true. */
12628 /* Convert (X + c) >= X to true. */
12634 }
12636 /* If we are comparing an ABS_EXPR with a constant, we can
12637 convert all the cases into explicit comparisons, but they may
12638 well not be faster than doing the ABS and one comparison.
12639 But ABS (X) <= C is a range comparison, which becomes a subtraction
12640 and a comparison, and is probably faster. */
12654 /* Convert ABS_EXPR<x> >= 0 to true. */
12661 {
12664 "when simplifying comparison of "
12666 WARN_STRICT_OVERFLOW_CONDITIONAL);
12669 arg0);
12670 }
12672 /* Convert ABS_EXPR<x> < 0 to false. */
12677 {
12680 "when simplifying comparison of "
12682 WARN_STRICT_OVERFLOW_CONDITIONAL);
12685 arg0);
12686 }
12688 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
12689 and similarly for >= into !=. */
12699 /* Similarly for X < (cast) (1 << Y). But cast can't be narrowing,
12700 otherwise Y might be >= # of bits in X's type and thus e.g.
12701 (unsigned char) (1 << Y) for Y 15 might be 0.
12702 If the cast is widening, then 1 << Y should have unsigned type,
12703 otherwise if Y is number of bits in the signed shift type minus 1,
12704 we can't optimize this. E.g. (unsigned long long) (1 << Y) for Y
12705 31 might be 0xffffffff80000000. */
12718 {
12724 }
12736 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
12737 {
12749 }
12754 /* When pedantic, a compound expression can be neither an lvalue
12755 nor an integer constant expression. */
12758 /* Don't let (0, 0) be null pointer constant. */
12766 }
12768 /* For constants M and N, if M == (1LL << cst) - 1 && (N & M) == M,
12769 ((A & N) + B) & M -> (A + B) & M
12770 Similarly if (N & M) == 0,
12771 ((A | N) + B) & M -> (A + B) & M
12772 and for - instead of + (or unary - instead of +)
12773 and/or ^ instead of |.
12774 If B is constant and (B & M) == 0, fold into A & M.
12776 This function is a helper for match.pd patterns. Return non-NULL
12777 type in which the simplified operation should be performed only
12778 if any optimization is possible.
12780 ARG1 is M above, ARG00 is left operand of +/-, if CODE00 is BIT_*_EXPR,
12781 then ARG00{0,1} are operands of that bitop, otherwise CODE00 is ERROR_MARK.
12782 Similarly for ARG01, CODE01 and ARG01{0,1}, just for the right operand of
12783 +/-. */
12784 tree
12789 {
12804 wide_int cst0;
12806 /* Now we know that arg0 is (C + D) or (C - D) or -C and
12807 arg1 (M) is == (1LL << cst) - 1.
12808 Store C into PMOP[0] and D into PMOP[1]. */
12815 {
12822 {
12825 }
12828 /* If C or D is of the form (A & N) where
12829 (N & M) == M, or of the form (A | N) or
12830 (A ^ N) where (N & M) == 0, replace it with A. */
12836 /* If C or D is a N where (N & M) == 0, it can be
12837 omitted (replaced with 0). */
12842 /* Similarly, with C - N where (-N & M) == 0. */
12850 }
12852 /* Only build anything new if we optimized one or both arguments above. */
12858 else
12860 }
12862 /* Used by contains_label_[p1]. */
12864 struct contains_label_data
12865 {
12868 };
12870 /* Callback for walk_tree, looking for LABEL_EXPR. Return *TP if it is
12871 a LABEL_EXPR or CASE_LABEL_EXPR not inside of another SWITCH_EXPR; otherwise
12872 return NULL_TREE. Do not check the subtrees of GOTO_EXPR. */
12874 static tree
12876 {
12879 {
12890 {
12898 }
12907 }
12908 }
12910 /* Return whether the sub-tree ST contains a label which is accessible from
12911 outside the sub-tree. */
12913 static bool
12915 {
12919 }
12921 /* Fold a ternary expression of code CODE and type TYPE with operands
12922 OP0, OP1, and OP2. Return the folded expression if folding is
12923 successful. Otherwise, return NULL_TREE. */
12925 tree
12928 {
12929 tree tem;
12936 /* If this is a commutative operation, and OP0 is a constant, move it
12937 to OP1 to reduce the number of tests below. */
12946 /* Strip any conversions that don't change the mode. This is safe
12947 for every expression, except for a comparison expression because
12948 its signedness is derived from its operands. So, in the latter
12949 case, only strip conversions that don't change the signedness.
12951 Note that this is done as an internal manipulation within the
12952 constant folder, in order to find the simplest representation of
12953 the arguments so that their form can be studied. In any cases,
12954 the appropriate type conversions should be put back in the tree
12955 that will get out of the constant folder. */
12957 {
12960 }
12963 {
12966 }
12969 {
12972 }
12975 {
12979 {
12985 }
12990 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
12991 so all simple results must be passed through pedantic_non_lvalue. */
12993 {
12996 /* Only optimize constant conditions when the selected branch
12997 has the same type as the COND_EXPR. This avoids optimizing
12998 away "c ? x : throw", where the throw has a void type.
12999 Avoid throwing away that operand which contains label. */
13006 }
13008 {
13015 {
13018 {
13026 }
13031 }
13032 }
13034 /* If we have A op B ? A : C, we may be able to convert this to a
13035 simpler expression, depending on the operation and the values
13036 of B and C. Signed zeros prevent all of these transformations,
13037 for reasons given above each one.
13039 Also try swapping the arguments and inverting the conditional. */
13043 {
13050 }
13055 {
13062 arg00,
13063 arg01,
13067 }
13069 /* If the second operand is simpler than the third, swap them
13070 since that produces better jump optimization results. */
13073 {
13075 /* See if this can be inverted. If it can't, possibly because
13076 it was a floating-point inequality comparison, don't do
13077 anything. */
13081 }
13083 /* Convert A ? 1 : 0 to simply A. */
13088 /* If we try to convert OP0 to our type, the
13089 call to fold will try to move the conversion inside
13090 a COND, which will recurse. In that case, the COND_EXPR
13091 is probably the best choice, so leave it alone. */
13095 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
13096 over COND_EXPR in cases such as floating point comparisons. */
13105 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
13110 {
13111 /* sign_bit_p looks through both zero and sign extensions,
13112 but for this optimization only sign extensions are
13113 usable. */
13116 {
13119 {
13122 }
13124 }
13125 /* sign_bit_p only checks ARG1 bits within A's precision.
13126 If <sign bit of A> has wider type than A, bits outside
13127 of A's precision in <sign bit of A> need to be checked.
13128 If they are all 0, this optimization needs to be done
13129 in unsigned A's type, if they are all 1 in signed A's type,
13130 otherwise this can't be done. */
13136 {
13138 tree tem_type;
13151 {
13154 }
13156 {
13159 }
13160 else
13162 }
13165 return
13171 arg1)));
13172 }
13174 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
13175 already handled above. */
13180 {
13190 op1);
13191 }
13193 /* A & N ? N : 0 is simply A & N if N is a power of two. This
13194 is probably obsolete because the first operand should be a
13195 truth value (that's why we have the two cases above), but let's
13196 leave it in until we can confirm this for all front-ends. */
13204 /* operand_equal_p compares just value, not precision, so e.g.
13205 arg1 could be 8-bit -128 and be power of two, but BIT_AND_EXPR
13206 second operand 32-bit -128, which is not a power of two (or vice
13207 versa. */
13211 /* Disable the transformations below for vectors, since
13212 fold_binary_op_with_conditional_arg may undo them immediately,
13213 yielding an infinite loop. */
13217 /* Convert A ? B : 0 into A && B if A and B are truth values. */
13226 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
13231 {
13233 /* Only perform transformation if ARG0 is easily inverted. */
13237 ? BIT_IOR_EXPR
13240 op1);
13241 }
13243 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
13248 {
13250 /* Only perform transformation if ARG0 is easily inverted. */
13256 op2);
13257 }
13259 /* Convert A ? 1 : B into A || B if A and B are truth values. */
13271 /* CALL_EXPRs used to be ternary exprs. Catch any mistaken uses
13272 of fold_ternary on them. */
13282 {
13284 unsigned HOST_WIDE_INT width
13295 {
13300 {
13302 {
13307 }
13313 }
13314 }
13315 }
13317 /* On constants we can use native encode/interpret to constant
13318 fold (nearly) all BIT_FIELD_REFs. */
13324 {
13327 /* Limit us to a reasonable amount of work. To relax the
13328 other limitations we need bit-shifting of the buffer
13329 and rounding up the size. */
13333 {
13335 unsigned HOST_WIDE_INT len
13340 {
13345 }
13346 }
13347 }
13352 /* Perform constant folding of BIT_INSERT_EXPR. */
13356 {
13357 /* Build a vector of integers from the tree mask. */
13358 vec_perm_builder builder;
13362 /* Create a vec_perm_indices for the integer vector. */
13367 }
13371 /* Perform (partial) constant folding of BIT_INSERT_EXPR. */
13374 {
13380 wide_int tem2
13384 }
13389 {
13391 unsigned HOST_WIDE_INT elsize
13394 {
13400 {
13406 }
13407 }
13408 }
13414 }
13416 /* Gets the element ACCESS_INDEX from CTOR, which must be a CONSTRUCTOR
13417 of an array (or vector). *CTOR_IDX if non-NULL is updated with the
13418 constructor element index of the value returned. If the element is
13419 not found NULL_TREE is returned and *CTOR_IDX is updated to
13420 the index of the element after the ACCESS_INDEX position (which
13421 may be outside of the CTOR array). */
13423 tree
13426 {
13432 {
13435 {
13436 /* Static constructors for variably sized objects makes no sense. */
13439 /* ??? When it is obvious that the range is signed, treat it so. */
13444 {
13446 low_bound
13448 SIGNED);
13449 }
13450 else
13451 {
13454 }
13455 }
13456 }
13460 index_sgn);
13472 {
13473 /* Array constructor might explicitly set index, or specify a range,
13474 or leave index NULL meaning that it is next index after previous
13475 one. */
13477 {
13479 max_index = index
13481 else
13482 {
13485 index_sgn);
13486 max_index
13488 index_sgn);
13490 }
13491 }
13493 {
13499 }
13500 else
13503 /* Do we have match? */
13505 {
13507 {
13511 }
13512 }
13514 /* We're past the element we search for. Note during parsing
13515 the elements might not be sorted.
13516 ??? We should use a binary search and a flag on the
13517 CONSTRUCTOR as to whether elements are sorted in declaration
13518 order. */
13520 }
13524 }
13526 /* Perform constant folding and related simplification of EXPR.
13527 The related simplifications include x*1 => x, x*0 => 0, etc.,
13528 and application of the associative law.
13529 NOP_EXPR conversions may be removed freely (as long as we
13530 are careful not to change the type of the overall expression).
13531 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
13532 but we can constant-fold them if they have constant operands. */
13534 #ifdef ENABLE_FOLD_CHECKING
13535 # define fold(x) fold_1 (x)
13537 static
13538 #endif
13539 tree
13541 {
13545 tree tem;
13548 /* Return right away if a constant. */
13552 /* CALL_EXPR-like objects with variable numbers of operands are
13553 treated specially. */
13555 {
13557 {
13560 }
13562 }
13565 {
13570 {
13588 }
13589 }
13592 {
13594 {
13601 {
13606 }
13609 }
13611 /* Return a VECTOR_CST if possible. */
13613 {
13619 tree val;
13625 }
13633 }
13635 #ifdef ENABLE_FOLD_CHECKING
13636 #undef fold
13643 /* When --enable-checking=fold, compute a digest of expr before
13644 and after actual fold call to see if fold did not accidentally
13645 change original expr. */
13647 tree
13649 {
13650 tree ret;
13670 }
13672 void
13674 {
13685 }
13687 static void
13689 {
13691 }
13693 static void
13696 {
13702 recursive_label:
13712 {
13713 /* Allow DECL_ASSEMBLER_NAME and symtab_node to be modified. */
13721 }
13729 {
13730 /* Allow these fields to be modified. */
13731 tree tmp;
13742 {
13745 }
13746 }
13748 {
13749 /* Allow the no-warning bit to be set. Perhaps we shouldn't allow
13750 that and change builtins.cc etc. instead - see PR89543. */
13756 }
13767 {
13770 {
13786 }
13790 {
13803 }
13820 {
13826 }
13829 {
13831 {
13834 }
13836 }
13847 {
13850 }
13858 }
13859 }
13861 /* Helper function for outputting the checksum of a tree T. When
13862 debugging with gdb, you can "define mynext" to be "next" followed
13863 by "call debug_fold_checksum (op0)", then just trace down till the
13864 outputs differ. */
13866 DEBUG_FUNCTION void
13868 {
13883 }
13885 #endif
13887 /* Fold a unary tree expression with code CODE of type TYPE with an
13888 operand OP0. LOC is the location of the resulting expression.
13889 Return a folded expression if successful. Otherwise, return a tree
13890 expression with code CODE of type TYPE with an operand OP0. */
13892 tree
13895 {
13896 tree tem;
13897 #ifdef ENABLE_FOLD_CHECKING
13906 #endif
13912 #ifdef ENABLE_FOLD_CHECKING
13919 #endif
13921 }
13923 /* Fold a binary tree expression with code CODE of type TYPE with
13924 operands OP0 and OP1. LOC is the location of the resulting
13925 expression. Return a folded expression if successful. Otherwise,
13926 return a tree expression with code CODE of type TYPE with operands
13927 OP0 and OP1. */
13929 tree
13932 MEM_STAT_DECL)
13933 {
13934 tree tem;
13935 #ifdef ENABLE_FOLD_CHECKING
13952 #endif
13958 #ifdef ENABLE_FOLD_CHECKING
13973 #endif
13975 }
13977 /* Fold a ternary tree expression with code CODE of type TYPE with
13978 operands OP0, OP1, and OP2. Return a folded expression if
13979 successful. Otherwise, return a tree expression with code CODE of
13980 type TYPE with operands OP0, OP1, and OP2. */
13982 tree
13985 {
13986 tree tem;
13987 #ifdef ENABLE_FOLD_CHECKING
14011 #endif
14018 #ifdef ENABLE_FOLD_CHECKING
14041 #endif
14043 }
14045 /* Fold a CALL_EXPR expression of type TYPE with operands FN and NARGS
14046 arguments in ARGARRAY, and a null static chain.
14047 Return a folded expression if successful. Otherwise, return a CALL_EXPR
14048 of type TYPE from the given operands as constructed by build_call_array. */
14050 tree
14053 {
14054 tree tem;
14055 #ifdef ENABLE_FOLD_CHECKING
14074 #endif
14080 #ifdef ENABLE_FOLD_CHECKING
14096 #endif
14098 }
14100 /* Perform constant folding and related simplification of initializer
14101 expression EXPR. These behave identically to "fold_buildN" but ignore
14102 potential run-time traps and exceptions that fold must preserve. */
14104 #define START_FOLD_INIT \
14105 int saved_signaling_nans = flag_signaling_nans;\
14106 int saved_trapping_math = flag_trapping_math;\
14107 int saved_rounding_math = flag_rounding_math;\
14108 int saved_trapv = flag_trapv;\
14109 int saved_folding_initializer = folding_initializer;\
14110 flag_signaling_nans = 0;\
14111 flag_trapping_math = 0;\
14112 flag_rounding_math = 0;\
14113 flag_trapv = 0;\
14114 folding_initializer = 1;
14116 #define END_FOLD_INIT \
14117 flag_signaling_nans = saved_signaling_nans;\
14118 flag_trapping_math = saved_trapping_math;\
14119 flag_rounding_math = saved_rounding_math;\
14120 flag_trapv = saved_trapv;\
14121 folding_initializer = saved_folding_initializer;
14123 tree
14125 {
14126 tree result;
14127 START_FOLD_INIT;
14131 END_FOLD_INIT;
14133 }
14135 tree
14138 {
14139 tree result;
14140 START_FOLD_INIT;
14144 END_FOLD_INIT;
14146 }
14148 tree
14151 {
14152 tree result;
14153 START_FOLD_INIT;
14157 END_FOLD_INIT;
14159 }
14161 tree
14164 {
14165 tree result;
14166 START_FOLD_INIT;
14170 END_FOLD_INIT;
14172 }
14174 tree
14177 {
14178 tree result;
14179 START_FOLD_INIT;
14183 END_FOLD_INIT;
14185 }
14187 #undef START_FOLD_INIT
14188 #undef END_FOLD_INIT
14190 /* Determine if first argument is a multiple of second argument. Return
14191 false if it is not, or we cannot easily determined it to be.
14193 An example of the sort of thing we care about (at this point; this routine
14194 could surely be made more general, and expanded to do what the *_DIV_EXPR's
14195 fold cases do now) is discovering that
14197 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14199 is a multiple of
14201 SAVE_EXPR (J * 8)
14203 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
14205 This code also handles discovering that
14207 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
14209 is a multiple of 8 so we don't have to worry about dealing with a
14210 possible remainder.
14212 Note that we *look* inside a SAVE_EXPR only to determine how it was
14213 calculated; it is not safe for fold to do much of anything else with the
14214 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
14215 at run time. For example, the latter example above *cannot* be implemented
14216 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
14217 evaluation time of the original SAVE_EXPR is not necessarily the same at
14218 the time the new expression is evaluated. The only optimization of this
14219 sort that would be valid is changing
14221 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
14223 divided by 8 to
14225 SAVE_EXPR (I) * SAVE_EXPR (J)
14227 (where the same SAVE_EXPR (J) is used in the original and the
14228 transformed version).
14230 NOWRAP specifies whether all outer operations in TYPE should
14231 be considered not wrapping. Any type conversion within TOP acts
14232 as a barrier and we will fall back to NOWRAP being false.
14233 NOWRAP is mostly used to treat expressions in TYPE_SIZE and friends
14234 as not wrapping even though they are generally using unsigned arithmetic. */
14236 bool
14238 {
14249 {
14251 /* Bitwise and provides a power of two multiple. If the mask is
14252 a multiple of BOTTOM then TOP is a multiple of BOTTOM. */
14259 /* If the multiplication can wrap we cannot recurse further unless
14260 the bottom is a power of two which is where wrapping does not
14261 matter. */
14267 {
14273 {
14276 /* Handle multiple_of_p ((x * 2 + 2) * 4, 8). */
14278 {
14282 {
14285 }
14286 }
14288 }
14289 }
14294 /* Handle X << CST as X * (1 << CST) and only process the constant. */
14296 {
14299 {
14300 wide_int mul_op
14304 nowrap);
14305 }
14306 }
14311 /* If the addition or subtraction can wrap we cannot recurse further
14312 unless bottom is a power of two which is where wrapping does not
14313 matter. */
14319 /* Handle cases like op0 + 0xfffffffd as op0 - 3 if the expression has
14320 unsigned type. For example, (X / 3) + 0xfffffffd is multiple of 3,
14321 but 0xfffffffd is not. */
14324 && nowrap
14329 /* It is impossible to prove if op0 +- op1 is multiple of bottom
14330 precisely, so be conservative here checking if both op0 and op1
14331 are multiple of bottom. Note we check the second operand first
14332 since it's usually simpler. */
14336 CASE_CONVERT:
14337 /* Can't handle conversions from non-integral or wider integral type. */
14342 /* NOWRAP only extends to operations in the outermost type so
14343 make sure to strip it off here. */
14358 SIGNED);
14364 {
14367 /* Check for special cases to see if top is defined as multiple
14368 of bottom:
14370 top = (X & ~(bottom - 1) ; bottom is power of 2
14372 or
14374 Y = X % bottom
14375 top = X - Y. */
14394 }
14396 /* fall through */
14404 }
14405 }
14407 /* Return true if expression X cannot be (or contain) a NaN or infinity.
14408 This function returns true for integer expressions, and returns
14409 false if uncertain. */
14411 bool
14413 {
14418 {
14441 {
14442 CASE_CFN_FABS:
14443 CASE_CFN_FABS_FN:
14445 CASE_CFN_FMAX:
14446 CASE_CFN_FMAX_FN:
14447 CASE_CFN_FMIN:
14448 CASE_CFN_FMIN_FN:
14453 }
14457 }
14458 }
14460 /* Return true if expression X evaluates to an infinity.
14461 This function returns false for integer expressions. */
14463 bool
14465 {
14469 {
14482 }
14483 }
14485 /* Return true if expression X could evaluate to an infinity.
14486 This function returns false for integer expressions, and returns
14487 true if uncertain. */
14489 bool
14491 {
14495 {
14508 }
14509 }
14511 /* Return true if expression X evaluates to a signaling NaN.
14512 This function returns false for integer expressions. */
14514 bool
14516 {
14520 {
14531 }
14532 }
14534 /* Return true if expression X could evaluate to a signaling NaN.
14535 This function returns false for integer expressions, and returns
14536 true if uncertain. */
14538 bool
14540 {
14544 {
14564 {
14565 CASE_CFN_FABS:
14566 CASE_CFN_FABS_FN:
14568 CASE_CFN_FMAX:
14569 CASE_CFN_FMAX_FN:
14570 CASE_CFN_FMIN:
14571 CASE_CFN_FMIN_FN:
14576 }
14579 }
14580 }
14582 /* Return true if expression X evaluates to a NaN.
14583 This function returns false for integer expressions. */
14585 bool
14587 {
14591 {
14602 }
14603 }
14605 /* Return true if expression X could evaluate to a NaN.
14606 This function returns false for integer expressions, and returns
14607 true if uncertain. */
14609 bool
14611 {
14615 {
14640 {
14641 CASE_CFN_FABS:
14642 CASE_CFN_FABS_FN:
14644 CASE_CFN_FMAX:
14645 CASE_CFN_FMAX_FN:
14646 CASE_CFN_FMIN:
14647 CASE_CFN_FMIN_FN:
14652 }
14655 }
14656 }
14658 /* Return true if expression X could evaluate to -0.0.
14659 This function returns true if uncertain. */
14661 bool
14663 {
14667 {
14682 {
14683 CASE_CFN_FABS:
14684 CASE_CFN_FABS_FN:
14688 }
14691 }
14692 /* Ideally !(tree_expr_nonzero_p (X) || tree_expr_nonnegative_p (X))
14693 * but currently those predicates require tree and not const_tree. */
14695 }
14697 #define tree_expr_nonnegative_warnv_p(X, Y) \
14700 #define RECURSE(X) \
14701 ((tree_expr_nonnegative_warnv_p) (X, strict_overflow_p, depth + 1))
14703 /* Return true if CODE or TYPE is known to be non-negative. */
14705 static bool
14707 {
14711 /* Truth values evaluate to 0 or 1, which is nonnegative unless we
14712 have a signed:1 type (where the value is -1 and 0). */
14715 }
14717 /* Return true if (CODE OP0) is known to be non-negative. If the return
14718 value is based on the assumption that signed overflow is undefined,
14719 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14720 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14722 bool
14725 {
14730 {
14732 /* We can't return 1 if flag_wrapv is set because
14733 ABS_EXPR<INT_MIN> = INT_MIN. */
14737 {
14740 }
14748 CASE_CONVERT:
14749 {
14754 {
14758 {
14762 }
14763 }
14765 {
14771 }
14772 }
14777 }
14779 /* We don't know sign of `t', so be conservative and return false. */
14781 }
14783 /* Return true if (CODE OP0 OP1) is known to be non-negative. If the return
14784 value is based on the assumption that signed overflow is undefined,
14785 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14786 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14788 bool
14792 {
14797 {
14803 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
14804 both unsigned and at least 2 bits shorter than the result. */
14808 {
14813 {
14817 }
14818 }
14823 {
14824 /* x * x is always non-negative for floating point x
14825 or without overflow. */
14828 {
14833 }
14834 }
14836 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
14837 both unsigned and their total bits is shorter than the result. */
14841 {
14860 {
14870 }
14871 }
14878 /* Usually RECURSE (op0) || RECURSE (op1) but NaNs complicate
14879 things. */
14904 }
14906 /* We don't know sign of `t', so be conservative and return false. */
14908 }
14910 /* Return true if T is known to be non-negative. If the return
14911 value is based on the assumption that signed overflow is undefined,
14912 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14913 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14915 bool
14917 {
14922 {
14936 /* Limit the depth of recursion to avoid quadratic behavior.
14937 This is expected to catch almost all occurrences in practice.
14938 If this code misses important cases that unbounded recursion
14939 would not, passes that need this information could be revised
14940 to provide it through dataflow propagation. */
14948 }
14949 }
14951 /* Return true if T is known to be non-negative. If the return
14952 value is based on the assumption that signed overflow is undefined,
14953 set *STRICT_OVERFLOW_P to true; otherwise, don't change
14954 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
14956 bool
14959 {
14961 {
14962 CASE_CFN_ACOS:
14963 CASE_CFN_ACOS_FN:
14964 CASE_CFN_ACOSH:
14965 CASE_CFN_ACOSH_FN:
14966 CASE_CFN_CABS:
14967 CASE_CFN_CABS_FN:
14968 CASE_CFN_COSH:
14969 CASE_CFN_COSH_FN:
14970 CASE_CFN_ERFC:
14971 CASE_CFN_ERFC_FN:
14972 CASE_CFN_EXP:
14973 CASE_CFN_EXP_FN:
14974 CASE_CFN_EXP10:
14975 CASE_CFN_EXP2:
14976 CASE_CFN_EXP2_FN:
14977 CASE_CFN_FABS:
14978 CASE_CFN_FABS_FN:
14979 CASE_CFN_FDIM:
14980 CASE_CFN_FDIM_FN:
14981 CASE_CFN_HYPOT:
14982 CASE_CFN_HYPOT_FN:
14983 CASE_CFN_POW10:
14984 CASE_CFN_FFS:
14985 CASE_CFN_PARITY:
14986 CASE_CFN_POPCOUNT:
14987 CASE_CFN_CLZ:
14988 CASE_CFN_CLRSB:
14993 /* Always true. */
14996 CASE_CFN_SQRT:
14997 CASE_CFN_SQRT_FN:
14998 /* sqrt(-0.0) is -0.0. */
15003 CASE_CFN_ASINH:
15004 CASE_CFN_ASINH_FN:
15005 CASE_CFN_ATAN:
15006 CASE_CFN_ATAN_FN:
15007 CASE_CFN_ATANH:
15008 CASE_CFN_ATANH_FN:
15009 CASE_CFN_CBRT:
15010 CASE_CFN_CBRT_FN:
15011 CASE_CFN_CEIL:
15012 CASE_CFN_CEIL_FN:
15013 CASE_CFN_ERF:
15014 CASE_CFN_ERF_FN:
15015 CASE_CFN_EXPM1:
15016 CASE_CFN_EXPM1_FN:
15017 CASE_CFN_FLOOR:
15018 CASE_CFN_FLOOR_FN:
15019 CASE_CFN_FMOD:
15020 CASE_CFN_FMOD_FN:
15021 CASE_CFN_FREXP:
15022 CASE_CFN_FREXP_FN:
15023 CASE_CFN_ICEIL:
15024 CASE_CFN_IFLOOR:
15025 CASE_CFN_IRINT:
15026 CASE_CFN_IROUND:
15027 CASE_CFN_LCEIL:
15028 CASE_CFN_LDEXP:
15029 CASE_CFN_LFLOOR:
15030 CASE_CFN_LLCEIL:
15031 CASE_CFN_LLFLOOR:
15032 CASE_CFN_LLRINT:
15033 CASE_CFN_LLRINT_FN:
15034 CASE_CFN_LLROUND:
15035 CASE_CFN_LLROUND_FN:
15036 CASE_CFN_LRINT:
15037 CASE_CFN_LRINT_FN:
15038 CASE_CFN_LROUND:
15039 CASE_CFN_LROUND_FN:
15040 CASE_CFN_MODF:
15041 CASE_CFN_MODF_FN:
15042 CASE_CFN_NEARBYINT:
15043 CASE_CFN_NEARBYINT_FN:
15044 CASE_CFN_RINT:
15045 CASE_CFN_RINT_FN:
15046 CASE_CFN_ROUND:
15047 CASE_CFN_ROUND_FN:
15048 CASE_CFN_ROUNDEVEN:
15049 CASE_CFN_ROUNDEVEN_FN:
15050 CASE_CFN_SCALB:
15051 CASE_CFN_SCALBLN:
15052 CASE_CFN_SCALBLN_FN:
15053 CASE_CFN_SCALBN:
15054 CASE_CFN_SCALBN_FN:
15055 CASE_CFN_SIGNBIT:
15056 CASE_CFN_SIGNIFICAND:
15057 CASE_CFN_SINH:
15058 CASE_CFN_SINH_FN:
15059 CASE_CFN_TANH:
15060 CASE_CFN_TANH_FN:
15061 CASE_CFN_TRUNC:
15062 CASE_CFN_TRUNC_FN:
15063 /* True if the 1st argument is nonnegative. */
15066 CASE_CFN_FMAX:
15067 CASE_CFN_FMAX_FN:
15068 /* Usually RECURSE (arg0) || RECURSE (arg1) but NaNs complicate
15069 things. In the presence of sNaNs, we're only guaranteed to be
15070 non-negative if both operands are non-negative. In the presence
15071 of qNaNs, we're non-negative if either operand is non-negative
15072 and can't be a qNaN, or if both operands are non-negative. */
15081 CASE_CFN_FMIN:
15082 CASE_CFN_FMIN_FN:
15083 /* True if the 1st AND 2nd arguments are nonnegative. */
15086 CASE_CFN_COPYSIGN:
15087 CASE_CFN_COPYSIGN_FN:
15088 /* True if the 2nd argument is nonnegative. */
15091 CASE_CFN_POWI:
15092 /* True if the 1st argument is nonnegative or the second
15093 argument is an even integer. */
15099 CASE_CFN_POW:
15100 CASE_CFN_POW_FN:
15101 /* True if the 1st argument is nonnegative or the second
15102 argument is an even integer valued real. */
15104 {
15105 REAL_VALUE_TYPE c;
15106 HOST_WIDE_INT n;
15111 {
15112 REAL_VALUE_TYPE cint;
15116 }
15117 }
15122 }
15124 }
15126 /* Return true if T is known to be non-negative. If the return
15127 value is based on the assumption that signed overflow is undefined,
15128 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15129 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15131 static bool
15133 {
15139 {
15141 {
15145 /* If the initializer is non-void, then it's a normal expression
15146 that will be assigned to the slot. */
15150 /* Otherwise, the initializer sets the slot in some way. One common
15151 way is an assignment statement at the end of the initializer. */
15153 {
15161 else
15163 }
15169 }
15172 {
15178 arg0,
15179 arg1,
15181 }
15194 }
15195 }
15197 #undef RECURSE
15198 #undef tree_expr_nonnegative_warnv_p
15200 /* Return true if T is known to be non-negative. If the return
15201 value is based on the assumption that signed overflow is undefined,
15202 set *STRICT_OVERFLOW_P to true; otherwise, don't change
15203 *STRICT_OVERFLOW_P. DEPTH is the current nesting depth of the query. */
15205 bool
15207 {
15214 {
15236 }
15239 {
15264 }
15265 }
15267 /* Return true if `t' is known to be non-negative. Handle warnings
15268 about undefined signed overflow. */
15270 bool
15272 {
15279 "determining that expression is always "
15281 WARN_STRICT_OVERFLOW_MISC);
15283 }
15286 /* Return true when (CODE OP0) is an address and is known to be nonzero.
15287 For floating point we further ensure that T is not denormal.
15288 Similar logic is present in nonzero_address in rtlanal.h.
15290 If the return value is based on the assumption that signed overflow
15291 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15292 change *STRICT_OVERFLOW_P. */
15294 bool
15297 {
15299 {
15302 strict_overflow_p);
15305 {
15311 strict_overflow_p));
15312 }
15317 strict_overflow_p);
15321 }
15324 }
15326 /* Return true when (CODE OP0 OP1) is an address and is known to be nonzero.
15327 For floating point we further ensure that T is not denormal.
15328 Similar logic is present in nonzero_address in rtlanal.h.
15330 If the return value is based on the assumption that signed overflow
15331 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15332 change *STRICT_OVERFLOW_P. */
15334 bool
15336 tree type,
15337 tree op0,
15339 {
15342 {
15346 {
15347 /* With the presence of negative values it is hard
15348 to say something. */
15355 /* One of operands must be positive and the other non-negative. */
15356 /* We don't set *STRICT_OVERFLOW_P here: even if this value
15357 overflows, on a twos-complement machine the sum of two
15358 nonnegative numbers can never be zero. */
15360 strict_overflow_p)
15362 strict_overflow_p));
15363 }
15368 {
15370 strict_overflow_p)
15372 strict_overflow_p))
15373 {
15376 }
15377 }
15386 {
15389 }
15396 {
15400 /* When both operands are nonzero, then MAX must be too. */
15402 strict_overflow_p))
15405 /* MAX where operand 0 is positive is positive. */
15407 strict_overflow_p);
15408 }
15409 /* MAX where operand 1 is positive is positive. */
15414 {
15418 }
15423 strict_overflow_p)
15425 strict_overflow_p));
15429 }
15432 }
15434 /* Return true when T is an address and is known to be nonzero.
15435 For floating point we further ensure that T is not denormal.
15436 Similar logic is present in nonzero_address in rtlanal.h.
15438 If the return value is based on the assumption that signed overflow
15439 is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't
15440 change *STRICT_OVERFLOW_P. */
15442 bool
15444 {
15447 {
15452 {
15464 /* For objects in symbol table check if we know they are non-zero.
15465 Don't do anything for variables and functions before symtab is built;
15466 it is quite possible that they will be declared weak later. */
15471 /* Constants are never weak. */
15476 }
15484 {
15488 }
15498 }
15500 }
15502 #define integer_valued_real_p(X) \
15505 #define RECURSE(X) \
15506 ((integer_valued_real_p) (X, depth + 1))
15508 /* Return true if the floating point result of (CODE OP0) has an
15509 integer value. We also allow +Inf, -Inf and NaN to be considered
15510 integer values. Return false for signaling NaN.
15512 DEPTH is the current nesting depth of the query. */
15514 bool
15516 {
15518 {
15525 CASE_CONVERT:
15526 {
15533 }
15537 }
15539 }
15541 /* Return true if the floating point result of (CODE OP0 OP1) has an
15542 integer value. We also allow +Inf, -Inf and NaN to be considered
15543 integer values. Return false for signaling NaN.
15545 DEPTH is the current nesting depth of the query. */
15547 bool
15549 {
15551 {
15561 }
15563 }
15565 /* Return true if the floating point result of calling FNDECL with arguments
15566 ARG0 and ARG1 has an integer value. We also allow +Inf, -Inf and NaN to be
15567 considered integer values. Return false for signaling NaN. If FNDECL
15568 takes fewer than 2 arguments, the remaining ARGn are null.
15570 DEPTH is the current nesting depth of the query. */
15572 bool
15574 {
15576 {
15577 CASE_CFN_CEIL:
15578 CASE_CFN_CEIL_FN:
15579 CASE_CFN_FLOOR:
15580 CASE_CFN_FLOOR_FN:
15581 CASE_CFN_NEARBYINT:
15582 CASE_CFN_NEARBYINT_FN:
15583 CASE_CFN_RINT:
15584 CASE_CFN_RINT_FN:
15585 CASE_CFN_ROUND:
15586 CASE_CFN_ROUND_FN:
15587 CASE_CFN_ROUNDEVEN:
15588 CASE_CFN_ROUNDEVEN_FN:
15589 CASE_CFN_TRUNC:
15590 CASE_CFN_TRUNC_FN:
15593 CASE_CFN_FMIN:
15594 CASE_CFN_FMIN_FN:
15595 CASE_CFN_FMAX:
15596 CASE_CFN_FMAX_FN:
15601 }
15603 }
15605 /* Return true if the floating point expression T (a GIMPLE_SINGLE_RHS)
15606 has an integer value. We also allow +Inf, -Inf and NaN to be
15607 considered integer values. Return false for signaling NaN.
15609 DEPTH is the current nesting depth of the query. */
15611 bool
15613 {
15615 {
15623 /* Limit the depth of recursion to avoid quadratic behavior.
15624 This is expected to catch almost all occurrences in practice.
15625 If this code misses important cases that unbounded recursion
15626 would not, passes that need this information could be revised
15627 to provide it through dataflow propagation. */
15631 depth));
15635 }
15637 }
15639 /* Return true if the floating point expression T (a GIMPLE_INVALID_RHS)
15640 has an integer value. We also allow +Inf, -Inf and NaN to be
15641 considered integer values. Return false for signaling NaN.
15643 DEPTH is the current nesting depth of the query. */
15645 static bool
15647 {
15649 {
15660 }
15662 }
15664 #undef RECURSE
15665 #undef integer_valued_real_p
15667 /* Return true if the floating point expression T has an integer value.
15668 We also allow +Inf, -Inf and NaN to be considered integer values.
15669 Return false for signaling NaN.
15671 DEPTH is the current nesting depth of the query. */
15673 bool
15675 {
15683 {
15699 }
15702 {
15708 {
15717 }
15721 }
15722 }
15724 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
15725 attempt to fold the expression to a constant without modifying TYPE,
15726 OP0 or OP1.
15728 If the expression could be simplified to a constant, then return
15729 the constant. If the expression would not be simplified to a
15730 constant, then return NULL_TREE. */
15732 tree
15734 {
15737 }
15739 /* Given the components of a unary expression CODE, TYPE and OP0,
15740 attempt to fold the expression to a constant without modifying
15741 TYPE or OP0.
15743 If the expression could be simplified to a constant, then return
15744 the constant. If the expression would not be simplified to a
15745 constant, then return NULL_TREE. */
15747 tree
15749 {
15752 }
15754 /* If EXP represents referencing an element in a constant string
15755 (either via pointer arithmetic or array indexing), return the
15756 tree representing the value accessed, otherwise return NULL. */
15758 tree
15760 {
15764 {
15766 tree index;
15767 tree string;
15772 else
15773 {
15777 /* Optimize the special-case of a zero lower bound.
15779 We convert the low_bound to sizetype to avoid some problems
15780 with constant folding. (E.g. suppose the lower bound is 1,
15781 and its mode is QI. Without the conversion,l (ARRAY
15782 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
15783 +INDEX), which becomes (ARRAY+255+INDEX). Oops!) */
15789 }
15791 scalar_int_mode char_mode;
15803 }
15805 }
15807 /* Folds a read from vector element at IDX of vector ARG. */
15809 tree
15811 {
15816 {
15820 {
15827 }
15828 }
15830 }
15832 /* Return the tree for neg (ARG0) when ARG0 is known to be either
15833 an integer constant, real, or fixed-point constant.
15835 TYPE is the type of the result. */
15837 static tree
15839 {
15843 {
15849 {
15850 FIXED_VALUE_TYPE f;
15855 /* Propagate overflow flags. */
15859 }
15863 {
15870 }
15873 }
15876 }
15878 /* Return the tree for abs (ARG0) when ARG0 is known to be either
15879 an integer constant or real constant.
15881 TYPE is the type of the result. */
15883 tree
15885 {
15889 {
15891 {
15892 /* If the value is unsigned or non-negative, then the absolute value
15893 is the same as the ordinary value. */
15897 ;
15899 /* If the value is negative, then the absolute value is
15900 its negation. */
15901 else
15904 /* Force to the destination type, set TREE_OVERFLOW for signed
15905 TYPE only. */
15907 }
15913 else
15919 }
15922 }
15924 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
15925 constant. TYPE is the type of the result. */
15927 static tree
15929 {
15933 }
15935 /* Given CODE, a relational operator, the target type, TYPE and two
15936 constant operands OP0 and OP1, return the result of the
15937 relational operation. If the result is not a compile time
15938 constant, then return NULL_TREE. */
15940 static tree
15942 {
15945 /* From here on, the only cases we handle are when the result is
15946 known to be a constant. */
15949 {
15953 /* Handle the cases where either operand is a NaN. */
15955 {
15957 {
15985 }
15988 }
15991 }
15994 {
15998 }
16000 /* Handle equality/inequality of complex constants. */
16002 {
16013 else
16015 }
16018 {
16020 {
16021 /* Have vector comparison with scalar boolean result. */
16029 {
16037 }
16039 }
16040 tree_vector_builder elts;
16045 {
16058 }
16061 }
16063 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
16065 To compute GT, swap the arguments and do LT.
16066 To compute GE, do LT and invert the result.
16067 To compute LE, swap the arguments, do LT and invert the result.
16068 To compute NE, do EQ and invert the result.
16070 Therefore, the code below must handle only EQ and LT. */
16073 {
16076 }
16078 /* Note that it is safe to invert for real values here because we
16079 have already handled the one case that it matters. */
16083 {
16086 }
16088 /* Compute a result for LT or EQ if args permit;
16089 Otherwise return T. */
16091 {
16094 else
16096 }
16097 else
16103 }
16105 /* If necessary, return a CLEANUP_POINT_EXPR for EXPR with the
16106 indicated TYPE. If no CLEANUP_POINT_EXPR is necessary, return EXPR
16107 itself. */
16109 tree
16111 {
16112 /* If the expression does not have side effects then we don't have to wrap
16113 it with a cleanup point expression. */
16117 /* If the expression is a return, check to see if the expression inside the
16118 return has no side effects or the right hand side of the modify expression
16119 inside the return. If either don't have side effects set we don't need to
16120 wrap the expression in a cleanup point expression. Note we don't check the
16121 left hand side of the modify because it should always be a return decl. */
16123 {
16130 }
16133 }
16135 /* Given a pointer value OP0 and a type TYPE, return a simplified version
16136 of an indirection through OP0, or NULL_TREE if no simplification is
16137 possible. */
16139 tree
16141 {
16143 tree subtype;
16144 poly_uint64 const_op01;
16153 {
16157 /* *&CONST_DECL -> to the value of the const decl. */
16160 /* *&p => p; make sure to handle *&"str"[cst] here. */
16162 {
16166 else
16168 }
16169 /* *(foo *)&fooarray => fooarray[0] */
16172 && (!in_gimple_form
16174 {
16184 }
16185 /* *(foo *)&complexfoo => __real__ complexfoo */
16189 /* *(foo *)&vectorfoo => BIT_FIELD_REF<vectorfoo,...> */
16192 {
16196 index);
16197 }
16198 }
16202 {
16208 {
16209 tree op00type;
16213 /* ((foo*)&vectorfoo)[1] => BIT_FIELD_REF<vectorfoo,...> */
16216 /* POINTER_PLUS_EXPR second operand is sizetype, unsigned,
16217 but we want to treat offsets with MSB set as negative.
16218 For the code below negative offsets are invalid and
16219 TYPE_SIZE of the element is something unsigned, so
16220 check whether op01 fits into poly_int64, which implies
16221 it is from 0 to INTTYPE_MAXIMUM (HOST_WIDE_INT), and
16222 then just use poly_uint64 because we want to treat the
16223 value as unsigned. */
16225 {
16227 poly_uint64 max_offset
16231 {
16236 }
16237 }
16238 /* ((foo*)&complexfoo)[1] => __imag__ complexfoo */
16241 {
16243 const_op01))
16245 }
16246 /* ((foo *)&fooarray)[1] => fooarray[1] */
16249 {
16258 {
16263 }
16264 }
16265 }
16266 }
16268 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
16271 && (!in_gimple_form
16273 {
16274 tree type_domain;
16284 NULL_TREE);
16285 }
16288 }
16290 /* Builds an expression for an indirection through T, simplifying some
16291 cases. */
16293 tree
16295 {
16303 }
16305 /* Given an INDIRECT_REF T, return either T or a simplified version. */
16307 tree
16309 {
16314 else
16316 }
16318 /* Strip non-trapping, non-side-effecting tree nodes from an expression
16319 whose result is ignored. The type of the returned tree need not be
16320 the same as the original expression. */
16322 tree
16324 {
16330 {
16341 else
16347 {
16363 }
16368 }
16369 }
16371 /* Return the value of VALUE, rounded up to a multiple of DIVISOR. */
16373 tree
16375 {
16381 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16382 have to do anything. Only do this when we are not given a const,
16383 because in that case, this check is more expensive than just
16384 doing it. */
16386 {
16391 }
16393 /* If divisor is a power of two, simplify this to bit manipulation. */
16395 {
16397 {
16411 }
16412 else
16413 {
16414 tree t;
16420 }
16421 }
16422 else
16423 {
16428 }
16431 }
16433 /* Likewise, but round down. */
16435 tree
16437 {
16444 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
16445 have to do anything. Only do this when we are not given a const,
16446 because in that case, this check is more expensive than just
16447 doing it. */
16449 {
16454 }
16456 /* If divisor is a power of two, simplify this to bit manipulation. */
16458 {
16459 tree t;
16463 }
16464 else
16465 {
16470 }
16473 }
16475 /* Returns the pointer to the base of the object addressed by EXP and
16476 extracts the information about the offset of the access, storing it
16477 to PBITPOS and POFFSET. */
16479 static tree
16482 {
16483 tree core;
16484 machine_mode mode;
16486 poly_int64 bitsize;
16495 {
16500 }
16502 {
16508 {
16509 poly_offset_int tem
16515 }
16516 }
16517 else
16518 {
16522 }
16525 }
16527 /* Returns true if addresses of E1 and E2 differ by a constant, false
16528 otherwise. If they do, E1 - E2 is stored in *DIFF. */
16530 bool
16532 {
16547 {
16557 }
16559 {
16560 /* If only one of the offsets is non-constant, the difference cannot
16561 be a constant. */
16563 }
16564 else
16569 }
16571 /* Return OFF converted to a pointer offset type suitable as offset for
16572 POINTER_PLUS_EXPR. Use location LOC for this conversion. */
16573 tree
16575 {
16579 }
16581 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16582 tree
16584 {
16587 }
16589 /* Build and fold a POINTER_PLUS_EXPR at LOC offsetting PTR by OFF. */
16590 tree
16592 {
16595 }
16597 /* Return a pointer to a NUL-terminated string containing the sequence
16598 of bytes corresponding to the representation of the object referred to
16599 by SRC (or a subsequence of such bytes within it if SRC is a reference
16600 to an initialized constant array plus some constant offset).
16601 Set *STRSIZE the number of bytes in the constant sequence including
16602 the terminating NUL byte. *STRSIZE is equal to sizeof(A) - OFFSET
16603 where A is the array that stores the constant sequence that SRC points
16604 to and OFFSET is the byte offset of SRC from the beginning of A. SRC
16605 need not point to a string or even an array of characters but may point
16606 to an object of any type. */
16610 {
16611 /* The offset into the array A storing the string, and A's byte size. */
16612 tree offset_node;
16613 tree mem_size;
16620 else
16627 {
16630 else
16632 }
16637 /* ARRAY_SIZE is the byte size of the array the constant sequence
16638 is stored in and equal to sizeof A. INIT_BYTES is the number
16639 of bytes in the constant sequence used to initialize the array,
16640 including any embedded NULs as well as the terminating NUL (for
16641 strings), but not including any trailing zeros/NULs past
16642 the terminating one appended implicitly to a string literal to
16643 zero out the remainder of the array it's stored in. For example,
16644 given:
16645 const char a[7] = "abc\0d";
16646 n = strlen (a + 1);
16647 ARRAY_SIZE is 7, INIT_BYTES is 6, and OFFSET is 1. For a valid
16648 (i.e., nul-terminated) string with no embedded nuls, INIT_BYTES
16649 is equal to strlen (A) + 1. */
16654 /* Ideally this would turn into a gcc_checking_assert over time. */
16662 {
16663 /* Compute and store the number of characters from the beginning
16664 of the substring at OFFSET to the end, including the terminating
16665 nul. Offsets past the initial length refer to null strings. */
16668 else
16670 }
16671 else
16672 {
16674 /* Support only properly NUL-terminated single byte strings. */
16679 }
16682 }
16684 /* Return a pointer to a NUL-terminated string corresponding to
16685 the expression STR referencing a constant string, possibly
16686 involving a constant offset. Return null if STR either doesn't
16687 reference a constant string or if it involves a nonconstant
16688 offset. */
16692 {
16694 }
16696 /* Given a tree T, compute which bits in T may be nonzero. */
16698 wide_int
16700 {
16702 {
16720 CASE_CONVERT:
16726 {
16731 }
16735 {
16743 }
16747 {
16755 }
16759 }
16762 }
16764 /* Helper function for address compare simplifications in match.pd.
16765 OP0 and OP1 are ADDR_EXPR operands being compared by CODE.
16766 TYPE is the type of comparison operands.
16767 BASE0, BASE1, OFF0 and OFF1 are set by the function.
16768 GENERIC is true if GENERIC folding and false for GIMPLE folding.
16769 Returns 0 if OP0 is known to be unequal to OP1 regardless of OFF{0,1},
16770 1 if bases are known to be equal and OP0 cmp OP1 depends on OFF0 cmp OFF1,
16771 and 2 if unknown. */
16773 int
16777 {
16787 {
16790 }
16792 {
16795 }
16800 /* Punt in GENERIC on variables with value expressions;
16801 the value expressions might point to fields/elements
16802 of other vars etc. */
16808 {
16812 }
16820 /* Assume different STRING_CSTs with the same content will be
16821 merged. */
16830 {
16833 /* If the offsets are equal we can ignore overflow. */
16836 /* Or if we compare using pointers to decls or strings. */
16841 }
16847 /* At this point we know (or assume) the two pointers point at
16848 different objects. */
16852 /* Punt on non-zero offsets from functions. */
16856 /* Or if the bases are neither decls nor string literals. */
16861 /* For initializers, assume addresses of different functions are
16862 different. */
16868 /* Compute whether one address points to the start of one
16869 object and another one to the end of another one. */
16872 {
16875 else
16877 }
16880 else
16881 {
16885 else
16887 }
16889 {
16892 else
16894 }
16897 else
16898 {
16902 else
16904 }
16906 {
16907 /* If one offset is pointing (or could be) to the beginning of one
16908 object and the other is pointing to one past the last byte of the
16909 other object, punt. */
16914 /* If both offsets are the same, there are some cases we know that are
16915 ok. Either if we know they aren't zero, or if we know both sizes
16916 are no zero. */
16922 }
16924 /* At this point, equal is 2 if either one or both pointers are out of
16925 bounds of their object, or one points to start of its object and the
16926 other points to end of its object. This is unspecified behavior
16927 e.g. in C++. Otherwise equal is 0. */
16931 /* When both pointers point to string literals, even when equal is 0,
16932 due to tail merging of string literals the pointers might be the same. */
16934 {
16941 /* If the bytes in the string literals starting at the pointers
16942 differ, the pointers need to be different. */
16947 {
16952 /* If even the bytes in the string literal before the
16953 pointers are the same, the string literals could be
16954 tail merged. */
16956 }
16958 }
16963 /* If this is a pointer comparison, ignore for now even
16964 valid equalities where one pointer is the offset zero
16965 of one object and the other to one past end of another one. */
16969 /* Assume that string literals can't be adjacent to variables
16970 (automatic or global). */
16974 /* Assume that automatic variables can't be adjacent to global
16975 variables. */
16980 }
16982 /* Return the single non-zero element of a CONSTRUCTOR or NULL_TREE. */
16983 tree
16985 {
16994 {
16998 }
17000 }
17002 #if CHECKING_P
17006 /* Helper functions for writing tests of folding trees. */
17008 /* Verify that the binary op (LHS CODE RHS) folds to CONSTANT. */
17010 static void
17012 tree constant)
17013 {
17015 }
17017 /* Verify that the binary op (LHS CODE RHS) folds to an NON_LVALUE_EXPR
17018 wrapping WRAPPED_EXPR. */
17020 static void
17022 tree wrapped_expr)
17023 {
17028 }
17030 /* Verify that various arithmetic binary operations are folded
17031 correctly. */
17033 static void
17035 {
17041 /* Addition. */
17042 /* 1 <-- (0 + 1) */
17044 one);
17046 one);
17048 /* (nonlvalue)x <-- (x + 0) */
17050 x);
17052 /* Subtraction. */
17053 /* 0 <-- (x - x) */
17055 zero);
17057 x);
17059 /* Multiplication. */
17060 /* 0 <-- (x * 0) */
17062 zero);
17064 /* (nonlvalue)x <-- (x * 1) */
17066 x);
17067 }
17071 /* Build a VECTOR_CST corresponding to VMODE, and has
17072 encoding given by NPATTERNS, NELTS_PER_PATTERN and STEP.
17073 Fill it with randomized elements, using rand() % THRESHOLD. */
17075 static tree
17079 {
17084 // Fill a0 for each pattern
17091 // Fill a1 for each pattern
17099 {
17104 }
17107 }
17109 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17110 when result is VLA. */
17112 static void
17115 {
17116 /* Actual npatterns and encoded_elts in res may be less than expected due
17117 to canonicalization. */
17124 }
17126 /* Validate result of VEC_PERM_EXPR folding for the unit-tests below,
17127 when the result is VLS. */
17129 static void
17131 {
17135 }
17137 /* Helper routine to push multiple elements into BUILDER. */
17141 {
17144 }
17146 #define ARG0(index) vector_cst_elt (arg0, index)
17147 #define ARG1(index) vector_cst_elt (arg1, index)
17149 /* Test cases where result is VNx4SI and input vectors are V4SI. */
17151 static void
17153 {
17155 {
17156 /* Case 1:
17157 sel = { 0, 4, 1, 5, ... }
17158 res = { arg[0], arg1[0], arg0[1], arg1[1], ...} // (4, 1) */
17159 {
17163 tree inner_type
17177 }
17179 /* Case 2: Same as case 1, but contains an out of bounds access which
17180 should wrap around.
17181 sel = {0, 8, 4, 12, ...} (4, 1)
17182 res = { arg0[0], arg0[0], arg1[0], arg1[0], ... } (4, 1). */
17183 {
17187 tree inner_type
17201 }
17202 }
17203 }
17205 /* Test cases where result is V4SI and input vectors are VNx4SI. */
17207 static void
17209 {
17211 {
17212 /* Case 1:
17213 sel = { 0, 1, 2, 3}
17214 res = { arg0[0], arg0[1], arg0[2], arg0[3] }. */
17215 {
17219 tree inner_type
17233 }
17235 /* Case 2: Same as Case 1, but crossing input vector.
17236 sel = {0, 2, 4, 6}
17237 In this case,the index 4 is ambiguous since len = 4 + 4x.
17238 Since we cannot determine, which vector to choose from during
17239 compile time, should return NULL_TREE. */
17240 {
17244 tree inner_type
17259 }
17260 }
17261 }
17263 /* Test all input vectors. */
17265 static void
17267 {
17268 /* Test with 10 different inputs. */
17270 {
17275 /* Case 1: mask = {0, ...} // (1, 1)
17276 res = { arg0[0], ... } // (1, 1) */
17277 {
17284 }
17286 /* Case 2: mask = {len, ...} // (1, 1)
17287 res = { arg1[0], ... } // (1, 1) */
17288 {
17296 }
17297 }
17298 }
17300 /* Test all vectors which contain at-least 2 elements. */
17302 static void
17304 {
17306 {
17307 /* Case 1: mask = { 0, len, ... } // (2, 1)
17308 res = { arg0[0], arg1[0], ... } // (2, 1) */
17309 {
17323 }
17325 /* Case 2: mask = { 0, len, 1, len+1, ... } // (2, 2)
17326 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2) */
17327 {
17341 }
17343 /* Case 4: mask = {0, 0, 1, ...} // (1, 3)
17344 Test that the stepped sequence of the pattern selects from
17345 same input pattern. Since input vectors have npatterns = 2,
17346 and step (a2 - a1) = 1, step is not a multiple of npatterns
17347 in input vector. So return NULL_TREE. */
17348 {
17363 }
17365 /* Case 5: mask = {len, 0, 1, ...} // (1, 3)
17366 Test that stepped sequence of the pattern selects from arg0.
17367 res = { arg1[0], arg0[0], arg0[1], ... } // (1, 3) */
17368 {
17382 }
17383 }
17384 }
17386 /* Test all vectors which contain at-least 4 elements. */
17388 static void
17390 {
17392 {
17393 /* Case 1: mask = { 0, len, 1, len+1, ... } // (4, 1)
17394 res: { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (4, 1) */
17395 {
17409 }
17411 /* Case 2: sel = {0, 1, 2, ...} // (1, 3)
17412 res: { arg0[0], arg0[1], arg0[2], ... } // (1, 3) */
17413 {
17426 }
17428 /* Case 3: sel = {len, len+1, len+2, ...} // (1, 3)
17429 res: { arg1[0], arg1[1], arg1[2], ... } // (1, 3) */
17430 {
17443 }
17445 /* Case 4:
17446 sel = { len, 0, 2, ... } // (1, 3)
17447 This should return NULL because we cross the input vectors.
17448 Because,
17449 Let's assume len = C + Cx
17450 a1 = 0
17451 S = 2
17452 esel = arg0_len / sel_npatterns = C + Cx
17453 ae = 0 + (esel - 2) * S
17454 = 0 + (C + Cx - 2) * 2
17455 = 2(C-2) + 2Cx
17457 For C >= 4:
17458 Let q1 = a1 / arg0_len = 0 / (C + Cx) = 0
17459 Let qe = ae / arg0_len = (2(C-2) + 2Cx) / (C + Cx) = 1
17460 Since q1 != qe, we cross input vectors.
17461 So return NULL_TREE. */
17462 {
17476 }
17478 /* Case 5: npatterns(arg0) = 4 > npatterns(sel) = 2
17479 mask = { 0, len, 1, len + 1, ...} // (2, 2)
17480 res = { arg0[0], arg1[0], arg0[1], arg1[1], ... } // (2, 2)
17482 Note that fold_vec_perm_cst will set
17483 res_npatterns = max(4, max(4, 2)) = 4
17484 However after canonicalizing, we will end up with shape (2, 2). */
17485 {
17498 }
17500 /* Case 6: Test combination in sel, where one pattern is dup and other
17501 is stepped sequence.
17502 sel = { 0, 0, 0, 1, 0, 2, ... } // (2, 3)
17503 res = { arg0[0], arg0[0], arg0[0],
17504 arg0[1], arg0[0], arg0[2], ... } // (2, 3) */
17505 {
17520 }
17521 }
17522 }
17524 /* Test all vectors which contain at-least 8 elements. */
17526 static void
17528 {
17530 {
17531 /* Case 1: sel_npatterns (4) > input npatterns (2)
17532 sel: { 0, 0, 1, len, 2, 0, 3, len, 4, 0, 5, len, ...} // (4, 3)
17533 res: { arg0[0], arg0[0], arg0[0], arg1[0],
17534 arg0[2], arg0[0], arg0[3], arg1[0],
17535 arg0[4], arg0[0], arg0[5], arg1[0], ... } // (4, 3) */
17536 {
17553 }
17554 }
17555 }
17557 /* Test vectors for which nunits[0] <= 4. */
17559 static void
17561 {
17562 /* Case 1: mask = {0, 4, ...} // (1, 2)
17563 This should return NULL_TREE because the index 4 may choose
17564 from either arg0 or arg1 depending on vector length. */
17565 {
17580 }
17581 }
17583 #undef ARG0
17584 #undef ARG1
17586 /* Return true if SIZE is of the form C + Cx and C is power of 2. */
17588 static bool
17590 {
17598 }
17600 /* Execute fold_vec_perm_cst unit tests. */
17602 static void
17604 {
17608 machine_mode vmode;
17610 {
17611 /* Obtain modes corresponding to VNx4SI and V4SI,
17612 to call mixed mode tests below.
17613 FIXME: Is there a better way to do this ? */
17615 {
17622 }
17639 }
17644 {
17647 }
17648 }
17651 /* Verify that various binary operations on vectors are folded
17652 correctly. */
17654 static void
17656 {
17663 /* Verify equality tests that return a scalar boolean result. */
17676 }
17678 /* Verify folding of VEC_DUPLICATE_EXPRs. */
17680 static void
17682 {
17685 /* This will be 1 if VEC_MODE isn't a vector mode. */
17692 }
17694 /* Run all of the selftests within this file. */
17696 void
17698 {
17703 }