| blob | 3ab360b0dd3a2850aedfdc59d0ef7c51eb2c4783 |
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, 2002,
3 1999, 2000, 2001, 2002 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 and prior overflow indicator, and
43 forces the value to fit the type. It returns an overflow indicator. */
62 HOST_WIDE_INT));
65 HOST_WIDE_INT *));
90 HOST_WIDE_INT *,
111 static tree fold_binary_op_with_conditional_arg
115 /* The following constants represent a bit based encoding of GCC's
116 comparison operators. This encoding simplifies transformations
117 on relational comparison operators, such as AND and OR. */
118 #define COMPCODE_FALSE 0
119 #define COMPCODE_LT 1
120 #define COMPCODE_EQ 2
121 #define COMPCODE_LE 3
122 #define COMPCODE_GT 4
123 #define COMPCODE_NE 5
124 #define COMPCODE_GE 6
125 #define COMPCODE_TRUE 7
127 /* We know that A1 + B1 = SUM1, using 2's complement arithmetic and ignoring
128 overflow. Suppose A, B and SUM have the same respective signs as A1, B1,
129 and SUM1. Then this yields nonzero if overflow occurred during the
130 addition.
132 Overflow occurs if A and B have the same sign, but A and SUM differ in
133 sign. Use `^' to test whether signs differ, and `< 0' to isolate the
134 sign. */
135 #define OVERFLOW_SUM_SIGN(a, b, sum) ((~((a) ^ (b)) & ((a) ^ (sum))) < 0)
136 \f
137 /* To do constant folding on INTEGER_CST nodes requires two-word arithmetic.
138 We do that by representing the two-word integer in 4 words, with only
139 HOST_BITS_PER_WIDE_INT / 2 bits stored in each word, as a positive
140 number. The value of the word is LOWPART + HIGHPART * BASE. */
142 #define LOWPART(x) \
143 ((x) & (((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) - 1))
144 #define HIGHPART(x) \
145 ((unsigned HOST_WIDE_INT) (x) >> HOST_BITS_PER_WIDE_INT / 2)
146 #define BASE ((unsigned HOST_WIDE_INT) 1 << HOST_BITS_PER_WIDE_INT / 2)
148 /* Unpack a two-word integer into 4 words.
149 LOW and HI are the integer, as two `HOST_WIDE_INT' pieces.
150 WORDS points to the array of HOST_WIDE_INTs. */
152 static void
156 HOST_WIDE_INT hi;
157 {
162 }
164 /* Pack an array of 4 words into a two-word integer.
165 WORDS points to the array of words.
166 The integer is stored into *LOW and *HI as two `HOST_WIDE_INT' pieces. */
168 static void
173 {
176 }
177 \f
178 /* Make the integer constant T valid for its type by setting to 0 or 1 all
179 the bits in the constant that don't belong in the type.
181 Return 1 if a signed overflow occurs, 0 otherwise. If OVERFLOW is
182 nonzero, a signed overflow has already occurred in calculating T, so
183 propagate it. */
185 int
187 tree t;
189 {
191 HOST_WIDE_INT high;
195 {
196 /* ??? Used to check for overflow here via CHECK_FLOAT_TYPE.
197 Consider doing it via real_convert now. */
199 }
209 else
212 /* First clear all bits that are beyond the type's precision. */
215 ;
219 else
220 {
224 }
226 /* Unsigned types do not suffer sign extension or overflow unless they
227 are a sizetype. */
233 /* If the value's sign bit is set, extend the sign. */
241 {
242 /* Value is negative:
243 set to 1 all the bits that are outside this type's precision. */
247 else
248 {
252 }
253 }
255 /* Return nonzero if signed overflow occurred. */
256 return
259 }
260 \f
261 /* Add two doubleword integers with doubleword result.
262 Each argument is given as two `HOST_WIDE_INT' pieces.
263 One argument is L1 and H1; the other, L2 and H2.
264 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
266 int
272 {
274 HOST_WIDE_INT h;
282 }
284 /* Negate a doubleword integer with doubleword result.
285 Return nonzero if the operation overflows, assuming it's signed.
286 The argument is given as two `HOST_WIDE_INT' pieces in L1 and H1.
287 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
289 int
292 HOST_WIDE_INT h1;
295 {
297 {
301 }
302 else
303 {
307 }
308 }
309 \f
310 /* Multiply two doubleword integers with doubleword result.
311 Return nonzero if the operation overflows, assuming it's signed.
312 Each argument is given as two `HOST_WIDE_INT' pieces.
313 One argument is L1 and H1; the other, L2 and H2.
314 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
316 int
322 {
337 {
340 {
342 /* This product is <= 0xFFFE0001, the sum <= 0xFFFF0000. */
344 /* Since prod[p] < 0xFFFF, this sum <= 0xFFFFFFFF. */
348 }
350 }
354 /* Check for overflow by calculating the top half of the answer in full;
355 it should agree with the low half's sign bit. */
358 {
361 }
363 {
366 }
368 }
369 \f
370 /* Shift the doubleword integer in L1, H1 left by COUNT places
371 keeping only PREC bits of result.
372 Shift right if COUNT is negative.
373 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
374 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
376 void
384 {
388 {
391 }
393 #ifdef SHIFT_COUNT_TRUNCATED
396 #endif
399 {
400 /* Shifting by the host word size is undefined according to the
401 ANSI standard, so we must handle this as a special case. */
404 }
406 {
409 }
410 else
411 {
415 }
417 /* Sign extend all bits that are beyond the precision. */
425 ;
427 {
430 }
431 else
432 {
436 }
437 }
439 /* Shift the doubleword integer in L1, H1 right by COUNT places
440 keeping only PREC bits of result. COUNT must be positive.
441 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
442 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
444 void
452 {
455 signmask = (arith
459 #ifdef SHIFT_COUNT_TRUNCATED
462 #endif
465 {
466 /* Shifting by the host word size is undefined according to the
467 ANSI standard, so we must handle this as a special case. */
470 }
472 {
475 }
476 else
477 {
481 }
483 /* Zero / sign extend all bits that are beyond the precision. */
486 {
489 }
491 ;
493 {
496 }
497 else
498 {
502 }
503 }
504 \f
505 /* Rotate the doubleword integer in L1, H1 left by COUNT places
506 keeping only PREC bits of result.
507 Rotate right if COUNT is negative.
508 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
510 void
517 {
529 }
531 /* Rotate the doubleword integer in L1, H1 left by COUNT places
532 keeping only PREC bits of result. COUNT must be positive.
533 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
535 void
542 {
554 }
555 \f
556 /* Divide doubleword integer LNUM, HNUM by doubleword integer LDEN, HDEN
557 for a quotient (stored in *LQUO, *HQUO) and remainder (in *LREM, *HREM).
558 CODE is a tree code for a kind of division, one of
559 TRUNC_DIV_EXPR, FLOOR_DIV_EXPR, CEIL_DIV_EXPR, ROUND_DIV_EXPR
560 or EXACT_DIV_EXPR
561 It controls how the quotient is rounded to an integer.
562 Return nonzero if the operation overflows.
563 UNS nonzero says do unsigned division. */
565 int
572 HOST_WIDE_INT hnum_orig;
574 HOST_WIDE_INT hden_orig;
577 {
593 /* calculate quotient sign and convert operands to unsigned. */
595 {
597 {
599 /* (minimum integer) / (-1) is the only overflow case. */
603 }
605 {
608 }
609 }
614 /* This unsigned division rounds toward zero. */
617 }
621 /* hden != 0 already checked. */
626 }
636 /* Special code for when the divisor < BASE. */
638 {
639 /* hnum != 0 already checked. */
641 {
645 }
646 }
647 else
648 {
649 /* Full double precision division,
650 with thanks to Don Knuth's "Seminumerical Algorithms". */
654 /* Find the highest nonzero divisor digit. */
657 {
660 }
662 /* Insure that the first digit of the divisor is at least BASE/2.
663 This is required by the quotient digit estimation algorithm. */
670 {
674 }
679 {
684 }
685 }
689 /* Main loop */
691 {
692 /* Guess the next quotient digit, quo_est, by dividing the first
693 two remaining dividend digits by the high order quotient digit.
694 quo_est is never low and is at most 2 high. */
701 else
704 /* Refine quo_est so it's usually correct, and at most one high. */
709 quo_est--;
711 /* Try QUO_EST as the quotient digit, by multiplying the
712 divisor by QUO_EST and subtracting from the remaining dividend.
713 Keep in mind that QUO_EST is the I - 1st digit. */
717 {
723 }
725 /* If quo_est was high by one, then num[i] went negative and
726 we need to correct things. */
728 {
729 quo_est--;
732 {
736 }
739 }
741 /* Store the quotient digit. */
743 }
744 }
748 finish_up:
749 /* if result is negative, make it so. */
753 /* compute trial remainder: rem = num - (quo * den) */
759 {
768 {
769 /* quo = quo - 1; */
772 }
773 else
780 {
783 }
784 else
790 {
796 /* Get absolute values */
802 /* If (2 * abs (lrem) >= abs (lden)) */
811 {
813 /* quo = quo - 1; */
816 else
817 /* quo = quo + 1; */
820 }
821 else
823 }
828 }
830 /* compute true remainder: rem = num - (quo * den) */
835 }
836 \f
837 /* Given T, an expression, return the negation of T. Allow for T to be
838 null, in which case return null. */
840 static tree
842 tree t;
843 {
844 tree type;
845 tree tem;
854 {
867 /* - (A - B) -> B - A */
877 }
880 }
881 \f
882 /* Split a tree IN into a constant, literal and variable parts that could be
883 combined with CODE to make IN. "constant" means an expression with
884 TREE_CONSTANT but that isn't an actual constant. CODE must be a
885 commutative arithmetic operation. Store the constant part into *CONP,
886 the literal in *LITP and return the variable part. If a part isn't
887 present, set it to null. If the tree does not decompose in this way,
888 return the entire tree as the variable part and the other parts as null.
890 If CODE is PLUS_EXPR we also split trees that use MINUS_EXPR. In that
891 case, we negate an operand that was subtracted. Except if it is a
892 literal for which we use *MINUS_LITP instead.
894 If NEGATE_P is true, we are negating all of IN, again except a literal
895 for which we use *MINUS_LITP instead.
897 If IN is itself a literal or constant, return it as appropriate.
899 Note that we do not guarantee that any of the three values will be the
900 same type as IN, but they will have the same signedness and mode. */
902 static tree
904 tree in;
908 {
915 /* Strip any conversions that don't change the machine mode or signedness. */
922 /* We can associate addition and subtraction together (even
923 though the C standard doesn't say so) for integers because
924 the value is not affected. For reals, the value might be
925 affected, so we can't. */
928 {
934 /* First see if either of the operands is a literal, then a constant. */
945 /* If we haven't dealt with either operand, this is not a case we can
946 decompose. Otherwise, VAR is either of the ones remaining, if any. */
951 else
954 /* Now do any needed negations. */
961 }
964 else
968 {
975 }
978 }
980 /* Re-associate trees split by the above function. T1 and T2 are either
981 expressions to associate or null. Return the new expression, if any. If
982 we build an operation, do it in TYPE and with CODE. */
984 static tree
988 tree type;
989 {
995 /* If either input is CODE, a PLUS_EXPR, or a MINUS_EXPR, don't
996 try to fold this since we will have infinite recursion. But do
997 deal with any NEGATE_EXPRs. */
1000 {
1002 {
1009 }
1011 }
1014 }
1015 \f
1016 /* Combine two integer constants ARG1 and ARG2 under operation CODE
1017 to produce a new constant.
1019 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1021 static tree
1026 {
1030 HOST_WIDE_INT hi;
1032 HOST_WIDE_INT garbageh;
1033 tree t;
1036 int is_sizetype
1047 {
1067 /* It's unclear from the C standard whether shifts can overflow.
1068 The following code ignores overflow; perhaps a C standard
1069 interpretation ruling is needed. */
1099 /* This is a shortcut for a common special case. */
1104 {
1110 }
1112 /* ... fall through ... */
1116 {
1119 }
1122 {
1125 }
1132 /* This is a shortcut for a common special case. */
1137 {
1142 }
1144 /* ... fall through ... */
1160 else
1166 else
1172 }
1174 /* If this is for a sizetype, can be represented as one (signed)
1175 HOST_WIDE_INT word, and doesn't overflow, use size_int since it caches
1176 constants. */
1182 else
1183 {
1186 }
1189 = ((notrunc
1196 /* If we're doing a size calculation, unsigned arithmetic does overflow.
1197 So check if force_fit_type truncated the value. */
1208 }
1210 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1211 constant. We assume ARG1 and ARG2 have the same data type, or at least
1212 are the same kind of constant and the same machine mode.
1214 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1216 static tree
1221 {
1229 {
1230 REAL_VALUE_TYPE d1;
1231 REAL_VALUE_TYPE d2;
1232 REAL_VALUE_TYPE value;
1233 tree t;
1238 /* If either operand is a NaN, just return it. Otherwise, set up
1239 for floating-point trap; we return an overflow. */
1249 value));
1259 }
1261 {
1267 tree t;
1270 {
1290 notrunc),
1296 notrunc));
1300 {
1301 tree magsquared
1305 notrunc);
1308 const_binop
1313 notrunc),
1315 notrunc),
1316 notrunc),
1318 const_binop
1323 notrunc),
1325 notrunc),
1326 notrunc),
1328 }
1333 }
1335 }
1337 }
1339 /* These are the hash table functions for the hash table of INTEGER_CST
1340 nodes of a sizetype. */
1342 /* Return the hash code code X, an INTEGER_CST. */
1344 static hashval_t
1347 {
1353 }
1355 /* Return nonzero if the value represented by *X (an INTEGER_CST tree node)
1356 is the same as that given by *Y, which is the same. */
1358 static int
1362 {
1370 }
1371 \f
1372 /* Return an INTEGER_CST with value whose low-order HOST_BITS_PER_WIDE_INT
1373 bits are given by NUMBER and of the sizetype represented by KIND. */
1375 tree
1377 HOST_WIDE_INT number;
1379 {
1381 }
1383 /* Likewise, but the desired type is specified explicitly. */
1387 htab_t size_htab;
1389 tree
1391 HOST_WIDE_INT number;
1392 tree type;
1393 {
1397 {
1400 }
1402 /* Adjust NEW_CONST to be the constant we want. If it's already in the
1403 hash table, we return the value from the hash table. Otherwise, we
1404 place that in the hash table and make a new node for the next time. */
1413 {
1419 }
1420 else
1422 }
1424 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
1425 is a tree code. The type of the result is taken from the operands.
1426 Both must be the same type integer type and it must be a size type.
1427 If the operands are constant, so is the result. */
1429 tree
1433 {
1440 /* Handle the special case of two integer constants faster. */
1442 {
1443 /* And some specific cases even faster than that. */
1452 /* Handle general case of two integer constants. */
1454 }
1460 }
1462 /* Given two values, either both of sizetype or both of bitsizetype,
1463 compute the difference between the two values. Return the value
1464 in signed type corresponding to the type of the operands. */
1466 tree
1469 {
1471 tree ctype;
1477 /* If the type is already signed, just do the simple thing. */
1484 /* If either operand is not a constant, do the conversions to the signed
1485 type and subtract. The hardware will do the right thing with any
1486 overflow in the subtraction. */
1491 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
1492 Otherwise, subtract the other way, convert to CTYPE (we know that can't
1493 overflow) and negate (which can't either). Special-case a result
1494 of zero while we're here. */
1499 else
1502 }
1503 \f
1505 /* Given T, a tree representing type conversion of ARG1, a constant,
1506 return a constant tree representing the result of conversion. */
1508 static tree
1510 tree t;
1511 tree arg1;
1512 {
1517 {
1519 {
1520 /* If we would build a constant wider than GCC supports,
1521 leave the conversion unfolded. */
1525 /* If we are trying to make a sizetype for a small integer, use
1526 size_int to pick up cached types to reduce duplicate nodes. */
1532 /* Given an integer constant, make new constant with new type,
1533 appropriately sign-extended or truncated. */
1537 /* Indicate an overflow if (1) ARG1 already overflowed,
1538 or (2) force_fit_type indicates an overflow.
1539 Tell force_fit_type that an overflow has already occurred
1540 if ARG1 is a too-large unsigned value and T is signed.
1541 But don't indicate an overflow if converting a pointer. */
1551 }
1553 {
1554 /* Don't initialize these, use assignments.
1555 Initialized local aggregates don't work on old compilers. */
1556 REAL_VALUE_TYPE x;
1557 REAL_VALUE_TYPE l;
1558 REAL_VALUE_TYPE u;
1569 /* See if X will be in range after truncation towards 0.
1570 To compensate for truncation, move the bounds away from 0,
1571 but reject if X exactly equals the adjusted bounds. */
1575 /* If X is a NaN, use zero instead and show we have an overflow.
1576 Otherwise, range check. */
1580 && !no_upper_bound
1584 {
1588 }
1594 }
1596 }
1598 {
1602 {
1604 {
1605 /* We make a copy of ARG1 so that we don't modify an
1606 existing constant tree. */
1610 }
1621 }
1622 }
1625 }
1626 \f
1627 /* Return an expr equal to X but certainly not valid as an lvalue. */
1629 tree
1631 tree x;
1632 {
1633 tree result;
1635 /* These things are certainly not lvalues. */
1646 }
1648 /* Nonzero means lvalues are limited to those valid in pedantic ANSI C.
1649 Zero means allow extended lvalues. */
1653 /* When pedantic, return an expr equal to X but certainly not valid as a
1654 pedantic lvalue. Otherwise, return X. */
1656 tree
1658 tree x;
1659 {
1662 else
1664 }
1665 \f
1666 /* Given a tree comparison code, return the code that is the logical inverse
1667 of the given code. It is not safe to do this for floating-point
1668 comparisons, except for NE_EXPR and EQ_EXPR. */
1670 static enum tree_code
1673 {
1675 {
1690 }
1691 }
1693 /* Similar, but return the comparison that results if the operands are
1694 swapped. This is safe for floating-point. */
1696 static enum tree_code
1699 {
1701 {
1715 }
1716 }
1719 /* Convert a comparison tree code from an enum tree_code representation
1720 into a compcode bit-based encoding. This function is the inverse of
1721 compcode_to_comparison. */
1723 static int
1726 {
1728 {
1743 }
1744 }
1746 /* Convert a compcode bit-based encoding of a comparison operator back
1747 to GCC's enum tree_code representation. This function is the
1748 inverse of comparison_to_compcode. */
1750 static enum tree_code
1753 {
1755 {
1770 }
1771 }
1773 /* Return nonzero if CODE is a tree code that represents a truth value. */
1775 static int
1778 {
1783 }
1784 \f
1785 /* Return nonzero if two operands are necessarily equal.
1786 If ONLY_CONST is nonzero, only return nonzero for constants.
1787 This function tests whether the operands are indistinguishable;
1788 it does not test whether they are equal using C's == operation.
1789 The distinction is important for IEEE floating point, because
1790 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
1791 (2) two NaNs may be indistinguishable, but NaN!=NaN. */
1793 int
1797 {
1798 /* If both types don't have the same signedness, then we can't consider
1799 them equal. We must check this before the STRIP_NOPS calls
1800 because they may change the signedness of the arguments. */
1808 /* This is needed for conversions and for COMPONENT_REF.
1809 Might as well play it safe and always test this. */
1815 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
1816 We don't care about side effects in that case because the SAVE_EXPR
1817 takes care of that for us. In all other cases, two expressions are
1818 equal if they have no side effects. If we have two identical
1819 expressions with side effects that should be treated the same due
1820 to the only side effects being identical SAVE_EXPR's, that will
1821 be detected in the recursive calls below. */
1827 /* Next handle constant cases, those for which we can return 1 even
1828 if ONLY_CONST is set. */
1831 {
1844 {
1854 {
1859 }
1862 }
1866 only_const)
1868 only_const));
1881 }
1887 {
1889 /* Two conversions are equal only if signedness and modes match. */
1905 /* For commutative ops, allow the other order. */
1918 /* If either of the pointer (or reference) expressions we are dereferencing
1919 contain a side effect, these cannot be equal. */
1925 {
1947 }
1956 }
1957 }
1958 \f
1959 /* Similar to operand_equal_p, but see if ARG0 might have been made by
1960 shorten_compare from ARG1 when ARG1 was being compared with OTHER.
1962 When in doubt, return 0. */
1964 static int
1967 tree other;
1968 {
1980 /* Discard any conversions that don't change the modes of ARG0 and ARG1
1981 and see if the inner values are the same. This removes any
1982 signedness comparison, which doesn't matter here. */
1989 /* Duplicate what shorten_compare does to ARG1 and see if that gives the
1990 actual comparison operand, ARG0.
1992 First throw away any conversions to wider types
1993 already present in the operands. */
2002 {
2005 /* Make sure shorter operand is extended the right way
2006 to match the longer operand. */
2012 }
2015 }
2016 \f
2017 /* See if ARG is an expression that is either a comparison or is performing
2018 arithmetic on comparisons. The comparisons must only be comparing
2019 two different values, which will be stored in *CVAL1 and *CVAL2; if
2020 they are nonzero it means that some operands have already been found.
2021 No variables may be used anywhere else in the expression except in the
2022 comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around
2023 the expression and save_expr needs to be called with CVAL1 and CVAL2.
2025 If this is true, return 1. Otherwise, return zero. */
2027 static int
2029 tree arg;
2032 {
2036 /* We can handle some of the 'e' cases here. */
2046 {
2047 /* If we've already found a CVAL1 or CVAL2, this expression is
2048 two complex to handle. */
2054 }
2057 {
2080 /* First see if we can handle the first operand, then the second. For
2081 the second operand, we know *CVAL1 can't be zero. It must be that
2082 one side of the comparison is each of the values; test for the
2083 case where this isn't true by failing if the two operands
2084 are the same. */
2093 ;
2097 ;
2098 else
2102 ;
2106 ;
2107 else
2114 }
2115 }
2116 \f
2117 /* ARG is a tree that is known to contain just arithmetic operations and
2118 comparisons. Evaluate the operations in the tree substituting NEW0 for
2119 any occurrence of OLD0 as an operand of a comparison and likewise for
2120 NEW1 and OLD1. */
2122 static tree
2124 tree arg;
2126 {
2131 /* We can handle some of the 'e' cases here. */
2139 {
2154 {
2171 }
2172 /* fall through - ??? */
2175 {
2179 /* We need to check both for exact equality and tree equality. The
2180 former will be true if the operand has a side-effect. In that
2181 case, we know the operand occurred exactly once. */
2194 }
2198 }
2199 }
2200 \f
2201 /* Return a tree for the case when the result of an expression is RESULT
2202 converted to TYPE and OMITTED was previously an operand of the expression
2203 but is now not needed (e.g., we folded OMITTED * 0).
2205 If OMITTED has side effects, we must evaluate it. Otherwise, just do
2206 the conversion of RESULT to TYPE. */
2208 static tree
2211 {
2218 }
2220 /* Similar, but call pedantic_non_lvalue instead of non_lvalue. */
2222 static tree
2225 {
2232 }
2233 \f
2234 /* Return a simplified tree node for the truth-negation of ARG. This
2235 never alters ARG itself. We assume that ARG is an operation that
2236 returns a truth value (0 or 1). */
2238 tree
2240 tree arg;
2241 {
2248 /* If this is a comparison, we can simply invert it, except for
2249 floating-point non-equality comparisons, in which case we just
2250 enclose a TRUTH_NOT_EXPR around what we have. */
2253 {
2255 && !flag_unsafe_math_optimizations
2259 else
2262 }
2265 {
2280 /* Here we can invert either operand. We invert the first operand
2281 unless the second operand is a TRUTH_NOT_EXPR in which case our
2282 result is the XOR of the first operand with the inside of the
2283 negation of the second operand. */
2288 else
2343 }
2347 }
2349 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
2350 operands are another bit-wise operation with a common input. If so,
2351 distribute the bit operations to save an operation and possibly two if
2352 constants are involved. For example, convert
2353 (A | B) & (A | C) into A | (B & C)
2354 Further simplification will occur if B and C are constants.
2356 If this optimization cannot be done, 0 will be returned. */
2358 static tree
2361 tree type;
2363 {
2364 tree common;
2374 {
2378 }
2380 {
2384 }
2386 {
2390 }
2392 {
2396 }
2397 else
2402 }
2403 \f
2404 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
2405 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero. */
2407 static tree
2409 tree inner;
2410 tree type;
2413 {
2420 }
2422 /* Optimize a bit-field compare.
2424 There are two cases: First is a compare against a constant and the
2425 second is a comparison of two items where the fields are at the same
2426 bit position relative to the start of a chunk (byte, halfword, word)
2427 large enough to contain it. In these cases we can avoid the shift
2428 implicit in bitfield extractions.
2430 For constants, we emit a compare of the shifted constant with the
2431 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
2432 compared. For two fields at the same position, we do the ANDs with the
2433 similar mask and compare the result of the ANDs.
2435 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
2436 COMPARE_TYPE is the type of the comparison, and LHS and RHS
2437 are the left and right operands of the comparison, respectively.
2439 If the optimization described above can be done, we return the resulting
2440 tree. Otherwise we return zero. */
2442 static tree
2445 tree compare_type;
2447 {
2456 tree mask;
2457 tree offset;
2459 /* Get all the information about the extractions being done. If the bit size
2460 if the same as the size of the underlying object, we aren't doing an
2461 extraction at all and so can do nothing. We also don't want to
2462 do anything if the inner expression is a PLACEHOLDER_EXPR since we
2463 then will no longer be able to replace it. */
2471 {
2472 /* If this is not a constant, we can only do something if bit positions,
2473 sizes, and signedness are the same. */
2481 }
2483 /* See if we can find a mode to refer to this field. We should be able to,
2484 but fail if we can't. */
2493 /* Set signed and unsigned types of the precision of this mode for the
2494 shifts below. */
2498 /* Compute the bit position and size for the new reference and our offset
2499 within it. If the new reference is the same size as the original, we
2500 won't optimize anything, so return zero. */
2510 /* Make the mask to be used against the extracted field. */
2520 /* If not comparing with constant, just rework the comparison
2521 and return. */
2526 mask),
2530 mask));
2532 /* Otherwise, we are handling the constant case. See if the constant is too
2533 big for the field. Warn and return a tree of for 0 (false) if so. We do
2534 this not only for its own sake, but to avoid having to test for this
2535 error case below. If we didn't, we might generate wrong code.
2537 For unsigned fields, the constant shifted right by the field length should
2538 be all zero. For signed fields, the high-order bits should agree with
2539 the sign bit. */
2542 {
2546 {
2552 }
2553 }
2554 else
2555 {
2559 {
2565 }
2566 }
2568 /* Single-bit compares should always be against zero. */
2570 {
2573 }
2575 /* Make a new bitfield reference, shift the constant over the
2576 appropriate number of bits and mask it with the computed mask
2577 (in case this was a signed field). If we changed it, make a new one. */
2580 {
2583 }
2593 rhs);
2594 }
2595 \f
2596 /* Subroutine for fold_truthop: decode a field reference.
2598 If EXP is a comparison reference, we return the innermost reference.
2600 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
2601 set to the starting bit number.
2603 If the innermost field can be completely contained in a mode-sized
2604 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
2606 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
2607 otherwise it is not changed.
2609 *PUNSIGNEDP is set to the signedness of the field.
2611 *PMASK is set to the mask used. This is either contained in a
2612 BIT_AND_EXPR or derived from the width of the field.
2614 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
2616 Return 0 if this is not a component reference or is one that we can't
2617 do anything with. */
2619 static tree
2622 tree exp;
2628 {
2631 tree unsigned_type;
2634 /* All the optimizations using this function assume integer fields.
2635 There are problems with FP fields since the type_for_size call
2636 below can fail for, e.g., XFmode. */
2643 {
2649 }
2658 /* Compute the mask to access the bitfield. */
2668 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
2676 }
2678 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
2679 bit positions. */
2681 static int
2683 tree mask;
2685 {
2688 tree tmask;
2693 return
2700 }
2702 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
2703 represents the sign bit of EXP's type. If EXP represents a sign
2704 or zero extension, also test VAL against the unextended type.
2705 The return value is the (sub)expression whose sign bit is VAL,
2706 or NULL_TREE otherwise. */
2708 static tree
2710 tree exp;
2711 tree val;
2712 {
2714 HOST_WIDE_INT hi;
2716 tree t;
2718 /* Tree EXP must have an integral type. */
2723 /* Tree VAL must be an integer constant. */
2730 {
2733 }
2734 else
2735 {
2738 }
2743 /* Handle extension from a narrower type. */
2749 }
2751 /* Subroutine for fold_truthop: determine if an operand is simple enough
2752 to be evaluated unconditionally. */
2754 static int
2756 tree exp;
2757 {
2758 /* Strip any conversions that don't change the machine mode. */
2770 /* Don't regard global variables as simple. They may be
2771 allocated in ways unknown to the compiler (shared memory,
2772 #pragma weak, etc). */
2775 /* Loading a static variable is unduly expensive, but global
2776 registers aren't expensive. */
2778 }
2779 \f
2780 /* The following functions are subroutines to fold_range_test and allow it to
2781 try to change a logical combination of comparisons into a range test.
2783 For example, both
2784 X == 2 || X == 3 || X == 4 || X == 5
2785 and
2786 X >= 2 && X <= 5
2787 are converted to
2788 (unsigned) (X - 2) <= 3
2790 We describe each set of comparisons as being either inside or outside
2791 a range, using a variable named like IN_P, and then describe the
2792 range with a lower and upper bound. If one of the bounds is omitted,
2793 it represents either the highest or lowest value of the type.
2795 In the comments below, we represent a range by two numbers in brackets
2796 preceded by a "+" to designate being inside that range, or a "-" to
2797 designate being outside that range, so the condition can be inverted by
2798 flipping the prefix. An omitted bound is represented by a "-". For
2799 example, "- [-, 10]" means being outside the range starting at the lowest
2800 possible value and ending at 10, in other words, being greater than 10.
2801 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
2802 always false.
2804 We set up things so that the missing bounds are handled in a consistent
2805 manner so neither a missing bound nor "true" and "false" need to be
2806 handled using a special case. */
2808 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
2809 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
2810 and UPPER1_P are nonzero if the respective argument is an upper bound
2811 and zero for a lower. TYPE, if nonzero, is the type of the result; it
2812 must be specified for a comparison. ARG1 will be converted to ARG0's
2813 type if both are specified. */
2815 static tree
2818 tree type;
2821 {
2822 tree tem;
2826 /* If neither arg represents infinity, do the normal operation.
2827 Else, if not a comparison, return infinity. Else handle the special
2828 comparison rules. Note that most of the cases below won't occur, but
2829 are handled for consistency. */
2832 {
2837 }
2842 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
2843 for neither. In real maths, we cannot assume open ended ranges are
2844 the same. But, this is computer arithmetic, where numbers are finite.
2845 We can therefore make the transformation of any unbounded range with
2846 the value Z, Z being greater than any representable number. This permits
2847 us to treat unbounded ranges as equal. */
2851 {
2872 }
2875 }
2876 \f
2877 /* Given EXP, a logical expression, set the range it is testing into
2878 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
2879 actually being tested. *PLOW and *PHIGH will be made of the same type
2880 as the returned expression. If EXP is not a comparison, we will most
2881 likely not be returning a useful value and range. */
2883 static tree
2885 tree exp;
2888 {
2895 /* Start with simply saying "EXP != 0" and then look at the code of EXP
2896 and see if we can refine the range. Some of the cases below may not
2897 happen, but it doesn't seem worth worrying about this. We "continue"
2898 the outer loop when we've changed something; otherwise we "break"
2899 the switch, which will "break" the while. */
2904 {
2908 {
2919 }
2921 /* Set ORIG_TYPE as soon as TYPE is non-null so that we do not
2922 lose a cast by accident. */
2927 {
2934 /* We can only do something if the range is testing for zero
2935 and if the second operand is an integer constant. Note that
2936 saying something is "in" the range we make is done by
2937 complementing IN_P since it will set in the initial case of
2938 being not equal to zero; "out" is leaving it alone. */
2945 {
2966 }
2970 /* If this is an unsigned comparison, we also know that EXP is
2971 greater than or equal to zero. We base the range tests we make
2972 on that fact, so we record it here so we can parse existing
2973 range tests. */
2975 {
2978 NULL_TREE))
2983 /* If the high bound is missing, but we
2984 have a low bound, reverse the range so
2985 it goes from zero to the low bound minus 1. */
2987 {
2992 }
2993 }
2997 /* (-x) IN [a,b] -> x in [-b, -a] */
3007 /* ~ X -> -X - 1 */
3016 /* If EXP is signed, any overflow in the computation is undefined,
3017 so we don't worry about it so long as our computations on
3018 the bounds don't overflow. For unsigned, overflow is defined
3019 and this is exactly the right thing. */
3028 /* Check for an unsigned range which has wrapped around the maximum
3029 value thus making n_high < n_low, and normalize it. */
3031 {
3037 /* If the range is of the form +/- [ x+1, x ], we won't
3038 be able to normalize it. But then, it represents the
3039 whole range or the empty set, so make it
3040 +/- [ -, - ]. */
3044 else
3046 }
3047 else
3070 /* If we're converting from an unsigned to a signed type,
3071 we will be doing the comparison as unsigned. The tests above
3072 have already verified that LOW and HIGH are both positive.
3074 So we have to make sure that the original unsigned value will
3075 be interpreted as positive. */
3077 {
3080 tree high_positive;
3082 /* A range without an upper bound is, naturally, unbounded.
3083 Since convert would have cropped a very large value, use
3084 the max value for the destination type. */
3085 high_positive
3094 /* If the low bound is specified, "and" the range with the
3095 range for which the original unsigned value will be
3096 positive. */
3098 {
3102 high_positive))
3106 }
3107 else
3108 {
3109 /* Otherwise, "or" the range with the range of the input
3110 that will be interpreted as negative. */
3114 high_positive))
3118 }
3119 }
3127 }
3130 }
3132 /* If EXP is a constant, we can evaluate whether this is true or false. */
3134 {
3141 }
3145 }
3146 \f
3147 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
3148 type, TYPE, return an expression to test if EXP is in (or out of, depending
3149 on IN_P) the range. */
3151 static tree
3153 tree type;
3154 tree exp;
3157 {
3159 tree value;
3178 {
3180 {
3184 }
3186 }
3188 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
3190 {
3192 HOST_WIDE_INT hi;
3197 {
3200 }
3201 else
3202 {
3205 }
3208 {
3210 {
3213 }
3216 }
3217 }
3226 }
3227 \f
3228 /* Given two ranges, see if we can merge them into one. Return 1 if we
3229 can, 0 if we can't. Set the output range into the specified parameters. */
3231 static int
3237 {
3241 tree tem;
3251 /* Make range 0 be the range that starts first, or ends last if they
3252 start at the same value. Swap them if it isn't. */
3255 || (lowequal
3258 {
3262 }
3264 /* Now flag two cases, whether the ranges are disjoint or whether the
3265 second range is totally subsumed in the first. Note that the tests
3266 below are simplified by the ones above. */
3272 /* We now have four cases, depending on whether we are including or
3273 excluding the two ranges. */
3275 {
3276 /* If they don't overlap, the result is false. If the second range
3277 is a subset it is the result. Otherwise, the range is from the start
3278 of the second to the end of the first. */
3283 else
3285 }
3288 {
3289 /* If they don't overlap, the result is the first range. If they are
3290 equal, the result is false. If the second range is a subset of the
3291 first, and the ranges begin at the same place, we go from just after
3292 the end of the first range to the end of the second. If the second
3293 range is not a subset of the first, or if it is a subset and both
3294 ranges end at the same place, the range starts at the start of the
3295 first range and ends just before the second range.
3296 Otherwise, we can't describe this as a single range. */
3302 {
3306 }
3308 {
3312 }
3313 else
3315 }
3318 {
3319 /* If they don't overlap, the result is the second range. If the second
3320 is a subset of the first, the result is false. Otherwise,
3321 the range starts just after the first range and ends at the
3322 end of the second. */
3327 else
3328 {
3332 }
3333 }
3335 else
3336 {
3337 /* The case where we are excluding both ranges. Here the complex case
3338 is if they don't overlap. In that case, the only time we have a
3339 range is if they are adjacent. If the second is a subset of the
3340 first, the result is the first. Otherwise, the range to exclude
3341 starts at the beginning of the first range and ends at the end of the
3342 second. */
3344 {
3351 else
3353 }
3356 else
3358 }
3362 }
3363 \f
3364 /* EXP is some logical combination of boolean tests. See if we can
3365 merge it into some range test. Return the new tree if so. */
3367 static tree
3369 tree exp;
3370 {
3377 tree tem;
3379 /* If this is an OR operation, invert both sides; we will invert
3380 again at the end. */
3384 /* If both expressions are the same, if we can merge the ranges, and we
3385 can build the range test, return it or it inverted. If one of the
3386 ranges is always true or always false, consider it to be the same
3387 expression as the other. */
3397 /* On machines where the branch cost is expensive, if this is a
3398 short-circuited branch and the underlying object on both sides
3399 is the same, make a non-short-circuit operation. */
3405 {
3406 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
3407 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
3408 which cases we can't do this. */
3417 {
3429 }
3430 }
3433 }
3434 \f
3435 /* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
3436 bit value. Arrange things so the extra bits will be set to zero if and
3437 only if C is signed-extended to its full width. If MASK is nonzero,
3438 it is an INTEGER_CST that should be AND'ed with the extra bits. */
3440 static tree
3442 tree c;
3445 tree mask;
3446 {
3449 tree temp;
3454 /* We work by getting just the sign bit into the low-order bit, then
3455 into the high-order bit, then sign-extend. We then XOR that value
3456 with C. */
3460 /* We must use a signed type in order to get an arithmetic right shift.
3461 However, we must also avoid introducing accidental overflows, so that
3462 a subsequent call to integer_zerop will work. Hence we must
3463 do the type conversion here. At this point, the constant is either
3464 zero or one, and the conversion to a signed type can never overflow.
3465 We could get an overflow if this conversion is done anywhere else. */
3473 /* If necessary, convert the type back to match the type of C. */
3478 }
3479 \f
3480 /* Find ways of folding logical expressions of LHS and RHS:
3481 Try to merge two comparisons to the same innermost item.
3482 Look for range tests like "ch >= '0' && ch <= '9'".
3483 Look for combinations of simple terms on machines with expensive branches
3484 and evaluate the RHS unconditionally.
3486 For example, if we have p->a == 2 && p->b == 4 and we can make an
3487 object large enough to span both A and B, we can do this with a comparison
3488 against the object ANDed with the a mask.
3490 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
3491 operations to do this with one comparison.
3493 We check for both normal comparisons and the BIT_AND_EXPRs made this by
3494 function and the one above.
3496 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
3497 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
3499 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
3500 two operands.
3502 We return the simplified tree or 0 if no optimization is possible. */
3504 static tree
3508 {
3509 /* If this is the "or" of two comparisons, we can do something if
3510 the comparisons are NE_EXPR. If this is the "and", we can do something
3511 if the comparisons are EQ_EXPR. I.e.,
3512 (a->b == 2 && a->c == 4) can become (a->new == NEW).
3514 WANTED_CODE is this operation code. For single bit fields, we can
3515 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
3516 comparison for one-bit fields. */
3536 /* Start by getting the comparison codes. Fail if anything is volatile.
3537 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
3538 it were surrounded with a NE_EXPR. */
3563 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
3567 {
3572 {
3580 }
3583 {
3592 }
3593 else
3603 }
3605 /* If the RHS can be evaluated unconditionally and its operands are
3606 simple, it wins to evaluate the RHS unconditionally on machines
3607 with expensive branches. In this case, this isn't a comparison
3608 that can be merged. Avoid doing this if the RHS is a floating-point
3609 comparison since those can trap. */
3615 {
3616 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
3624 integer_zero_node);
3626 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
3634 integer_zero_node);
3637 }
3639 /* See if the comparisons can be merged. Then get all the parameters for
3640 each side. */
3664 /* It must be true that the inner operation on the lhs of each
3665 comparison must be the same if we are to be able to do anything.
3666 Then see if we have constants. If not, the same must be true for
3667 the rhs's. */
3678 else
3681 /* If either comparison code is not correct for our logical operation,
3682 fail. However, we can convert a one-bit comparison against zero into
3683 the opposite comparison against that bit being set in the field. */
3687 {
3689 {
3690 /* Make the left operand unsigned, since we are only interested
3691 in the value of one bit. Otherwise we are doing the wrong
3692 thing below. */
3695 }
3696 else
3698 }
3700 /* This is analogous to the code for l_const above. */
3702 {
3704 {
3707 }
3708 else
3710 }
3712 /* After this point all optimizations will generate bit-field
3713 references, which we might not want. */
3717 /* See if we can find a mode that contains both fields being compared on
3718 the left. If we can't, fail. Otherwise, update all constants and masks
3719 to be relative to a field of that size. */
3724 volatilep);
3734 {
3737 }
3745 {
3753 {
3757 wanted_code == NE_EXPR
3759 }
3760 }
3762 {
3770 {
3774 wanted_code == NE_EXPR
3776 }
3777 }
3779 /* If the right sides are not constant, do the same for it. Also,
3780 disallow this optimization if a size or signedness mismatch occurs
3781 between the left and right sides. */
3783 {
3786 /* Make sure the two fields on the right
3787 correspond to the left without being swapped. */
3795 volatilep);
3805 {
3808 }
3815 /* Make a mask that corresponds to both fields being compared.
3816 Do this for both items being compared. If the operands are the
3817 same size and the bits being compared are in the same position
3818 then we can do this by masking both and comparing the masked
3819 results. */
3823 {
3835 }
3837 /* There is still another way we can do something: If both pairs of
3838 fields being compared are adjacent, we may be able to make a wider
3839 field containing them both.
3841 Note that we still must mask the lhs/rhs expressions. Furthermore,
3842 the mask must be shifted to account for the shift done by
3843 make_bit_field_ref. */
3848 {
3849 tree type;
3861 /* Convert to the smaller type before masking out unwanted bits. */
3864 {
3866 {
3870 }
3872 {
3876 }
3877 }
3886 }
3889 }
3891 /* Handle the case of comparisons with constants. If there is something in
3892 common between the masks, those bits of the constants must be the same.
3893 If not, the condition is always false. Test for this to avoid generating
3894 incorrect code below. */
3899 {
3901 {
3904 }
3905 else
3906 {
3909 }
3910 }
3912 /* Construct the expression we will return. First get the component
3913 reference we will make. Unless the mask is all ones the width of
3914 that field, perform the mask operation. Then compare with the
3915 merged constant. */
3925 }
3926 \f
3927 /* Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a
3928 constant. */
3930 static tree
3932 tree t;
3933 {
3938 tree minmax_const;
3940 tree inner;
3950 /* If something does not permit us to optimize, return the original tree. */
3958 /* Now handle all the various comparison codes. We only handle EQ_EXPR
3959 and GT_EXPR, doing the rest with recursive calls using logical
3960 simplifications. */
3962 {
3964 return
3968 return
3970 optimize_minmax_comparison
3972 optimize_minmax_comparison
3977 /* MAX (X, 0) == 0 -> X <= 0 */
3981 /* MAX (X, 0) == 5 -> X == 5 */
3985 /* MAX (X, 0) == -1 -> false */
3989 /* MIN (X, 0) == 0 -> X >= 0 */
3993 /* MIN (X, 0) == 5 -> false */
3996 else
3997 /* MIN (X, 0) == -1 -> X == -1 */
4002 /* MAX (X, 0) > 0 -> X > 0
4003 MAX (X, 0) > 5 -> X > 5 */
4007 /* MAX (X, 0) > -1 -> true */
4011 /* MIN (X, 0) > 0 -> false
4012 MIN (X, 0) > 5 -> false */
4015 else
4016 /* MIN (X, 0) > -1 -> X > -1 */
4021 }
4022 }
4023 \f
4024 /* T is an integer expression that is being multiplied, divided, or taken a
4025 modulus (CODE says which and what kind of divide or modulus) by a
4026 constant C. See if we can eliminate that operation by folding it with
4027 other operations already in T. WIDE_TYPE, if non-null, is a type that
4028 should be used for the computation if wider than our type.
4030 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
4031 (X * 2) + (Y * 4). We must, however, be assured that either the original
4032 expression would not overflow or that overflow is undefined for the type
4033 in the language in question.
4035 We also canonicalize (X + 7) * 4 into X * 4 + 28 in the hope that either
4036 the machine has a multiply-accumulate insn or that this is part of an
4037 addressing calculation.
4039 If we return a non-null expression, it is an equivalent form of the
4040 original computation, but need not be in the original type. */
4042 static tree
4044 tree t;
4045 tree c;
4047 tree wide_type;
4048 {
4058 /* Don't deal with constants of zero here; they confuse the code below. */
4068 /* Note that we need not handle conditional operations here since fold
4069 already handles those cases. So just do arithmetic here. */
4071 {
4073 /* For a constant, we can always simplify if we are a multiply
4074 or (for divide and modulus) if it is a multiple of our constant. */
4081 /* If op0 is an expression ... */
4086 /* ... and is unsigned, and its type is smaller than ctype,
4087 then we cannot pass through as widening. */
4093 /* ... or its type is larger than ctype,
4094 then we cannot pass through this truncation. */
4099 /* Pass the constant down and see if we can make a simplification. If
4100 we can, replace this expression with the inner simplification for
4101 possible later conversion to our or some other type. */
4113 /* If widening the type changes the signedness, then we can't perform
4114 this optimization as that changes the result. */
4118 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
4121 {
4127 }
4137 /* If this has not been evaluated and the operand has no side effects,
4138 we can see if we can do something inside it and make a new one.
4139 Note that this test is overly conservative since we can do this
4140 if the only reason it had side effects is that it was another
4141 similar SAVE_EXPR, but that isn't worth bothering with. */
4144 wide_type)))
4145 {
4152 }
4156 /* If the second operand is constant, this is a multiplication
4157 or floor division, by a power of two, so we can treat it that
4158 way unless the multiplier or divisor overflows. */
4160 /* const_binop may not detect overflow correctly,
4161 so check for it explicitly here. */
4175 /* See if we can eliminate the operation on both sides. If we can, we
4176 can return a new PLUS or MINUS. If we can't, the only remaining
4177 cases where we can do anything are if the second operand is a
4178 constant. */
4183 /* If not multiplication, we can only do this if both operands
4184 are divisible by c. */
4190 /* If this was a subtraction, negate OP1 and set it to be an addition.
4191 This simplifies the logic below. */
4198 /* If either OP1 or C are negative, this optimization is not safe for
4199 some of the division and remainder types while for others we need
4200 to change the code. */
4202 {
4210 }
4212 /* If it's a multiply or a division/modulus operation of a multiple
4213 of our constant, do the operation and verify it doesn't overflow. */
4216 {
4220 }
4221 else
4224 /* If we have an unsigned type is not a sizetype, we cannot widen
4225 the operation since it will change the result if the original
4226 computation overflowed. */
4232 /* If we were able to eliminate our operation from the first side,
4233 apply our operation to the second side and reform the PLUS. */
4237 /* The last case is if we are a multiply. In that case, we can
4238 apply the distributive law to commute the multiply and addition
4239 if the multiplication of the constants doesn't overflow. */
4244 op1));
4249 /* We have a special case here if we are doing something like
4250 (C * 8) % 4 since we know that's zero. */
4257 /* ... fall through ... */
4261 /* If we can extract our operation from the LHS, do so and return a
4262 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
4263 do something only if the second operand is a constant. */
4275 /* If these are the same operation types, we can associate them
4276 assuming no overflow. */
4283 /* If these operations "cancel" each other, we have the main
4284 optimizations of this pass, which occur when either constant is a
4285 multiple of the other, in which case we replace this with either an
4286 operation or CODE or TCODE.
4288 If we have an unsigned type that is not a sizetype, we cannot do
4289 this since it will change the result if the original computation
4290 overflowed. */
4297 {
4308 }
4313 }
4316 }
4317 \f
4318 /* If T contains a COMPOUND_EXPR which was inserted merely to evaluate
4319 S, a SAVE_EXPR, return the expression actually being evaluated. Note
4320 that we may sometimes modify the tree. */
4322 static tree
4324 tree t;
4325 tree s;
4326 {
4329 /* See if this is the COMPOUND_EXPR we want to eliminate. */
4334 /* See if this is a COND_EXPR or a simple arithmetic operator. We
4335 don't bother handling any other types. */
4337 {
4341 }
4346 {
4349 }
4352 }
4353 \f
4354 /* Return a node which has the indicated constant VALUE (either 0 or
4355 1), and is of the indicated TYPE. */
4357 static tree
4360 tree type;
4361 {
4366 integer_zero_node);
4367 else
4368 {
4373 }
4374 }
4376 /* Utility function for the following routine, to see how complex a nesting of
4377 COND_EXPRs can be. EXPR is the expression and LIMIT is a count beyond which
4378 we don't care (to avoid spending too much time on complex expressions.). */
4380 static int
4382 tree expr;
4384 {
4395 }
4397 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
4398 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
4399 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
4400 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
4401 COND is the first argument to CODE; otherwise (as in the example
4402 given here), it is the second argument. TYPE is the type of the
4403 original expression. */
4405 static tree
4408 tree type;
4409 tree cond;
4410 tree arg;
4412 {
4416 /* In the end, we'll produce a COND_EXPR. Both arms of the
4417 conditional expression will be binary operations. The left-hand
4418 side of the expression to be executed if the condition is true
4419 will be pointed to by TRUE_LHS. Similarly, the right-hand side
4420 of the expression to be executed if the condition is true will be
4421 pointed to by TRUE_RHS. FALSE_LHS and FALSE_RHS are analogous --
4422 but apply to the expression to be executed if the conditional is
4423 false. */
4428 /* These are the codes to use for the left-hand side and right-hand
4429 side of the COND_EXPR. Normally, they are the same as CODE. */
4432 /* And these are the types of the expressions. */
4438 {
4442 }
4443 else
4444 {
4448 }
4451 {
4455 /* If this operand throws an expression, then it does not make
4456 sense to try to perform a logical or arithmetic operation
4457 involving it. Instead of building `a + throw 3' for example,
4458 we simply build `a, throw 3'. */
4460 {
4462 {
4465 }
4466 else
4468 }
4470 {
4472 {
4475 }
4476 else
4478 }
4479 }
4480 else
4481 {
4486 }
4488 /* If ARG is complex we want to make sure we only evaluate
4489 it once. Though this is only required if it is volatile, it
4490 might be more efficient even if it is not. However, if we
4491 succeed in folding one part to a constant, we do not need
4492 to make this SAVE_EXPR. Since we do this optimization
4493 primarily to see if we do end up with constant and this
4494 SAVE_EXPR interferes with later optimizations, suppressing
4495 it when we can is important.
4497 If we are not in a function, we can't make a SAVE_EXPR, so don't
4498 try to do so. Don't try to see if the result is a constant
4499 if an arm is a COND_EXPR since we get exponential behavior
4500 in that case. */
4509 {
4518 {
4522 }
4523 }
4536 else
4538 }
4540 \f
4541 /* Subroutine of fold() that checks for the addition of +/- 0.0.
4543 If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type
4544 TYPE, X + ADDEND is the same as X. If NEGATE, return true if X -
4545 ADDEND is the same as X.
4547 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
4548 and finite. The problematic cases are when X is zero, and its mode
4549 has signed zeros. In the case of rounding towards -infinity,
4550 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
4551 modes, X + 0 is not the same as X because -0 + 0 is 0. */
4553 static bool
4557 {
4561 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
4565 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
4570 /* The mode has signed zeros, and we have to honor their sign.
4571 In this situation, there is only one case we can return true for.
4572 X - 0 is the same as X unless rounding towards -infinity is
4573 supported. */
4575 }
4578 /* Perform constant folding and related simplification of EXPR.
4579 The related simplifications include x*1 => x, x*0 => 0, etc.,
4580 and application of the associative law.
4581 NOP_EXPR conversions may be removed freely (as long as we
4582 are careful not to change the C type of the overall expression)
4583 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
4584 but we can constant-fold them if they have constant operands. */
4586 tree
4588 tree expr;
4589 {
4592 tree tem;
4598 /* WINS will be nonzero when the switch is done
4599 if all operands are constant. */
4602 /* Don't try to process an RTL_EXPR since its operands aren't trees.
4603 Likewise for a SAVE_EXPR that's already been evaluated. */
4607 /* Return right away if a constant. */
4611 #ifdef MAX_INTEGER_COMPUTATION_MODE
4613 #endif
4616 {
4617 tree subop;
4619 /* Special case for conversion ops that can have fixed point args. */
4622 /* Don't use STRIP_NOPS, because signedness of argument type matters. */
4628 else
4633 )
4634 /* Note that TREE_CONSTANT isn't enough:
4635 static var addresses are constant but we can't
4636 do arithmetic on them. */
4638 }
4640 {
4644 {
4646 tree subop;
4652 {
4653 /* Signedness matters here. Perhaps we can refine this
4654 later. */
4656 }
4657 else
4658 /* Strip any conversions that don't change the mode. */
4663 else
4668 /* Note that TREE_CONSTANT isn't enough:
4669 static var addresses are constant but we can't
4670 do arithmetic on them. */
4677 }
4678 }
4680 /* If this is a commutative operation, and ARG0 is a constant, move it
4681 to ARG1 to reduce the number of tests below. */
4686 {
4691 }
4693 /* Now WINS is set as described above,
4694 ARG0 is the first operand of EXPR,
4695 and ARG1 is the second operand (if it has more than one operand).
4697 First check for cases where an arithmetic operation is applied to a
4698 compound, conditional, or comparison operation. Push the arithmetic
4699 operation inside the compound or conditional to see if any folding
4700 can then be done. Convert comparison to conditional for this purpose.
4701 The also optimizes non-constant cases that used to be done in
4702 expand_expr.
4704 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
4705 one of the operands is a comparison and the other is a comparison, a
4706 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
4707 code below would make the expression more complex. Change it to a
4708 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
4709 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
4721 {
4731 }
4734 {
4739 {
4749 /* If this was a conversion, and all we did was to move into
4750 inside the COND_EXPR, bring it back out. But leave it if
4751 it is a conversion from integer to integer and the
4752 result precision is no wider than a word since such a
4753 conversion is cheap and may be optimized away by combine,
4754 while it couldn't if it were outside the COND_EXPR. Then return
4755 so we don't get into an infinite recursion loop taking the
4756 conversion out and then back in. */
4768 && (INTEGRAL_TYPE_P
4779 }
4784 }
4787 {
4802 return
4816 return
4819 }
4830 {
4846 /* Other kinds of FIX are not handled properly by fold_convert. */
4851 /* Handle cases of two conversions in a row. */
4854 {
4874 /* In addition to the cases of two conversions in a row
4875 handled below, if we are converting something to its own
4876 type via an object of identical or wider precision, neither
4877 conversion is needed. */
4883 /* Likewise, if the intermediate and final types are either both
4884 float or both integer, we don't need the middle conversion if
4885 it is wider than the final type and doesn't change the signedness
4886 (for integers). Avoid this if the final type is a pointer
4887 since then we sometimes need the inner conversion. Likewise if
4888 the outer has a precision not equal to the size of its mode. */
4898 /* If we have a sign-extension of a zero-extended value, we can
4899 replace that by a single zero-extension. */
4905 /* Two conversions in a row are not needed unless:
4906 - some conversion is floating-point (overstrict for now), or
4907 - the intermediate type is narrower than both initial and
4908 final, or
4909 - the intermediate type and innermost type differ in signedness,
4910 and the outermost type is wider than the intermediate, or
4911 - the initial type is a pointer type and the precisions of the
4912 intermediate and final types differ, or
4913 - the final type is a pointer type and the precisions of the
4914 initial and intermediate types differ. */
4928 }
4932 /* Detect assigning a bitfield. */
4935 {
4936 /* Don't leave an assignment inside a conversion
4937 unless assigning a bitfield. */
4940 /* First do the assignment, then return converted constant. */
4944 }
4946 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
4947 constants (if x has signed type, the sign bit cannot be set
4948 in c). This folds extension into the BIT_AND_EXPR. */
4953 {
4963 <= HOST_BITS_PER_WIDE_INT
4965 {
4972 #ifdef LOAD_EXTEND_OP
4976 {
4980 }
4981 #endif
4982 }
4987 }
4990 {
4993 }
5004 {
5008 }
5017 {
5019 {
5021 HOST_WIDE_INT high;
5032 }
5035 }
5038 /* Convert -((double)float) into (double)(-float). */
5041 {
5046 }
5048 /* Convert - (a - b) to (b - a) for non-floating-point. */
5058 {
5060 {
5061 /* If the value is unsigned, then the absolute value is
5062 the same as the ordinary value. */
5065 /* Similarly, if the value is non-negative. */
5068 /* If the value is negative, then the absolute value is
5069 its negation. */
5070 else
5071 {
5073 HOST_WIDE_INT high;
5084 }
5085 }
5087 {
5091 }
5092 }
5095 /* Convert fabs((double)float) into (double)fabsf(float). */
5098 {
5103 }
5104 else
5105 {
5106 /* fabs(sqrt(x)) = sqrt(x) and fabs(exp(x)) = exp(x). */
5115 }
5140 {
5147 }
5153 /* A + (-B) -> A - B */
5156 /* (-A) + B -> B - A */
5160 {
5164 /* If we are adding two BIT_AND_EXPR's, both of which are and'ing
5165 with a constant, and the two constants have no bits in common,
5166 we should treat this as a BIT_IOR_EXPR since this may produce more
5167 simplifications. */
5175 {
5178 }
5180 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
5181 (plus (plus (mult) (mult)) (foo)) so that we can
5182 take advantage of the factoring cases below. */
5187 {
5192 else
5203 parg1));
5208 parg0));
5209 }
5212 {
5216 /* (A * C) + (B * C) -> (A+B) * C.
5217 We are most concerned about the case where C is a constant,
5218 but other combinations show up during loop reduction. Since
5219 it is not difficult, try all four possibilities. */
5236 /* No identical multiplicands; see if we can find a common
5237 power-of-two factor in non-power-of-two multiplies. This
5238 can help in multi-dimensional array access. */
5243 {
5248 /* Move min of absolute values to int11. */
5251 {
5255 }
5258 {
5263 }
5264 }
5269 same));
5270 }
5271 }
5273 /* See if ARG1 is zero and X + ARG1 reduces to X. */
5277 /* Likewise if the operands are reversed. */
5281 bit_rotate:
5282 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
5283 is a rotate of A by C1 bits. */
5284 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
5285 is a rotate of A by B bits. */
5286 {
5295 {
5314 {
5322 TYPE_PRECISION
5327 ? LROTATE_EXPR
5330 }
5332 {
5340 TYPE_PRECISION
5345 ? LROTATE_EXPR
5348 }
5349 }
5350 }
5352 associate:
5353 /* In most languages, can't associate operations on floats through
5354 parentheses. Rather than remember where the parentheses were, we
5355 don't associate floats at all. It shouldn't matter much. However,
5356 associating multiplications is only very slightly inaccurate, so do
5357 that if -funsafe-math-optimizations is specified. */
5362 {
5366 /* Split both trees into variables, constants, and literals. Then
5367 associate each group together, the constants with literals,
5368 then the result with variables. This increases the chances of
5369 literals being recombined later and of generating relocatable
5370 expressions for the sum of a constant and literal. */
5375 /* Only do something if we found more than two objects. Otherwise,
5376 nothing has changed and we risk infinite recursion. */
5381 {
5382 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
5391 /* Preserve the MINUS_EXPR if the negative part of the literal is
5392 greater than the positive part. Otherwise, the multiplicative
5393 folding code (i.e extract_muldiv) may be fooled in case
5394 unsigned constants are substracted, like in the following
5395 example: ((X*2 + 4) - 8U)/2. */
5397 {
5399 {
5403 }
5404 else
5405 {
5409 }
5410 }
5412 {
5416 else
5417 {
5422 }
5423 }
5427 }
5428 }
5430 binary:
5434 {
5435 /* The return value should always have
5436 the same type as the original expression. */
5441 }
5445 /* A - (-B) -> A + B */
5448 /* (-A) - CST -> (-CST) - A for floating point (what about ints ?) */
5450 return
5457 {
5463 /* (A * C) - (B * C) -> (A-B) * C. Since we are most concerned
5464 about the case where C is a constant, just try one of the
5465 four possibilities. */
5475 }
5477 /* See if ARG1 is zero and X - ARG1 reduces to X. */
5481 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
5482 ARG0 is zero and X + ARG0 reduces to X, since that would mean
5483 (-ARG1 + ARG0) reduces to -ARG1. */
5487 /* Fold &x - &x. This can happen from &x.foo - &x.
5488 This is unsafe for certain floats even in non-IEEE formats.
5489 In IEEE, it is unsafe because it does wrong for NaNs.
5490 Also note that operand_equal_p is always false if an operand
5491 is volatile. */
5500 /* (-A) * (-B) -> A * B */
5506 {
5512 /* (a * (1 << b)) is (a << b) */
5527 }
5528 else
5529 {
5530 /* Maybe fold x * 0 to 0. The expressions aren't the same
5531 when x is NaN, since x * 0 is also NaN. Nor are they the
5532 same in modes with signed zeros, since multiplying a
5533 negative value by 0 gives -0, not +0. */
5538 /* In IEEE floating point, x*1 is not equivalent to x for snans. */
5543 /* Transform x * -1.0 into -x. */
5548 /* x*2 is x+x */
5552 {
5555 }
5558 {
5562 /* Optimize sqrt(x)*sqrt(y) as sqrt(x*y). */
5566 {
5573 }
5575 /* Optimize exp(x)*exp(y) as exp(x+y). */
5579 {
5586 }
5587 }
5588 }
5592 bit_ior:
5601 /* Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
5603 This results in more efficient code for machines without a NAND
5604 instruction. Combine will canonicalize to the first form
5605 which will allow use of NAND instructions provided by the
5606 backend if they exist. */
5609 {
5614 }
5616 /* See if this can be simplified into a rotate first. If that
5617 is unsuccessful continue in the association code. */
5626 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
5627 with a constant, and the two constants have no bits in common,
5628 we should treat this as a BIT_IOR_EXPR since this may produce more
5629 simplifications. */
5637 {
5640 }
5642 /* See if this can be simplified into a rotate first. If that
5643 is unsuccessful continue in the association code. */
5647 bit_and:
5655 /* Simplify ((int)c & 0x377) into (int)c, if c is unsigned char. */
5658 {
5659 unsigned int prec
5666 }
5668 /* Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
5670 This results in more efficient code for machines without a NOR
5671 instruction. Combine will canonicalize to the first form
5672 which will allow use of NOR instructions provided by the
5673 backend if they exist. */
5676 {
5681 }
5691 {
5695 }
5699 /* Don't touch a floating-point divide by zero unless the mode
5700 of the constant can represent infinity. */
5706 /* (-A) / (-B) -> A / B */
5711 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
5716 /* If ARG1 is a constant, we can convert this to a multiply by the
5717 reciprocal. This does not have the same rounding properties,
5718 so only do this if -funsafe-math-optimizations. We can actually
5719 always safely do it if ARG1 is a power of two, but it's hard to
5720 tell if it is or not in a portable manner. */
5722 {
5727 /* Find the reciprocal if optimizing and the result is exact. */
5729 {
5730 REAL_VALUE_TYPE r;
5733 {
5736 }
5737 }
5738 }
5739 /* Convert A/B/C to A/(B*C). */
5742 {
5745 arg1)));
5746 }
5747 /* Convert A/(B/C) to (A/B)*C. */
5750 {
5755 }
5757 /* Optimize x/exp(y) into x*exp(-y). */
5759 {
5764 {
5771 }
5772 }
5785 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
5786 operation, EXACT_DIV_EXPR.
5788 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
5789 At one time others generated faster code, it's not clear if they do
5790 after the last round to changes to the DIV code in expmed.c. */
5825 /* Optimize -1 >> x for arithmetic right shifts. */
5828 /* ... fall through ... */
5831 shift:
5837 /* Since negative shift count is not well-defined,
5838 don't try to compute it in the compiler. */
5841 /* Rewrite an LROTATE_EXPR by a constant into an
5842 RROTATE_EXPR by a new constant. */
5844 {
5848 = const_binop
5855 }
5857 /* If we have a rotate of a bit operation with the rotate count and
5858 the second operand of the bit operation both constant,
5859 permute the two operations. */
5872 /* Two consecutive rotates adding up to the width of the mode can
5873 be ignored. */
5904 /* Note that the operand of this must be an int
5905 and its values must be 0 or 1.
5906 ("true" is a fixed value perhaps depending on the language,
5907 but we don't handle values other than 1 correctly yet.) */
5909 /* Avoid infinite recursion. */
5915 /* Note that the operands of this must be ints
5916 and their values must be 0 or 1.
5917 ("true" is a fixed value perhaps depending on the language.) */
5918 /* If first arg is constant zero, return it. */
5922 /* If either arg is constant true, drop it. */
5926 /* Preserve sequence points. */
5929 /* If second arg is constant zero, result is zero, but first arg
5930 must be evaluated. */
5933 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
5934 case will be handled here. */
5938 truth_andor:
5939 /* We only do these simplifications if we are optimizing. */
5943 /* Check for things like (A || B) && (A || C). We can convert this
5944 to A || (B && C). Note that either operator can be any of the four
5945 truth and/or operations and the transformation will still be
5946 valid. Also note that we only care about order for the
5947 ANDIF and ORIF operators. If B contains side effects, this
5948 might change the truth-value of A. */
5955 {
5975 /* This case if tricky because we must either have commutative
5976 operators or else A10 must not have side-effects. */
5982 a01));
5983 }
5985 /* See if we can build a range comparison. */
5989 /* Check for the possibility of merging component references. If our
5990 lhs is another similar operation, try to merge its rhs with our
5991 rhs. Then try to merge our lhs and rhs. */
6003 /* Note that the operands of this must be ints
6004 and their values must be 0 or true.
6005 ("true" is a fixed value perhaps depending on the language.) */
6006 /* If first arg is constant true, return it. */
6010 /* If either arg is constant zero, drop it. */
6014 /* Preserve sequence points. */
6017 /* If second arg is constant true, result is true, but we must
6018 evaluate first arg. */
6021 /* Likewise for first arg, but note this only occurs here for
6022 TRUTH_OR_EXPR. */
6028 /* If either arg is constant zero, drop it. */
6033 /* If either arg is constant true, this is a logical inversion. */
6046 /* If one arg is a real or integer constant, put it last. */
6051 {
6058 }
6061 {
6069 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
6074 /* (-a) CMP (-b) -> b CMP a */
6079 /* (-a) CMP CST -> a swap(CMP) (-CST) */
6081 return
6087 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
6088 /* a CMP (-0) -> a CMP 0 */
6094 /* If this is a comparison of a real constant with a PLUS_EXPR
6095 or a MINUS_EXPR of a real constant, we can convert it into a
6096 comparison with a revised real constant as long as no overflow
6097 occurs when unsafe_math_optimizations are enabled. */
6108 }
6110 /* Convert foo++ == CONST into ++foo == CONST + INCR.
6111 First, see if one arg is constant; find the constant arg
6112 and the other one. */
6113 {
6123 {
6124 /* This optimization is invalid for ordered comparisons
6125 if CONST+INCR overflows or if foo+incr might overflow.
6126 This optimization is invalid for floating point due to rounding.
6127 For pointer types we assume overflow doesn't happen. */
6131 {
6132 tree newconst
6136 /* Do not overwrite the current varop to be a preincrement,
6137 create a new node so that we won't confuse our caller who
6138 might create trees and throw them away, reusing the
6139 arguments that they passed to build. This shows up in
6140 the THEN or ELSE parts of ?: being postincrements. */
6145 /* If VAROP is a reference to a bitfield, we must mask
6146 the constant by the width of the field. */
6150 {
6151 int size
6153 (TREE_OPERAND
6157 tree folded_compare;
6159 /* First check whether the comparison would come out
6160 always the same. If we don't do that we would
6161 change the meaning with the masking. */
6165 else
6168 constop));
6183 mask)));
6184 }
6190 }
6191 }
6193 {
6197 {
6198 tree newconst
6202 /* Do not overwrite the current varop to be a predecrement,
6203 create a new node so that we won't confuse our caller who
6204 might create trees and throw them away, reusing the
6205 arguments that they passed to build. This shows up in
6206 the THEN or ELSE parts of ?: being postdecrements. */
6214 {
6215 int size
6217 (TREE_OPERAND
6221 tree folded_compare;
6226 else
6229 constop));
6244 mask)));
6245 }
6251 }
6252 }
6253 }
6255 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
6256 This transformation affects the cases which are handled in later
6257 optimizations involving comparisons with non-negative constants. */
6261 {
6263 {
6278 }
6279 }
6281 /* Comparisons with the highest or lowest possible integer of
6282 the specified size will have known values. */
6283 {
6291 {
6298 {
6301 }
6302 else
6303 {
6306 }
6311 {
6315 arg0);
6323 arg0);
6329 /* The GE_EXPR and LT_EXPR cases above are not normally
6330 reached because of previous transformations. */
6334 }
6338 {
6351 }
6355 {
6359 arg0);
6368 arg0);
6376 }
6380 {
6393 }
6398 /* signed_type does not work on pointer types. */
6400 {
6401 /* The following case also applies to X < signed_max+1
6402 and X >= signed_max+1 because previous transformations. */
6404 {
6408 return fold
6412 }
6413 }
6414 }
6415 }
6417 /* If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
6418 a MINUS_EXPR of a constant, we can convert it into a comparison with
6419 a revised constant as long as no overflow occurs. */
6431 /* Similarly for a NEGATE_EXPR. */
6440 /* If we have X - Y == 0, we can convert that to X == Y and similarly
6441 for !=. Don't do this for ordered comparisons due to overflow. */
6447 /* If we are widening one operand of an integer comparison,
6448 see if the other operand is similarly being widened. Perhaps we
6449 can do the comparison in the narrower type. */
6459 /* If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a
6460 constant, we can simplify it. */
6467 /* If we are comparing an ABS_EXPR with a constant, we can
6468 convert all the cases into explicit comparisons, but they may
6469 well not be faster than doing the ABS and one comparison.
6470 But ABS (X) <= C is a range comparison, which becomes a subtraction
6471 and a comparison, and is probably faster. */
6483 /* If this is an EQ or NE comparison with zero and ARG0 is
6484 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
6485 two operations, but the latter can be done in one less insn
6486 on machines that have only two-operand insns or on which a
6487 constant cannot be the first operand. */
6490 {
6493 return
6501 integer_one_node)),
6502 arg1));
6505 return
6513 integer_one_node)),
6514 arg1));
6515 }
6517 /* If this is an NE or EQ comparison of zero against the result of a
6518 signed MOD operation whose second operand is a power of 2, make
6519 the MOD operation unsigned since it is simpler and equivalent. */
6528 {
6535 }
6537 /* If this is an NE comparison of zero with an AND of one, remove the
6538 comparison since the AND will give the correct value. */
6544 /* If we have (A & C) == C where C is a power of 2, convert this into
6545 (A & C) != 0. Similarly for NE_EXPR. */
6553 /* If we have (A & C) != 0 where C is the sign bit of A, convert
6554 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
6558 {
6562 {
6567 }
6568 }
6570 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
6571 and similarly for >= into !=. */
6587 return
6594 /* Simplify comparison of something with itself. (For IEEE
6595 floating-point, we can only do some of these simplifications.) */
6597 {
6599 {
6610 /* For NE, we can only do this simplification if integer. */
6613 /* ... fall through ... */
6619 }
6620 }
6622 /* If we are comparing an expression that just has comparisons
6623 of two integer values, arithmetic expressions of those comparisons,
6624 and constants, we can simplify it. There are only three cases
6625 to check: the two values can either be equal, the first can be
6626 greater, or the second can be greater. Fold the expression for
6627 those three values. Since each value must be 0 or 1, we have
6628 eight possibilities, each of which corresponds to the constant 0
6629 or 1 or one of the six possible comparisons.
6631 This handles common cases like (a > b) == 0 but also handles
6632 expressions like ((x > y) - (y > x)) > 0, which supposedly
6633 occur in macroized code. */
6636 {
6641 /* Don't handle degenerate cases here; they should already
6642 have been handled anyway. */
6651 {
6655 /* We can't just pass T to eval_subst in case cval1 or cval2
6656 was the same as ARG1. */
6658 tree high_result
6661 arg1));
6662 tree equal_result
6665 arg1));
6666 tree low_result
6669 arg1));
6671 /* All three of these results should be 0 or 1. Confirm they
6672 are. Then use those values to select the proper code
6673 to use. */
6681 {
6682 /* Make a 3-bit mask with the high-order bit being the
6683 value for `>', the next for '=', and the low for '<'. */
6687 {
6689 /* Always false. */
6710 /* Always true. */
6712 }
6717 else
6719 }
6720 }
6721 }
6723 /* If this is a comparison of a field, we may be able to simplify it. */
6728 /* Handle the constant case even without -O
6729 to make sure the warnings are given. */
6731 {
6734 }
6736 /* If this is a comparison of complex values and either or both sides
6737 are a COMPLEX_EXPR or COMPLEX_CST, it is best to split up the
6738 comparisons and join them with a TRUTH_ANDIF_EXPR or TRUTH_ORIF_EXPR.
6739 This may prevent needless evaluations. */
6746 {
6759 type,
6762 }
6764 /* Optimize comparisons of strlen vs zero to a compare of the
6765 first character of the string vs zero. To wit,
6766 strlen(ptr) == 0 => *ptr == 0
6767 strlen(ptr) != 0 => *ptr != 0
6768 Other cases should reduce to one of these two (or a constant)
6769 due to the return value of strlen being unsigned. */
6774 {
6776 tree arglist;
6788 integer_zero_node));
6789 }
6791 /* From here on, the only cases we handle are when the result is
6792 known to be a constant.
6794 To compute GT, swap the arguments and do LT.
6795 To compute GE, do LT and invert the result.
6796 To compute LE, swap the arguments, do LT and invert the result.
6797 To compute NE, do EQ and invert the result.
6799 Therefore, the code below must handle only EQ and LT. */
6802 {
6805 }
6807 /* Note that it is safe to invert for real values here because we
6808 will check below in the one case that it matters. */
6813 {
6816 }
6818 /* Compute a result for LT or EQ if args permit;
6819 otherwise return T. */
6821 {
6824 else
6829 }
6832 /* Assume a nonexplicit constant cannot equal an explicit one,
6833 since such code would be undefined anyway.
6834 Exception: on sysvr4, using #pragma weak,
6835 a label can come out as 0. */
6842 #endif
6843 /* Two real constants can be compared explicitly. */
6845 {
6846 /* If either operand is a NaN, the result is false with two
6847 exceptions: First, an NE_EXPR is true on NaNs, but that case
6848 is already handled correctly since we will be inverting the
6849 result for NE_EXPR. Second, if we had inverted a LE_EXPR
6850 or a GE_EXPR into a LT_EXPR, we must return true so that it
6851 will be inverted into false. */
6861 else
6865 }
6879 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
6880 so all simple results must be passed through pedantic_non_lvalue. */
6882 return pedantic_non_lvalue
6887 /* If the second operand is zero, invert the comparison and swap
6888 the second and third operands. Likewise if the second operand
6889 is constant and the third is not or if the third operand is
6890 equivalent to the first operand of the comparison. */
6898 {
6899 /* See if this can be inverted. If it can't, possibly because
6900 it was a floating-point inequality comparison, don't do
6901 anything. */
6905 {
6909 /* arg1 should be the first argument of the new T. */
6912 }
6913 }
6915 /* If we have A op B ? A : C, we may be able to convert this to a
6916 simpler expression, depending on the operation and the values
6917 of B and C. Signed zeros prevent all of these transformations,
6918 for reasons given above each one. */
6924 {
6930 /* If we have A op 0 ? A : -A, consider applying the following
6931 transformations:
6933 A == 0? A : -A same as -A
6934 A != 0? A : -A same as A
6935 A >= 0? A : -A same as abs (A)
6936 A > 0? A : -A same as abs (A)
6937 A <= 0? A : -A same as -abs (A)
6938 A < 0? A : -A same as -abs (A)
6940 None of these transformations work for modes with signed
6941 zeros. If A is +/-0, the first two transformations will
6942 change the sign of the result (from +0 to -0, or vice
6943 versa). The last four will fix the sign of the result,
6944 even though the original expressions could be positive or
6945 negative, depending on the sign of A.
6947 Note that all these transformations are correct if A is
6948 NaN, since the two alternatives (A and -A) are also NaNs. */
6955 {
6957 return
6958 pedantic_non_lvalue
6960 negate_expr
6962 arg1))));
6970 return pedantic_non_lvalue
6978 return pedantic_non_lvalue
6982 arg1)))));
6985 }
6987 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
6988 A == 0 ? A : 0 is always 0 unless A is -0. Note that
6989 both transformations are correct when A is NaN: A != 0
6990 is then true, and A == 0 is false. */
6993 {
6998 }
7000 /* Try some transformations of A op B ? A : B.
7002 A == B? A : B same as B
7003 A != B? A : B same as A
7004 A >= B? A : B same as max (A, B)
7005 A > B? A : B same as max (B, A)
7006 A <= B? A : B same as min (A, B)
7007 A < B? A : B same as min (B, A)
7009 As above, these transformations don't work in the presence
7010 of signed zeros. For example, if A and B are zeros of
7011 opposite sign, the first two transformations will change
7012 the sign of the result. In the last four, the original
7013 expressions give different results for (A=+0, B=-0) and
7014 (A=-0, B=+0), but the transformed expressions do not.
7016 The first two transformations are correct if either A or B
7017 is a NaN. In the first transformation, the condition will
7018 be false, and B will indeed be chosen. In the case of the
7019 second transformation, the condition A != B will be true,
7020 and A will be chosen.
7022 The conversions to max() and min() are not correct if B is
7023 a number and A is not. The conditions in the original
7024 expressions will be false, so all four give B. The min()
7025 and max() versions would give a NaN instead. */
7028 {
7033 /* Avoid adding NOP_EXPRs in case this is an lvalue. */
7035 {
7039 }
7042 {
7049 /* In C++ a ?: expression can be an lvalue, so put the
7050 operand which will be used if they are equal first
7051 so that we can convert this back to the
7052 corresponding COND_EXPR. */
7054 return pedantic_non_lvalue
7064 return pedantic_non_lvalue
7073 }
7074 }
7076 /* If this is A op C1 ? A : C2 with C1 and C2 constant integers,
7077 we might still be able to simplify this. For example,
7078 if C1 is one less or one more than C2, this might have started
7079 out as a MIN or MAX and been transformed by this function.
7080 Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE. */
7086 {
7088 /* We can replace A with C1 in this case. */
7095 /* If C1 is C2 + 1, this is min(A, C2). */
7100 return pedantic_non_lvalue
7105 /* If C1 is C2 - 1, this is min(A, C2). */
7110 return pedantic_non_lvalue
7115 /* If C1 is C2 - 1, this is max(A, C2). */
7120 return pedantic_non_lvalue
7125 /* If C1 is C2 + 1, this is max(A, C2). */
7130 return pedantic_non_lvalue
7137 }
7138 }
7140 /* If the second operand is simpler than the third, swap them
7141 since that produces better jump optimization results. */
7147 {
7148 /* See if this can be inverted. If it can't, possibly because
7149 it was a floating-point inequality comparison, don't do
7150 anything. */
7154 {
7158 /* arg1 should be the first argument of the new T. */
7161 }
7162 }
7164 /* Convert A ? 1 : 0 to simply A. */
7167 /* If we try to convert TREE_OPERAND (t, 0) to our type, the
7168 call to fold will try to move the conversion inside
7169 a COND, which will recurse. In that case, the COND_EXPR
7170 is probably the best choice, so leave it alone. */
7174 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
7175 over COND_EXPR in cases such as floating point comparisons. */
7182 /* Look for expressions of the form A & 2 ? 2 : 0. The result of this
7183 operation is simply A & 2. */
7194 /* Convert A ? B : 0 into A && B if A and B are truth values. */
7201 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
7205 {
7206 /* Only perform transformation if ARG0 is easily inverted. */
7211 }
7216 /* When pedantic, a compound expression can be neither an lvalue
7217 nor an integer constant expression. */
7220 /* Don't let (0, 0) be null pointer constant. */
7262 /* Pull arithmetic ops out of the CLEANUP_POINT_EXPR where
7263 appropriate. */
7268 {
7272 tree arg01;
7283 {
7297 arg01));
7298 }
7301 }
7304 /* Check for a built-in function. */
7309 {
7313 }
7319 }
7321 /* Determine if first argument is a multiple of second argument. Return 0 if
7322 it is not, or we cannot easily determined it to be.
7324 An example of the sort of thing we care about (at this point; this routine
7325 could surely be made more general, and expanded to do what the *_DIV_EXPR's
7326 fold cases do now) is discovering that
7328 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
7330 is a multiple of
7332 SAVE_EXPR (J * 8)
7334 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
7336 This code also handles discovering that
7338 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
7340 is a multiple of 8 so we don't have to worry about dealing with a
7341 possible remainder.
7343 Note that we *look* inside a SAVE_EXPR only to determine how it was
7344 calculated; it is not safe for fold to do much of anything else with the
7345 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
7346 at run time. For example, the latter example above *cannot* be implemented
7347 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
7348 evaluation time of the original SAVE_EXPR is not necessarily the same at
7349 the time the new expression is evaluated. The only optimization of this
7350 sort that would be valid is changing
7352 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
7354 divided by 8 to
7356 SAVE_EXPR (I) * SAVE_EXPR (J)
7358 (where the same SAVE_EXPR (J) is used in the original and the
7359 transformed version). */
7361 static int
7363 tree type;
7364 tree top;
7365 tree bottom;
7366 {
7374 {
7386 {
7390 /* const_binop may not detect overflow correctly,
7391 so check for it explicitly here. */
7400 }
7404 /* Can't handle conversions from non-integral or wider integral type. */
7410 /* .. fall through ... */
7426 }
7427 }
7429 /* Return true if `t' is known to be non-negative. */
7431 int
7433 tree t;
7434 {
7436 {
7477 /* Truth values evaluate to 0 or 1, which is nonnegative. */
7479 else
7480 /* We don't know sign of `t', so be conservative and return false. */
7482 }
7483 }
7485 /* Return true if `r' is known to be non-negative.
7486 Only handles constants at the moment. */
7488 int
7490 rtx r;
7491 {
7493 {
7503 {
7505 rtx elt;
7510 {
7514 }
7517 }
7521 /* These are always nonnegative. */
7526 }
7527 }