| blob | ab81bf4913140510f1e44dbdcb26f2ec2493196f |
1 /* Fold a constant sub-tree into a single node for C-compiler
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
3 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
22 /*@@ This file should be rewritten to use an arbitrary precision
23 @@ representation for "struct tree_int_cst" and "struct tree_real_cst".
24 @@ Perhaps the routines could also be used for bc/dc, and made a lib.
25 @@ The routines that translate from the ap rep should
26 @@ warn if precision et. al. is lost.
27 @@ This would also make life easier when this technology is used
28 @@ for cross-compilers. */
30 /* The entry points in this file are fold, size_int_wide, size_binop
31 and force_fit_type.
33 fold takes a tree as argument and returns a simplified tree.
35 size_binop takes a tree code for an arithmetic operation
36 and two operands that are trees, and produces a tree for the
37 result, assuming the type comes from `sizetype'.
39 size_int takes an integer value, and creates a tree constant
40 with type from `sizetype'.
42 force_fit_type takes a constant, an overflowable flag and prior
43 overflow indicators. It forces the value to fit the type and sets
44 TREE_OVERFLOW and TREE_CONSTANT_OVERFLOW as appropriate. */
62 /* The following constants represent a bit based encoding of GCC's
63 comparison operators. This encoding simplifies transformations
64 on relational comparison operators, such as AND and OR. */
82 };
117 tree);
141 /* We know that A1 + B1 = SUM1, using 2's complement arithmetic and ignoring
142 overflow. Suppose A, B and SUM have the same respective signs as A1, B1,
143 and SUM1. Then this yields nonzero if overflow occurred during the
144 addition.
146 Overflow occurs if A and B have the same sign, but A and SUM differ in
147 sign. Use `^' to test whether signs differ, and `< 0' to isolate the
148 sign. */
149 #define OVERFLOW_SUM_SIGN(a, b, sum) ((~((a) ^ (b)) & ((a) ^ (sum))) < 0)
150 \f
151 /* To do constant folding on INTEGER_CST nodes requires two-word arithmetic.
152 We do that by representing the two-word integer in 4 words, with only
153 HOST_BITS_PER_WIDE_INT / 2 bits stored in each word, as a positive
154 number. The value of the word is LOWPART + HIGHPART * BASE. */
156 #define LOWPART(x) \
157 ((x) & (((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) - 1))
158 #define HIGHPART(x) \
159 ((unsigned HOST_WIDE_INT) (x) >> HOST_BITS_PER_WIDE_INT / 2)
160 #define BASE ((unsigned HOST_WIDE_INT) 1 << HOST_BITS_PER_WIDE_INT / 2)
162 /* Unpack a two-word integer into 4 words.
163 LOW and HI are the integer, as two `HOST_WIDE_INT' pieces.
164 WORDS points to the array of HOST_WIDE_INTs. */
166 static void
168 {
173 }
175 /* Pack an array of 4 words into a two-word integer.
176 WORDS points to the array of words.
177 The integer is stored into *LOW and *HI as two `HOST_WIDE_INT' pieces. */
179 static void
182 {
185 }
186 \f
187 /* T is an INT_CST node. OVERFLOWABLE indicates if we are interested
188 in overflow of the value, when >0 we are only interested in signed
189 overflow, for <0 we are interested in any overflow. OVERFLOWED
190 indicates whether overflow has already occurred. CONST_OVERFLOWED
191 indicates whether constant overflow has already occurred. We force
192 T's value to be within range of T's type (by setting to 0 or 1 all
193 the bits outside the type's range). We set TREE_OVERFLOWED if,
194 OVERFLOWED is nonzero,
195 or OVERFLOWABLE is >0 and signed overflow occurs
196 or OVERFLOWABLE is <0 and any overflow occurs
197 We set TREE_CONSTANT_OVERFLOWED if,
198 CONST_OVERFLOWED is nonzero
199 or we set TREE_OVERFLOWED.
200 We return either the original T, or a copy. */
202 tree
205 {
207 HOST_WIDE_INT high;
219 else
221 /* Size types *are* sign extended. */
226 /* First clear all bits that are beyond the type's precision. */
229 ;
232 else
233 {
237 }
244 {
245 /* Sign extend top half? */
249 }
251 {
254 }
255 else
256 {
257 /* Sign extend bottom half? */
259 {
262 }
263 }
265 /* If the value changed, return a new node. */
268 {
274 {
278 }
280 {
283 }
284 }
287 }
288 \f
289 /* Add two doubleword integers with doubleword result.
290 Each argument is given as two `HOST_WIDE_INT' pieces.
291 One argument is L1 and H1; the other, L2 and H2.
292 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
294 int
298 {
300 HOST_WIDE_INT h;
308 }
310 /* Negate a doubleword integer with doubleword result.
311 Return nonzero if the operation overflows, assuming it's signed.
312 The argument is given as two `HOST_WIDE_INT' pieces in L1 and H1.
313 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
315 int
318 {
320 {
324 }
325 else
326 {
330 }
331 }
332 \f
333 /* Multiply two doubleword integers with doubleword result.
334 Return nonzero if the operation overflows, assuming it's signed.
335 Each argument is given as two `HOST_WIDE_INT' pieces.
336 One argument is L1 and H1; the other, L2 and H2.
337 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
339 int
343 {
358 {
361 {
363 /* This product is <= 0xFFFE0001, the sum <= 0xFFFF0000. */
365 /* Since prod[p] < 0xFFFF, this sum <= 0xFFFFFFFF. */
369 }
371 }
375 /* Check for overflow by calculating the top half of the answer in full;
376 it should agree with the low half's sign bit. */
379 {
382 }
384 {
387 }
389 }
390 \f
391 /* Shift the doubleword integer in L1, H1 left by COUNT places
392 keeping only PREC bits of result.
393 Shift right if COUNT is negative.
394 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
395 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
397 void
401 {
405 {
408 }
414 {
415 /* Shifting by the host word size is undefined according to the
416 ANSI standard, so we must handle this as a special case. */
419 }
421 {
424 }
425 else
426 {
430 }
432 /* Sign extend all bits that are beyond the precision. */
440 ;
442 {
445 }
446 else
447 {
451 }
452 }
454 /* Shift the doubleword integer in L1, H1 right by COUNT places
455 keeping only PREC bits of result. COUNT must be positive.
456 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
457 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
459 void
464 {
467 signmask = (arith
475 {
476 /* Shifting by the host word size is undefined according to the
477 ANSI standard, so we must handle this as a special case. */
480 }
482 {
485 }
486 else
487 {
491 }
493 /* Zero / sign extend all bits that are beyond the precision. */
496 {
499 }
501 ;
503 {
506 }
507 else
508 {
512 }
513 }
514 \f
515 /* Rotate the doubleword integer in L1, H1 left by COUNT places
516 keeping only PREC bits of result.
517 Rotate right if COUNT is negative.
518 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
520 void
524 {
536 }
538 /* Rotate the doubleword integer in L1, H1 left by COUNT places
539 keeping only PREC bits of result. COUNT must be positive.
540 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
542 void
546 {
558 }
559 \f
560 /* Divide doubleword integer LNUM, HNUM by doubleword integer LDEN, HDEN
561 for a quotient (stored in *LQUO, *HQUO) and remainder (in *LREM, *HREM).
562 CODE is a tree code for a kind of division, one of
563 TRUNC_DIV_EXPR, FLOOR_DIV_EXPR, CEIL_DIV_EXPR, ROUND_DIV_EXPR
564 or EXACT_DIV_EXPR
565 It controls how the quotient is rounded to an integer.
566 Return nonzero if the operation overflows.
567 UNS nonzero says do unsigned division. */
569 int
572 HOST_WIDE_INT hnum_orig,
574 HOST_WIDE_INT hden_orig,
578 {
594 /* Calculate quotient sign and convert operands to unsigned. */
596 {
598 {
600 /* (minimum integer) / (-1) is the only overflow case. */
604 }
606 {
609 }
610 }
615 /* This unsigned division rounds toward zero. */
618 }
622 /* hden != 0 already checked. */
627 }
637 /* Special code for when the divisor < BASE. */
639 {
640 /* hnum != 0 already checked. */
642 {
646 }
647 }
648 else
649 {
650 /* Full double precision division,
651 with thanks to Don Knuth's "Seminumerical Algorithms". */
655 /* Find the highest nonzero divisor digit. */
658 {
661 }
663 /* Insure that the first digit of the divisor is at least BASE/2.
664 This is required by the quotient digit estimation algorithm. */
671 {
675 }
680 {
685 }
686 }
690 /* Main loop */
692 {
693 /* Guess the next quotient digit, quo_est, by dividing the first
694 two remaining dividend digits by the high order quotient digit.
695 quo_est is never low and is at most 2 high. */
702 else
705 /* Refine quo_est so it's usually correct, and at most one high. */
710 quo_est--;
712 /* Try QUO_EST as the quotient digit, by multiplying the
713 divisor by QUO_EST and subtracting from the remaining dividend.
714 Keep in mind that QUO_EST is the I - 1st digit. */
718 {
724 }
726 /* If quo_est was high by one, then num[i] went negative and
727 we need to correct things. */
729 {
730 quo_est--;
733 {
737 }
740 }
742 /* Store the quotient digit. */
744 }
745 }
749 finish_up:
750 /* If result is negative, make it so. */
754 /* Compute trial remainder: rem = num - (quo * den) */
760 {
769 {
770 /* quo = quo - 1; */
773 }
774 else
781 {
784 }
785 else
791 {
797 /* Get absolute values. */
803 /* If (2 * abs (lrem) >= abs (lden)) */
812 {
814 /* quo = quo - 1; */
817 else
818 /* quo = quo + 1; */
821 }
822 else
824 }
829 }
831 /* Compute true remainder: rem = num - (quo * den) */
836 }
837 \f
838 /* Return true if built-in mathematical function specified by CODE
839 preserves the sign of it argument, i.e. -f(x) == f(-x). */
841 static bool
843 {
845 {
862 }
864 }
866 /* Check whether we may negate an integer constant T without causing
867 overflow. */
869 bool
871 {
874 tree type;
884 {
889 }
890 else
895 }
897 /* Determine whether an expression T can be cheaply negated using
898 the function negate_expr. */
900 static bool
902 {
903 tree type;
912 {
917 /* Check that -CST will not overflow type. */
931 /* -(A + B) -> (-B) - A. */
936 /* -(A + B) -> (-A) - B. */
940 /* We can't turn -(A-B) into B-A when we honor signed zeros. */
949 /* Fall through. */
958 /* Negate -((double)float) as (double)(-float). */
960 {
964 }
968 /* Negate -f(x) as f(-x). */
974 /* Optimize -((int)x >> 31) into (unsigned)x >> 31. */
976 {
982 }
987 }
989 }
991 /* Given T, an expression, return the negation of T. Allow for T to be
992 null, in which case return null. */
994 static tree
996 {
997 tree type;
998 tree tem;
1007 {
1018 /* Two's complement FP formats, such as c4x, may overflow. */
1024 {
1033 }
1041 {
1042 /* -(A + B) -> (-B) - A. */
1046 {
1051 }
1053 /* -(A + B) -> (-A) - B. */
1055 {
1060 }
1061 }
1065 /* - (A - B) -> B - A */
1078 /* Fall through. */
1082 {
1095 }
1099 /* Convert -((double)float) into (double)(-float). */
1101 {
1105 }
1109 /* Negate -f(x) as f(-x). */
1112 {
1119 }
1123 /* Optimize -((int)x >> 31) into (unsigned)x >> 31. */
1125 {
1130 {
1137 }
1138 }
1143 }
1147 }
1148 \f
1149 /* Split a tree IN into a constant, literal and variable parts that could be
1150 combined with CODE to make IN. "constant" means an expression with
1151 TREE_CONSTANT but that isn't an actual constant. CODE must be a
1152 commutative arithmetic operation. Store the constant part into *CONP,
1153 the literal in *LITP and return the variable part. If a part isn't
1154 present, set it to null. If the tree does not decompose in this way,
1155 return the entire tree as the variable part and the other parts as null.
1157 If CODE is PLUS_EXPR we also split trees that use MINUS_EXPR. In that
1158 case, we negate an operand that was subtracted. Except if it is a
1159 literal for which we use *MINUS_LITP instead.
1161 If NEGATE_P is true, we are negating all of IN, again except a literal
1162 for which we use *MINUS_LITP instead.
1164 If IN is itself a literal or constant, return it as appropriate.
1166 Note that we do not guarantee that any of the three values will be the
1167 same type as IN, but they will have the same signedness and mode. */
1169 static tree
1172 {
1179 /* Strip any conversions that don't change the machine mode or signedness. */
1186 /* We can associate addition and subtraction together (even
1187 though the C standard doesn't say so) for integers because
1188 the value is not affected. For reals, the value might be
1189 affected, so we can't. */
1192 {
1198 /* First see if either of the operands is a literal, then a constant. */
1209 /* If we haven't dealt with either operand, this is not a case we can
1210 decompose. Otherwise, VAR is either of the ones remaining, if any. */
1215 else
1218 /* Now do any needed negations. */
1225 }
1228 else
1232 {
1239 }
1242 }
1244 /* Re-associate trees split by the above function. T1 and T2 are either
1245 expressions to associate or null. Return the new expression, if any. If
1246 we build an operation, do it in TYPE and with CODE. */
1248 static tree
1250 {
1256 /* If either input is CODE, a PLUS_EXPR, or a MINUS_EXPR, don't
1257 try to fold this since we will have infinite recursion. But do
1258 deal with any NEGATE_EXPRs. */
1261 {
1263 {
1272 }
1274 {
1277 }
1281 }
1285 }
1286 \f
1287 /* Combine two integer constants ARG1 and ARG2 under operation CODE
1288 to produce a new constant.
1290 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1292 tree
1294 {
1298 HOST_WIDE_INT hi;
1300 HOST_WIDE_INT garbageh;
1301 tree t;
1304 int is_sizetype
1315 {
1331 /* It's unclear from the C standard whether shifts can overflow.
1332 The following code ignores overflow; perhaps a C standard
1333 interpretation ruling is needed. */
1363 /* This is a shortcut for a common special case. */
1368 {
1374 }
1376 /* ... fall through ... */
1380 {
1383 }
1386 {
1389 }
1396 /* This is a shortcut for a common special case. */
1401 {
1406 }
1408 /* ... fall through ... */
1424 else
1430 else
1436 }
1441 {
1442 /* Propagate overflow flags ourselves. */
1445 {
1449 }
1451 {
1454 }
1455 }
1456 else
1464 }
1466 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1467 constant. We assume ARG1 and ARG2 have the same data type, or at least
1468 are the same kind of constant and the same machine mode.
1470 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1472 static tree
1474 {
1482 {
1484 REAL_VALUE_TYPE d1;
1485 REAL_VALUE_TYPE d2;
1486 REAL_VALUE_TYPE value;
1495 /* Don't perform operation if we honor signaling NaNs and
1496 either operand is a NaN. */
1501 /* Don't perform operation if it would raise a division
1502 by zero exception. */
1508 /* If either operand is a NaN, just return it. Otherwise, set up
1509 for floating-point trap; we return an overflow. */
1525 }
1527 {
1533 tree t;
1536 {
1556 notrunc),
1562 notrunc));
1566 {
1567 tree magsquared
1571 notrunc);
1574 const_binop
1579 notrunc),
1581 notrunc),
1582 notrunc),
1584 const_binop
1589 notrunc),
1591 notrunc),
1592 notrunc),
1594 }
1599 }
1601 }
1603 }
1605 /* Create a size type INT_CST node with NUMBER sign extended. KIND
1606 indicates which particular sizetype to create. */
1608 tree
1610 {
1612 }
1613 \f
1614 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
1615 is a tree code. The type of the result is taken from the operands.
1616 Both must be the same type integer type and it must be a size type.
1617 If the operands are constant, so is the result. */
1619 tree
1621 {
1627 /* Handle the special case of two integer constants faster. */
1629 {
1630 /* And some specific cases even faster than that. */
1639 /* Handle general case of two integer constants. */
1641 }
1647 }
1649 /* Given two values, either both of sizetype or both of bitsizetype,
1650 compute the difference between the two values. Return the value
1651 in signed type corresponding to the type of the operands. */
1653 tree
1655 {
1657 tree ctype;
1662 /* If the type is already signed, just do the simple thing. */
1668 /* If either operand is not a constant, do the conversions to the signed
1669 type and subtract. The hardware will do the right thing with any
1670 overflow in the subtraction. */
1675 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
1676 Otherwise, subtract the other way, convert to CTYPE (we know that can't
1677 overflow) and negate (which can't either). Special-case a result
1678 of zero while we're here. */
1683 else
1687 }
1688 \f
1689 /* Construct a vector of zero elements of vector type TYPE. */
1691 static tree
1693 {
1704 }
1707 /* Attempt to fold type conversion operation CODE of expression ARG1 to
1708 type TYPE. If no simplification can be done return NULL_TREE. */
1710 static tree
1712 {
1714 tree t;
1720 {
1722 {
1723 /* If we would build a constant wider than GCC supports,
1724 leave the conversion unfolded. */
1728 /* Given an integer constant, make new constant with new type,
1729 appropriately sign-extended or truncated. */
1734 /* Don't set the overflow when
1735 converting a pointer */
1743 }
1745 {
1746 /* The following code implements the floating point to integer
1747 conversion rules required by the Java Language Specification,
1748 that IEEE NaNs are mapped to zero and values that overflow
1749 the target precision saturate, i.e. values greater than
1750 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
1751 are mapped to INT_MIN. These semantics are allowed by the
1752 C and C++ standards that simply state that the behavior of
1753 FP-to-integer conversion is unspecified upon overflow. */
1756 REAL_VALUE_TYPE r;
1760 {
1779 }
1781 /* If R is NaN, return zero and show we have an overflow. */
1783 {
1787 }
1789 /* See if R is less than the lower bound or greater than the
1790 upper bound. */
1793 {
1797 {
1801 }
1802 }
1805 {
1808 {
1811 {
1815 }
1816 }
1817 }
1827 }
1828 }
1830 {
1834 {
1836 {
1837 /* We make a copy of ARG1 so that we don't modify an
1838 existing constant tree. */
1842 }
1852 }
1853 }
1855 }
1857 /* Convert expression ARG to type TYPE. Used by the middle-end for
1858 simple conversions in preference to calling the front-end's convert. */
1860 tree
1862 {
1864 tree tem;
1880 {
1885 {
1889 }
1894 {
1897 }
1904 {
1908 }
1910 {
1914 }
1917 {
1933 }
1937 {
1946 {
1950 {
1954 }
1962 }
1966 }
1981 }
1982 }
1983 \f
1984 /* Return an expr equal to X but certainly not valid as an lvalue. */
1986 tree
1988 {
1989 /* While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to
1990 us. */
1994 /* We only need to wrap lvalue tree codes. */
1996 {
2030 /* Assume the worst for front-end tree codes. */
2034 }
2036 }
2038 /* Nonzero means lvalues are limited to those valid in pedantic ANSI C.
2039 Zero means allow extended lvalues. */
2043 /* When pedantic, return an expr equal to X but certainly not valid as a
2044 pedantic lvalue. Otherwise, return X. */
2046 static tree
2048 {
2051 else
2053 }
2054 \f
2055 /* Given a tree comparison code, return the code that is the logical inverse
2056 of the given code. It is not safe to do this for floating-point
2057 comparisons, except for NE_EXPR and EQ_EXPR, so we receive a machine mode
2058 as well: if reversing the comparison is unsafe, return ERROR_MARK. */
2060 static enum tree_code
2062 {
2067 {
2098 }
2099 }
2101 /* Similar, but return the comparison that results if the operands are
2102 swapped. This is safe for floating-point. */
2104 enum tree_code
2106 {
2108 {
2122 }
2123 }
2126 /* Convert a comparison tree code from an enum tree_code representation
2127 into a compcode bit-based encoding. This function is the inverse of
2128 compcode_to_comparison. */
2130 static enum comparison_code
2132 {
2134 {
2165 }
2166 }
2168 /* Convert a compcode bit-based encoding of a comparison operator back
2169 to GCC's enum tree_code representation. This function is the
2170 inverse of comparison_to_compcode. */
2172 static enum tree_code
2174 {
2176 {
2207 }
2208 }
2210 /* Return a tree for the comparison which is the combination of
2211 doing the AND or OR (depending on CODE) of the two operations LCODE
2212 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2213 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2214 if this makes the transformation invalid. */
2216 tree
2220 {
2227 {
2238 }
2241 {
2242 /* Eliminate unordered comparisons, as well as LTGT and ORD
2243 which are not used unless the mode has NaNs. */
2249 }
2251 {
2252 /* Check that the original operation and the optimized ones will trap
2253 under the same condition. */
2264 /* In a short-circuited boolean expression the LHS might be
2265 such that the RHS, if evaluated, will never trap. For
2266 example, in ORD (x, y) && (x < y), we evaluate the RHS only
2267 if neither x nor y is NaN. (This is a mixed blessing: for
2268 example, the expression above will never trap, hence
2269 optimizing it to x < y would be invalid). */
2274 /* If the comparison was short-circuited, and only the RHS
2275 trapped, we may now generate a spurious trap. */
2280 /* If we changed the conditions that cause a trap, we lose. */
2283 }
2289 else
2292 }
2294 /* Return nonzero if CODE is a tree code that represents a truth value. */
2296 static int
2298 {
2303 }
2304 \f
2305 /* Return nonzero if two operands (typically of the same tree node)
2306 are necessarily equal. If either argument has side-effects this
2307 function returns zero. FLAGS modifies behavior as follows:
2309 If OEP_ONLY_CONST is set, only return nonzero for constants.
2310 This function tests whether the operands are indistinguishable;
2311 it does not test whether they are equal using C's == operation.
2312 The distinction is important for IEEE floating point, because
2313 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
2314 (2) two NaNs may be indistinguishable, but NaN!=NaN.
2316 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
2317 even though it may hold multiple values during a function.
2318 This is because a GCC tree node guarantees that nothing else is
2319 executed between the evaluation of its "operands" (which may often
2320 be evaluated in arbitrary order). Hence if the operands themselves
2321 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
2322 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
2323 unset means assuming isochronic (or instantaneous) tree equivalence.
2324 Unless comparing arbitrary expression trees, such as from different
2325 statements, this flag can usually be left unset.
2327 If OEP_PURE_SAME is set, then pure functions with identical arguments
2328 are considered the same. It is used when the caller has other ways
2329 to ensure that global memory is unchanged in between. */
2331 int
2333 {
2334 /* If either is ERROR_MARK, they aren't equal. */
2338 /* If both types don't have the same signedness, then we can't consider
2339 them equal. We must check this before the STRIP_NOPS calls
2340 because they may change the signedness of the arguments. */
2348 /* This is needed for conversions and for COMPONENT_REF.
2349 Might as well play it safe and always test this. */
2355 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
2356 We don't care about side effects in that case because the SAVE_EXPR
2357 takes care of that for us. In all other cases, two expressions are
2358 equal if they have no side effects. If we have two identical
2359 expressions with side effects that should be treated the same due
2360 to the only side effects being identical SAVE_EXPR's, that will
2361 be detected in the recursive calls below. */
2367 /* Next handle constant cases, those for which we can return 1 even
2368 if ONLY_CONST is set. */
2371 {
2384 {
2394 {
2396 flags))
2400 }
2403 }
2407 flags)
2409 flags));
2422 }
2427 /* Define macros to test an operand from arg0 and arg1 for equality and a
2428 variant that allows null and views null as being different from any
2429 non-null value. In the latter case, if either is null, the both
2430 must be; otherwise, do the normal comparison. */
2431 #define OP_SAME(N) operand_equal_p (TREE_OPERAND (arg0, N), \
2432 TREE_OPERAND (arg1, N), flags)
2434 #define OP_SAME_WITH_NULL(N) \
2435 ((!TREE_OPERAND (arg0, N) || !TREE_OPERAND (arg1, N)) \
2436 ? TREE_OPERAND (arg0, N) == TREE_OPERAND (arg1, N) : OP_SAME (N))
2439 {
2441 /* Two conversions are equal only if signedness and modes match. */
2443 {
2456 }
2466 /* For commutative ops, allow the other order. */
2474 /* If either of the pointer (or reference) expressions we are
2475 dereferencing contain a side effect, these cannot be equal. */
2481 {
2491 /* Operands 2 and 3 may be null. */
2498 /* Handle operand 2 the same as for ARRAY_REF. */
2506 }
2510 {
2525 /* Otherwise take into account this is a commutative operation. */
2532 /* If the CALL_EXPRs call different functions, then they
2533 clearly can not be equal. */
2537 {
2541 else
2545 }
2547 /* Now see if all the arguments are the same. operand_equal_p
2548 does not handle TREE_LIST, so we walk the operands here
2549 feeding them to operand_equal_p. */
2553 {
2555 flags))
2560 }
2562 /* If we get here and both argument lists are exhausted
2563 then the CALL_EXPRs are equal. */
2568 }
2571 /* Consider __builtin_sqrt equal to sqrt. */
2579 }
2581 #undef OP_SAME
2582 #undef OP_SAME_WITH_NULL
2583 }
2584 \f
2585 /* Similar to operand_equal_p, but see if ARG0 might have been made by
2586 shorten_compare from ARG1 when ARG1 was being compared with OTHER.
2588 When in doubt, return 0. */
2590 static int
2592 {
2604 /* Discard any conversions that don't change the modes of ARG0 and ARG1
2605 and see if the inner values are the same. This removes any
2606 signedness comparison, which doesn't matter here. */
2613 /* Duplicate what shorten_compare does to ARG1 and see if that gives the
2614 actual comparison operand, ARG0.
2616 First throw away any conversions to wider types
2617 already present in the operands. */
2626 {
2629 /* Make sure shorter operand is extended the right way
2630 to match the longer operand. */
2636 }
2639 }
2640 \f
2641 /* See if ARG is an expression that is either a comparison or is performing
2642 arithmetic on comparisons. The comparisons must only be comparing
2643 two different values, which will be stored in *CVAL1 and *CVAL2; if
2644 they are nonzero it means that some operands have already been found.
2645 No variables may be used anywhere else in the expression except in the
2646 comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around
2647 the expression and save_expr needs to be called with CVAL1 and CVAL2.
2649 If this is true, return 1. Otherwise, return zero. */
2651 static int
2653 {
2657 /* We can handle some of the tcc_expression cases here. */
2667 {
2668 /* If we've already found a CVAL1 or CVAL2, this expression is
2669 two complex to handle. */
2675 }
2678 {
2701 /* First see if we can handle the first operand, then the second. For
2702 the second operand, we know *CVAL1 can't be zero. It must be that
2703 one side of the comparison is each of the values; test for the
2704 case where this isn't true by failing if the two operands
2705 are the same. */
2714 ;
2718 ;
2719 else
2723 ;
2727 ;
2728 else
2735 }
2736 }
2737 \f
2738 /* ARG is a tree that is known to contain just arithmetic operations and
2739 comparisons. Evaluate the operations in the tree substituting NEW0 for
2740 any occurrence of OLD0 as an operand of a comparison and likewise for
2741 NEW1 and OLD1. */
2743 static tree
2745 {
2750 /* We can handle some of the tcc_expression cases here. */
2758 {
2773 {
2790 }
2791 /* Fall through - ??? */
2794 {
2798 /* We need to check both for exact equality and tree equality. The
2799 former will be true if the operand has a side-effect. In that
2800 case, we know the operand occurred exactly once. */
2813 }
2817 }
2818 }
2819 \f
2820 /* Return a tree for the case when the result of an expression is RESULT
2821 converted to TYPE and OMITTED was previously an operand of the expression
2822 but is now not needed (e.g., we folded OMITTED * 0).
2824 If OMITTED has side effects, we must evaluate it. Otherwise, just do
2825 the conversion of RESULT to TYPE. */
2827 tree
2829 {
2836 }
2838 /* Similar, but call pedantic_non_lvalue instead of non_lvalue. */
2840 static tree
2842 {
2849 }
2851 /* Return a tree for the case when the result of an expression is RESULT
2852 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
2853 of the expression but are now not needed.
2855 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
2856 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
2857 evaluated before OMITTED2. Otherwise, if neither has side effects,
2858 just do the conversion of RESULT to TYPE. */
2860 tree
2862 {
2871 }
2873 \f
2874 /* Return a simplified tree node for the truth-negation of ARG. This
2875 never alters ARG itself. We assume that ARG is an operation that
2876 returns a truth value (0 or 1).
2878 FIXME: one would think we would fold the result, but it causes
2879 problems with the dominator optimizer. */
2880 tree
2882 {
2889 /* If this is a comparison, we can simply invert it, except for
2890 floating-point non-equality comparisons, in which case we just
2891 enclose a TRUTH_NOT_EXPR around what we have. */
2894 {
2897 && flag_trapping_math
2901 else
2902 {
2907 else
2910 }
2911 }
2914 {
2930 /* Here we can invert either operand. We invert the first operand
2931 unless the second operand is a TRUTH_NOT_EXPR in which case our
2932 result is the XOR of the first operand with the inside of the
2933 negation of the second operand. */
2938 else
2992 }
2995 }
2997 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
2998 operands are another bit-wise operation with a common input. If so,
2999 distribute the bit operations to save an operation and possibly two if
3000 constants are involved. For example, convert
3001 (A | B) & (A | C) into A | (B & C)
3002 Further simplification will occur if B and C are constants.
3004 If this optimization cannot be done, 0 will be returned. */
3006 static tree
3008 {
3009 tree common;
3019 {
3023 }
3025 {
3029 }
3031 {
3035 }
3037 {
3041 }
3042 else
3047 }
3048 \f
3049 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
3050 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero. */
3052 static tree
3055 {
3062 }
3064 /* Optimize a bit-field compare.
3066 There are two cases: First is a compare against a constant and the
3067 second is a comparison of two items where the fields are at the same
3068 bit position relative to the start of a chunk (byte, halfword, word)
3069 large enough to contain it. In these cases we can avoid the shift
3070 implicit in bitfield extractions.
3072 For constants, we emit a compare of the shifted constant with the
3073 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
3074 compared. For two fields at the same position, we do the ANDs with the
3075 similar mask and compare the result of the ANDs.
3077 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
3078 COMPARE_TYPE is the type of the comparison, and LHS and RHS
3079 are the left and right operands of the comparison, respectively.
3081 If the optimization described above can be done, we return the resulting
3082 tree. Otherwise we return zero. */
3084 static tree
3087 {
3096 tree mask;
3097 tree offset;
3099 /* Get all the information about the extractions being done. If the bit size
3100 if the same as the size of the underlying object, we aren't doing an
3101 extraction at all and so can do nothing. We also don't want to
3102 do anything if the inner expression is a PLACEHOLDER_EXPR since we
3103 then will no longer be able to replace it. */
3111 {
3112 /* If this is not a constant, we can only do something if bit positions,
3113 sizes, and signedness are the same. */
3121 }
3123 /* See if we can find a mode to refer to this field. We should be able to,
3124 but fail if we can't. */
3133 /* Set signed and unsigned types of the precision of this mode for the
3134 shifts below. */
3138 /* Compute the bit position and size for the new reference and our offset
3139 within it. If the new reference is the same size as the original, we
3140 won't optimize anything, so return zero. */
3150 /* Make the mask to be used against the extracted field. */
3159 /* If not comparing with constant, just rework the comparison
3160 and return. */
3165 mask),
3169 mask));
3171 /* Otherwise, we are handling the constant case. See if the constant is too
3172 big for the field. Warn and return a tree of for 0 (false) if so. We do
3173 this not only for its own sake, but to avoid having to test for this
3174 error case below. If we didn't, we might generate wrong code.
3176 For unsigned fields, the constant shifted right by the field length should
3177 be all zero. For signed fields, the high-order bits should agree with
3178 the sign bit. */
3181 {
3185 {
3189 }
3190 }
3191 else
3192 {
3196 {
3200 }
3201 }
3203 /* Single-bit compares should always be against zero. */
3205 {
3208 }
3210 /* Make a new bitfield reference, shift the constant over the
3211 appropriate number of bits and mask it with the computed mask
3212 (in case this was a signed field). If we changed it, make a new one. */
3215 {
3218 }
3228 rhs);
3229 }
3230 \f
3231 /* Subroutine for fold_truthop: decode a field reference.
3233 If EXP is a comparison reference, we return the innermost reference.
3235 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
3236 set to the starting bit number.
3238 If the innermost field can be completely contained in a mode-sized
3239 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
3241 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
3242 otherwise it is not changed.
3244 *PUNSIGNEDP is set to the signedness of the field.
3246 *PMASK is set to the mask used. This is either contained in a
3247 BIT_AND_EXPR or derived from the width of the field.
3249 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
3251 Return 0 if this is not a component reference or is one that we can't
3252 do anything with. */
3254 static tree
3259 {
3263 tree unsigned_type;
3266 /* All the optimizations using this function assume integer fields.
3267 There are problems with FP fields since the type_for_size call
3268 below can fail for, e.g., XFmode. */
3272 /* We are interested in the bare arrangement of bits, so strip everything
3273 that doesn't affect the machine mode. However, record the type of the
3274 outermost expression if it may matter below. */
3282 {
3288 }
3297 /* If the number of bits in the reference is the same as the bitsize of
3298 the outer type, then the outer type gives the signedness. Otherwise
3299 (in case of a small bitfield) the signedness is unchanged. */
3303 /* Compute the mask to access the bitfield. */
3313 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
3321 }
3323 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
3324 bit positions. */
3326 static int
3328 {
3331 tree tmask;
3336 return
3343 }
3345 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
3346 represents the sign bit of EXP's type. If EXP represents a sign
3347 or zero extension, also test VAL against the unextended type.
3348 The return value is the (sub)expression whose sign bit is VAL,
3349 or NULL_TREE otherwise. */
3351 static tree
3353 {
3357 tree t;
3359 /* Tree EXP must have an integral type. */
3364 /* Tree VAL must be an integer constant. */
3371 {
3378 }
3379 else
3380 {
3387 }
3389 /* We mask off those bits beyond TREE_TYPE (exp) so that we can
3390 treat VAL as if it were unsigned. */
3395 /* Handle extension from a narrower type. */
3401 }
3403 /* Subroutine for fold_truthop: determine if an operand is simple enough
3404 to be evaluated unconditionally. */
3406 static int
3408 {
3409 /* Strip any conversions that don't change the machine mode. */
3418 /* Don't regard global variables as simple. They may be
3419 allocated in ways unknown to the compiler (shared memory,
3420 #pragma weak, etc). */
3423 /* Loading a static variable is unduly expensive, but global
3424 registers aren't expensive. */
3426 }
3427 \f
3428 /* The following functions are subroutines to fold_range_test and allow it to
3429 try to change a logical combination of comparisons into a range test.
3431 For example, both
3432 X == 2 || X == 3 || X == 4 || X == 5
3433 and
3434 X >= 2 && X <= 5
3435 are converted to
3436 (unsigned) (X - 2) <= 3
3438 We describe each set of comparisons as being either inside or outside
3439 a range, using a variable named like IN_P, and then describe the
3440 range with a lower and upper bound. If one of the bounds is omitted,
3441 it represents either the highest or lowest value of the type.
3443 In the comments below, we represent a range by two numbers in brackets
3444 preceded by a "+" to designate being inside that range, or a "-" to
3445 designate being outside that range, so the condition can be inverted by
3446 flipping the prefix. An omitted bound is represented by a "-". For
3447 example, "- [-, 10]" means being outside the range starting at the lowest
3448 possible value and ending at 10, in other words, being greater than 10.
3449 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
3450 always false.
3452 We set up things so that the missing bounds are handled in a consistent
3453 manner so neither a missing bound nor "true" and "false" need to be
3454 handled using a special case. */
3456 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
3457 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
3458 and UPPER1_P are nonzero if the respective argument is an upper bound
3459 and zero for a lower. TYPE, if nonzero, is the type of the result; it
3460 must be specified for a comparison. ARG1 will be converted to ARG0's
3461 type if both are specified. */
3463 static tree
3466 {
3467 tree tem;
3471 /* If neither arg represents infinity, do the normal operation.
3472 Else, if not a comparison, return infinity. Else handle the special
3473 comparison rules. Note that most of the cases below won't occur, but
3474 are handled for consistency. */
3477 {
3482 }
3487 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
3488 for neither. In real maths, we cannot assume open ended ranges are
3489 the same. But, this is computer arithmetic, where numbers are finite.
3490 We can therefore make the transformation of any unbounded range with
3491 the value Z, Z being greater than any representable number. This permits
3492 us to treat unbounded ranges as equal. */
3496 {
3517 }
3520 }
3521 \f
3522 /* Given EXP, a logical expression, set the range it is testing into
3523 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
3524 actually being tested. *PLOW and *PHIGH will be made of the same type
3525 as the returned expression. If EXP is not a comparison, we will most
3526 likely not be returning a useful value and range. */
3528 static tree
3530 {
3537 /* Start with simply saying "EXP != 0" and then look at the code of EXP
3538 and see if we can refine the range. Some of the cases below may not
3539 happen, but it doesn't seem worth worrying about this. We "continue"
3540 the outer loop when we've changed something; otherwise we "break"
3541 the switch, which will "break" the while. */
3547 {
3552 {
3564 }
3567 {
3574 /* We can only do something if the range is testing for zero
3575 and if the second operand is an integer constant. Note that
3576 saying something is "in" the range we make is done by
3577 complementing IN_P since it will set in the initial case of
3578 being not equal to zero; "out" is leaving it alone. */
3585 {
3606 }
3608 /* If this is an unsigned comparison, we also know that EXP is
3609 greater than or equal to zero. We base the range tests we make
3610 on that fact, so we record it here so we can parse existing
3611 range tests. We test arg0_type since often the return type
3612 of, e.g. EQ_EXPR, is boolean. */
3614 {
3618 NULL_TREE))
3623 /* If the high bound is missing, but we have a nonzero low
3624 bound, reverse the range so it goes from zero to the low bound
3625 minus 1. */
3627 {
3632 }
3633 }
3639 /* (-x) IN [a,b] -> x in [-b, -a] */
3651 /* ~ X -> -X - 1 */
3660 /* If EXP is signed, any overflow in the computation is undefined,
3661 so we don't worry about it so long as our computations on
3662 the bounds don't overflow. For unsigned, overflow is defined
3663 and this is exactly the right thing. */
3672 /* Check for an unsigned range which has wrapped around the maximum
3673 value thus making n_high < n_low, and normalize it. */
3675 {
3681 /* If the range is of the form +/- [ x+1, x ], we won't
3682 be able to normalize it. But then, it represents the
3683 whole range or the empty set, so make it
3684 +/- [ -, - ]. */
3688 else
3690 }
3691 else
3715 /* If we're converting arg0 from an unsigned type, to exp,
3716 a signed type, we will be doing the comparison as unsigned.
3717 The tests above have already verified that LOW and HIGH
3718 are both positive.
3720 So we have to ensure that we will handle large unsigned
3721 values the same way that the current signed bounds treat
3722 negative values. */
3725 {
3726 tree high_positive;
3730 /* A range without an upper bound is, naturally, unbounded.
3731 Since convert would have cropped a very large value, use
3732 the max value for the destination type. */
3733 high_positive
3740 high_positive),
3742 integer_one_node)));
3744 /* If the low bound is specified, "and" the range with the
3745 range for which the original unsigned value will be
3746 positive. */
3748 {
3752 integer_zero_node),
3753 high_positive))
3757 }
3758 else
3759 {
3760 /* Otherwise, "or" the range with the range of the input
3761 that will be interpreted as negative. */
3765 integer_zero_node),
3766 high_positive))
3770 }
3771 }
3779 }
3782 }
3784 /* If EXP is a constant, we can evaluate whether this is true or false. */
3786 {
3793 }
3797 }
3798 \f
3799 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
3800 type, TYPE, return an expression to test if EXP is in (or out of, depending
3801 on IN_P) the range. Return 0 if the test couldn't be created. */
3803 static tree
3805 {
3807 tree value;
3810 {
3816 }
3831 {
3833 {
3837 }
3839 }
3841 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
3843 {
3845 HOST_WIDE_INT hi;
3850 {
3853 }
3854 else
3855 {
3858 }
3861 {
3863 {
3866 }
3869 }
3870 }
3874 {
3877 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
3878 for the type in question, as we rely on this here. */
3880 {
3891 {
3897 }
3901 }
3902 }
3908 value);
3911 }
3912 \f
3913 /* Given two ranges, see if we can merge them into one. Return 1 if we
3914 can, 0 if we can't. Set the output range into the specified parameters. */
3916 static int
3919 {
3923 tree tem;
3933 /* Make range 0 be the range that starts first, or ends last if they
3934 start at the same value. Swap them if it isn't. */
3937 || (lowequal
3940 {
3944 }
3946 /* Now flag two cases, whether the ranges are disjoint or whether the
3947 second range is totally subsumed in the first. Note that the tests
3948 below are simplified by the ones above. */
3954 /* We now have four cases, depending on whether we are including or
3955 excluding the two ranges. */
3957 {
3958 /* If they don't overlap, the result is false. If the second range
3959 is a subset it is the result. Otherwise, the range is from the start
3960 of the second to the end of the first. */
3965 else
3967 }
3970 {
3971 /* If they don't overlap, the result is the first range. If they are
3972 equal, the result is false. If the second range is a subset of the
3973 first, and the ranges begin at the same place, we go from just after
3974 the end of the first range to the end of the second. If the second
3975 range is not a subset of the first, or if it is a subset and both
3976 ranges end at the same place, the range starts at the start of the
3977 first range and ends just before the second range.
3978 Otherwise, we can't describe this as a single range. */
3984 {
3988 }
3990 {
3994 }
3995 else
3997 }
4000 {
4001 /* If they don't overlap, the result is the second range. If the second
4002 is a subset of the first, the result is false. Otherwise,
4003 the range starts just after the first range and ends at the
4004 end of the second. */
4009 else
4010 {
4014 }
4015 }
4017 else
4018 {
4019 /* The case where we are excluding both ranges. Here the complex case
4020 is if they don't overlap. In that case, the only time we have a
4021 range is if they are adjacent. If the second is a subset of the
4022 first, the result is the first. Otherwise, the range to exclude
4023 starts at the beginning of the first range and ends at the end of the
4024 second. */
4026 {
4033 else
4034 {
4035 /* Canonicalize - [min, x] into - [-, x]. */
4038 {
4043 /* FALLTHROUGH */
4057 }
4059 /* Canonicalize - [x, max] into - [x, -]. */
4062 {
4067 /* FALLTHROUGH */
4083 }
4085 /* The ranges might be also adjacent between the maximum and
4086 minimum values of the given type. For
4087 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
4088 return + [x + 1, y - 1]. */
4090 {
4099 }
4100 else
4102 }
4103 }
4106 else
4108 }
4112 }
4113 \f
4115 /* Subroutine of fold, looking inside expressions of the form
4116 A op B ? A : C, where ARG0, ARG1 and ARG2 are the three operands
4117 of the COND_EXPR. This function is being used also to optimize
4118 A op B ? C : A, by reversing the comparison first.
4120 Return a folded expression whose code is not a COND_EXPR
4121 anymore, or NULL_TREE if no folding opportunity is found. */
4123 static tree
4125 {
4130 tree tem;
4135 /* If we have A op 0 ? A : -A, consider applying the following
4136 transformations:
4138 A == 0? A : -A same as -A
4139 A != 0? A : -A same as A
4140 A >= 0? A : -A same as abs (A)
4141 A > 0? A : -A same as abs (A)
4142 A <= 0? A : -A same as -abs (A)
4143 A < 0? A : -A same as -abs (A)
4145 None of these transformations work for modes with signed
4146 zeros. If A is +/-0, the first two transformations will
4147 change the sign of the result (from +0 to -0, or vice
4148 versa). The last four will fix the sign of the result,
4149 even though the original expressions could be positive or
4150 negative, depending on the sign of A.
4152 Note that all these transformations are correct if A is
4153 NaN, since the two alternatives (A and -A) are also NaNs. */
4160 {
4172 /* Fall through. */
4194 }
4196 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
4197 A == 0 ? A : 0 is always 0 unless A is -0. Note that
4198 both transformations are correct when A is NaN: A != 0
4199 is then true, and A == 0 is false. */
4202 {
4207 }
4209 /* Try some transformations of A op B ? A : B.
4211 A == B? A : B same as B
4212 A != B? A : B same as A
4213 A >= B? A : B same as max (A, B)
4214 A > B? A : B same as max (B, A)
4215 A <= B? A : B same as min (A, B)
4216 A < B? A : B same as min (B, A)
4218 As above, these transformations don't work in the presence
4219 of signed zeros. For example, if A and B are zeros of
4220 opposite sign, the first two transformations will change
4221 the sign of the result. In the last four, the original
4222 expressions give different results for (A=+0, B=-0) and
4223 (A=-0, B=+0), but the transformed expressions do not.
4225 The first two transformations are correct if either A or B
4226 is a NaN. In the first transformation, the condition will
4227 be false, and B will indeed be chosen. In the case of the
4228 second transformation, the condition A != B will be true,
4229 and A will be chosen.
4231 The conversions to max() and min() are not correct if B is
4232 a number and A is not. The conditions in the original
4233 expressions will be false, so all four give B. The min()
4234 and max() versions would give a NaN instead. */
4236 {
4241 /* Avoid adding NOP_EXPRs in case this is an lvalue. */
4243 {
4247 }
4250 {
4259 /* In C++ a ?: expression can be an lvalue, so put the
4260 operand which will be used if they are equal first
4261 so that we can convert this back to the
4262 corresponding COND_EXPR. */
4264 {
4271 }
4278 {
4285 }
4298 }
4299 }
4301 /* If this is A op C1 ? A : C2 with C1 and C2 constant integers,
4302 we might still be able to simplify this. For example,
4303 if C1 is one less or one more than C2, this might have started
4304 out as a MIN or MAX and been transformed by this function.
4305 Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE. */
4311 {
4313 /* We can replace A with C1 in this case. */
4318 /* If C1 is C2 + 1, this is min(A, C2). */
4320 OEP_ONLY_CONST)
4324 OEP_ONLY_CONST))
4330 /* If C1 is C2 - 1, this is min(A, C2). */
4332 OEP_ONLY_CONST)
4336 OEP_ONLY_CONST))
4342 /* If C1 is C2 - 1, this is max(A, C2). */
4344 OEP_ONLY_CONST)
4348 OEP_ONLY_CONST))
4354 /* If C1 is C2 + 1, this is max(A, C2). */
4356 OEP_ONLY_CONST)
4360 OEP_ONLY_CONST))
4368 }
4371 }
4374 \f
4375 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
4376 #define LOGICAL_OP_NON_SHORT_CIRCUIT (BRANCH_COST >= 2)
4377 #endif
4379 /* EXP is some logical combination of boolean tests. See if we can
4380 merge it into some range test. Return the new tree if so. */
4382 static tree
4384 {
4391 tree tem;
4393 /* If this is an OR operation, invert both sides; we will invert
4394 again at the end. */
4398 /* If both expressions are the same, if we can merge the ranges, and we
4399 can build the range test, return it or it inverted. If one of the
4400 ranges is always true or always false, consider it to be the same
4401 expression as the other. */
4411 /* On machines where the branch cost is expensive, if this is a
4412 short-circuited branch and the underlying object on both sides
4413 is the same, make a non-short-circuit operation. */
4419 {
4420 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
4421 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
4422 which cases we can't do this. */
4431 {
4443 }
4444 }
4447 }
4448 \f
4449 /* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
4450 bit value. Arrange things so the extra bits will be set to zero if and
4451 only if C is signed-extended to its full width. If MASK is nonzero,
4452 it is an INTEGER_CST that should be AND'ed with the extra bits. */
4454 static tree
4456 {
4459 tree temp;
4464 /* We work by getting just the sign bit into the low-order bit, then
4465 into the high-order bit, then sign-extend. We then XOR that value
4466 with C. */
4470 /* We must use a signed type in order to get an arithmetic right shift.
4471 However, we must also avoid introducing accidental overflows, so that
4472 a subsequent call to integer_zerop will work. Hence we must
4473 do the type conversion here. At this point, the constant is either
4474 zero or one, and the conversion to a signed type can never overflow.
4475 We could get an overflow if this conversion is done anywhere else. */
4484 /* If necessary, convert the type back to match the type of C. */
4489 }
4490 \f
4491 /* Find ways of folding logical expressions of LHS and RHS:
4492 Try to merge two comparisons to the same innermost item.
4493 Look for range tests like "ch >= '0' && ch <= '9'".
4494 Look for combinations of simple terms on machines with expensive branches
4495 and evaluate the RHS unconditionally.
4497 For example, if we have p->a == 2 && p->b == 4 and we can make an
4498 object large enough to span both A and B, we can do this with a comparison
4499 against the object ANDed with the a mask.
4501 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
4502 operations to do this with one comparison.
4504 We check for both normal comparisons and the BIT_AND_EXPRs made this by
4505 function and the one above.
4507 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
4508 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
4510 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
4511 two operands.
4513 We return the simplified tree or 0 if no optimization is possible. */
4515 static tree
4517 {
4518 /* If this is the "or" of two comparisons, we can do something if
4519 the comparisons are NE_EXPR. If this is the "and", we can do something
4520 if the comparisons are EQ_EXPR. I.e.,
4521 (a->b == 2 && a->c == 4) can become (a->new == NEW).
4523 WANTED_CODE is this operation code. For single bit fields, we can
4524 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
4525 comparison for one-bit fields. */
4545 /* Start by getting the comparison codes. Fail if anything is volatile.
4546 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
4547 it were surrounded with a NE_EXPR. */
4556 {
4560 }
4563 {
4567 }
4578 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
4581 {
4582 tree result;
4585 {
4590 }
4593 {
4599 }
4600 }
4605 /* If the RHS can be evaluated unconditionally and its operands are
4606 simple, it wins to evaluate the RHS unconditionally on machines
4607 with expensive branches. In this case, this isn't a comparison
4608 that can be merged. Avoid doing this if the RHS is a floating-point
4609 comparison since those can trap. */
4615 {
4616 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
4626 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
4638 }
4640 /* See if the comparisons can be merged. Then get all the parameters for
4641 each side. */
4665 /* It must be true that the inner operation on the lhs of each
4666 comparison must be the same if we are to be able to do anything.
4667 Then see if we have constants. If not, the same must be true for
4668 the rhs's. */
4679 else
4682 /* If either comparison code is not correct for our logical operation,
4683 fail. However, we can convert a one-bit comparison against zero into
4684 the opposite comparison against that bit being set in the field. */
4688 {
4690 {
4691 /* Make the left operand unsigned, since we are only interested
4692 in the value of one bit. Otherwise we are doing the wrong
4693 thing below. */
4696 }
4697 else
4699 }
4701 /* This is analogous to the code for l_const above. */
4703 {
4705 {
4708 }
4709 else
4711 }
4713 /* After this point all optimizations will generate bit-field
4714 references, which we might not want. */
4718 /* See if we can find a mode that contains both fields being compared on
4719 the left. If we can't, fail. Otherwise, update all constants and masks
4720 to be relative to a field of that size. */
4725 volatilep);
4735 {
4738 }
4746 {
4754 {
4758 }
4759 }
4761 {
4769 {
4773 }
4774 }
4776 /* If the right sides are not constant, do the same for it. Also,
4777 disallow this optimization if a size or signedness mismatch occurs
4778 between the left and right sides. */
4780 {
4783 /* Make sure the two fields on the right
4784 correspond to the left without being swapped. */
4792 volatilep);
4802 {
4805 }
4812 /* Make a mask that corresponds to both fields being compared.
4813 Do this for both items being compared. If the operands are the
4814 same size and the bits being compared are in the same position
4815 then we can do this by masking both and comparing the masked
4816 results. */
4820 {
4832 }
4834 /* There is still another way we can do something: If both pairs of
4835 fields being compared are adjacent, we may be able to make a wider
4836 field containing them both.
4838 Note that we still must mask the lhs/rhs expressions. Furthermore,
4839 the mask must be shifted to account for the shift done by
4840 make_bit_field_ref. */
4845 {
4846 tree type;
4858 /* Convert to the smaller type before masking out unwanted bits. */
4861 {
4863 {
4867 }
4869 {
4873 }
4874 }
4883 }
4886 }
4888 /* Handle the case of comparisons with constants. If there is something in
4889 common between the masks, those bits of the constants must be the same.
4890 If not, the condition is always false. Test for this to avoid generating
4891 incorrect code below. */
4896 {
4898 {
4901 }
4902 else
4903 {
4906 }
4907 }
4909 /* Construct the expression we will return. First get the component
4910 reference we will make. Unless the mask is all ones the width of
4911 that field, perform the mask operation. Then compare with the
4912 merged constant. */
4922 }
4923 \f
4924 /* Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a
4925 constant. */
4927 static tree
4929 {
4934 tree minmax_const;
4936 tree inner;
4946 /* If something does not permit us to optimize, return the original tree. */
4954 /* Now handle all the various comparison codes. We only handle EQ_EXPR
4955 and GT_EXPR, doing the rest with recursive calls using logical
4956 simplifications. */
4958 {
4960 return
4964 return
4966 optimize_minmax_comparison
4968 optimize_minmax_comparison
4973 /* MAX (X, 0) == 0 -> X <= 0 */
4977 /* MAX (X, 0) == 5 -> X == 5 */
4981 /* MAX (X, 0) == -1 -> false */
4985 /* MIN (X, 0) == 0 -> X >= 0 */
4989 /* MIN (X, 0) == 5 -> false */
4992 else
4993 /* MIN (X, 0) == -1 -> X == -1 */
4998 /* MAX (X, 0) > 0 -> X > 0
4999 MAX (X, 0) > 5 -> X > 5 */
5003 /* MAX (X, 0) > -1 -> true */
5007 /* MIN (X, 0) > 0 -> false
5008 MIN (X, 0) > 5 -> false */
5011 else
5012 /* MIN (X, 0) > -1 -> X > -1 */
5017 }
5018 }
5019 \f
5020 /* T is an integer expression that is being multiplied, divided, or taken a
5021 modulus (CODE says which and what kind of divide or modulus) by a
5022 constant C. See if we can eliminate that operation by folding it with
5023 other operations already in T. WIDE_TYPE, if non-null, is a type that
5024 should be used for the computation if wider than our type.
5026 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
5027 (X * 2) + (Y * 4). We must, however, be assured that either the original
5028 expression would not overflow or that overflow is undefined for the type
5029 in the language in question.
5031 We also canonicalize (X + 7) * 4 into X * 4 + 28 in the hope that either
5032 the machine has a multiply-accumulate insn or that this is part of an
5033 addressing calculation.
5035 If we return a non-null expression, it is an equivalent form of the
5036 original computation, but need not be in the original type. */
5038 static tree
5040 {
5041 /* To avoid exponential search depth, refuse to allow recursion past
5042 three levels. Beyond that (1) it's highly unlikely that we'll find
5043 something interesting and (2) we've probably processed it before
5044 when we built the inner expression. */
5047 tree ret;
5052 depth++;
5054 depth--;
5057 }
5059 static tree
5061 {
5071 /* Don't deal with constants of zero here; they confuse the code below. */
5081 /* Note that we need not handle conditional operations here since fold
5082 already handles those cases. So just do arithmetic here. */
5084 {
5086 /* For a constant, we can always simplify if we are a multiply
5087 or (for divide and modulus) if it is a multiple of our constant. */
5095 /* If op0 is an expression ... */
5100 /* ... and is unsigned, and its type is smaller than ctype,
5101 then we cannot pass through as widening. */
5107 /* ... or this is a truncation (t is narrower than op0),
5108 then we cannot pass through this narrowing. */
5111 /* ... or signedness changes for division or modulus,
5112 then we cannot pass through this conversion. */
5118 /* Pass the constant down and see if we can make a simplification. If
5119 we can, replace this expression with the inner simplification for
5120 possible later conversion to our or some other type. */
5125 code == MULT_EXPR
5131 /* If widening the type changes it from signed to unsigned, then we
5132 must avoid building ABS_EXPR itself as unsigned. */
5134 {
5137 {
5140 }
5142 }
5143 /* FALLTHROUGH */
5150 /* If widening the type changes the signedness, then we can't perform
5151 this optimization as that changes the result. */
5155 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
5158 {
5164 }
5168 /* If the second operand is constant, this is a multiplication
5169 or floor division, by a power of two, so we can treat it that
5170 way unless the multiplier or divisor overflows. Signed
5171 left-shift overflow is implementation-defined rather than
5172 undefined in C90, so do not convert signed left shift into
5173 multiplication. */
5176 /* const_binop may not detect overflow correctly,
5177 so check for it explicitly here. */
5182 size_one_node,
5192 /* See if we can eliminate the operation on both sides. If we can, we
5193 can return a new PLUS or MINUS. If we can't, the only remaining
5194 cases where we can do anything are if the second operand is a
5195 constant. */
5200 /* If not multiplication, we can only do this if both operands
5201 are divisible by c. */
5207 /* If this was a subtraction, negate OP1 and set it to be an addition.
5208 This simplifies the logic below. */
5215 /* If either OP1 or C are negative, this optimization is not safe for
5216 some of the division and remainder types while for others we need
5217 to change the code. */
5219 {
5227 }
5229 /* If it's a multiply or a division/modulus operation of a multiple
5230 of our constant, do the operation and verify it doesn't overflow. */
5233 {
5236 /* We allow the constant to overflow with wrapping semantics. */
5240 }
5241 else
5244 /* If we have an unsigned type is not a sizetype, we cannot widen
5245 the operation since it will change the result if the original
5246 computation overflowed. */
5252 /* If we were able to eliminate our operation from the first side,
5253 apply our operation to the second side and reform the PLUS. */
5257 /* The last case is if we are a multiply. In that case, we can
5258 apply the distributive law to commute the multiply and addition
5259 if the multiplication of the constants doesn't overflow. */
5265 op1));
5270 /* We have a special case here if we are doing something like
5271 (C * 8) % 4 since we know that's zero. */
5278 /* ... fall through ... */
5282 /* If we can extract our operation from the LHS, do so and return a
5283 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
5284 do something only if the second operand is a constant. */
5296 /* If these are the same operation types, we can associate them
5297 assuming no overflow. */
5304 /* If these operations "cancel" each other, we have the main
5305 optimizations of this pass, which occur when either constant is a
5306 multiple of the other, in which case we replace this with either an
5307 operation or CODE or TCODE.
5309 If we have an unsigned type that is not a sizetype, we cannot do
5310 this since it will change the result if the original computation
5311 overflowed. */
5314 && ! flag_wrapv
5319 {
5330 }
5335 }
5338 }
5339 \f
5340 /* Return a node which has the indicated constant VALUE (either 0 or
5341 1), and is of the indicated TYPE. */
5343 tree
5345 {
5353 else
5355 }
5357 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
5358 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
5359 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
5360 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
5361 COND is the first argument to CODE; otherwise (as in the example
5362 given here), it is the second argument. TYPE is the type of the
5363 original expression. Return NULL_TREE if no simplification is
5364 possible. */
5366 static tree
5369 {
5374 /* This transformation is only worthwhile if we don't have to wrap
5375 arg in a SAVE_EXPR, and the operation can be simplified on atleast
5376 one of the branches once its pushed inside the COND_EXPR. */
5381 {
5385 /* If this operand throws an expression, then it does not make
5386 sense to try to perform a logical or arithmetic operation
5387 involving it. */
5392 }
5393 else
5394 {
5399 }
5410 }
5412 \f
5413 /* Subroutine of fold() that checks for the addition of +/- 0.0.
5415 If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type
5416 TYPE, X + ADDEND is the same as X. If NEGATE, return true if X -
5417 ADDEND is the same as X.
5419 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
5420 and finite. The problematic cases are when X is zero, and its mode
5421 has signed zeros. In the case of rounding towards -infinity,
5422 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
5423 modes, X + 0 is not the same as X because -0 + 0 is 0. */
5425 static bool
5427 {
5431 /* Don't allow the fold with -fsignaling-nans. */
5435 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
5439 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
5444 /* The mode has signed zeros, and we have to honor their sign.
5445 In this situation, there is only one case we can return true for.
5446 X - 0 is the same as X unless rounding towards -infinity is
5447 supported. */
5449 }
5451 /* Subroutine of fold() that checks comparisons of built-in math
5452 functions against real constants.
5454 FCODE is the DECL_FUNCTION_CODE of the built-in, CODE is the comparison
5455 operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE
5456 is the type of the result and ARG0 and ARG1 are the operands of the
5457 comparison. ARG1 must be a TREE_REAL_CST.
5459 The function returns the constant folded tree if a simplification
5460 can be made, and NULL_TREE otherwise. */
5462 static tree
5465 {
5466 REAL_VALUE_TYPE c;
5469 {
5475 {
5476 /* sqrt(x) < y is always false, if y is negative. */
5480 /* sqrt(x) > y is always true, if y is negative and we
5481 don't care about NaNs, i.e. negative values of x. */
5485 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
5488 }
5490 {
5491 REAL_VALUE_TYPE c2;
5497 {
5498 /* sqrt(x) > y is x == +Inf, when y is very large. */
5503 /* sqrt(x) > y is always false, when y is very large
5504 and we don't care about infinities. */
5506 }
5508 /* sqrt(x) > c is the same as x > c*c. */
5511 }
5513 {
5514 REAL_VALUE_TYPE c2;
5520 {
5521 /* sqrt(x) < y is always true, when y is a very large
5522 value and we don't care about NaNs or Infinities. */
5526 /* sqrt(x) < y is x != +Inf when y is very large and we
5527 don't care about NaNs. */
5532 /* sqrt(x) < y is x >= 0 when y is very large and we
5533 don't care about Infinities. */
5538 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
5547 dconst0))),
5550 c2)))));
5551 }
5553 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
5558 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
5561 {
5566 dconst0))),
5569 c2)))));
5570 }
5571 }
5572 }
5575 }
5577 /* Subroutine of fold() that optimizes comparisons against Infinities,
5578 either +Inf or -Inf.
5580 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5581 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5582 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5584 The function returns the constant folded tree if a simplification
5585 can be made, and NULL_TREE otherwise. */
5587 static tree
5589 {
5591 REAL_VALUE_TYPE max;
5592 tree temp;
5597 /* For negative infinity swap the sense of the comparison. */
5603 {
5605 /* x > +Inf is always false, if with ignore sNANs. */
5611 /* x <= +Inf is always true, if we don't case about NaNs. */
5615 /* x <= +Inf is the same as x == x, i.e. isfinite(x). */
5618 {
5621 }
5626 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
5632 /* x < +Inf is always equal to x <= DBL_MAX. */
5638 /* x != +Inf is always equal to !(x > DBL_MAX). */
5644 /* The transformation below creates non-gimple code and thus is
5645 not appropriate if we are in gimple form. */
5655 }
5658 }
5660 /* Subroutine of fold() that optimizes comparisons of a division by
5661 a nonzero integer constant against an integer constant, i.e.
5662 X/C1 op C2.
5664 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5665 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5666 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5668 The function returns the constant folded tree if a simplification
5669 can be made, and NULL_TREE otherwise. */
5671 static tree
5673 {
5678 HOST_WIDE_INT hpart;
5681 /* We have to do this the hard way to detect unsigned overflow.
5682 prod = int_const_binop (MULT_EXPR, arg01, arg1, 0); */
5691 {
5695 /* Likewise hi = int_const_binop (PLUS_EXPR, prod, tmp, 0). */
5704 }
5706 {
5709 {
5727 }
5728 }
5729 else
5730 {
5731 /* A negative divisor reverses the relational operators. */
5736 {
5754 }
5755 }
5758 {
5799 }
5802 }
5805 /* If CODE with arguments ARG0 and ARG1 represents a single bit
5806 equality/inequality test, then return a simplified form of
5807 the test using shifts and logical operations. Otherwise return
5808 NULL. TYPE is the desired result type. */
5810 tree
5812 tree result_type)
5813 {
5814 /* If this is a TRUTH_NOT_EXPR, it may have a single bit test inside
5815 operand 0. */
5817 {
5822 /* Extract the arguments of the EQ/NE. */
5826 /* This requires us to invert the code. */
5828 }
5830 /* If this is testing a single bit, we can optimize the test. */
5834 {
5841 tree arg00;
5843 /* If we have (A & C) != 0 where C is the sign bit of A, convert
5844 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
5847 /* This is only a win if casting to a signed type is cheap,
5848 i.e. when arg00's type is not a partial mode. */
5851 {
5856 }
5858 /* Otherwise we have (A & C) != 0 where C is a single bit,
5859 convert that into ((A >> C2) & 1). Where C2 = log2(C).
5860 Similarly for (A & C) == 0. */
5862 /* If INNER is a right shift of a constant and it plus BITNUM does
5863 not overflow, adjust BITNUM and INNER. */
5870 {
5873 }
5875 /* If we are going to be able to omit the AND below, we must do our
5876 operations as unsigned. If we must use the AND, we have a choice.
5877 Normally unsigned is faster, but for some machines signed is. */
5878 #ifdef LOAD_EXTEND_OP
5880 #else
5882 #endif
5897 /* Put the AND last so it can combine with more things. */
5901 /* Make sure to return the proper type. */
5905 }
5907 }
5909 /* Check whether we are allowed to reorder operands arg0 and arg1,
5910 such that the evaluation of arg1 occurs before arg0. */
5912 static bool
5914 {
5921 }
5923 /* Test whether it is preferable two swap two operands, ARG0 and
5924 ARG1, for example because ARG0 is an integer constant and ARG1
5925 isn't. If REORDER is true, only recommend swapping if we can
5926 evaluate the operands in reverse order. */
5928 bool
5930 {
5966 /* It is preferable to swap two SSA_NAME to ensure a canonical form
5967 for commutative and comparison operators. Ensuring a canonical
5968 form allows the optimizers to find additional redundancies without
5969 having to explicitly check for both orderings. */
5976 }
5978 /* Fold comparison ARG0 CODE ARG1 (with result in TYPE), where
5979 ARG0 is extended to a wider type. */
5981 static tree
5983 {
5985 tree arg1_unw;
5998 /* If possible, express the comparison in the shorter mode. */
6010 /* If we are comparing with the integer that does not fit into the range
6011 of the shorter type, the result is known. */
6022 {
6049 }
6052 }
6054 /* Fold comparison ARG0 CODE ARG1 (with result in TYPE), where for
6055 ARG0 just the signedness is changed. */
6057 static tree
6060 {
6085 {
6092 }
6093 else
6097 }
6099 /* Tries to replace &a[idx] CODE s * delta with &a[idx CODE delta], if s is
6100 step of the array. TYPE is the type of the expression. ADDR is the address.
6101 MULT is the multiplicative expression. If the function succeeds, the new
6102 address expression is returned. Otherwise NULL_TREE is returned. */
6104 static tree
6106 {
6111 tree itype;
6117 {
6120 }
6122 {
6125 }
6126 else
6130 {
6132 {
6140 /* If the type sizes do not match, we might run into problems
6141 when one of them would overflow. */
6150 }
6154 }
6156 /* We found the suitable array reference. So copy everything up to it,
6157 and replace the index. */
6164 {
6168 }
6172 delta));
6175 }
6178 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
6179 means A >= Y && A != MAX, but in this case we know that
6180 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
6182 static tree
6184 {
6191 else
6200 {
6203 }
6205 {
6208 }
6209 else
6220 }
6222 /* Perform constant folding and related simplification of EXPR.
6223 The related simplifications include x*1 => x, x*0 => 0, etc.,
6224 and application of the associative law.
6225 NOP_EXPR conversions may be removed freely (as long as we
6226 are careful not to change the type of the overall expression).
6227 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
6228 but we can constant-fold them if they have constant operands. */
6230 #ifdef ENABLE_FOLD_CHECKING
6231 # define fold(x) fold_1 (x)
6233 static
6234 #endif
6235 tree
6237 {
6241 tree tem;
6246 /* WINS will be nonzero when the switch is done
6247 if all operands are constant. */
6250 /* Return right away if a constant. */
6255 {
6256 tree subop;
6258 /* Special case for conversion ops that can have fixed point args. */
6261 /* Don't use STRIP_NOPS, because signedness of argument type matters. */
6267 else
6272 /* Note that TREE_CONSTANT isn't enough:
6273 static var addresses are constant but we can't
6274 do arithmetic on them. */
6276 }
6278 {
6282 {
6284 tree subop;
6289 /* Strip any conversions that don't change the mode. This is
6290 safe for every expression, except for a comparison expression
6291 because its signedness is derived from its operands. So, in
6292 the latter case, only strip conversions that don't change the
6293 signedness.
6295 Note that this is done as an internal manipulation within the
6296 constant folder, in order to find the simplest representation
6297 of the arguments so that their form can be studied. In any
6298 cases, the appropriate type conversions should be put back in
6299 the tree that will get out of the constant folder. */
6302 else
6307 else
6312 /* Note that TREE_CONSTANT isn't enough:
6313 static var addresses are constant but we can't
6314 do arithmetic on them. */
6321 }
6322 }
6324 /* If this is a commutative operation, and ARG0 is a constant, move it
6325 to ARG1 to reduce the number of tests below. */
6331 /* Now WINS is set as described above,
6332 ARG0 is the first operand of EXPR,
6333 and ARG1 is the second operand (if it has more than one operand).
6335 First check for cases where an arithmetic operation is applied to a
6336 compound, conditional, or comparison operation. Push the arithmetic
6337 operation inside the compound or conditional to see if any folding
6338 can then be done. Convert comparison to conditional for this purpose.
6339 The also optimizes non-constant cases that used to be done in
6340 expand_expr.
6342 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
6343 one of the operands is a comparison and the other is a comparison, a
6344 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
6345 code below would make the expression more complex. Change it to a
6346 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
6347 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
6359 {
6370 }
6373 {
6378 {
6388 /* If this was a conversion, and all we did was to move into
6389 inside the COND_EXPR, bring it back out. But leave it if
6390 it is a conversion from integer to integer and the
6391 result precision is no wider than a word since such a
6392 conversion is cheap and may be optimized away by combine,
6393 while it couldn't if it were outside the COND_EXPR. Then return
6394 so we don't get into an infinite recursion loop taking the
6395 conversion out and then back in. */
6407 && (INTEGRAL_TYPE_P
6418 }
6420 {
6422 {
6426 }
6430 integer_one_node)),
6432 integer_zero_node))));
6433 }
6434 }
6445 {
6449 arg1)));
6457 {
6462 }
6465 {
6470 }
6471 }
6474 {
6488 /* Handle cases of two conversions in a row. */
6491 {
6510 /* In addition to the cases of two conversions in a row
6511 handled below, if we are converting something to its own
6512 type via an object of identical or wider precision, neither
6513 conversion is needed. */
6520 /* Likewise, if the intermediate and final types are either both
6521 float or both integer, we don't need the middle conversion if
6522 it is wider than the final type and doesn't change the signedness
6523 (for integers). Avoid this if the final type is a pointer
6524 since then we sometimes need the inner conversion. Likewise if
6525 the outer has a precision not equal to the size of its mode. */
6536 /* If we have a sign-extension of a zero-extended value, we can
6537 replace that by a single zero-extension. */
6544 /* Two conversions in a row are not needed unless:
6545 - some conversion is floating-point (overstrict for now), or
6546 - the intermediate type is narrower than both initial and
6547 final, or
6548 - the intermediate type and innermost type differ in signedness,
6549 and the outermost type is wider than the intermediate, or
6550 - the initial type is a pointer type and the precisions of the
6551 intermediate and final types differ, or
6552 - the final type is a pointer type and the precisions of the
6553 initial and intermediate types differ. */
6568 }
6572 /* Detect assigning a bitfield. */
6575 {
6576 /* Don't leave an assignment inside a conversion
6577 unless assigning a bitfield. */
6581 /* First do the assignment, then return converted constant. */
6586 }
6588 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
6589 constants (if x has signed type, the sign bit cannot be set
6590 in c). This folds extension into the BIT_AND_EXPR. */
6595 {
6605 <= HOST_BITS_PER_WIDE_INT
6607 {
6614 #ifdef LOAD_EXTEND_OP
6618 {
6622 }
6623 #endif
6624 }
6629 }
6631 /* Convert (T1)((T2)X op Y) into (T1)X op Y, for pointer types T1 and
6632 T2 being pointers to types of the same size. */
6637 {
6649 }
6663 {
6667 }
6672 {
6677 }
6690 /* Convert fabs((double)float) into (double)fabsf(float). */
6693 {
6698 targ0)));
6699 }
6732 /* A + (-B) -> A - B */
6735 /* (-A) + B -> B - A */
6740 {
6744 /* If we are adding two BIT_AND_EXPR's, both of which are and'ing
6745 with a constant, and the two constants have no bits in common,
6746 we should treat this as a BIT_IOR_EXPR since this may produce more
6747 simplifications. */
6755 {
6758 }
6760 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
6761 (plus (plus (mult) (mult)) (foo)) so that we can
6762 take advantage of the factoring cases below. */
6769 {
6775 else
6797 parg1)))));
6798 }
6801 {
6805 /* (A * C) + (B * C) -> (A+B) * C.
6806 We are most concerned about the case where C is a constant,
6807 but other combinations show up during loop reduction. Since
6808 it is not difficult, try all four possibilities. */
6825 /* No identical multiplicands; see if we can find a common
6826 power-of-two factor in non-power-of-two multiplies. This
6827 can help in multi-dimensional array access. */
6832 {
6837 /* Move min of absolute values to int11. */
6840 {
6844 }
6847 {
6853 }
6854 }
6861 same));
6862 }
6864 /* Try replacing &a[i1] + c * i2 with &a[i1 + i2], if c is step
6865 of the array. Loop optimizer sometimes produce this type of
6866 expressions. */
6869 {
6873 }
6876 {
6880 }
6881 }
6882 else
6883 {
6884 /* See if ARG1 is zero and X + ARG1 reduces to X. */
6888 /* Likewise if the operands are reversed. */
6892 /* Convert X + -C into X - C. */
6895 {
6901 }
6903 /* Convert x+x into x*2.0. */
6909 /* Convert x*c+x into x*(c+1). */
6915 {
6916 REAL_VALUE_TYPE c;
6922 }
6924 /* Convert x+x*c into x*(c+1). */
6930 {
6931 REAL_VALUE_TYPE c;
6937 }
6939 /* Convert x*c1+x*c2 into x*(c1+c2). */
6949 {
6958 }
6959 /* Convert a + (b*c + d*e) into (a + b*c) + d*e. */
6963 {
6968 {
6969 tree tree0;
6972 }
6973 }
6974 /* Convert (b*c + d*e) + a into b*c + (d*e +a). */
6978 {
6983 {
6984 tree tree0;
6987 }
6988 }
6989 }
6991 bit_rotate:
6992 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
6993 is a rotate of A by C1 bits. */
6994 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
6995 is a rotate of A by B bits. */
6996 {
7005 {
7024 {
7032 TYPE_PRECISION
7037 ? LROTATE_EXPR
7040 }
7042 {
7050 TYPE_PRECISION
7055 ? LROTATE_EXPR
7058 }
7059 }
7060 }
7062 associate:
7063 /* In most languages, can't associate operations on floats through
7064 parentheses. Rather than remember where the parentheses were, we
7065 don't associate floats at all, unless the user has specified
7066 -funsafe-math-optimizations. */
7070 {
7074 /* Split both trees into variables, constants, and literals. Then
7075 associate each group together, the constants with literals,
7076 then the result with variables. This increases the chances of
7077 literals being recombined later and of generating relocatable
7078 expressions for the sum of a constant and literal. */
7083 /* Only do something if we found more than two objects. Otherwise,
7084 nothing has changed and we risk infinite recursion. */
7089 {
7090 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
7099 /* Preserve the MINUS_EXPR if the negative part of the literal is
7100 greater than the positive part. Otherwise, the multiplicative
7101 folding code (i.e extract_muldiv) may be fooled in case
7102 unsigned constants are subtracted, like in the following
7103 example: ((X*2 + 4) - 8U)/2. */
7105 {
7109 {
7113 }
7114 else
7115 {
7119 }
7120 }
7122 {
7127 else
7128 {
7134 }
7135 }
7140 }
7141 }
7143 binary:
7147 {
7148 /* The return value should always have
7149 the same type as the original expression. */
7154 }
7158 /* A - (-B) -> A + B */
7161 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
7171 {
7177 /* Fold A - (A & B) into ~B & A. */
7180 {
7185 arg0));
7190 arg0));
7191 }
7193 /* Fold (A & ~B) - (A & B) into (A ^ B) - B, where B is
7194 any power of 2 minus 1. */
7199 {
7205 {
7209 }
7210 }
7211 }
7213 /* See if ARG1 is zero and X - ARG1 reduces to X. */
7217 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
7218 ARG0 is zero and X + ARG0 reduces to X, since that would mean
7219 (-ARG1 + ARG0) reduces to -ARG1. */
7223 /* Fold &x - &x. This can happen from &x.foo - &x.
7224 This is unsafe for certain floats even in non-IEEE formats.
7225 In IEEE, it is unsafe because it does wrong for NaNs.
7226 Also note that operand_equal_p is always false if an operand
7227 is volatile. */
7233 /* A - B -> A + (-B) if B is easily negatable. */
7236 /* Avoid this transformation if B is a positive REAL_CST. */
7242 /* Try folding difference of addresses. */
7243 {
7244 HOST_WIDE_INT diff;
7250 }
7252 /* Try replacing &a[i1] - c * i2 with &a[i1 - i2], if c is step
7253 of the array. Loop optimizer sometimes produce this type of
7254 expressions. */
7257 {
7261 }
7266 {
7267 /* (A * C) - (B * C) -> (A-B) * C. */
7275 /* (A * C1) - (A * C2) -> A * (C1-C2). */
7283 }
7288 /* (-A) * (-B) -> A * B */
7299 {
7305 /* (a * (1 << b)) is (a << b) */
7321 }
7322 else
7323 {
7324 /* Maybe fold x * 0 to 0. The expressions aren't the same
7325 when x is NaN, since x * 0 is also NaN. Nor are they the
7326 same in modes with signed zeros, since multiplying a
7327 negative value by 0 gives -0, not +0. */
7332 /* In IEEE floating point, x*1 is not equivalent to x for snans. */
7337 /* Transform x * -1.0 into -x. */
7342 /* Convert (C1/X)*C2 into (C1*C2)/X. */
7347 {
7353 }
7356 {
7360 /* Optimizations of root(...)*root(...). */
7362 {
7367 /* Optimize sqrt(x)*sqrt(x) as x. */
7373 /* Optimize root(x)*root(y) as root(x*y). */
7378 }
7380 /* Optimize expN(x)*expN(y) as expN(x+y). */
7382 {
7389 }
7391 /* Optimizations of pow(...)*pow(...). */
7395 {
7403 /* Optimize pow(x,y)*pow(z,y) as pow(x*z,y). */
7405 {
7410 arg01));
7412 }
7414 /* Optimize pow(x,y)*pow(x,z) as pow(x,y+z). */
7416 {
7421 arg));
7423 }
7424 }
7426 /* Optimize tan(x)*cos(x) as sin(x). */
7435 {
7441 }
7443 /* Optimize x*pow(x,c) as pow(x,c+1). */
7447 {
7454 {
7456 REAL_VALUE_TYPE c;
7465 }
7466 }
7468 /* Optimize pow(x,c)*x as pow(x,c+1). */
7472 {
7479 {
7481 REAL_VALUE_TYPE c;
7490 }
7491 }
7493 /* Optimize x*x as pow(x,2.0), which is expanded as x*x. */
7496 {
7500 {
7505 }
7506 }
7507 }
7508 }
7512 bit_ior:
7520 /* ~X | X is -1. */
7523 {
7527 }
7529 /* X | ~X is -1. */
7532 {
7536 }
7542 /* Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
7544 This results in more efficient code for machines without a NAND
7545 instruction. Combine will canonicalize to the first form
7546 which will allow use of NAND instructions provided by the
7547 backend if they exist. */
7550 {
7555 }
7557 /* See if this can be simplified into a rotate first. If that
7558 is unsuccessful continue in the association code. */
7569 /* ~X ^ X is -1. */
7572 {
7576 }
7578 /* X ^ ~X is -1. */
7581 {
7585 }
7587 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
7588 with a constant, and the two constants have no bits in common,
7589 we should treat this as a BIT_IOR_EXPR since this may produce more
7590 simplifications. */
7598 {
7601 }
7603 /* See if this can be simplified into a rotate first. If that
7604 is unsuccessful continue in the association code. */
7615 /* ~X & X is always zero. */
7620 /* X & ~X is always zero. */
7628 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
7631 {
7632 unsigned int prec
7639 }
7641 /* Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
7643 This results in more efficient code for machines without a NOR
7644 instruction. Combine will canonicalize to the first form
7645 which will allow use of NOR instructions provided by the
7646 backend if they exist. */
7649 {
7654 }
7659 /* Don't touch a floating-point divide by zero unless the mode
7660 of the constant can represent infinity. */
7666 /* (-A) / (-B) -> A / B */
7676 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
7681 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
7686 /* If ARG1 is a constant, we can convert this to a multiply by the
7687 reciprocal. This does not have the same rounding properties,
7688 so only do this if -funsafe-math-optimizations. We can actually
7689 always safely do it if ARG1 is a power of two, but it's hard to
7690 tell if it is or not in a portable manner. */
7692 {
7697 /* Find the reciprocal if optimizing and the result is exact. */
7699 {
7700 REAL_VALUE_TYPE r;
7703 {
7706 }
7707 }
7708 }
7709 /* Convert A/B/C to A/(B*C). */
7716 /* Convert A/(B/C) to (A/B)*C. */
7724 /* Convert C1/(X*C2) into (C1/C2)/X. */
7729 {
7735 }
7738 {
7740 /* Optimize x/expN(y) into x*expN(-y). */
7742 {
7749 }
7751 /* Optimize x/pow(y,z) into x*pow(y,-z). */
7755 {
7764 }
7765 }
7768 {
7772 /* Optimize sin(x)/cos(x) as tan(x). */
7778 {
7784 }
7786 /* Optimize cos(x)/sin(x) as 1.0/tan(x). */
7792 {
7796 {
7801 }
7802 }
7804 /* Optimize pow(x,c)/x as pow(x,c-1). */
7808 {
7814 {
7816 REAL_VALUE_TYPE c;
7825 }
7826 }
7827 }
7839 /* X / -1 is -X. */
7846 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
7847 operation, EXACT_DIV_EXPR.
7849 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
7850 At one time others generated faster code, it's not clear if they do
7851 after the last round to changes to the DIV code in expmed.c. */
7872 /* X % -1 is zero. */
7879 /* Optimize unsigned TRUNC_MOD_EXPR by a power of two into a
7880 BIT_AND_EXPR, i.e. "X % C" into "X & C2". */
7884 {
7886 tree mask;
7891 {
7895 }
7896 else
7897 {
7900 }
7905 }
7907 /* X % -C is the same as X % C. */
7912 && !flag_trapv
7913 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
7918 /* X % -Y is the same as X % Y. */
7940 /* Optimize -1 >> x for arithmetic right shifts. */
7943 /* ... fall through ... */
7946 shift:
7952 /* Since negative shift count is not well-defined,
7953 don't try to compute it in the compiler. */
7956 /* Rewrite an LROTATE_EXPR by a constant into an
7957 RROTATE_EXPR by a new constant. */
7959 {
7965 }
7967 /* If we have a rotate of a bit operation with the rotate count and
7968 the second operand of the bit operation both constant,
7969 permute the two operations. */
7981 /* Two consecutive rotates adding up to the width of the mode can
7982 be ignored. */
8013 /* The argument to invert_truthvalue must have Boolean type. */
8017 /* Note that the operand of this must be an int
8018 and its values must be 0 or 1.
8019 ("true" is a fixed value perhaps depending on the language,
8020 but we don't handle values other than 1 correctly yet.) */
8022 /* Avoid infinite recursion. */
8024 {
8029 }
8033 /* Note that the operands of this must be ints
8034 and their values must be 0 or 1.
8035 ("true" is a fixed value perhaps depending on the language.) */
8036 /* If first arg is constant zero, return it. */
8040 /* If either arg is constant true, drop it. */
8044 /* Preserve sequence points. */
8047 /* If second arg is constant zero, result is zero, but first arg
8048 must be evaluated. */
8051 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
8052 case will be handled here. */
8056 /* !X && X is always false. */
8060 /* X && !X is always false. */
8065 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
8066 means A >= Y && A != MAX, but in this case we know that
8067 A < X <= MAX. */
8071 {
8079 }
8081 truth_andor:
8082 /* We only do these simplifications if we are optimizing. */
8086 /* Check for things like (A || B) && (A || C). We can convert this
8087 to A || (B && C). Note that either operator can be any of the four
8088 truth and/or operations and the transformation will still be
8089 valid. Also note that we only care about order for the
8090 ANDIF and ORIF operators. If B contains side effects, this
8091 might change the truth-value of A. */
8098 {
8118 /* This case if tricky because we must either have commutative
8119 operators or else A10 must not have side-effects. */
8125 a01));
8126 }
8128 /* See if we can build a range comparison. */
8132 /* Check for the possibility of merging component references. If our
8133 lhs is another similar operation, try to merge its rhs with our
8134 rhs. Then try to merge our lhs and rhs. */
8146 /* Note that the operands of this must be ints
8147 and their values must be 0 or true.
8148 ("true" is a fixed value perhaps depending on the language.) */
8149 /* If first arg is constant true, return it. */
8153 /* If either arg is constant zero, drop it. */
8157 /* Preserve sequence points. */
8160 /* If second arg is constant true, result is true, but we must
8161 evaluate first arg. */
8164 /* Likewise for first arg, but note this only occurs here for
8165 TRUTH_OR_EXPR. */
8169 /* !X || X is always true. */
8173 /* X || !X is always true. */
8181 /* If the second arg is constant zero, drop it. */
8184 /* If the second arg is constant true, this is a logical inversion. */
8187 /* Identical arguments cancel to zero. */
8191 /* !X ^ X is always true. */
8196 /* X ^ !X is always true. */
8209 /* If one arg is a real or integer constant, put it last. */
8213 /* If this is an equality comparison of the address of a non-weak
8214 object against zero, then we know the result. */
8222 /* If this is an equality comparison of the address of two non-weak,
8223 unaliased symbols neither of which are extern (since we do not
8224 have access to attributes for externs), then we know the result. */
8240 type);
8243 {
8251 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
8256 /* (-a) CMP (-b) -> b CMP a */
8263 {
8264 REAL_VALUE_TYPE cst;
8267 /* (-a) CMP CST -> a swap(CMP) (-CST) */
8269 return
8275 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
8276 /* a CMP (-0) -> a CMP 0 */
8281 /* x != NaN is always true, other ops are always false. */
8284 {
8287 }
8289 /* Fold comparisons against infinity. */
8291 {
8295 }
8296 }
8298 /* If this is a comparison of a real constant with a PLUS_EXPR
8299 or a MINUS_EXPR of a real constant, we can convert it into a
8300 comparison with a revised real constant as long as no overflow
8301 occurs when unsafe_math_optimizations are enabled. */
8313 /* Likewise, we can simplify a comparison of a real constant with
8314 a MINUS_EXPR whose first operand is also a real constant, i.e.
8315 (c1 - x) < c2 becomes x > c1-c2. */
8326 /* Fold comparisons against built-in math functions. */
8328 && flag_unsafe_math_optimizations
8330 {
8334 {
8338 }
8339 }
8340 }
8342 /* Convert foo++ == CONST into ++foo == CONST + INCR. */
8346 /* This optimization is invalid for ordered comparisons
8347 if CONST+INCR overflows or if foo+incr might overflow.
8348 This optimization is invalid for floating point due to rounding.
8349 For pointer types we assume overflow doesn't happen. */
8353 {
8357 {
8363 }
8364 else
8365 {
8371 }
8374 /* If VAROP is a reference to a bitfield, we must mask
8375 the constant by the width of the field. */
8380 {
8385 /* First check whether the comparison would come out
8386 always the same. If we don't do that we would
8387 change the meaning with the masking. */
8401 }
8404 }
8406 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
8407 This transformation affects the cases which are handled in later
8408 optimizations involving comparisons with non-negative constants. */
8412 {
8414 {
8425 }
8426 }
8428 /* Comparisons with the highest or lowest possible integer of
8429 the specified size will have known values.
8431 This is quite similar to fold_relational_hi_lo; however, my
8432 attempts to share the code have been nothing but trouble.
8433 I give up for now. */
8434 {
8442 {
8449 {
8452 }
8453 else
8454 {
8457 }
8462 {
8475 /* The GE_EXPR and LT_EXPR cases above are not normally
8476 reached because of previous transformations. */
8480 }
8484 {
8493 }
8497 {
8512 }
8516 {
8525 }
8531 /* signed_type does not work on pointer types. */
8533 {
8534 /* The following case also applies to X < signed_max+1
8535 and X >= signed_max+1 because previous transformations. */
8537 {
8541 return fold
8545 }
8546 }
8547 }
8548 }
8550 /* If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
8551 a MINUS_EXPR of a constant, we can convert it into a comparison with
8552 a revised constant as long as no overflow occurs. */
8564 /* Similarly for a NEGATE_EXPR. */
8573 /* If we have X - Y == 0, we can convert that to X == Y and similarly
8574 for !=. Don't do this for ordered comparisons due to overflow. */
8582 {
8583 /* If we are widening one operand of an integer comparison,
8584 see if the other operand is similarly being widened. Perhaps we
8585 can do the comparison in the narrower type. */
8590 /* Or if we are changing signedness. */
8594 }
8596 /* If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a
8597 constant, we can simplify it. */
8604 /* If we are comparing an ABS_EXPR with a constant, we can
8605 convert all the cases into explicit comparisons, but they may
8606 well not be faster than doing the ABS and one comparison.
8607 But ABS (X) <= C is a range comparison, which becomes a subtraction
8608 and a comparison, and is probably faster. */
8621 /* If this is an EQ or NE comparison with zero and ARG0 is
8622 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
8623 two operations, but the latter can be done in one less insn
8624 on machines that have only two-operand insns or on which a
8625 constant cannot be the first operand. */
8628 {
8633 return
8639 integer_one_node)),
8640 arg1));
8643 return
8649 integer_one_node)),
8650 arg1));
8651 }
8653 /* If this is an NE or EQ comparison of zero against the result of a
8654 signed MOD operation whose second operand is a power of 2, make
8655 the MOD operation unsigned since it is simpler and equivalent. */
8664 {
8674 }
8676 /* If this is an NE comparison of zero with an AND of one, remove the
8677 comparison since the AND will give the correct value. */
8683 /* If we have (A & C) == C where C is a power of 2, convert this into
8684 (A & C) != 0. Similarly for NE_EXPR. */
8691 integer_zero_node)));
8693 /* If we have (A & C) != 0 or (A & C) == 0 and C is a power of
8694 2, then fold the expression into shifts and logical operations. */
8699 /* If we have (A & C) == D where D & ~C != 0, convert this into 0.
8700 Similarly for NE_EXPR. */
8705 {
8714 }
8716 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
8717 Similarly for NE_EXPR. */
8722 {
8729 }
8731 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
8732 and similarly for >= into !=. */
8748 return
8756 /* Simplify comparison of something with itself. (For IEEE
8757 floating-point, we can only do some of these simplifications.) */
8759 {
8761 {
8776 /* For NE, we can only do this simplification if integer
8777 or we don't honor IEEE floating point NaNs. */
8781 /* ... fall through ... */
8787 }
8788 }
8790 /* If we are comparing an expression that just has comparisons
8791 of two integer values, arithmetic expressions of those comparisons,
8792 and constants, we can simplify it. There are only three cases
8793 to check: the two values can either be equal, the first can be
8794 greater, or the second can be greater. Fold the expression for
8795 those three values. Since each value must be 0 or 1, we have
8796 eight possibilities, each of which corresponds to the constant 0
8797 or 1 or one of the six possible comparisons.
8799 This handles common cases like (a > b) == 0 but also handles
8800 expressions like ((x > y) - (y > x)) > 0, which supposedly
8801 occur in macroized code. */
8804 {
8809 /* Don't handle degenerate cases here; they should already
8810 have been handled anyway. */
8819 {
8823 /* We can't just pass T to eval_subst in case cval1 or cval2
8824 was the same as ARG1. */
8826 tree high_result
8830 arg1));
8831 tree equal_result
8835 arg1));
8836 tree low_result
8840 arg1));
8842 /* All three of these results should be 0 or 1. Confirm they
8843 are. Then use those values to select the proper code
8844 to use. */
8852 {
8853 /* Make a 3-bit mask with the high-order bit being the
8854 value for `>', the next for '=', and the low for '<'. */
8858 {
8860 /* Always false. */
8881 /* Always true. */
8883 }
8888 else
8890 }
8891 }
8892 }
8894 /* If this is a comparison of a field, we may be able to simplify it. */
8899 /* Handle the constant case even without -O
8900 to make sure the warnings are given. */
8902 {
8906 }
8908 /* If this is a comparison of complex values and either or both sides
8909 are a COMPLEX_EXPR or COMPLEX_CST, it is best to split up the
8910 comparisons and join them with a TRUTH_ANDIF_EXPR or TRUTH_ORIF_EXPR.
8911 This may prevent needless evaluations. */
8918 {
8931 type,
8934 }
8936 /* Optimize comparisons of strlen vs zero to a compare of the
8937 first character of the string vs zero. To wit,
8938 strlen(ptr) == 0 => *ptr == 0
8939 strlen(ptr) != 0 => *ptr != 0
8940 Other cases should reduce to one of these two (or a constant)
8941 due to the return value of strlen being unsigned. */
8945 {
8947 tree arglist;
8960 integer_zero_node)));
8961 }
8963 /* We can fold X/C1 op C2 where C1 and C2 are integer constants
8964 into a single range test. */
8971 {
8975 }
8995 {
8999 }
9001 /* If the first operand is NaN, the result is constant. */
9005 {
9007 ? integer_zero_node
9010 }
9012 /* If the second operand is NaN, the result is constant. */
9016 {
9018 ? integer_zero_node
9021 }
9023 /* Simplify unordered comparison of something with itself. */
9029 && !flag_trapping_math
9033 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
9034 {
9045 }
9050 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
9051 so all simple results must be passed through pedantic_non_lvalue. */
9053 {
9055 /* Only optimize constant conditions when the selected branch
9056 has the same type as the COND_EXPR. This avoids optimizing
9057 away "c ? x : throw", where the throw has a void type. */
9062 }
9066 /* If we have A op B ? A : C, we may be able to convert this to a
9067 simpler expression, depending on the operation and the values
9068 of B and C. Signed zeros prevent all of these transformations,
9069 for reasons given above each one.
9071 Also try swapping the arguments and inverting the conditional. */
9076 {
9082 }
9089 {
9092 {
9098 }
9099 }
9101 /* If the second operand is simpler than the third, swap them
9102 since that produces better jump optimization results. */
9105 {
9106 /* See if this can be inverted. If it can't, possibly because
9107 it was a floating-point inequality comparison, don't do
9108 anything. */
9114 }
9116 /* Convert A ? 1 : 0 to simply A. */
9119 /* If we try to convert TREE_OPERAND (t, 0) to our type, the
9120 call to fold will try to move the conversion inside
9121 a COND, which will recurse. In that case, the COND_EXPR
9122 is probably the best choice, so leave it alone. */
9126 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
9127 over COND_EXPR in cases such as floating point comparisons. */
9134 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
9142 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
9143 already handled above. */
9148 {
9157 }
9159 /* A & N ? N : 0 is simply A & N if N is a power of two. This
9160 is probably obsolete because the first operand should be a
9161 truth value (that's why we have the two cases above), but let's
9162 leave it in until we can confirm this for all front-ends. */
9173 /* Convert A ? B : 0 into A && B if A and B are truth values. */
9179 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
9183 {
9184 /* Only perform transformation if ARG0 is easily inverted. */
9188 }
9190 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
9194 {
9195 /* Only perform transformation if ARG0 is easily inverted. */
9200 }
9202 /* Convert A ? 1 : B into A || B if A and B are truth values. */
9212 /* When pedantic, a compound expression can be neither an lvalue
9213 nor an integer constant expression. */
9216 /* Don't let (0, 0) be null pointer constant. */
9259 /* Check for a built-in function. */
9264 {
9268 }
9274 }
9276 #ifdef ENABLE_FOLD_CHECKING
9277 #undef fold
9283 /* When --enable-checking=fold, compute a digest of expr before
9284 and after actual fold call to see if fold did not accidentally
9285 change original expr. */
9287 tree
9289 {
9290 tree ret;
9293 htab_t ht;
9312 }
9314 void
9316 {
9319 htab_t ht;
9329 }
9331 static void
9333 {
9335 }
9337 static void
9339 {
9357 {
9358 /* Allow DECL_ASSEMBLER_NAME to be modified. */
9362 }
9366 {
9367 /* Allow these fields to be modified. */
9374 }
9381 {
9384 {
9398 }
9402 {
9413 }
9447 {
9450 }
9460 }
9461 }
9463 #endif
9465 /* Perform constant folding and related simplification of initializer
9466 expression EXPR. This behaves identically to "fold" but ignores
9467 potential run-time traps and exceptions that fold must preserve. */
9469 tree
9471 {
9475 tree result;
9488 }
9490 /* Determine if first argument is a multiple of second argument. Return 0 if
9491 it is not, or we cannot easily determined it to be.
9493 An example of the sort of thing we care about (at this point; this routine
9494 could surely be made more general, and expanded to do what the *_DIV_EXPR's
9495 fold cases do now) is discovering that
9497 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9499 is a multiple of
9501 SAVE_EXPR (J * 8)
9503 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
9505 This code also handles discovering that
9507 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9509 is a multiple of 8 so we don't have to worry about dealing with a
9510 possible remainder.
9512 Note that we *look* inside a SAVE_EXPR only to determine how it was
9513 calculated; it is not safe for fold to do much of anything else with the
9514 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
9515 at run time. For example, the latter example above *cannot* be implemented
9516 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
9517 evaluation time of the original SAVE_EXPR is not necessarily the same at
9518 the time the new expression is evaluated. The only optimization of this
9519 sort that would be valid is changing
9521 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
9523 divided by 8 to
9525 SAVE_EXPR (I) * SAVE_EXPR (J)
9527 (where the same SAVE_EXPR (J) is used in the original and the
9528 transformed version). */
9530 static int
9532 {
9540 {
9552 {
9556 /* const_binop may not detect overflow correctly,
9557 so check for it explicitly here. */
9563 size_one_node,
9567 }
9571 /* Can't handle conversions from non-integral or wider integral type. */
9577 /* .. fall through ... */
9593 }
9594 }
9596 /* Return true if `t' is known to be non-negative. */
9598 int
9600 {
9602 {
9617 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
9618 both unsigned and at least 2 bits shorter than the result. */
9622 {
9627 {
9631 }
9632 }
9637 {
9638 /* x * x for floating point x is always non-negative. */
9643 }
9645 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
9646 both unsigned and their total bits is shorter than the result. */
9650 {
9657 }
9686 {
9691 {
9695 {
9699 }
9700 }
9702 {
9708 }
9709 }
9735 {
9739 /* If the initializer is non-void, then it's a normal expression
9740 that will be assigned to the slot. */
9744 /* Otherwise, the initializer sets the slot in some way. One common
9745 way is an assignment statement at the end of the initializer. */
9747 {
9755 else
9757 }
9763 }
9766 {
9773 {
9774 #define CASE_BUILTIN_F(BUILT_IN_FN) \
9775 case BUILT_IN_FN: case BUILT_IN_FN##F: case BUILT_IN_FN##L:
9776 #define CASE_BUILTIN_I(BUILT_IN_FN) \
9777 case BUILT_IN_FN: case BUILT_IN_FN##L: case BUILT_IN_FN##LL:
9795 /* Always true. */
9799 /* sqrt(-0.0) is -0.0. */
9827 /* True if the 1st argument is nonnegative. */
9831 /* True if the 1st OR 2nd arguments are nonnegative. */
9836 /* True if the 1st AND 2nd arguments are nonnegative. */
9841 /* True if the 2nd argument is nonnegative. */
9846 #undef CASE_BUILTIN_F
9847 #undef CASE_BUILTIN_I
9848 }
9849 }
9851 /* ... fall through ... */
9855 /* Truth values evaluate to 0 or 1, which is nonnegative. */
9857 }
9859 /* We don't know sign of `t', so be conservative and return false. */
9861 }
9863 /* Return true when T is an address and is known to be nonzero.
9864 For floating point we further ensure that T is not denormal.
9865 Similar logic is present in nonzero_address in rtlanal.h. */
9867 static bool
9869 {
9872 /* Doing something useful for floating point would need more work. */
9877 {
9883 /* We used to test for !integer_zerop here. This does not work correctly
9884 if TREE_CONSTANT_OVERFLOW (t). */
9890 {
9891 /* With the presence of negative values it is hard
9892 to say something. */
9896 /* One of operands must be positive and the other non-negative. */
9899 }
9904 {
9907 }
9911 {
9917 }
9921 {
9927 /* Weak declarations may link to NULL. */
9931 /* Constants are never weak. */
9936 }
9948 {
9949 /* When both operands are nonzero, then MAX must be too. */
9953 /* MAX where operand 0 is positive is positive. */
9955 }
9956 /* MAX where operand 1 is positive is positive. */
9977 }
9979 }
9981 /* See if we are applying CODE, a relational to the highest or lowest
9982 possible integer of TYPE. If so, then the result is a compile
9983 time constant. */
9985 static tree
9988 {
9999 {
10006 {
10009 }
10010 else
10011 {
10014 }
10019 {
10033 /* The GE_EXPR and LT_EXPR cases above are not normally
10034 reached because of previous transformations. */
10038 }
10042 {
10053 }
10057 {
10074 }
10078 {
10089 }
10094 /* signed_type does not work on pointer types. */
10096 {
10097 /* The following case also applies to X < signed_max+1
10098 and X >= signed_max+1 because previous transformations. */
10100 {
10106 type,
10115 /* If we are in gimple form, then returning EXP would create
10116 non-gimple expressions. Clearing it is safe and insures
10117 we do not allow a non-gimple expression to escape. */
10122 }
10123 }
10124 }
10127 }
10130 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
10131 attempt to fold the expression to a constant without modifying TYPE,
10132 OP0 or OP1.
10134 If the expression could be simplified to a constant, then return
10135 the constant. If the expression would not be simplified to a
10136 constant, then return NULL_TREE.
10138 Note this is primarily designed to be called after gimplification
10139 of the tree structures and when at least one operand is a constant.
10140 As a result of those simplifying assumptions this routine is far
10141 simpler than the generic fold routine. */
10143 tree
10145 {
10147 tree subop0;
10148 tree subop1;
10149 tree tem;
10151 /* If this is a commutative operation, and ARG0 is a constant, move it
10152 to ARG1 to reduce the number of tests below. */
10155 {
10159 }
10161 /* If either operand is a complex type, extract its real component. */
10164 else
10169 else
10172 /* Note if either argument is not a real or integer constant.
10173 With a few exceptions, simplification is limited to cases
10174 where both arguments are constants. */
10182 {
10184 /* (plus (address) (const_int)) is a constant. */
10192 {
10196 }
10199 binary:
10203 /* Both arguments are constants. Simplify. */
10206 {
10207 /* The return value should always have the same type as
10208 the original expression. */
10213 }
10217 /* Fold &x - &x. This can happen from &x.foo - &x.
10218 This is unsafe for certain floats even in non-IEEE formats.
10219 In IEEE, it is unsafe because it does wrong for NaNs.
10220 Also note that operand_equal_p is always false if an
10221 operand is volatile. */
10229 /* Special case multiplication or bitwise AND where one argument
10230 is zero. */
10233 else
10242 /* Special case when we know the result will be all ones. */
10258 /* Division by zero is undefined. */
10285 /* Optimize -1 >> x for arithmetic right shifts. */
10288 /* ... fall through ... */
10294 /* Since negative shift count is not well-defined, don't
10295 try to compute it in the compiler. */
10303 /* -1 rotated either direction by any amount is still -1. */
10307 /* 0 rotated either direction by any amount is still zero. */
10324 /* If one arg is a real or integer constant, put it last. */
10329 {
10330 tree temp;
10336 }
10338 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
10339 This transformation affects the cases which are handled in later
10340 optimizations involving comparisons with non-negative constants. */
10344 {
10346 {
10359 }
10360 }
10366 /* Fall through. */
10382 /* This could probably be handled. */
10386 /* If second arg is constant zero, result is zero, but first arg
10387 must be evaluated. */
10390 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
10391 case will be handled here. */
10399 /* If second arg is constant true, result is true, but we must
10400 evaluate first arg. */
10403 /* Likewise for first arg, but note this only occurs here for
10404 TRUTH_OR_EXPR. */
10413 {
10416 }
10421 }
10422 }
10424 /* Given the components of a unary expression CODE, TYPE and OP0,
10425 attempt to fold the expression to a constant without modifying
10426 TYPE or OP0.
10428 If the expression could be simplified to a constant, then return
10429 the constant. If the expression would not be simplified to a
10430 constant, then return NULL_TREE.
10432 Note this is primarily designed to be called after gimplification
10433 of the tree structures and when op0 is a constant. As a result
10434 of those simplifying assumptions this routine is far simpler than
10435 the generic fold routine. */
10437 tree
10439 {
10440 /* Make sure we have a suitable constant argument. */
10442 {
10443 tree subop;
10447 else
10452 }
10455 {
10467 else
10473 else
10479 else
10485 else
10491 else
10503 }
10504 }
10506 /* If EXP represents referencing an element in a constant string
10507 (either via pointer arithmetic or array indexing), return the
10508 tree representing the value accessed, otherwise return NULL. */
10510 tree
10512 {
10514 {
10516 tree index;
10517 tree string;
10521 else
10522 {
10526 /* Optimize the special-case of a zero lower bound.
10528 We convert the low_bound to sizetype to avoid some problems
10529 with constant folding. (E.g. suppose the lower bound is 1,
10530 and its mode is QI. Without the conversion,l (ARRAY
10531 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
10532 +INDEX), which becomes (ARRAY+255+INDEX). Opps!) */
10537 }
10551 }
10553 }
10555 /* Return the tree for neg (ARG0) when ARG0 is known to be either
10556 an integer constant or real constant.
10558 TYPE is the type of the result. */
10560 static tree
10562 {
10566 {
10568 {
10570 HOST_WIDE_INT high;
10580 }
10588 }
10591 }
10593 /* Return the tree for abs (ARG0) when ARG0 is known to be either
10594 an integer constant or real constant.
10596 TYPE is the type of the result. */
10598 tree
10600 {
10604 {
10606 /* If the value is unsigned, then the absolute value is
10607 the same as the ordinary value. */
10610 /* Similarly, if the value is non-negative. */
10613 /* If the value is negative, then the absolute value is
10614 its negation. */
10615 else
10616 {
10618 HOST_WIDE_INT high;
10625 }
10631 else
10637 }
10640 }
10642 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
10643 constant. TYPE is the type of the result. */
10645 static tree
10647 {
10659 }
10661 /* Given CODE, a relational operator, the target type, TYPE and two
10662 constant operands OP0 and OP1, return the result of the
10663 relational operation. If the result is not a compile time
10664 constant, then return NULL_TREE. */
10666 static tree
10668 {
10671 /* From here on, the only cases we handle are when the result is
10672 known to be a constant. */
10675 {
10679 /* Handle the cases where either operand is a NaN. */
10681 {
10683 {
10711 }
10714 }
10717 }
10719 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
10721 To compute GT, swap the arguments and do LT.
10722 To compute GE, do LT and invert the result.
10723 To compute LE, swap the arguments, do LT and invert the result.
10724 To compute NE, do EQ and invert the result.
10726 Therefore, the code below must handle only EQ and LT. */
10729 {
10734 }
10736 /* Note that it is safe to invert for real values here because we
10737 have already handled the one case that it matters. */
10741 {
10744 }
10746 /* Compute a result for LT or EQ if args permit;
10747 Otherwise return T. */
10749 {
10754 else
10756 }
10757 else
10763 }
10765 /* Build an expression for the a clean point containing EXPR with type TYPE.
10766 Don't build a cleanup point expression for EXPR which don't have side
10767 effects. */
10769 tree
10771 {
10772 /* If the expression does not have side effects then we don't have to wrap
10773 it with a cleanup point expression. */
10778 }
10780 /* Build an expression for the address of T. Folds away INDIRECT_REF to
10781 avoid confusing the gimplify process. */
10783 tree
10785 {
10786 /* The size of the object is not relevant when talking about its address. */
10790 /* Note: doesn't apply to ALIGN_INDIRECT_REF */
10793 {
10797 }
10798 else
10799 {
10810 }
10813 }
10815 tree
10817 {
10819 }
10821 /* Builds an expression for an indirection through T, simplifying some
10822 cases. */
10824 tree
10826 {
10829 tree subtype;
10833 {
10836 /* *&p => p */
10839 /* *(foo *)&fooarray => fooarray[0] */
10843 }
10845 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
10849 {
10852 }
10855 }
10857 /* Strip non-trapping, non-side-effecting tree nodes from an expression
10858 whose result is ignored. The type of the returned tree need not be
10859 the same as the original expression. */
10861 tree
10863 {
10869 {
10880 else
10886 {
10902 }
10907 }
10908 }
10910 /* Return the value of VALUE, rounded up to a multiple of DIVISOR.
10911 This can only be applied to objects of a sizetype. */
10913 tree
10915 {
10922 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10923 have to do anything. Only do this when we are not given a const,
10924 because in that case, this check is more expensive than just
10925 doing it. */
10927 {
10932 }
10934 /* If divisor is a power of two, simplify this to bit manipulation. */
10936 {
10937 tree t;
10943 }
10944 else
10945 {
10950 }
10953 }
10955 /* Likewise, but round down. */
10957 tree
10959 {
10966 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10967 have to do anything. Only do this when we are not given a const,
10968 because in that case, this check is more expensive than just
10969 doing it. */
10971 {
10976 }
10978 /* If divisor is a power of two, simplify this to bit manipulation. */
10980 {
10981 tree t;
10985 }
10986 else
10987 {
10992 }
10995 }
10997 /* Returns the pointer to the base of the object addressed by EXP and
10998 extracts the information about the offset of the access, storing it
10999 to PBITPOS and POFFSET. */
11001 static tree
11004 {
11005 tree core;
11008 HOST_WIDE_INT bitsize;
11011 {
11017 }
11018 else
11019 {
11023 }
11026 }
11028 /* Returns true if addresses of E1 and E2 differ by a constant, false
11029 otherwise. If they do, E1 - E2 is stored in *DIFF. */
11031 bool
11033 {
11047 {
11057 }
11059 {
11060 /* If only one of the offsets is non-constant, the difference cannot
11061 be a constant. */
11063 }
11064 else
11069 }