| blob | 0f9f3d419a57074e4bd8eb59dfea0859e744b71c |
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 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 *,
112 static tree fold_binary_op_with_conditional_arg
118 /* The following constants represent a bit based encoding of GCC's
119 comparison operators. This encoding simplifies transformations
120 on relational comparison operators, such as AND and OR. */
121 #define COMPCODE_FALSE 0
122 #define COMPCODE_LT 1
123 #define COMPCODE_EQ 2
124 #define COMPCODE_LE 3
125 #define COMPCODE_GT 4
126 #define COMPCODE_NE 5
127 #define COMPCODE_GE 6
128 #define COMPCODE_TRUE 7
130 /* We know that A1 + B1 = SUM1, using 2's complement arithmetic and ignoring
131 overflow. Suppose A, B and SUM have the same respective signs as A1, B1,
132 and SUM1. Then this yields nonzero if overflow occurred during the
133 addition.
135 Overflow occurs if A and B have the same sign, but A and SUM differ in
136 sign. Use `^' to test whether signs differ, and `< 0' to isolate the
137 sign. */
138 #define OVERFLOW_SUM_SIGN(a, b, sum) ((~((a) ^ (b)) & ((a) ^ (sum))) < 0)
139 \f
140 /* To do constant folding on INTEGER_CST nodes requires two-word arithmetic.
141 We do that by representing the two-word integer in 4 words, with only
142 HOST_BITS_PER_WIDE_INT / 2 bits stored in each word, as a positive
143 number. The value of the word is LOWPART + HIGHPART * BASE. */
145 #define LOWPART(x) \
146 ((x) & (((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) - 1))
147 #define HIGHPART(x) \
148 ((unsigned HOST_WIDE_INT) (x) >> HOST_BITS_PER_WIDE_INT / 2)
149 #define BASE ((unsigned HOST_WIDE_INT) 1 << HOST_BITS_PER_WIDE_INT / 2)
151 /* Unpack a two-word integer into 4 words.
152 LOW and HI are the integer, as two `HOST_WIDE_INT' pieces.
153 WORDS points to the array of HOST_WIDE_INTs. */
155 static void
159 HOST_WIDE_INT hi;
160 {
165 }
167 /* Pack an array of 4 words into a two-word integer.
168 WORDS points to the array of words.
169 The integer is stored into *LOW and *HI as two `HOST_WIDE_INT' pieces. */
171 static void
176 {
179 }
180 \f
181 /* Make the integer constant T valid for its type by setting to 0 or 1 all
182 the bits in the constant that don't belong in the type.
184 Return 1 if a signed overflow occurs, 0 otherwise. If OVERFLOW is
185 nonzero, a signed overflow has already occurred in calculating T, so
186 propagate it. */
188 int
190 tree t;
192 {
194 HOST_WIDE_INT high;
198 {
199 /* ??? Used to check for overflow here via CHECK_FLOAT_TYPE.
200 Consider doing it via real_convert now. */
202 }
212 else
215 /* First clear all bits that are beyond the type's precision. */
218 ;
222 else
223 {
227 }
229 /* Unsigned types do not suffer sign extension or overflow unless they
230 are a sizetype. */
236 /* If the value's sign bit is set, extend the sign. */
244 {
245 /* Value is negative:
246 set to 1 all the bits that are outside this type's precision. */
250 else
251 {
255 }
256 }
258 /* Return nonzero if signed overflow occurred. */
259 return
262 }
263 \f
264 /* Add two doubleword integers with doubleword result.
265 Each argument is given as two `HOST_WIDE_INT' pieces.
266 One argument is L1 and H1; the other, L2 and H2.
267 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
269 int
275 {
277 HOST_WIDE_INT h;
285 }
287 /* Negate a doubleword integer with doubleword result.
288 Return nonzero if the operation overflows, assuming it's signed.
289 The argument is given as two `HOST_WIDE_INT' pieces in L1 and H1.
290 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
292 int
295 HOST_WIDE_INT h1;
298 {
300 {
304 }
305 else
306 {
310 }
311 }
312 \f
313 /* Multiply two doubleword integers with doubleword result.
314 Return nonzero if the operation overflows, assuming it's signed.
315 Each argument is given as two `HOST_WIDE_INT' pieces.
316 One argument is L1 and H1; the other, L2 and H2.
317 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
319 int
325 {
340 {
343 {
345 /* This product is <= 0xFFFE0001, the sum <= 0xFFFF0000. */
347 /* Since prod[p] < 0xFFFF, this sum <= 0xFFFFFFFF. */
351 }
353 }
357 /* Check for overflow by calculating the top half of the answer in full;
358 it should agree with the low half's sign bit. */
361 {
364 }
366 {
369 }
371 }
372 \f
373 /* Shift the doubleword integer in L1, H1 left by COUNT places
374 keeping only PREC bits of result.
375 Shift right if COUNT is negative.
376 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
377 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
379 void
387 {
391 {
394 }
396 #ifdef SHIFT_COUNT_TRUNCATED
399 #endif
402 {
403 /* Shifting by the host word size is undefined according to the
404 ANSI standard, so we must handle this as a special case. */
407 }
409 {
412 }
413 else
414 {
418 }
420 /* Sign extend all bits that are beyond the precision. */
428 ;
430 {
433 }
434 else
435 {
439 }
440 }
442 /* Shift the doubleword integer in L1, H1 right by COUNT places
443 keeping only PREC bits of result. COUNT must be positive.
444 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
445 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
447 void
455 {
458 signmask = (arith
462 #ifdef SHIFT_COUNT_TRUNCATED
465 #endif
468 {
469 /* Shifting by the host word size is undefined according to the
470 ANSI standard, so we must handle this as a special case. */
473 }
475 {
478 }
479 else
480 {
484 }
486 /* Zero / sign extend all bits that are beyond the precision. */
489 {
492 }
494 ;
496 {
499 }
500 else
501 {
505 }
506 }
507 \f
508 /* Rotate the doubleword integer in L1, H1 left by COUNT places
509 keeping only PREC bits of result.
510 Rotate right if COUNT is negative.
511 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
513 void
520 {
532 }
534 /* Rotate the doubleword integer in L1, H1 left by COUNT places
535 keeping only PREC bits of result. COUNT must be positive.
536 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
538 void
545 {
557 }
558 \f
559 /* Divide doubleword integer LNUM, HNUM by doubleword integer LDEN, HDEN
560 for a quotient (stored in *LQUO, *HQUO) and remainder (in *LREM, *HREM).
561 CODE is a tree code for a kind of division, one of
562 TRUNC_DIV_EXPR, FLOOR_DIV_EXPR, CEIL_DIV_EXPR, ROUND_DIV_EXPR
563 or EXACT_DIV_EXPR
564 It controls how the quotient is rounded to an integer.
565 Return nonzero if the operation overflows.
566 UNS nonzero says do unsigned division. */
568 int
575 HOST_WIDE_INT hnum_orig;
577 HOST_WIDE_INT hden_orig;
580 {
596 /* calculate quotient sign and convert operands to unsigned. */
598 {
600 {
602 /* (minimum integer) / (-1) is the only overflow case. */
606 }
608 {
611 }
612 }
617 /* This unsigned division rounds toward zero. */
620 }
624 /* hden != 0 already checked. */
629 }
639 /* Special code for when the divisor < BASE. */
641 {
642 /* hnum != 0 already checked. */
644 {
648 }
649 }
650 else
651 {
652 /* Full double precision division,
653 with thanks to Don Knuth's "Seminumerical Algorithms". */
657 /* Find the highest nonzero divisor digit. */
660 {
663 }
665 /* Insure that the first digit of the divisor is at least BASE/2.
666 This is required by the quotient digit estimation algorithm. */
673 {
677 }
682 {
687 }
688 }
692 /* Main loop */
694 {
695 /* Guess the next quotient digit, quo_est, by dividing the first
696 two remaining dividend digits by the high order quotient digit.
697 quo_est is never low and is at most 2 high. */
704 else
707 /* Refine quo_est so it's usually correct, and at most one high. */
712 quo_est--;
714 /* Try QUO_EST as the quotient digit, by multiplying the
715 divisor by QUO_EST and subtracting from the remaining dividend.
716 Keep in mind that QUO_EST is the I - 1st digit. */
720 {
726 }
728 /* If quo_est was high by one, then num[i] went negative and
729 we need to correct things. */
731 {
732 quo_est--;
735 {
739 }
742 }
744 /* Store the quotient digit. */
746 }
747 }
751 finish_up:
752 /* if result is negative, make it so. */
756 /* compute trial remainder: rem = num - (quo * den) */
762 {
771 {
772 /* quo = quo - 1; */
775 }
776 else
783 {
786 }
787 else
793 {
799 /* Get absolute values */
805 /* If (2 * abs (lrem) >= abs (lden)) */
814 {
816 /* quo = quo - 1; */
819 else
820 /* quo = quo + 1; */
823 }
824 else
826 }
831 }
833 /* compute true remainder: rem = num - (quo * den) */
838 }
839 \f
840 /* Determine whether an expression T can be cheaply negated using
841 the function negate_expr. */
843 static bool
845 tree t;
846 {
849 tree type;
858 {
863 /* Check that -CST will not overflow type. */
866 {
871 }
872 else
885 }
887 }
889 /* Given T, an expression, return the negation of T. Allow for T to be
890 null, in which case return null. */
892 static tree
894 tree t;
895 {
896 tree type;
897 tree tem;
906 {
919 /* - (A - B) -> B - A */
929 }
932 }
933 \f
934 /* Split a tree IN into a constant, literal and variable parts that could be
935 combined with CODE to make IN. "constant" means an expression with
936 TREE_CONSTANT but that isn't an actual constant. CODE must be a
937 commutative arithmetic operation. Store the constant part into *CONP,
938 the literal in *LITP and return the variable part. If a part isn't
939 present, set it to null. If the tree does not decompose in this way,
940 return the entire tree as the variable part and the other parts as null.
942 If CODE is PLUS_EXPR we also split trees that use MINUS_EXPR. In that
943 case, we negate an operand that was subtracted. Except if it is a
944 literal for which we use *MINUS_LITP instead.
946 If NEGATE_P is true, we are negating all of IN, again except a literal
947 for which we use *MINUS_LITP instead.
949 If IN is itself a literal or constant, return it as appropriate.
951 Note that we do not guarantee that any of the three values will be the
952 same type as IN, but they will have the same signedness and mode. */
954 static tree
956 tree in;
960 {
967 /* Strip any conversions that don't change the machine mode or signedness. */
974 /* We can associate addition and subtraction together (even
975 though the C standard doesn't say so) for integers because
976 the value is not affected. For reals, the value might be
977 affected, so we can't. */
980 {
986 /* First see if either of the operands is a literal, then a constant. */
997 /* If we haven't dealt with either operand, this is not a case we can
998 decompose. Otherwise, VAR is either of the ones remaining, if any. */
1003 else
1006 /* Now do any needed negations. */
1013 }
1016 else
1020 {
1027 }
1030 }
1032 /* Re-associate trees split by the above function. T1 and T2 are either
1033 expressions to associate or null. Return the new expression, if any. If
1034 we build an operation, do it in TYPE and with CODE. */
1036 static tree
1040 tree type;
1041 {
1047 /* If either input is CODE, a PLUS_EXPR, or a MINUS_EXPR, don't
1048 try to fold this since we will have infinite recursion. But do
1049 deal with any NEGATE_EXPRs. */
1052 {
1054 {
1061 }
1063 }
1066 }
1067 \f
1068 /* Combine two integer constants ARG1 and ARG2 under operation CODE
1069 to produce a new constant.
1071 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1073 static tree
1078 {
1082 HOST_WIDE_INT hi;
1084 HOST_WIDE_INT garbageh;
1085 tree t;
1088 int is_sizetype
1099 {
1119 /* It's unclear from the C standard whether shifts can overflow.
1120 The following code ignores overflow; perhaps a C standard
1121 interpretation ruling is needed. */
1151 /* This is a shortcut for a common special case. */
1156 {
1162 }
1164 /* ... fall through ... */
1168 {
1171 }
1174 {
1177 }
1184 /* This is a shortcut for a common special case. */
1189 {
1194 }
1196 /* ... fall through ... */
1212 else
1218 else
1224 }
1226 /* If this is for a sizetype, can be represented as one (signed)
1227 HOST_WIDE_INT word, and doesn't overflow, use size_int since it caches
1228 constants. */
1234 else
1235 {
1238 }
1241 = ((notrunc
1248 /* If we're doing a size calculation, unsigned arithmetic does overflow.
1249 So check if force_fit_type truncated the value. */
1260 }
1262 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1263 constant. We assume ARG1 and ARG2 have the same data type, or at least
1264 are the same kind of constant and the same machine mode.
1266 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1268 static tree
1273 {
1281 {
1282 REAL_VALUE_TYPE d1;
1283 REAL_VALUE_TYPE d2;
1284 REAL_VALUE_TYPE value;
1285 tree t;
1290 /* If either operand is a NaN, just return it. Otherwise, set up
1291 for floating-point trap; we return an overflow. */
1301 value));
1311 }
1313 {
1319 tree t;
1322 {
1342 notrunc),
1348 notrunc));
1352 {
1353 tree magsquared
1357 notrunc);
1360 const_binop
1365 notrunc),
1367 notrunc),
1368 notrunc),
1370 const_binop
1375 notrunc),
1377 notrunc),
1378 notrunc),
1380 }
1385 }
1387 }
1389 }
1391 /* These are the hash table functions for the hash table of INTEGER_CST
1392 nodes of a sizetype. */
1394 /* Return the hash code code X, an INTEGER_CST. */
1396 static hashval_t
1399 {
1405 }
1407 /* Return nonzero if the value represented by *X (an INTEGER_CST tree node)
1408 is the same as that given by *Y, which is the same. */
1410 static int
1414 {
1422 }
1423 \f
1424 /* Return an INTEGER_CST with value whose low-order HOST_BITS_PER_WIDE_INT
1425 bits are given by NUMBER and of the sizetype represented by KIND. */
1427 tree
1429 HOST_WIDE_INT number;
1431 {
1433 }
1435 /* Likewise, but the desired type is specified explicitly. */
1439 htab_t size_htab;
1441 tree
1443 HOST_WIDE_INT number;
1444 tree type;
1445 {
1449 {
1452 }
1454 /* Adjust NEW_CONST to be the constant we want. If it's already in the
1455 hash table, we return the value from the hash table. Otherwise, we
1456 place that in the hash table and make a new node for the next time. */
1465 {
1471 }
1472 else
1474 }
1476 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
1477 is a tree code. The type of the result is taken from the operands.
1478 Both must be the same type integer type and it must be a size type.
1479 If the operands are constant, so is the result. */
1481 tree
1485 {
1492 /* Handle the special case of two integer constants faster. */
1494 {
1495 /* And some specific cases even faster than that. */
1504 /* Handle general case of two integer constants. */
1506 }
1512 }
1514 /* Given two values, either both of sizetype or both of bitsizetype,
1515 compute the difference between the two values. Return the value
1516 in signed type corresponding to the type of the operands. */
1518 tree
1521 {
1523 tree ctype;
1529 /* If the type is already signed, just do the simple thing. */
1536 /* If either operand is not a constant, do the conversions to the signed
1537 type and subtract. The hardware will do the right thing with any
1538 overflow in the subtraction. */
1543 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
1544 Otherwise, subtract the other way, convert to CTYPE (we know that can't
1545 overflow) and negate (which can't either). Special-case a result
1546 of zero while we're here. */
1551 else
1554 }
1555 \f
1557 /* Given T, a tree representing type conversion of ARG1, a constant,
1558 return a constant tree representing the result of conversion. */
1560 static tree
1562 tree t;
1563 tree arg1;
1564 {
1569 {
1571 {
1572 /* If we would build a constant wider than GCC supports,
1573 leave the conversion unfolded. */
1577 /* If we are trying to make a sizetype for a small integer, use
1578 size_int to pick up cached types to reduce duplicate nodes. */
1584 /* Given an integer constant, make new constant with new type,
1585 appropriately sign-extended or truncated. */
1589 /* Indicate an overflow if (1) ARG1 already overflowed,
1590 or (2) force_fit_type indicates an overflow.
1591 Tell force_fit_type that an overflow has already occurred
1592 if ARG1 is a too-large unsigned value and T is signed.
1593 But don't indicate an overflow if converting a pointer. */
1603 }
1605 {
1606 /* Don't initialize these, use assignments.
1607 Initialized local aggregates don't work on old compilers. */
1608 REAL_VALUE_TYPE x;
1609 REAL_VALUE_TYPE l;
1610 REAL_VALUE_TYPE u;
1621 /* See if X will be in range after truncation towards 0.
1622 To compensate for truncation, move the bounds away from 0,
1623 but reject if X exactly equals the adjusted bounds. */
1627 /* If X is a NaN, use zero instead and show we have an overflow.
1628 Otherwise, range check. */
1632 && !no_upper_bound
1636 {
1640 }
1646 }
1648 }
1650 {
1654 {
1656 {
1657 /* We make a copy of ARG1 so that we don't modify an
1658 existing constant tree. */
1662 }
1673 }
1674 }
1677 }
1678 \f
1679 /* Return an expr equal to X but certainly not valid as an lvalue. */
1681 tree
1683 tree x;
1684 {
1685 tree result;
1687 /* These things are certainly not lvalues. */
1698 }
1700 /* Nonzero means lvalues are limited to those valid in pedantic ANSI C.
1701 Zero means allow extended lvalues. */
1705 /* When pedantic, return an expr equal to X but certainly not valid as a
1706 pedantic lvalue. Otherwise, return X. */
1708 tree
1710 tree x;
1711 {
1714 else
1716 }
1717 \f
1718 /* Given a tree comparison code, return the code that is the logical inverse
1719 of the given code. It is not safe to do this for floating-point
1720 comparisons, except for NE_EXPR and EQ_EXPR. */
1722 static enum tree_code
1725 {
1727 {
1742 }
1743 }
1745 /* Similar, but return the comparison that results if the operands are
1746 swapped. This is safe for floating-point. */
1748 static enum tree_code
1751 {
1753 {
1767 }
1768 }
1771 /* Convert a comparison tree code from an enum tree_code representation
1772 into a compcode bit-based encoding. This function is the inverse of
1773 compcode_to_comparison. */
1775 static int
1778 {
1780 {
1795 }
1796 }
1798 /* Convert a compcode bit-based encoding of a comparison operator back
1799 to GCC's enum tree_code representation. This function is the
1800 inverse of comparison_to_compcode. */
1802 static enum tree_code
1805 {
1807 {
1822 }
1823 }
1825 /* Return nonzero if CODE is a tree code that represents a truth value. */
1827 static int
1830 {
1835 }
1836 \f
1837 /* Return nonzero if two operands are necessarily equal.
1838 If ONLY_CONST is nonzero, only return nonzero for constants.
1839 This function tests whether the operands are indistinguishable;
1840 it does not test whether they are equal using C's == operation.
1841 The distinction is important for IEEE floating point, because
1842 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
1843 (2) two NaNs may be indistinguishable, but NaN!=NaN. */
1845 int
1849 {
1850 /* If both types don't have the same signedness, then we can't consider
1851 them equal. We must check this before the STRIP_NOPS calls
1852 because they may change the signedness of the arguments. */
1860 /* This is needed for conversions and for COMPONENT_REF.
1861 Might as well play it safe and always test this. */
1867 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
1868 We don't care about side effects in that case because the SAVE_EXPR
1869 takes care of that for us. In all other cases, two expressions are
1870 equal if they have no side effects. If we have two identical
1871 expressions with side effects that should be treated the same due
1872 to the only side effects being identical SAVE_EXPR's, that will
1873 be detected in the recursive calls below. */
1879 /* Next handle constant cases, those for which we can return 1 even
1880 if ONLY_CONST is set. */
1883 {
1896 {
1906 {
1911 }
1914 }
1918 only_const)
1920 only_const));
1933 }
1939 {
1941 /* Two conversions are equal only if signedness and modes match. */
1957 /* For commutative ops, allow the other order. */
1970 /* If either of the pointer (or reference) expressions we are dereferencing
1971 contain a side effect, these cannot be equal. */
1977 {
1999 }
2008 }
2009 }
2010 \f
2011 /* Similar to operand_equal_p, but see if ARG0 might have been made by
2012 shorten_compare from ARG1 when ARG1 was being compared with OTHER.
2014 When in doubt, return 0. */
2016 static int
2019 tree other;
2020 {
2032 /* Discard any conversions that don't change the modes of ARG0 and ARG1
2033 and see if the inner values are the same. This removes any
2034 signedness comparison, which doesn't matter here. */
2041 /* Duplicate what shorten_compare does to ARG1 and see if that gives the
2042 actual comparison operand, ARG0.
2044 First throw away any conversions to wider types
2045 already present in the operands. */
2054 {
2057 /* Make sure shorter operand is extended the right way
2058 to match the longer operand. */
2064 }
2067 }
2068 \f
2069 /* See if ARG is an expression that is either a comparison or is performing
2070 arithmetic on comparisons. The comparisons must only be comparing
2071 two different values, which will be stored in *CVAL1 and *CVAL2; if
2072 they are nonzero it means that some operands have already been found.
2073 No variables may be used anywhere else in the expression except in the
2074 comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around
2075 the expression and save_expr needs to be called with CVAL1 and CVAL2.
2077 If this is true, return 1. Otherwise, return zero. */
2079 static int
2081 tree arg;
2084 {
2088 /* We can handle some of the 'e' cases here. */
2098 {
2099 /* If we've already found a CVAL1 or CVAL2, this expression is
2100 two complex to handle. */
2106 }
2109 {
2132 /* First see if we can handle the first operand, then the second. For
2133 the second operand, we know *CVAL1 can't be zero. It must be that
2134 one side of the comparison is each of the values; test for the
2135 case where this isn't true by failing if the two operands
2136 are the same. */
2145 ;
2149 ;
2150 else
2154 ;
2158 ;
2159 else
2166 }
2167 }
2168 \f
2169 /* ARG is a tree that is known to contain just arithmetic operations and
2170 comparisons. Evaluate the operations in the tree substituting NEW0 for
2171 any occurrence of OLD0 as an operand of a comparison and likewise for
2172 NEW1 and OLD1. */
2174 static tree
2176 tree arg;
2178 {
2183 /* We can handle some of the 'e' cases here. */
2191 {
2206 {
2223 }
2224 /* fall through - ??? */
2227 {
2231 /* We need to check both for exact equality and tree equality. The
2232 former will be true if the operand has a side-effect. In that
2233 case, we know the operand occurred exactly once. */
2246 }
2250 }
2251 }
2252 \f
2253 /* Return a tree for the case when the result of an expression is RESULT
2254 converted to TYPE and OMITTED was previously an operand of the expression
2255 but is now not needed (e.g., we folded OMITTED * 0).
2257 If OMITTED has side effects, we must evaluate it. Otherwise, just do
2258 the conversion of RESULT to TYPE. */
2260 tree
2263 {
2270 }
2272 /* Similar, but call pedantic_non_lvalue instead of non_lvalue. */
2274 static tree
2277 {
2284 }
2285 \f
2286 /* Return a simplified tree node for the truth-negation of ARG. This
2287 never alters ARG itself. We assume that ARG is an operation that
2288 returns a truth value (0 or 1). */
2290 tree
2292 tree arg;
2293 {
2300 /* If this is a comparison, we can simply invert it, except for
2301 floating-point non-equality comparisons, in which case we just
2302 enclose a TRUTH_NOT_EXPR around what we have. */
2305 {
2307 && !flag_unsafe_math_optimizations
2311 else
2314 }
2317 {
2332 /* Here we can invert either operand. We invert the first operand
2333 unless the second operand is a TRUTH_NOT_EXPR in which case our
2334 result is the XOR of the first operand with the inside of the
2335 negation of the second operand. */
2340 else
2395 }
2399 }
2401 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
2402 operands are another bit-wise operation with a common input. If so,
2403 distribute the bit operations to save an operation and possibly two if
2404 constants are involved. For example, convert
2405 (A | B) & (A | C) into A | (B & C)
2406 Further simplification will occur if B and C are constants.
2408 If this optimization cannot be done, 0 will be returned. */
2410 static tree
2413 tree type;
2415 {
2416 tree common;
2426 {
2430 }
2432 {
2436 }
2438 {
2442 }
2444 {
2448 }
2449 else
2454 }
2455 \f
2456 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
2457 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero. */
2459 static tree
2461 tree inner;
2462 tree type;
2465 {
2472 }
2474 /* Optimize a bit-field compare.
2476 There are two cases: First is a compare against a constant and the
2477 second is a comparison of two items where the fields are at the same
2478 bit position relative to the start of a chunk (byte, halfword, word)
2479 large enough to contain it. In these cases we can avoid the shift
2480 implicit in bitfield extractions.
2482 For constants, we emit a compare of the shifted constant with the
2483 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
2484 compared. For two fields at the same position, we do the ANDs with the
2485 similar mask and compare the result of the ANDs.
2487 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
2488 COMPARE_TYPE is the type of the comparison, and LHS and RHS
2489 are the left and right operands of the comparison, respectively.
2491 If the optimization described above can be done, we return the resulting
2492 tree. Otherwise we return zero. */
2494 static tree
2497 tree compare_type;
2499 {
2508 tree mask;
2509 tree offset;
2511 /* Get all the information about the extractions being done. If the bit size
2512 if the same as the size of the underlying object, we aren't doing an
2513 extraction at all and so can do nothing. We also don't want to
2514 do anything if the inner expression is a PLACEHOLDER_EXPR since we
2515 then will no longer be able to replace it. */
2523 {
2524 /* If this is not a constant, we can only do something if bit positions,
2525 sizes, and signedness are the same. */
2533 }
2535 /* See if we can find a mode to refer to this field. We should be able to,
2536 but fail if we can't. */
2545 /* Set signed and unsigned types of the precision of this mode for the
2546 shifts below. */
2550 /* Compute the bit position and size for the new reference and our offset
2551 within it. If the new reference is the same size as the original, we
2552 won't optimize anything, so return zero. */
2562 /* Make the mask to be used against the extracted field. */
2572 /* If not comparing with constant, just rework the comparison
2573 and return. */
2578 mask),
2582 mask));
2584 /* Otherwise, we are handling the constant case. See if the constant is too
2585 big for the field. Warn and return a tree of for 0 (false) if so. We do
2586 this not only for its own sake, but to avoid having to test for this
2587 error case below. If we didn't, we might generate wrong code.
2589 For unsigned fields, the constant shifted right by the field length should
2590 be all zero. For signed fields, the high-order bits should agree with
2591 the sign bit. */
2594 {
2598 {
2604 }
2605 }
2606 else
2607 {
2611 {
2617 }
2618 }
2620 /* Single-bit compares should always be against zero. */
2622 {
2625 }
2627 /* Make a new bitfield reference, shift the constant over the
2628 appropriate number of bits and mask it with the computed mask
2629 (in case this was a signed field). If we changed it, make a new one. */
2632 {
2635 }
2645 rhs);
2646 }
2647 \f
2648 /* Subroutine for fold_truthop: decode a field reference.
2650 If EXP is a comparison reference, we return the innermost reference.
2652 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
2653 set to the starting bit number.
2655 If the innermost field can be completely contained in a mode-sized
2656 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
2658 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
2659 otherwise it is not changed.
2661 *PUNSIGNEDP is set to the signedness of the field.
2663 *PMASK is set to the mask used. This is either contained in a
2664 BIT_AND_EXPR or derived from the width of the field.
2666 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
2668 Return 0 if this is not a component reference or is one that we can't
2669 do anything with. */
2671 static tree
2674 tree exp;
2680 {
2683 tree unsigned_type;
2686 /* All the optimizations using this function assume integer fields.
2687 There are problems with FP fields since the type_for_size call
2688 below can fail for, e.g., XFmode. */
2695 {
2701 }
2710 /* Compute the mask to access the bitfield. */
2720 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
2728 }
2730 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
2731 bit positions. */
2733 static int
2735 tree mask;
2737 {
2740 tree tmask;
2745 return
2752 }
2754 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
2755 represents the sign bit of EXP's type. If EXP represents a sign
2756 or zero extension, also test VAL against the unextended type.
2757 The return value is the (sub)expression whose sign bit is VAL,
2758 or NULL_TREE otherwise. */
2760 static tree
2762 tree exp;
2763 tree val;
2764 {
2766 HOST_WIDE_INT hi;
2768 tree t;
2770 /* Tree EXP must have an integral type. */
2775 /* Tree VAL must be an integer constant. */
2782 {
2785 }
2786 else
2787 {
2790 }
2795 /* Handle extension from a narrower type. */
2801 }
2803 /* Subroutine for fold_truthop: determine if an operand is simple enough
2804 to be evaluated unconditionally. */
2806 static int
2808 tree exp;
2809 {
2810 /* Strip any conversions that don't change the machine mode. */
2822 /* Don't regard global variables as simple. They may be
2823 allocated in ways unknown to the compiler (shared memory,
2824 #pragma weak, etc). */
2827 /* Loading a static variable is unduly expensive, but global
2828 registers aren't expensive. */
2830 }
2831 \f
2832 /* The following functions are subroutines to fold_range_test and allow it to
2833 try to change a logical combination of comparisons into a range test.
2835 For example, both
2836 X == 2 || X == 3 || X == 4 || X == 5
2837 and
2838 X >= 2 && X <= 5
2839 are converted to
2840 (unsigned) (X - 2) <= 3
2842 We describe each set of comparisons as being either inside or outside
2843 a range, using a variable named like IN_P, and then describe the
2844 range with a lower and upper bound. If one of the bounds is omitted,
2845 it represents either the highest or lowest value of the type.
2847 In the comments below, we represent a range by two numbers in brackets
2848 preceded by a "+" to designate being inside that range, or a "-" to
2849 designate being outside that range, so the condition can be inverted by
2850 flipping the prefix. An omitted bound is represented by a "-". For
2851 example, "- [-, 10]" means being outside the range starting at the lowest
2852 possible value and ending at 10, in other words, being greater than 10.
2853 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
2854 always false.
2856 We set up things so that the missing bounds are handled in a consistent
2857 manner so neither a missing bound nor "true" and "false" need to be
2858 handled using a special case. */
2860 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
2861 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
2862 and UPPER1_P are nonzero if the respective argument is an upper bound
2863 and zero for a lower. TYPE, if nonzero, is the type of the result; it
2864 must be specified for a comparison. ARG1 will be converted to ARG0's
2865 type if both are specified. */
2867 static tree
2870 tree type;
2873 {
2874 tree tem;
2878 /* If neither arg represents infinity, do the normal operation.
2879 Else, if not a comparison, return infinity. Else handle the special
2880 comparison rules. Note that most of the cases below won't occur, but
2881 are handled for consistency. */
2884 {
2889 }
2894 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
2895 for neither. In real maths, we cannot assume open ended ranges are
2896 the same. But, this is computer arithmetic, where numbers are finite.
2897 We can therefore make the transformation of any unbounded range with
2898 the value Z, Z being greater than any representable number. This permits
2899 us to treat unbounded ranges as equal. */
2903 {
2924 }
2927 }
2928 \f
2929 /* Given EXP, a logical expression, set the range it is testing into
2930 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
2931 actually being tested. *PLOW and *PHIGH will be made of the same type
2932 as the returned expression. If EXP is not a comparison, we will most
2933 likely not be returning a useful value and range. */
2935 static tree
2937 tree exp;
2940 {
2947 /* Start with simply saying "EXP != 0" and then look at the code of EXP
2948 and see if we can refine the range. Some of the cases below may not
2949 happen, but it doesn't seem worth worrying about this. We "continue"
2950 the outer loop when we've changed something; otherwise we "break"
2951 the switch, which will "break" the while. */
2956 {
2960 {
2971 }
2973 /* Set ORIG_TYPE as soon as TYPE is non-null so that we do not
2974 lose a cast by accident. */
2979 {
2986 /* We can only do something if the range is testing for zero
2987 and if the second operand is an integer constant. Note that
2988 saying something is "in" the range we make is done by
2989 complementing IN_P since it will set in the initial case of
2990 being not equal to zero; "out" is leaving it alone. */
2997 {
3018 }
3022 /* If this is an unsigned comparison, we also know that EXP is
3023 greater than or equal to zero. We base the range tests we make
3024 on that fact, so we record it here so we can parse existing
3025 range tests. */
3027 {
3030 NULL_TREE))
3035 /* If the high bound is missing, but we
3036 have a low bound, reverse the range so
3037 it goes from zero to the low bound minus 1. */
3039 {
3044 }
3045 }
3049 /* (-x) IN [a,b] -> x in [-b, -a] */
3059 /* ~ X -> -X - 1 */
3068 /* If EXP is signed, any overflow in the computation is undefined,
3069 so we don't worry about it so long as our computations on
3070 the bounds don't overflow. For unsigned, overflow is defined
3071 and this is exactly the right thing. */
3080 /* Check for an unsigned range which has wrapped around the maximum
3081 value thus making n_high < n_low, and normalize it. */
3083 {
3089 /* If the range is of the form +/- [ x+1, x ], we won't
3090 be able to normalize it. But then, it represents the
3091 whole range or the empty set, so make it
3092 +/- [ -, - ]. */
3096 else
3098 }
3099 else
3122 /* If we're converting from an unsigned to a signed type,
3123 we will be doing the comparison as unsigned. The tests above
3124 have already verified that LOW and HIGH are both positive.
3126 So we have to make sure that the original unsigned value will
3127 be interpreted as positive. */
3129 {
3132 tree high_positive;
3134 /* A range without an upper bound is, naturally, unbounded.
3135 Since convert would have cropped a very large value, use
3136 the max value for the destination type. */
3137 high_positive
3146 /* If the low bound is specified, "and" the range with the
3147 range for which the original unsigned value will be
3148 positive. */
3150 {
3154 high_positive))
3158 }
3159 else
3160 {
3161 /* Otherwise, "or" the range with the range of the input
3162 that will be interpreted as negative. */
3166 high_positive))
3170 }
3171 }
3179 }
3182 }
3184 /* If EXP is a constant, we can evaluate whether this is true or false. */
3186 {
3193 }
3197 }
3198 \f
3199 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
3200 type, TYPE, return an expression to test if EXP is in (or out of, depending
3201 on IN_P) the range. */
3203 static tree
3205 tree type;
3206 tree exp;
3209 {
3211 tree value;
3230 {
3232 {
3236 }
3238 }
3240 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
3242 {
3244 HOST_WIDE_INT hi;
3249 {
3252 }
3253 else
3254 {
3257 }
3260 {
3262 {
3265 }
3268 }
3269 }
3278 }
3279 \f
3280 /* Given two ranges, see if we can merge them into one. Return 1 if we
3281 can, 0 if we can't. Set the output range into the specified parameters. */
3283 static int
3289 {
3293 tree tem;
3303 /* Make range 0 be the range that starts first, or ends last if they
3304 start at the same value. Swap them if it isn't. */
3307 || (lowequal
3310 {
3314 }
3316 /* Now flag two cases, whether the ranges are disjoint or whether the
3317 second range is totally subsumed in the first. Note that the tests
3318 below are simplified by the ones above. */
3324 /* We now have four cases, depending on whether we are including or
3325 excluding the two ranges. */
3327 {
3328 /* If they don't overlap, the result is false. If the second range
3329 is a subset it is the result. Otherwise, the range is from the start
3330 of the second to the end of the first. */
3335 else
3337 }
3340 {
3341 /* If they don't overlap, the result is the first range. If they are
3342 equal, the result is false. If the second range is a subset of the
3343 first, and the ranges begin at the same place, we go from just after
3344 the end of the first range to the end of the second. If the second
3345 range is not a subset of the first, or if it is a subset and both
3346 ranges end at the same place, the range starts at the start of the
3347 first range and ends just before the second range.
3348 Otherwise, we can't describe this as a single range. */
3354 {
3358 }
3360 {
3364 }
3365 else
3367 }
3370 {
3371 /* If they don't overlap, the result is the second range. If the second
3372 is a subset of the first, the result is false. Otherwise,
3373 the range starts just after the first range and ends at the
3374 end of the second. */
3379 else
3380 {
3384 }
3385 }
3387 else
3388 {
3389 /* The case where we are excluding both ranges. Here the complex case
3390 is if they don't overlap. In that case, the only time we have a
3391 range is if they are adjacent. If the second is a subset of the
3392 first, the result is the first. Otherwise, the range to exclude
3393 starts at the beginning of the first range and ends at the end of the
3394 second. */
3396 {
3403 else
3405 }
3408 else
3410 }
3414 }
3415 \f
3416 /* EXP is some logical combination of boolean tests. See if we can
3417 merge it into some range test. Return the new tree if so. */
3419 static tree
3421 tree exp;
3422 {
3429 tree tem;
3431 /* If this is an OR operation, invert both sides; we will invert
3432 again at the end. */
3436 /* If both expressions are the same, if we can merge the ranges, and we
3437 can build the range test, return it or it inverted. If one of the
3438 ranges is always true or always false, consider it to be the same
3439 expression as the other. */
3449 /* On machines where the branch cost is expensive, if this is a
3450 short-circuited branch and the underlying object on both sides
3451 is the same, make a non-short-circuit operation. */
3457 {
3458 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
3459 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
3460 which cases we can't do this. */
3469 {
3481 }
3482 }
3485 }
3486 \f
3487 /* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
3488 bit value. Arrange things so the extra bits will be set to zero if and
3489 only if C is signed-extended to its full width. If MASK is nonzero,
3490 it is an INTEGER_CST that should be AND'ed with the extra bits. */
3492 static tree
3494 tree c;
3497 tree mask;
3498 {
3501 tree temp;
3506 /* We work by getting just the sign bit into the low-order bit, then
3507 into the high-order bit, then sign-extend. We then XOR that value
3508 with C. */
3512 /* We must use a signed type in order to get an arithmetic right shift.
3513 However, we must also avoid introducing accidental overflows, so that
3514 a subsequent call to integer_zerop will work. Hence we must
3515 do the type conversion here. At this point, the constant is either
3516 zero or one, and the conversion to a signed type can never overflow.
3517 We could get an overflow if this conversion is done anywhere else. */
3525 /* If necessary, convert the type back to match the type of C. */
3530 }
3531 \f
3532 /* Find ways of folding logical expressions of LHS and RHS:
3533 Try to merge two comparisons to the same innermost item.
3534 Look for range tests like "ch >= '0' && ch <= '9'".
3535 Look for combinations of simple terms on machines with expensive branches
3536 and evaluate the RHS unconditionally.
3538 For example, if we have p->a == 2 && p->b == 4 and we can make an
3539 object large enough to span both A and B, we can do this with a comparison
3540 against the object ANDed with the a mask.
3542 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
3543 operations to do this with one comparison.
3545 We check for both normal comparisons and the BIT_AND_EXPRs made this by
3546 function and the one above.
3548 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
3549 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
3551 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
3552 two operands.
3554 We return the simplified tree or 0 if no optimization is possible. */
3556 static tree
3560 {
3561 /* If this is the "or" of two comparisons, we can do something if
3562 the comparisons are NE_EXPR. If this is the "and", we can do something
3563 if the comparisons are EQ_EXPR. I.e.,
3564 (a->b == 2 && a->c == 4) can become (a->new == NEW).
3566 WANTED_CODE is this operation code. For single bit fields, we can
3567 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
3568 comparison for one-bit fields. */
3588 /* Start by getting the comparison codes. Fail if anything is volatile.
3589 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
3590 it were surrounded with a NE_EXPR. */
3615 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
3619 {
3624 {
3632 }
3635 {
3644 }
3645 else
3655 }
3657 /* If the RHS can be evaluated unconditionally and its operands are
3658 simple, it wins to evaluate the RHS unconditionally on machines
3659 with expensive branches. In this case, this isn't a comparison
3660 that can be merged. Avoid doing this if the RHS is a floating-point
3661 comparison since those can trap. */
3667 {
3668 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
3676 integer_zero_node);
3678 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
3686 integer_zero_node);
3689 }
3691 /* See if the comparisons can be merged. Then get all the parameters for
3692 each side. */
3716 /* It must be true that the inner operation on the lhs of each
3717 comparison must be the same if we are to be able to do anything.
3718 Then see if we have constants. If not, the same must be true for
3719 the rhs's. */
3730 else
3733 /* If either comparison code is not correct for our logical operation,
3734 fail. However, we can convert a one-bit comparison against zero into
3735 the opposite comparison against that bit being set in the field. */
3739 {
3741 {
3742 /* Make the left operand unsigned, since we are only interested
3743 in the value of one bit. Otherwise we are doing the wrong
3744 thing below. */
3747 }
3748 else
3750 }
3752 /* This is analogous to the code for l_const above. */
3754 {
3756 {
3759 }
3760 else
3762 }
3764 /* After this point all optimizations will generate bit-field
3765 references, which we might not want. */
3769 /* See if we can find a mode that contains both fields being compared on
3770 the left. If we can't, fail. Otherwise, update all constants and masks
3771 to be relative to a field of that size. */
3776 volatilep);
3786 {
3789 }
3797 {
3805 {
3809 wanted_code == NE_EXPR
3811 }
3812 }
3814 {
3822 {
3826 wanted_code == NE_EXPR
3828 }
3829 }
3831 /* If the right sides are not constant, do the same for it. Also,
3832 disallow this optimization if a size or signedness mismatch occurs
3833 between the left and right sides. */
3835 {
3838 /* Make sure the two fields on the right
3839 correspond to the left without being swapped. */
3847 volatilep);
3857 {
3860 }
3867 /* Make a mask that corresponds to both fields being compared.
3868 Do this for both items being compared. If the operands are the
3869 same size and the bits being compared are in the same position
3870 then we can do this by masking both and comparing the masked
3871 results. */
3875 {
3887 }
3889 /* There is still another way we can do something: If both pairs of
3890 fields being compared are adjacent, we may be able to make a wider
3891 field containing them both.
3893 Note that we still must mask the lhs/rhs expressions. Furthermore,
3894 the mask must be shifted to account for the shift done by
3895 make_bit_field_ref. */
3900 {
3901 tree type;
3913 /* Convert to the smaller type before masking out unwanted bits. */
3916 {
3918 {
3922 }
3924 {
3928 }
3929 }
3938 }
3941 }
3943 /* Handle the case of comparisons with constants. If there is something in
3944 common between the masks, those bits of the constants must be the same.
3945 If not, the condition is always false. Test for this to avoid generating
3946 incorrect code below. */
3951 {
3953 {
3956 }
3957 else
3958 {
3961 }
3962 }
3964 /* Construct the expression we will return. First get the component
3965 reference we will make. Unless the mask is all ones the width of
3966 that field, perform the mask operation. Then compare with the
3967 merged constant. */
3977 }
3978 \f
3979 /* Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a
3980 constant. */
3982 static tree
3984 tree t;
3985 {
3990 tree minmax_const;
3992 tree inner;
4002 /* If something does not permit us to optimize, return the original tree. */
4010 /* Now handle all the various comparison codes. We only handle EQ_EXPR
4011 and GT_EXPR, doing the rest with recursive calls using logical
4012 simplifications. */
4014 {
4016 return
4020 return
4022 optimize_minmax_comparison
4024 optimize_minmax_comparison
4029 /* MAX (X, 0) == 0 -> X <= 0 */
4033 /* MAX (X, 0) == 5 -> X == 5 */
4037 /* MAX (X, 0) == -1 -> false */
4041 /* MIN (X, 0) == 0 -> X >= 0 */
4045 /* MIN (X, 0) == 5 -> false */
4048 else
4049 /* MIN (X, 0) == -1 -> X == -1 */
4054 /* MAX (X, 0) > 0 -> X > 0
4055 MAX (X, 0) > 5 -> X > 5 */
4059 /* MAX (X, 0) > -1 -> true */
4063 /* MIN (X, 0) > 0 -> false
4064 MIN (X, 0) > 5 -> false */
4067 else
4068 /* MIN (X, 0) > -1 -> X > -1 */
4073 }
4074 }
4075 \f
4076 /* T is an integer expression that is being multiplied, divided, or taken a
4077 modulus (CODE says which and what kind of divide or modulus) by a
4078 constant C. See if we can eliminate that operation by folding it with
4079 other operations already in T. WIDE_TYPE, if non-null, is a type that
4080 should be used for the computation if wider than our type.
4082 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
4083 (X * 2) + (Y * 4). We must, however, be assured that either the original
4084 expression would not overflow or that overflow is undefined for the type
4085 in the language in question.
4087 We also canonicalize (X + 7) * 4 into X * 4 + 28 in the hope that either
4088 the machine has a multiply-accumulate insn or that this is part of an
4089 addressing calculation.
4091 If we return a non-null expression, it is an equivalent form of the
4092 original computation, but need not be in the original type. */
4094 static tree
4096 tree t;
4097 tree c;
4099 tree wide_type;
4100 {
4101 /* To avoid exponential search depth, refuse to allow recursion past
4102 three levels. Beyond that (1) it's highly unlikely that we'll find
4103 something interesting and (2) we've probably processed it before
4104 when we built the inner expression. */
4107 tree ret;
4112 depth++;
4114 depth--;
4117 }
4119 static tree
4121 tree t;
4122 tree c;
4124 tree wide_type;
4125 {
4135 /* Don't deal with constants of zero here; they confuse the code below. */
4145 /* Note that we need not handle conditional operations here since fold
4146 already handles those cases. So just do arithmetic here. */
4148 {
4150 /* For a constant, we can always simplify if we are a multiply
4151 or (for divide and modulus) if it is a multiple of our constant. */
4158 /* If op0 is an expression ... */
4163 /* ... and is unsigned, and its type is smaller than ctype,
4164 then we cannot pass through as widening. */
4170 /* ... or its type is larger than ctype,
4171 then we cannot pass through this truncation. */
4174 /* ... or signedness changes for division or modulus,
4175 then we cannot pass through this conversion. */
4181 /* Pass the constant down and see if we can make a simplification. If
4182 we can, replace this expression with the inner simplification for
4183 possible later conversion to our or some other type. */
4195 /* If widening the type changes the signedness, then we can't perform
4196 this optimization as that changes the result. */
4200 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
4203 {
4209 }
4219 /* If this has not been evaluated and the operand has no side effects,
4220 we can see if we can do something inside it and make a new one.
4221 Note that this test is overly conservative since we can do this
4222 if the only reason it had side effects is that it was another
4223 similar SAVE_EXPR, but that isn't worth bothering with. */
4226 wide_type)))
4227 {
4234 }
4238 /* If the second operand is constant, this is a multiplication
4239 or floor division, by a power of two, so we can treat it that
4240 way unless the multiplier or divisor overflows. */
4242 /* const_binop may not detect overflow correctly,
4243 so check for it explicitly here. */
4257 /* See if we can eliminate the operation on both sides. If we can, we
4258 can return a new PLUS or MINUS. If we can't, the only remaining
4259 cases where we can do anything are if the second operand is a
4260 constant. */
4265 /* If not multiplication, we can only do this if both operands
4266 are divisible by c. */
4272 /* If this was a subtraction, negate OP1 and set it to be an addition.
4273 This simplifies the logic below. */
4280 /* If either OP1 or C are negative, this optimization is not safe for
4281 some of the division and remainder types while for others we need
4282 to change the code. */
4284 {
4292 }
4294 /* If it's a multiply or a division/modulus operation of a multiple
4295 of our constant, do the operation and verify it doesn't overflow. */
4298 {
4302 }
4303 else
4306 /* If we have an unsigned type is not a sizetype, we cannot widen
4307 the operation since it will change the result if the original
4308 computation overflowed. */
4314 /* If we were able to eliminate our operation from the first side,
4315 apply our operation to the second side and reform the PLUS. */
4319 /* The last case is if we are a multiply. In that case, we can
4320 apply the distributive law to commute the multiply and addition
4321 if the multiplication of the constants doesn't overflow. */
4326 op1));
4331 /* We have a special case here if we are doing something like
4332 (C * 8) % 4 since we know that's zero. */
4339 /* ... fall through ... */
4343 /* If we can extract our operation from the LHS, do so and return a
4344 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
4345 do something only if the second operand is a constant. */
4357 /* If these are the same operation types, we can associate them
4358 assuming no overflow. */
4365 /* If these operations "cancel" each other, we have the main
4366 optimizations of this pass, which occur when either constant is a
4367 multiple of the other, in which case we replace this with either an
4368 operation or CODE or TCODE.
4370 If we have an unsigned type that is not a sizetype, we cannot do
4371 this since it will change the result if the original computation
4372 overflowed. */
4379 {
4390 }
4395 }
4398 }
4399 \f
4400 /* If T contains a COMPOUND_EXPR which was inserted merely to evaluate
4401 S, a SAVE_EXPR, return the expression actually being evaluated. Note
4402 that we may sometimes modify the tree. */
4404 static tree
4406 tree t;
4407 tree s;
4408 {
4411 /* See if this is the COMPOUND_EXPR we want to eliminate. */
4416 /* See if this is a COND_EXPR or a simple arithmetic operator. We
4417 don't bother handling any other types. */
4419 {
4423 }
4428 {
4431 }
4434 }
4435 \f
4436 /* Return a node which has the indicated constant VALUE (either 0 or
4437 1), and is of the indicated TYPE. */
4439 static tree
4442 tree type;
4443 {
4448 integer_zero_node);
4449 else
4450 {
4455 }
4456 }
4458 /* Utility function for the following routine, to see how complex a nesting of
4459 COND_EXPRs can be. EXPR is the expression and LIMIT is a count beyond which
4460 we don't care (to avoid spending too much time on complex expressions.). */
4462 static int
4464 tree expr;
4466 {
4477 }
4479 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
4480 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
4481 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
4482 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
4483 COND is the first argument to CODE; otherwise (as in the example
4484 given here), it is the second argument. TYPE is the type of the
4485 original expression. */
4487 static tree
4490 tree type;
4491 tree cond;
4492 tree arg;
4494 {
4498 /* In the end, we'll produce a COND_EXPR. Both arms of the
4499 conditional expression will be binary operations. The left-hand
4500 side of the expression to be executed if the condition is true
4501 will be pointed to by TRUE_LHS. Similarly, the right-hand side
4502 of the expression to be executed if the condition is true will be
4503 pointed to by TRUE_RHS. FALSE_LHS and FALSE_RHS are analogous --
4504 but apply to the expression to be executed if the conditional is
4505 false. */
4510 /* These are the codes to use for the left-hand side and right-hand
4511 side of the COND_EXPR. Normally, they are the same as CODE. */
4514 /* And these are the types of the expressions. */
4520 {
4524 }
4525 else
4526 {
4530 }
4533 {
4537 /* If this operand throws an expression, then it does not make
4538 sense to try to perform a logical or arithmetic operation
4539 involving it. Instead of building `a + throw 3' for example,
4540 we simply build `a, throw 3'. */
4542 {
4544 {
4547 }
4548 else
4550 }
4552 {
4554 {
4557 }
4558 else
4560 }
4561 }
4562 else
4563 {
4568 }
4570 /* If ARG is complex we want to make sure we only evaluate
4571 it once. Though this is only required if it is volatile, it
4572 might be more efficient even if it is not. However, if we
4573 succeed in folding one part to a constant, we do not need
4574 to make this SAVE_EXPR. Since we do this optimization
4575 primarily to see if we do end up with constant and this
4576 SAVE_EXPR interferes with later optimizations, suppressing
4577 it when we can is important.
4579 If we are not in a function, we can't make a SAVE_EXPR, so don't
4580 try to do so. Don't try to see if the result is a constant
4581 if an arm is a COND_EXPR since we get exponential behavior
4582 in that case. */
4591 {
4600 {
4604 }
4605 }
4618 else
4620 }
4622 \f
4623 /* Subroutine of fold() that checks for the addition of +/- 0.0.
4625 If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type
4626 TYPE, X + ADDEND is the same as X. If NEGATE, return true if X -
4627 ADDEND is the same as X.
4629 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
4630 and finite. The problematic cases are when X is zero, and its mode
4631 has signed zeros. In the case of rounding towards -infinity,
4632 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
4633 modes, X + 0 is not the same as X because -0 + 0 is 0. */
4635 static bool
4639 {
4643 /* Don't allow the fold with -fsignaling-nans. */
4647 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
4651 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
4656 /* The mode has signed zeros, and we have to honor their sign.
4657 In this situation, there is only one case we can return true for.
4658 X - 0 is the same as X unless rounding towards -infinity is
4659 supported. */
4661 }
4663 /* Subroutine of fold() that checks comparisons of built-in math
4664 functions against real constants.
4666 FCODE is the DECL_FUNCTION_CODE of the built-in, CODE is the comparison
4667 operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE
4668 is the type of the result and ARG0 and ARG1 are the operands of the
4669 comparison. ARG1 must be a TREE_REAL_CST.
4671 The function returns the constant folded tree if a simplification
4672 can be made, and NULL_TREE otherwise. */
4674 static tree
4679 {
4680 REAL_VALUE_TYPE c;
4685 {
4691 {
4692 /* sqrt(x) < y is always false, if y is negative. */
4696 arg);
4698 /* sqrt(x) > y is always true, if y is negative and we
4699 don't care about NaNs, i.e. negative values of x. */
4703 arg);
4705 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
4708 }
4710 {
4711 REAL_VALUE_TYPE c2;
4717 {
4718 /* sqrt(x) > y is x == +Inf, when y is very large. */
4723 /* sqrt(x) > y is always false, when y is very large
4724 and we don't care about infinities. */
4727 arg);
4728 }
4730 /* sqrt(x) > c is the same as x > c*c. */
4733 }
4735 {
4736 REAL_VALUE_TYPE c2;
4742 {
4743 /* sqrt(x) < y is always true, when y is a very large
4744 value and we don't care about NaNs or Infinities. */
4748 arg);
4750 /* sqrt(x) < y is x != +Inf when y is very large and we
4751 don't care about NaNs. */
4756 /* sqrt(x) < y is x >= 0 when y is very large and we
4757 don't care about Infinities. */
4762 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
4771 dconst0))),
4774 c2)))));
4775 }
4777 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
4782 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
4785 {
4790 dconst0))),
4793 c2)))));
4794 }
4795 }
4796 }
4799 }
4802 /* Perform constant folding and related simplification of EXPR.
4803 The related simplifications include x*1 => x, x*0 => 0, etc.,
4804 and application of the associative law.
4805 NOP_EXPR conversions may be removed freely (as long as we
4806 are careful not to change the C type of the overall expression)
4807 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
4808 but we can constant-fold them if they have constant operands. */
4810 tree
4812 tree expr;
4813 {
4816 tree tem;
4822 /* WINS will be nonzero when the switch is done
4823 if all operands are constant. */
4826 /* Don't try to process an RTL_EXPR since its operands aren't trees.
4827 Likewise for a SAVE_EXPR that's already been evaluated. */
4831 /* Return right away if a constant. */
4835 #ifdef MAX_INTEGER_COMPUTATION_MODE
4837 #endif
4840 {
4841 tree subop;
4843 /* Special case for conversion ops that can have fixed point args. */
4846 /* Don't use STRIP_NOPS, because signedness of argument type matters. */
4852 else
4857 )
4858 /* Note that TREE_CONSTANT isn't enough:
4859 static var addresses are constant but we can't
4860 do arithmetic on them. */
4862 }
4864 {
4868 {
4870 tree subop;
4876 {
4877 /* Signedness matters here. Perhaps we can refine this
4878 later. */
4880 }
4881 else
4882 /* Strip any conversions that don't change the mode. */
4887 else
4892 /* Note that TREE_CONSTANT isn't enough:
4893 static var addresses are constant but we can't
4894 do arithmetic on them. */
4901 }
4902 }
4904 /* If this is a commutative operation, and ARG0 is a constant, move it
4905 to ARG1 to reduce the number of tests below. */
4910 {
4915 }
4917 /* Now WINS is set as described above,
4918 ARG0 is the first operand of EXPR,
4919 and ARG1 is the second operand (if it has more than one operand).
4921 First check for cases where an arithmetic operation is applied to a
4922 compound, conditional, or comparison operation. Push the arithmetic
4923 operation inside the compound or conditional to see if any folding
4924 can then be done. Convert comparison to conditional for this purpose.
4925 The also optimizes non-constant cases that used to be done in
4926 expand_expr.
4928 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
4929 one of the operands is a comparison and the other is a comparison, a
4930 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
4931 code below would make the expression more complex. Change it to a
4932 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
4933 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
4945 {
4955 }
4958 {
4963 {
4973 /* If this was a conversion, and all we did was to move into
4974 inside the COND_EXPR, bring it back out. But leave it if
4975 it is a conversion from integer to integer and the
4976 result precision is no wider than a word since such a
4977 conversion is cheap and may be optimized away by combine,
4978 while it couldn't if it were outside the COND_EXPR. Then return
4979 so we don't get into an infinite recursion loop taking the
4980 conversion out and then back in. */
4992 && (INTEGRAL_TYPE_P
5003 }
5008 }
5011 {
5026 return
5040 return
5043 }
5054 {
5070 /* Other kinds of FIX are not handled properly by fold_convert. */
5075 /* Handle cases of two conversions in a row. */
5078 {
5098 /* In addition to the cases of two conversions in a row
5099 handled below, if we are converting something to its own
5100 type via an object of identical or wider precision, neither
5101 conversion is needed. */
5107 /* Likewise, if the intermediate and final types are either both
5108 float or both integer, we don't need the middle conversion if
5109 it is wider than the final type and doesn't change the signedness
5110 (for integers). Avoid this if the final type is a pointer
5111 since then we sometimes need the inner conversion. Likewise if
5112 the outer has a precision not equal to the size of its mode. */
5122 /* If we have a sign-extension of a zero-extended value, we can
5123 replace that by a single zero-extension. */
5129 /* Two conversions in a row are not needed unless:
5130 - some conversion is floating-point (overstrict for now), or
5131 - the intermediate type is narrower than both initial and
5132 final, or
5133 - the intermediate type and innermost type differ in signedness,
5134 and the outermost type is wider than the intermediate, or
5135 - the initial type is a pointer type and the precisions of the
5136 intermediate and final types differ, or
5137 - the final type is a pointer type and the precisions of the
5138 initial and intermediate types differ. */
5152 }
5156 /* Detect assigning a bitfield. */
5159 {
5160 /* Don't leave an assignment inside a conversion
5161 unless assigning a bitfield. */
5164 /* First do the assignment, then return converted constant. */
5168 }
5170 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
5171 constants (if x has signed type, the sign bit cannot be set
5172 in c). This folds extension into the BIT_AND_EXPR. */
5177 {
5187 <= HOST_BITS_PER_WIDE_INT
5189 {
5196 #ifdef LOAD_EXTEND_OP
5200 {
5204 }
5205 #endif
5206 }
5211 }
5214 {
5217 }
5228 {
5232 }
5241 {
5243 {
5245 HOST_WIDE_INT high;
5256 }
5259 }
5262 /* Convert -((double)float) into (double)(-float). */
5265 {
5270 }
5272 /* Convert - (a - b) to (b - a) for non-floating-point. */
5282 {
5284 {
5285 /* If the value is unsigned, then the absolute value is
5286 the same as the ordinary value. */
5289 /* Similarly, if the value is non-negative. */
5292 /* If the value is negative, then the absolute value is
5293 its negation. */
5294 else
5295 {
5297 HOST_WIDE_INT high;
5308 }
5309 }
5311 {
5315 }
5316 }
5319 /* Convert fabs((double)float) into (double)fabsf(float). */
5322 {
5327 }
5328 else
5329 {
5330 /* fabs(sqrt(x)) = sqrt(x) and fabs(exp(x)) = exp(x). */
5339 }
5364 {
5371 }
5377 /* A + (-B) -> A - B */
5380 /* (-A) + B -> B - A */
5384 {
5388 /* If we are adding two BIT_AND_EXPR's, both of which are and'ing
5389 with a constant, and the two constants have no bits in common,
5390 we should treat this as a BIT_IOR_EXPR since this may produce more
5391 simplifications. */
5399 {
5402 }
5404 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
5405 (plus (plus (mult) (mult)) (foo)) so that we can
5406 take advantage of the factoring cases below. */
5411 {
5416 else
5427 parg1));
5432 parg0));
5433 }
5436 {
5440 /* (A * C) + (B * C) -> (A+B) * C.
5441 We are most concerned about the case where C is a constant,
5442 but other combinations show up during loop reduction. Since
5443 it is not difficult, try all four possibilities. */
5460 /* No identical multiplicands; see if we can find a common
5461 power-of-two factor in non-power-of-two multiplies. This
5462 can help in multi-dimensional array access. */
5467 {
5472 /* Move min of absolute values to int11. */
5475 {
5479 }
5482 {
5487 }
5488 }
5493 same));
5494 }
5495 }
5497 /* See if ARG1 is zero and X + ARG1 reduces to X. */
5501 /* Likewise if the operands are reversed. */
5505 bit_rotate:
5506 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
5507 is a rotate of A by C1 bits. */
5508 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
5509 is a rotate of A by B bits. */
5510 {
5519 {
5538 {
5546 TYPE_PRECISION
5551 ? LROTATE_EXPR
5554 }
5556 {
5564 TYPE_PRECISION
5569 ? LROTATE_EXPR
5572 }
5573 }
5574 }
5576 associate:
5577 /* In most languages, can't associate operations on floats through
5578 parentheses. Rather than remember where the parentheses were, we
5579 don't associate floats at all. It shouldn't matter much. However,
5580 associating multiplications is only very slightly inaccurate, so do
5581 that if -funsafe-math-optimizations is specified. */
5586 {
5590 /* Split both trees into variables, constants, and literals. Then
5591 associate each group together, the constants with literals,
5592 then the result with variables. This increases the chances of
5593 literals being recombined later and of generating relocatable
5594 expressions for the sum of a constant and literal. */
5599 /* Only do something if we found more than two objects. Otherwise,
5600 nothing has changed and we risk infinite recursion. */
5605 {
5606 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
5615 /* Preserve the MINUS_EXPR if the negative part of the literal is
5616 greater than the positive part. Otherwise, the multiplicative
5617 folding code (i.e extract_muldiv) may be fooled in case
5618 unsigned constants are substracted, like in the following
5619 example: ((X*2 + 4) - 8U)/2. */
5621 {
5623 {
5627 }
5628 else
5629 {
5633 }
5634 }
5636 {
5640 else
5641 {
5646 }
5647 }
5651 }
5652 }
5654 binary:
5658 {
5659 /* The return value should always have
5660 the same type as the original expression. */
5665 }
5669 /* A - (-B) -> A + B */
5672 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
5682 {
5688 /* (A * C) - (B * C) -> (A-B) * C. Since we are most concerned
5689 about the case where C is a constant, just try one of the
5690 four possibilities. */
5701 /* Fold A - (A & B) into ~B & A. */
5704 {
5709 arg0));
5714 arg0));
5715 }
5716 }
5718 /* See if ARG1 is zero and X - ARG1 reduces to X. */
5722 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
5723 ARG0 is zero and X + ARG0 reduces to X, since that would mean
5724 (-ARG1 + ARG0) reduces to -ARG1. */
5728 /* Fold &x - &x. This can happen from &x.foo - &x.
5729 This is unsafe for certain floats even in non-IEEE formats.
5730 In IEEE, it is unsafe because it does wrong for NaNs.
5731 Also note that operand_equal_p is always false if an operand
5732 is volatile. */
5741 /* (-A) * (-B) -> A * B */
5747 {
5753 /* (a * (1 << b)) is (a << b) */
5768 }
5769 else
5770 {
5771 /* Maybe fold x * 0 to 0. The expressions aren't the same
5772 when x is NaN, since x * 0 is also NaN. Nor are they the
5773 same in modes with signed zeros, since multiplying a
5774 negative value by 0 gives -0, not +0. */
5779 /* In IEEE floating point, x*1 is not equivalent to x for snans. */
5784 /* Transform x * -1.0 into -x. */
5789 /* x*2 is x+x */
5793 {
5796 }
5799 {
5803 /* Optimize sqrt(x)*sqrt(y) as sqrt(x*y). */
5807 {
5814 }
5816 /* Optimize exp(x)*exp(y) as exp(x+y). */
5820 {
5827 }
5828 }
5829 }
5833 bit_ior:
5842 /* Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
5844 This results in more efficient code for machines without a NAND
5845 instruction. Combine will canonicalize to the first form
5846 which will allow use of NAND instructions provided by the
5847 backend if they exist. */
5850 {
5855 }
5857 /* See if this can be simplified into a rotate first. If that
5858 is unsuccessful continue in the association code. */
5867 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
5868 with a constant, and the two constants have no bits in common,
5869 we should treat this as a BIT_IOR_EXPR since this may produce more
5870 simplifications. */
5878 {
5881 }
5883 /* See if this can be simplified into a rotate first. If that
5884 is unsuccessful continue in the association code. */
5888 bit_and:
5896 /* Simplify ((int)c & 0x377) into (int)c, if c is unsigned char. */
5899 {
5900 unsigned int prec
5907 }
5909 /* Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
5911 This results in more efficient code for machines without a NOR
5912 instruction. Combine will canonicalize to the first form
5913 which will allow use of NOR instructions provided by the
5914 backend if they exist. */
5917 {
5922 }
5932 {
5936 }
5940 /* Don't touch a floating-point divide by zero unless the mode
5941 of the constant can represent infinity. */
5947 /* (-A) / (-B) -> A / B */
5952 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
5957 /* If ARG1 is a constant, we can convert this to a multiply by the
5958 reciprocal. This does not have the same rounding properties,
5959 so only do this if -funsafe-math-optimizations. We can actually
5960 always safely do it if ARG1 is a power of two, but it's hard to
5961 tell if it is or not in a portable manner. */
5963 {
5968 /* Find the reciprocal if optimizing and the result is exact. */
5970 {
5971 REAL_VALUE_TYPE r;
5974 {
5977 }
5978 }
5979 }
5980 /* Convert A/B/C to A/(B*C). */
5983 {
5986 arg1)));
5987 }
5988 /* Convert A/(B/C) to (A/B)*C. */
5991 {
5996 }
5998 /* Optimize x/exp(y) into x*exp(-y). */
6000 {
6005 {
6012 }
6013 }
6026 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
6027 operation, EXACT_DIV_EXPR.
6029 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
6030 At one time others generated faster code, it's not clear if they do
6031 after the last round to changes to the DIV code in expmed.c. */
6066 /* Optimize -1 >> x for arithmetic right shifts. */
6069 /* ... fall through ... */
6072 shift:
6078 /* Since negative shift count is not well-defined,
6079 don't try to compute it in the compiler. */
6082 /* Rewrite an LROTATE_EXPR by a constant into an
6083 RROTATE_EXPR by a new constant. */
6085 {
6089 = const_binop
6096 }
6098 /* If we have a rotate of a bit operation with the rotate count and
6099 the second operand of the bit operation both constant,
6100 permute the two operations. */
6113 /* Two consecutive rotates adding up to the width of the mode can
6114 be ignored. */
6145 /* Note that the operand of this must be an int
6146 and its values must be 0 or 1.
6147 ("true" is a fixed value perhaps depending on the language,
6148 but we don't handle values other than 1 correctly yet.) */
6150 /* Avoid infinite recursion. */
6156 /* Note that the operands of this must be ints
6157 and their values must be 0 or 1.
6158 ("true" is a fixed value perhaps depending on the language.) */
6159 /* If first arg is constant zero, return it. */
6163 /* If either arg is constant true, drop it. */
6167 /* Preserve sequence points. */
6170 /* If second arg is constant zero, result is zero, but first arg
6171 must be evaluated. */
6174 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
6175 case will be handled here. */
6179 truth_andor:
6180 /* We only do these simplifications if we are optimizing. */
6184 /* Check for things like (A || B) && (A || C). We can convert this
6185 to A || (B && C). Note that either operator can be any of the four
6186 truth and/or operations and the transformation will still be
6187 valid. Also note that we only care about order for the
6188 ANDIF and ORIF operators. If B contains side effects, this
6189 might change the truth-value of A. */
6196 {
6216 /* This case if tricky because we must either have commutative
6217 operators or else A10 must not have side-effects. */
6223 a01));
6224 }
6226 /* See if we can build a range comparison. */
6230 /* Check for the possibility of merging component references. If our
6231 lhs is another similar operation, try to merge its rhs with our
6232 rhs. Then try to merge our lhs and rhs. */
6244 /* Note that the operands of this must be ints
6245 and their values must be 0 or true.
6246 ("true" is a fixed value perhaps depending on the language.) */
6247 /* If first arg is constant true, return it. */
6251 /* If either arg is constant zero, drop it. */
6255 /* Preserve sequence points. */
6258 /* If second arg is constant true, result is true, but we must
6259 evaluate first arg. */
6262 /* Likewise for first arg, but note this only occurs here for
6263 TRUTH_OR_EXPR. */
6269 /* If either arg is constant zero, drop it. */
6274 /* If either arg is constant true, this is a logical inversion. */
6287 /* If one arg is a real or integer constant, put it last. */
6292 {
6299 }
6302 {
6310 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
6315 /* (-a) CMP (-b) -> b CMP a */
6320 /* (-a) CMP CST -> a swap(CMP) (-CST) */
6322 return
6328 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
6329 /* a CMP (-0) -> a CMP 0 */
6335 /* If this is a comparison of a real constant with a PLUS_EXPR
6336 or a MINUS_EXPR of a real constant, we can convert it into a
6337 comparison with a revised real constant as long as no overflow
6338 occurs when unsafe_math_optimizations are enabled. */
6350 /* Fold comparisons against built-in math functions. */
6352 && flag_unsafe_math_optimizations
6354 {
6358 {
6362 }
6363 }
6364 }
6366 /* Convert foo++ == CONST into ++foo == CONST + INCR.
6367 First, see if one arg is constant; find the constant arg
6368 and the other one. */
6369 {
6379 {
6380 /* This optimization is invalid for ordered comparisons
6381 if CONST+INCR overflows or if foo+incr might overflow.
6382 This optimization is invalid for floating point due to rounding.
6383 For pointer types we assume overflow doesn't happen. */
6387 {
6388 tree newconst
6392 /* Do not overwrite the current varop to be a preincrement,
6393 create a new node so that we won't confuse our caller who
6394 might create trees and throw them away, reusing the
6395 arguments that they passed to build. This shows up in
6396 the THEN or ELSE parts of ?: being postincrements. */
6401 /* If VAROP is a reference to a bitfield, we must mask
6402 the constant by the width of the field. */
6406 {
6407 int size
6409 (TREE_OPERAND
6413 tree folded_compare;
6415 /* First check whether the comparison would come out
6416 always the same. If we don't do that we would
6417 change the meaning with the masking. */
6421 else
6424 constop));
6439 mask)));
6440 }
6446 }
6447 }
6449 {
6453 {
6454 tree newconst
6458 /* Do not overwrite the current varop to be a predecrement,
6459 create a new node so that we won't confuse our caller who
6460 might create trees and throw them away, reusing the
6461 arguments that they passed to build. This shows up in
6462 the THEN or ELSE parts of ?: being postdecrements. */
6470 {
6471 int size
6473 (TREE_OPERAND
6477 tree folded_compare;
6482 else
6485 constop));
6500 mask)));
6501 }
6507 }
6508 }
6509 }
6511 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
6512 This transformation affects the cases which are handled in later
6513 optimizations involving comparisons with non-negative constants. */
6517 {
6519 {
6534 }
6535 }
6537 /* Comparisons with the highest or lowest possible integer of
6538 the specified size will have known values. */
6539 {
6547 {
6554 {
6557 }
6558 else
6559 {
6562 }
6567 {
6571 arg0);
6579 arg0);
6585 /* The GE_EXPR and LT_EXPR cases above are not normally
6586 reached because of previous transformations. */
6590 }
6594 {
6607 }
6611 {
6615 arg0);
6624 arg0);
6632 }
6636 {
6649 }
6654 /* signed_type does not work on pointer types. */
6656 {
6657 /* The following case also applies to X < signed_max+1
6658 and X >= signed_max+1 because previous transformations. */
6660 {
6664 return fold
6668 }
6669 }
6670 }
6671 }
6673 /* If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
6674 a MINUS_EXPR of a constant, we can convert it into a comparison with
6675 a revised constant as long as no overflow occurs. */
6687 /* Similarly for a NEGATE_EXPR. */
6696 /* If we have X - Y == 0, we can convert that to X == Y and similarly
6697 for !=. Don't do this for ordered comparisons due to overflow. */
6703 /* If we are widening one operand of an integer comparison,
6704 see if the other operand is similarly being widened. Perhaps we
6705 can do the comparison in the narrower type. */
6715 /* If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a
6716 constant, we can simplify it. */
6723 /* If we are comparing an ABS_EXPR with a constant, we can
6724 convert all the cases into explicit comparisons, but they may
6725 well not be faster than doing the ABS and one comparison.
6726 But ABS (X) <= C is a range comparison, which becomes a subtraction
6727 and a comparison, and is probably faster. */
6739 /* If this is an EQ or NE comparison with zero and ARG0 is
6740 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
6741 two operations, but the latter can be done in one less insn
6742 on machines that have only two-operand insns or on which a
6743 constant cannot be the first operand. */
6746 {
6749 return
6757 integer_one_node)),
6758 arg1));
6761 return
6769 integer_one_node)),
6770 arg1));
6771 }
6773 /* If this is an NE or EQ comparison of zero against the result of a
6774 signed MOD operation whose second operand is a power of 2, make
6775 the MOD operation unsigned since it is simpler and equivalent. */
6784 {
6791 }
6793 /* If this is an NE comparison of zero with an AND of one, remove the
6794 comparison since the AND will give the correct value. */
6800 /* If we have (A & C) == C where C is a power of 2, convert this into
6801 (A & C) != 0. Similarly for NE_EXPR. */
6809 /* If we have (A & C) != 0 where C is the sign bit of A, convert
6810 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
6814 {
6818 {
6823 }
6824 }
6826 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
6827 and similarly for >= into !=. */
6843 return
6850 /* Simplify comparison of something with itself. (For IEEE
6851 floating-point, we can only do some of these simplifications.) */
6853 {
6855 {
6866 /* For NE, we can only do this simplification if integer. */
6869 /* ... fall through ... */
6875 }
6876 }
6878 /* If we are comparing an expression that just has comparisons
6879 of two integer values, arithmetic expressions of those comparisons,
6880 and constants, we can simplify it. There are only three cases
6881 to check: the two values can either be equal, the first can be
6882 greater, or the second can be greater. Fold the expression for
6883 those three values. Since each value must be 0 or 1, we have
6884 eight possibilities, each of which corresponds to the constant 0
6885 or 1 or one of the six possible comparisons.
6887 This handles common cases like (a > b) == 0 but also handles
6888 expressions like ((x > y) - (y > x)) > 0, which supposedly
6889 occur in macroized code. */
6892 {
6897 /* Don't handle degenerate cases here; they should already
6898 have been handled anyway. */
6907 {
6911 /* We can't just pass T to eval_subst in case cval1 or cval2
6912 was the same as ARG1. */
6914 tree high_result
6917 arg1));
6918 tree equal_result
6921 arg1));
6922 tree low_result
6925 arg1));
6927 /* All three of these results should be 0 or 1. Confirm they
6928 are. Then use those values to select the proper code
6929 to use. */
6937 {
6938 /* Make a 3-bit mask with the high-order bit being the
6939 value for `>', the next for '=', and the low for '<'. */
6943 {
6945 /* Always false. */
6966 /* Always true. */
6968 }
6973 else
6975 }
6976 }
6977 }
6979 /* If this is a comparison of a field, we may be able to simplify it. */
6984 /* Handle the constant case even without -O
6985 to make sure the warnings are given. */
6987 {
6990 }
6992 /* If this is a comparison of complex values and either or both sides
6993 are a COMPLEX_EXPR or COMPLEX_CST, it is best to split up the
6994 comparisons and join them with a TRUTH_ANDIF_EXPR or TRUTH_ORIF_EXPR.
6995 This may prevent needless evaluations. */
7002 {
7015 type,
7018 }
7020 /* Optimize comparisons of strlen vs zero to a compare of the
7021 first character of the string vs zero. To wit,
7022 strlen(ptr) == 0 => *ptr == 0
7023 strlen(ptr) != 0 => *ptr != 0
7024 Other cases should reduce to one of these two (or a constant)
7025 due to the return value of strlen being unsigned. */
7030 {
7032 tree arglist;
7044 integer_zero_node));
7045 }
7047 /* From here on, the only cases we handle are when the result is
7048 known to be a constant.
7050 To compute GT, swap the arguments and do LT.
7051 To compute GE, do LT and invert the result.
7052 To compute LE, swap the arguments, do LT and invert the result.
7053 To compute NE, do EQ and invert the result.
7055 Therefore, the code below must handle only EQ and LT. */
7058 {
7061 }
7063 /* Note that it is safe to invert for real values here because we
7064 will check below in the one case that it matters. */
7069 {
7072 }
7074 /* Compute a result for LT or EQ if args permit;
7075 otherwise return T. */
7077 {
7080 else
7085 }
7088 /* Assume a nonexplicit constant cannot equal an explicit one,
7089 since such code would be undefined anyway.
7090 Exception: on sysvr4, using #pragma weak,
7091 a label can come out as 0. */
7098 #endif
7099 /* Two real constants can be compared explicitly. */
7101 {
7102 /* If either operand is a NaN, the result is false with two
7103 exceptions: First, an NE_EXPR is true on NaNs, but that case
7104 is already handled correctly since we will be inverting the
7105 result for NE_EXPR. Second, if we had inverted a LE_EXPR
7106 or a GE_EXPR into a LT_EXPR, we must return true so that it
7107 will be inverted into false. */
7117 else
7121 }
7135 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
7136 so all simple results must be passed through pedantic_non_lvalue. */
7138 return pedantic_non_lvalue
7143 /* If the second operand is zero, invert the comparison and swap
7144 the second and third operands. Likewise if the second operand
7145 is constant and the third is not or if the third operand is
7146 equivalent to the first operand of the comparison. */
7154 {
7155 /* See if this can be inverted. If it can't, possibly because
7156 it was a floating-point inequality comparison, don't do
7157 anything. */
7161 {
7165 /* arg1 should be the first argument of the new T. */
7168 }
7169 }
7171 /* If we have A op B ? A : C, we may be able to convert this to a
7172 simpler expression, depending on the operation and the values
7173 of B and C. Signed zeros prevent all of these transformations,
7174 for reasons given above each one. */
7180 {
7186 /* If we have A op 0 ? A : -A, consider applying the following
7187 transformations:
7189 A == 0? A : -A same as -A
7190 A != 0? A : -A same as A
7191 A >= 0? A : -A same as abs (A)
7192 A > 0? A : -A same as abs (A)
7193 A <= 0? A : -A same as -abs (A)
7194 A < 0? A : -A same as -abs (A)
7196 None of these transformations work for modes with signed
7197 zeros. If A is +/-0, the first two transformations will
7198 change the sign of the result (from +0 to -0, or vice
7199 versa). The last four will fix the sign of the result,
7200 even though the original expressions could be positive or
7201 negative, depending on the sign of A.
7203 Note that all these transformations are correct if A is
7204 NaN, since the two alternatives (A and -A) are also NaNs. */
7211 {
7213 return
7214 pedantic_non_lvalue
7216 negate_expr
7218 arg1))));
7226 return pedantic_non_lvalue
7234 return pedantic_non_lvalue
7238 arg1)))));
7241 }
7243 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
7244 A == 0 ? A : 0 is always 0 unless A is -0. Note that
7245 both transformations are correct when A is NaN: A != 0
7246 is then true, and A == 0 is false. */
7249 {
7254 }
7256 /* Try some transformations of A op B ? A : B.
7258 A == B? A : B same as B
7259 A != B? A : B same as A
7260 A >= B? A : B same as max (A, B)
7261 A > B? A : B same as max (B, A)
7262 A <= B? A : B same as min (A, B)
7263 A < B? A : B same as min (B, A)
7265 As above, these transformations don't work in the presence
7266 of signed zeros. For example, if A and B are zeros of
7267 opposite sign, the first two transformations will change
7268 the sign of the result. In the last four, the original
7269 expressions give different results for (A=+0, B=-0) and
7270 (A=-0, B=+0), but the transformed expressions do not.
7272 The first two transformations are correct if either A or B
7273 is a NaN. In the first transformation, the condition will
7274 be false, and B will indeed be chosen. In the case of the
7275 second transformation, the condition A != B will be true,
7276 and A will be chosen.
7278 The conversions to max() and min() are not correct if B is
7279 a number and A is not. The conditions in the original
7280 expressions will be false, so all four give B. The min()
7281 and max() versions would give a NaN instead. */
7284 {
7289 /* Avoid adding NOP_EXPRs in case this is an lvalue. */
7291 {
7295 }
7298 {
7305 /* In C++ a ?: expression can be an lvalue, so put the
7306 operand which will be used if they are equal first
7307 so that we can convert this back to the
7308 corresponding COND_EXPR. */
7310 return pedantic_non_lvalue
7320 return pedantic_non_lvalue
7329 }
7330 }
7332 /* If this is A op C1 ? A : C2 with C1 and C2 constant integers,
7333 we might still be able to simplify this. For example,
7334 if C1 is one less or one more than C2, this might have started
7335 out as a MIN or MAX and been transformed by this function.
7336 Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE. */
7342 {
7344 /* We can replace A with C1 in this case. */
7351 /* If C1 is C2 + 1, this is min(A, C2). */
7356 return pedantic_non_lvalue
7361 /* If C1 is C2 - 1, this is min(A, C2). */
7366 return pedantic_non_lvalue
7371 /* If C1 is C2 - 1, this is max(A, C2). */
7376 return pedantic_non_lvalue
7381 /* If C1 is C2 + 1, this is max(A, C2). */
7386 return pedantic_non_lvalue
7393 }
7394 }
7396 /* If the second operand is simpler than the third, swap them
7397 since that produces better jump optimization results. */
7403 {
7404 /* See if this can be inverted. If it can't, possibly because
7405 it was a floating-point inequality comparison, don't do
7406 anything. */
7410 {
7414 /* arg1 should be the first argument of the new T. */
7417 }
7418 }
7420 /* Convert A ? 1 : 0 to simply A. */
7423 /* If we try to convert TREE_OPERAND (t, 0) to our type, the
7424 call to fold will try to move the conversion inside
7425 a COND, which will recurse. In that case, the COND_EXPR
7426 is probably the best choice, so leave it alone. */
7430 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
7431 over COND_EXPR in cases such as floating point comparisons. */
7438 /* Look for expressions of the form A & 2 ? 2 : 0. The result of this
7439 operation is simply A & 2. */
7450 /* Convert A ? B : 0 into A && B if A and B are truth values. */
7457 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
7461 {
7462 /* Only perform transformation if ARG0 is easily inverted. */
7467 }
7472 /* When pedantic, a compound expression can be neither an lvalue
7473 nor an integer constant expression. */
7476 /* Don't let (0, 0) be null pointer constant. */
7518 /* Pull arithmetic ops out of the CLEANUP_POINT_EXPR where
7519 appropriate. */
7524 {
7528 tree arg01;
7539 {
7553 arg01));
7554 }
7557 }
7560 /* Check for a built-in function. */
7565 {
7569 }
7575 }
7577 /* Determine if first argument is a multiple of second argument. Return 0 if
7578 it is not, or we cannot easily determined it to be.
7580 An example of the sort of thing we care about (at this point; this routine
7581 could surely be made more general, and expanded to do what the *_DIV_EXPR's
7582 fold cases do now) is discovering that
7584 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
7586 is a multiple of
7588 SAVE_EXPR (J * 8)
7590 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
7592 This code also handles discovering that
7594 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
7596 is a multiple of 8 so we don't have to worry about dealing with a
7597 possible remainder.
7599 Note that we *look* inside a SAVE_EXPR only to determine how it was
7600 calculated; it is not safe for fold to do much of anything else with the
7601 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
7602 at run time. For example, the latter example above *cannot* be implemented
7603 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
7604 evaluation time of the original SAVE_EXPR is not necessarily the same at
7605 the time the new expression is evaluated. The only optimization of this
7606 sort that would be valid is changing
7608 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
7610 divided by 8 to
7612 SAVE_EXPR (I) * SAVE_EXPR (J)
7614 (where the same SAVE_EXPR (J) is used in the original and the
7615 transformed version). */
7617 static int
7619 tree type;
7620 tree top;
7621 tree bottom;
7622 {
7630 {
7642 {
7646 /* const_binop may not detect overflow correctly,
7647 so check for it explicitly here. */
7656 }
7660 /* Can't handle conversions from non-integral or wider integral type. */
7666 /* .. fall through ... */
7682 }
7683 }
7685 /* Return true if `t' is known to be non-negative. */
7687 int
7689 tree t;
7690 {
7692 {
7701 /* These are undefined at zero. This is true even if
7702 C[LT]Z_DEFINED_VALUE_AT_ZERO is set, since what we're
7703 computing here is a user-visible property. */
7743 /* Truth values evaluate to 0 or 1, which is nonnegative. */
7745 else
7746 /* We don't know sign of `t', so be conservative and return false. */
7748 }
7749 }
7751 /* Return true if `r' is known to be non-negative.
7752 Only handles constants at the moment. */
7754 int
7756 rtx r;
7757 {
7759 {
7769 {
7771 rtx elt;
7776 {
7780 }
7783 }
7787 /* These are always nonnegative. */
7792 }
7793 }