| blob | 0a7b7d226a314351e679f8552c4fcf8e247744df |
1 /* Fold a constant sub-tree into a single node for C-compiler
2 Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
3 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
22 /*@@ This file should be rewritten to use an arbitrary precision
23 @@ representation for "struct tree_int_cst" and "struct tree_real_cst".
24 @@ Perhaps the routines could also be used for bc/dc, and made a lib.
25 @@ The routines that translate from the ap rep should
26 @@ warn if precision et. al. is lost.
27 @@ This would also make life easier when this technology is used
28 @@ for cross-compilers. */
30 /* The entry points in this file are fold, size_int_wide, size_binop
31 and force_fit_type.
33 fold takes a tree as argument and returns a simplified tree.
35 size_binop takes a tree code for an arithmetic operation
36 and two operands that are trees, and produces a tree for the
37 result, assuming the type comes from `sizetype'.
39 size_int takes an integer value, and creates a tree constant
40 with type from `sizetype'.
42 force_fit_type takes a constant, an overflowable flag and prior
43 overflow indicators. It forces the value to fit the type and sets
44 TREE_OVERFLOW and TREE_CONSTANT_OVERFLOW as appropriate. */
62 /* The following constants represent a bit based encoding of GCC's
63 comparison operators. This encoding simplifies transformations
64 on relational comparison operators, such as AND and OR. */
82 };
117 tree);
141 /* We know that A1 + B1 = SUM1, using 2's complement arithmetic and ignoring
142 overflow. Suppose A, B and SUM have the same respective signs as A1, B1,
143 and SUM1. Then this yields nonzero if overflow occurred during the
144 addition.
146 Overflow occurs if A and B have the same sign, but A and SUM differ in
147 sign. Use `^' to test whether signs differ, and `< 0' to isolate the
148 sign. */
149 #define OVERFLOW_SUM_SIGN(a, b, sum) ((~((a) ^ (b)) & ((a) ^ (sum))) < 0)
150 \f
151 /* To do constant folding on INTEGER_CST nodes requires two-word arithmetic.
152 We do that by representing the two-word integer in 4 words, with only
153 HOST_BITS_PER_WIDE_INT / 2 bits stored in each word, as a positive
154 number. The value of the word is LOWPART + HIGHPART * BASE. */
156 #define LOWPART(x) \
157 ((x) & (((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT / 2)) - 1))
158 #define HIGHPART(x) \
159 ((unsigned HOST_WIDE_INT) (x) >> HOST_BITS_PER_WIDE_INT / 2)
160 #define BASE ((unsigned HOST_WIDE_INT) 1 << HOST_BITS_PER_WIDE_INT / 2)
162 /* Unpack a two-word integer into 4 words.
163 LOW and HI are the integer, as two `HOST_WIDE_INT' pieces.
164 WORDS points to the array of HOST_WIDE_INTs. */
166 static void
168 {
173 }
175 /* Pack an array of 4 words into a two-word integer.
176 WORDS points to the array of words.
177 The integer is stored into *LOW and *HI as two `HOST_WIDE_INT' pieces. */
179 static void
182 {
185 }
186 \f
187 /* T is an INT_CST node. OVERFLOWABLE indicates if we are interested
188 in overflow of the value, when >0 we are only interested in signed
189 overflow, for <0 we are interested in any overflow. OVERFLOWED
190 indicates whether overflow has already occurred. CONST_OVERFLOWED
191 indicates whether constant overflow has already occurred. We force
192 T's value to be within range of T's type (by setting to 0 or 1 all
193 the bits outside the type's range). We set TREE_OVERFLOWED if,
194 OVERFLOWED is nonzero,
195 or OVERFLOWABLE is >0 and signed overflow occurs
196 or OVERFLOWABLE is <0 and any overflow occurs
197 We set TREE_CONSTANT_OVERFLOWED if,
198 CONST_OVERFLOWED is nonzero
199 or we set TREE_OVERFLOWED.
200 We return either the original T, or a copy. */
202 tree
205 {
207 HOST_WIDE_INT high;
219 else
221 /* Size types *are* sign extended. */
226 /* First clear all bits that are beyond the type's precision. */
229 ;
232 else
233 {
237 }
244 {
245 /* Sign extend top half? */
249 }
251 {
254 }
255 else
256 {
257 /* Sign extend bottom half? */
259 {
262 }
263 }
265 /* If the value changed, return a new node. */
268 {
274 {
278 }
280 {
283 }
284 }
287 }
288 \f
289 /* Add two doubleword integers with doubleword result.
290 Each argument is given as two `HOST_WIDE_INT' pieces.
291 One argument is L1 and H1; the other, L2 and H2.
292 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
294 int
298 {
300 HOST_WIDE_INT h;
308 }
310 /* Negate a doubleword integer with doubleword result.
311 Return nonzero if the operation overflows, assuming it's signed.
312 The argument is given as two `HOST_WIDE_INT' pieces in L1 and H1.
313 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
315 int
318 {
320 {
324 }
325 else
326 {
330 }
331 }
332 \f
333 /* Multiply two doubleword integers with doubleword result.
334 Return nonzero if the operation overflows, assuming it's signed.
335 Each argument is given as two `HOST_WIDE_INT' pieces.
336 One argument is L1 and H1; the other, L2 and H2.
337 The value is stored as two `HOST_WIDE_INT' pieces in *LV and *HV. */
339 int
343 {
358 {
361 {
363 /* This product is <= 0xFFFE0001, the sum <= 0xFFFF0000. */
365 /* Since prod[p] < 0xFFFF, this sum <= 0xFFFFFFFF. */
369 }
371 }
375 /* Check for overflow by calculating the top half of the answer in full;
376 it should agree with the low half's sign bit. */
379 {
382 }
384 {
387 }
389 }
390 \f
391 /* Shift the doubleword integer in L1, H1 left by COUNT places
392 keeping only PREC bits of result.
393 Shift right if COUNT is negative.
394 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
395 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
397 void
401 {
405 {
408 }
414 {
415 /* Shifting by the host word size is undefined according to the
416 ANSI standard, so we must handle this as a special case. */
419 }
421 {
424 }
425 else
426 {
430 }
432 /* Sign extend all bits that are beyond the precision. */
440 ;
442 {
445 }
446 else
447 {
451 }
452 }
454 /* Shift the doubleword integer in L1, H1 right by COUNT places
455 keeping only PREC bits of result. COUNT must be positive.
456 ARITH nonzero specifies arithmetic shifting; otherwise use logical shift.
457 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
459 void
464 {
467 signmask = (arith
475 {
476 /* Shifting by the host word size is undefined according to the
477 ANSI standard, so we must handle this as a special case. */
480 }
482 {
485 }
486 else
487 {
491 }
493 /* Zero / sign extend all bits that are beyond the precision. */
496 {
499 }
501 ;
503 {
506 }
507 else
508 {
512 }
513 }
514 \f
515 /* Rotate the doubleword integer in L1, H1 left by COUNT places
516 keeping only PREC bits of result.
517 Rotate right if COUNT is negative.
518 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
520 void
524 {
536 }
538 /* Rotate the doubleword integer in L1, H1 left by COUNT places
539 keeping only PREC bits of result. COUNT must be positive.
540 Store the value as two `HOST_WIDE_INT' pieces in *LV and *HV. */
542 void
546 {
558 }
559 \f
560 /* Divide doubleword integer LNUM, HNUM by doubleword integer LDEN, HDEN
561 for a quotient (stored in *LQUO, *HQUO) and remainder (in *LREM, *HREM).
562 CODE is a tree code for a kind of division, one of
563 TRUNC_DIV_EXPR, FLOOR_DIV_EXPR, CEIL_DIV_EXPR, ROUND_DIV_EXPR
564 or EXACT_DIV_EXPR
565 It controls how the quotient is rounded to an integer.
566 Return nonzero if the operation overflows.
567 UNS nonzero says do unsigned division. */
569 int
572 HOST_WIDE_INT hnum_orig,
574 HOST_WIDE_INT hden_orig,
578 {
594 /* Calculate quotient sign and convert operands to unsigned. */
596 {
598 {
600 /* (minimum integer) / (-1) is the only overflow case. */
604 }
606 {
609 }
610 }
615 /* This unsigned division rounds toward zero. */
618 }
622 /* hden != 0 already checked. */
627 }
637 /* Special code for when the divisor < BASE. */
639 {
640 /* hnum != 0 already checked. */
642 {
646 }
647 }
648 else
649 {
650 /* Full double precision division,
651 with thanks to Don Knuth's "Seminumerical Algorithms". */
655 /* Find the highest nonzero divisor digit. */
658 {
661 }
663 /* Insure that the first digit of the divisor is at least BASE/2.
664 This is required by the quotient digit estimation algorithm. */
671 {
675 }
680 {
685 }
686 }
690 /* Main loop */
692 {
693 /* Guess the next quotient digit, quo_est, by dividing the first
694 two remaining dividend digits by the high order quotient digit.
695 quo_est is never low and is at most 2 high. */
702 else
705 /* Refine quo_est so it's usually correct, and at most one high. */
710 quo_est--;
712 /* Try QUO_EST as the quotient digit, by multiplying the
713 divisor by QUO_EST and subtracting from the remaining dividend.
714 Keep in mind that QUO_EST is the I - 1st digit. */
718 {
724 }
726 /* If quo_est was high by one, then num[i] went negative and
727 we need to correct things. */
729 {
730 quo_est--;
733 {
737 }
740 }
742 /* Store the quotient digit. */
744 }
745 }
749 finish_up:
750 /* If result is negative, make it so. */
754 /* Compute trial remainder: rem = num - (quo * den) */
760 {
769 {
770 /* quo = quo - 1; */
773 }
774 else
781 {
784 }
785 else
791 {
797 /* Get absolute values. */
803 /* If (2 * abs (lrem) >= abs (lden)) */
812 {
814 /* quo = quo - 1; */
817 else
818 /* quo = quo + 1; */
821 }
822 else
824 }
829 }
831 /* Compute true remainder: rem = num - (quo * den) */
836 }
837 \f
838 /* Return true if built-in mathematical function specified by CODE
839 preserves the sign of it argument, i.e. -f(x) == f(-x). */
841 static bool
843 {
845 {
862 }
864 }
866 /* Check whether we may negate an integer constant T without causing
867 overflow. */
869 bool
871 {
874 tree type;
884 {
889 }
890 else
895 }
897 /* Determine whether an expression T can be cheaply negated using
898 the function negate_expr. */
900 static bool
902 {
903 tree type;
912 {
917 /* Check that -CST will not overflow type. */
931 /* -(A + B) -> (-B) - A. */
936 /* -(A + B) -> (-A) - B. */
940 /* We can't turn -(A-B) into B-A when we honor signed zeros. */
949 /* Fall through. */
958 /* Negate -((double)float) as (double)(-float). */
960 {
964 }
968 /* Negate -f(x) as f(-x). */
974 /* Optimize -((int)x >> 31) into (unsigned)x >> 31. */
976 {
982 }
987 }
989 }
991 /* Given T, an expression, return the negation of T. Allow for T to be
992 null, in which case return null. */
994 static tree
996 {
997 tree type;
998 tree tem;
1007 {
1018 /* Two's complement FP formats, such as c4x, may overflow. */
1024 {
1033 }
1041 {
1042 /* -(A + B) -> (-B) - A. */
1046 {
1051 }
1053 /* -(A + B) -> (-A) - B. */
1055 {
1060 }
1061 }
1065 /* - (A - B) -> B - A */
1078 /* Fall through. */
1082 {
1095 }
1099 /* Convert -((double)float) into (double)(-float). */
1101 {
1105 }
1109 /* Negate -f(x) as f(-x). */
1112 {
1119 }
1123 /* Optimize -((int)x >> 31) into (unsigned)x >> 31. */
1125 {
1130 {
1137 }
1138 }
1143 }
1147 }
1148 \f
1149 /* Split a tree IN into a constant, literal and variable parts that could be
1150 combined with CODE to make IN. "constant" means an expression with
1151 TREE_CONSTANT but that isn't an actual constant. CODE must be a
1152 commutative arithmetic operation. Store the constant part into *CONP,
1153 the literal in *LITP and return the variable part. If a part isn't
1154 present, set it to null. If the tree does not decompose in this way,
1155 return the entire tree as the variable part and the other parts as null.
1157 If CODE is PLUS_EXPR we also split trees that use MINUS_EXPR. In that
1158 case, we negate an operand that was subtracted. Except if it is a
1159 literal for which we use *MINUS_LITP instead.
1161 If NEGATE_P is true, we are negating all of IN, again except a literal
1162 for which we use *MINUS_LITP instead.
1164 If IN is itself a literal or constant, return it as appropriate.
1166 Note that we do not guarantee that any of the three values will be the
1167 same type as IN, but they will have the same signedness and mode. */
1169 static tree
1172 {
1179 /* Strip any conversions that don't change the machine mode or signedness. */
1186 /* We can associate addition and subtraction together (even
1187 though the C standard doesn't say so) for integers because
1188 the value is not affected. For reals, the value might be
1189 affected, so we can't. */
1192 {
1198 /* First see if either of the operands is a literal, then a constant. */
1209 /* If we haven't dealt with either operand, this is not a case we can
1210 decompose. Otherwise, VAR is either of the ones remaining, if any. */
1215 else
1218 /* Now do any needed negations. */
1225 }
1228 else
1232 {
1239 }
1242 }
1244 /* Re-associate trees split by the above function. T1 and T2 are either
1245 expressions to associate or null. Return the new expression, if any. If
1246 we build an operation, do it in TYPE and with CODE. */
1248 static tree
1250 {
1256 /* If either input is CODE, a PLUS_EXPR, or a MINUS_EXPR, don't
1257 try to fold this since we will have infinite recursion. But do
1258 deal with any NEGATE_EXPRs. */
1261 {
1263 {
1272 }
1274 {
1277 }
1281 }
1285 }
1286 \f
1287 /* Combine two integer constants ARG1 and ARG2 under operation CODE
1288 to produce a new constant.
1290 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1292 tree
1294 {
1298 HOST_WIDE_INT hi;
1300 HOST_WIDE_INT garbageh;
1301 tree t;
1304 int is_sizetype
1315 {
1331 /* It's unclear from the C standard whether shifts can overflow.
1332 The following code ignores overflow; perhaps a C standard
1333 interpretation ruling is needed. */
1363 /* This is a shortcut for a common special case. */
1368 {
1374 }
1376 /* ... fall through ... */
1380 {
1383 }
1386 {
1389 }
1396 /* This is a shortcut for a common special case. */
1401 {
1406 }
1408 /* ... fall through ... */
1424 else
1430 else
1436 }
1441 {
1442 /* Propagate overflow flags ourselves. */
1445 {
1449 }
1451 {
1454 }
1455 }
1456 else
1464 }
1466 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1467 constant. We assume ARG1 and ARG2 have the same data type, or at least
1468 are the same kind of constant and the same machine mode.
1470 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1472 static tree
1474 {
1482 {
1484 REAL_VALUE_TYPE d1;
1485 REAL_VALUE_TYPE d2;
1486 REAL_VALUE_TYPE value;
1495 /* Don't perform operation if we honor signaling NaNs and
1496 either operand is a NaN. */
1501 /* Don't perform operation if it would raise a division
1502 by zero exception. */
1508 /* If either operand is a NaN, just return it. Otherwise, set up
1509 for floating-point trap; we return an overflow. */
1525 }
1527 {
1533 tree t;
1536 {
1556 notrunc),
1562 notrunc));
1566 {
1567 tree magsquared
1571 notrunc);
1574 const_binop
1579 notrunc),
1581 notrunc),
1582 notrunc),
1584 const_binop
1589 notrunc),
1591 notrunc),
1592 notrunc),
1594 }
1599 }
1601 }
1603 }
1605 /* Create a size type INT_CST node with NUMBER sign extended. KIND
1606 indicates which particular sizetype to create. */
1608 tree
1610 {
1612 }
1613 \f
1614 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
1615 is a tree code. The type of the result is taken from the operands.
1616 Both must be the same type integer type and it must be a size type.
1617 If the operands are constant, so is the result. */
1619 tree
1621 {
1627 /* Handle the special case of two integer constants faster. */
1629 {
1630 /* And some specific cases even faster than that. */
1639 /* Handle general case of two integer constants. */
1641 }
1647 }
1649 /* Given two values, either both of sizetype or both of bitsizetype,
1650 compute the difference between the two values. Return the value
1651 in signed type corresponding to the type of the operands. */
1653 tree
1655 {
1657 tree ctype;
1662 /* If the type is already signed, just do the simple thing. */
1668 /* If either operand is not a constant, do the conversions to the signed
1669 type and subtract. The hardware will do the right thing with any
1670 overflow in the subtraction. */
1675 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
1676 Otherwise, subtract the other way, convert to CTYPE (we know that can't
1677 overflow) and negate (which can't either). Special-case a result
1678 of zero while we're here. */
1683 else
1687 }
1688 \f
1689 /* Construct a vector of zero elements of vector type TYPE. */
1691 static tree
1693 {
1704 }
1707 /* Attempt to fold type conversion operation CODE of expression ARG1 to
1708 type TYPE. If no simplification can be done return NULL_TREE. */
1710 static tree
1712 {
1714 tree t;
1720 {
1722 {
1723 /* If we would build a constant wider than GCC supports,
1724 leave the conversion unfolded. */
1728 /* Given an integer constant, make new constant with new type,
1729 appropriately sign-extended or truncated. */
1734 /* Don't set the overflow when
1735 converting a pointer */
1743 }
1745 {
1746 /* The following code implements the floating point to integer
1747 conversion rules required by the Java Language Specification,
1748 that IEEE NaNs are mapped to zero and values that overflow
1749 the target precision saturate, i.e. values greater than
1750 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
1751 are mapped to INT_MIN. These semantics are allowed by the
1752 C and C++ standards that simply state that the behavior of
1753 FP-to-integer conversion is unspecified upon overflow. */
1756 REAL_VALUE_TYPE r;
1760 {
1779 }
1781 /* If R is NaN, return zero and show we have an overflow. */
1783 {
1787 }
1789 /* See if R is less than the lower bound or greater than the
1790 upper bound. */
1793 {
1797 {
1801 }
1802 }
1805 {
1808 {
1811 {
1815 }
1816 }
1817 }
1827 }
1828 }
1830 {
1834 {
1836 {
1837 /* We make a copy of ARG1 so that we don't modify an
1838 existing constant tree. */
1842 }
1852 }
1853 }
1855 }
1857 /* Convert expression ARG to type TYPE. Used by the middle-end for
1858 simple conversions in preference to calling the front-end's convert. */
1860 tree
1862 {
1864 tree tem;
1880 {
1885 {
1889 }
1894 {
1897 }
1904 {
1908 }
1910 {
1914 }
1917 {
1933 }
1937 {
1946 {
1950 {
1954 }
1962 }
1966 }
1981 }
1982 }
1983 \f
1984 /* Return an expr equal to X but certainly not valid as an lvalue. */
1986 tree
1988 {
1989 /* While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to
1990 us. */
1994 /* We only need to wrap lvalue tree codes. */
1996 {
2030 /* Assume the worst for front-end tree codes. */
2034 }
2036 }
2038 /* Nonzero means lvalues are limited to those valid in pedantic ANSI C.
2039 Zero means allow extended lvalues. */
2043 /* When pedantic, return an expr equal to X but certainly not valid as a
2044 pedantic lvalue. Otherwise, return X. */
2046 static tree
2048 {
2051 else
2053 }
2054 \f
2055 /* Given a tree comparison code, return the code that is the logical inverse
2056 of the given code. It is not safe to do this for floating-point
2057 comparisons, except for NE_EXPR and EQ_EXPR, so we receive a machine mode
2058 as well: if reversing the comparison is unsafe, return ERROR_MARK. */
2060 static enum tree_code
2062 {
2067 {
2098 }
2099 }
2101 /* Similar, but return the comparison that results if the operands are
2102 swapped. This is safe for floating-point. */
2104 enum tree_code
2106 {
2108 {
2122 }
2123 }
2126 /* Convert a comparison tree code from an enum tree_code representation
2127 into a compcode bit-based encoding. This function is the inverse of
2128 compcode_to_comparison. */
2130 static enum comparison_code
2132 {
2134 {
2165 }
2166 }
2168 /* Convert a compcode bit-based encoding of a comparison operator back
2169 to GCC's enum tree_code representation. This function is the
2170 inverse of comparison_to_compcode. */
2172 static enum tree_code
2174 {
2176 {
2207 }
2208 }
2210 /* Return a tree for the comparison which is the combination of
2211 doing the AND or OR (depending on CODE) of the two operations LCODE
2212 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2213 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2214 if this makes the transformation invalid. */
2216 tree
2220 {
2227 {
2238 }
2241 {
2242 /* Eliminate unordered comparisons, as well as LTGT and ORD
2243 which are not used unless the mode has NaNs. */
2249 }
2251 {
2252 /* Check that the original operation and the optimized ones will trap
2253 under the same condition. */
2264 /* In a short-circuited boolean expression the LHS might be
2265 such that the RHS, if evaluated, will never trap. For
2266 example, in ORD (x, y) && (x < y), we evaluate the RHS only
2267 if neither x nor y is NaN. (This is a mixed blessing: for
2268 example, the expression above will never trap, hence
2269 optimizing it to x < y would be invalid). */
2274 /* If the comparison was short-circuited, and only the RHS
2275 trapped, we may now generate a spurious trap. */
2280 /* If we changed the conditions that cause a trap, we lose. */
2283 }
2289 else
2292 }
2294 /* Return nonzero if CODE is a tree code that represents a truth value. */
2296 static int
2298 {
2303 }
2304 \f
2305 /* Return nonzero if two operands (typically of the same tree node)
2306 are necessarily equal. If either argument has side-effects this
2307 function returns zero. FLAGS modifies behavior as follows:
2309 If OEP_ONLY_CONST is set, only return nonzero for constants.
2310 This function tests whether the operands are indistinguishable;
2311 it does not test whether they are equal using C's == operation.
2312 The distinction is important for IEEE floating point, because
2313 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
2314 (2) two NaNs may be indistinguishable, but NaN!=NaN.
2316 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
2317 even though it may hold multiple values during a function.
2318 This is because a GCC tree node guarantees that nothing else is
2319 executed between the evaluation of its "operands" (which may often
2320 be evaluated in arbitrary order). Hence if the operands themselves
2321 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
2322 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
2323 unset means assuming isochronic (or instantaneous) tree equivalence.
2324 Unless comparing arbitrary expression trees, such as from different
2325 statements, this flag can usually be left unset.
2327 If OEP_PURE_SAME is set, then pure functions with identical arguments
2328 are considered the same. It is used when the caller has other ways
2329 to ensure that global memory is unchanged in between. */
2331 int
2333 {
2334 /* If one is specified and the other isn't, they aren't equal and if
2335 neither is specified, they are.
2337 ??? This is temporary and is meant only to handle the cases of the
2338 optional operands for COMPONENT_REF and ARRAY_REF. */
2343 /* If either is ERROR_MARK, they aren't equal. */
2347 /* If both types don't have the same signedness, then we can't consider
2348 them equal. We must check this before the STRIP_NOPS calls
2349 because they may change the signedness of the arguments. */
2357 /* This is needed for conversions and for COMPONENT_REF.
2358 Might as well play it safe and always test this. */
2364 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
2365 We don't care about side effects in that case because the SAVE_EXPR
2366 takes care of that for us. In all other cases, two expressions are
2367 equal if they have no side effects. If we have two identical
2368 expressions with side effects that should be treated the same due
2369 to the only side effects being identical SAVE_EXPR's, that will
2370 be detected in the recursive calls below. */
2376 /* Next handle constant cases, those for which we can return 1 even
2377 if ONLY_CONST is set. */
2380 {
2393 {
2403 {
2405 flags))
2409 }
2412 }
2416 flags)
2418 flags));
2431 }
2437 {
2439 /* Two conversions are equal only if signedness and modes match. */
2441 {
2454 }
2467 /* For commutative ops, allow the other order. */
2475 /* If either of the pointer (or reference) expressions we are
2476 dereferencing contain a side effect, these cannot be equal. */
2482 {
2521 }
2525 {
2551 /* If the CALL_EXPRs call different functions, then they
2552 clearly can not be equal. */
2557 {
2561 else
2565 }
2567 /* Now see if all the arguments are the same. operand_equal_p
2568 does not handle TREE_LIST, so we walk the operands here
2569 feeding them to operand_equal_p. */
2573 {
2575 flags))
2580 }
2582 /* If we get here and both argument lists are exhausted
2583 then the CALL_EXPRs are equal. */
2588 }
2591 /* Consider __builtin_sqrt equal to sqrt. */
2599 }
2600 }
2601 \f
2602 /* Similar to operand_equal_p, but see if ARG0 might have been made by
2603 shorten_compare from ARG1 when ARG1 was being compared with OTHER.
2605 When in doubt, return 0. */
2607 static int
2609 {
2621 /* Discard any conversions that don't change the modes of ARG0 and ARG1
2622 and see if the inner values are the same. This removes any
2623 signedness comparison, which doesn't matter here. */
2630 /* Duplicate what shorten_compare does to ARG1 and see if that gives the
2631 actual comparison operand, ARG0.
2633 First throw away any conversions to wider types
2634 already present in the operands. */
2643 {
2646 /* Make sure shorter operand is extended the right way
2647 to match the longer operand. */
2653 }
2656 }
2657 \f
2658 /* See if ARG is an expression that is either a comparison or is performing
2659 arithmetic on comparisons. The comparisons must only be comparing
2660 two different values, which will be stored in *CVAL1 and *CVAL2; if
2661 they are nonzero it means that some operands have already been found.
2662 No variables may be used anywhere else in the expression except in the
2663 comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around
2664 the expression and save_expr needs to be called with CVAL1 and CVAL2.
2666 If this is true, return 1. Otherwise, return zero. */
2668 static int
2670 {
2674 /* We can handle some of the tcc_expression cases here. */
2684 {
2685 /* If we've already found a CVAL1 or CVAL2, this expression is
2686 two complex to handle. */
2692 }
2695 {
2718 /* First see if we can handle the first operand, then the second. For
2719 the second operand, we know *CVAL1 can't be zero. It must be that
2720 one side of the comparison is each of the values; test for the
2721 case where this isn't true by failing if the two operands
2722 are the same. */
2731 ;
2735 ;
2736 else
2740 ;
2744 ;
2745 else
2752 }
2753 }
2754 \f
2755 /* ARG is a tree that is known to contain just arithmetic operations and
2756 comparisons. Evaluate the operations in the tree substituting NEW0 for
2757 any occurrence of OLD0 as an operand of a comparison and likewise for
2758 NEW1 and OLD1. */
2760 static tree
2762 {
2767 /* We can handle some of the tcc_expression cases here. */
2775 {
2790 {
2807 }
2808 /* Fall through - ??? */
2811 {
2815 /* We need to check both for exact equality and tree equality. The
2816 former will be true if the operand has a side-effect. In that
2817 case, we know the operand occurred exactly once. */
2830 }
2834 }
2835 }
2836 \f
2837 /* Return a tree for the case when the result of an expression is RESULT
2838 converted to TYPE and OMITTED was previously an operand of the expression
2839 but is now not needed (e.g., we folded OMITTED * 0).
2841 If OMITTED has side effects, we must evaluate it. Otherwise, just do
2842 the conversion of RESULT to TYPE. */
2844 tree
2846 {
2853 }
2855 /* Similar, but call pedantic_non_lvalue instead of non_lvalue. */
2857 static tree
2859 {
2866 }
2868 /* Return a tree for the case when the result of an expression is RESULT
2869 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
2870 of the expression but are now not needed.
2872 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
2873 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
2874 evaluated before OMITTED2. Otherwise, if neither has side effects,
2875 just do the conversion of RESULT to TYPE. */
2877 tree
2879 {
2888 }
2890 \f
2891 /* Return a simplified tree node for the truth-negation of ARG. This
2892 never alters ARG itself. We assume that ARG is an operation that
2893 returns a truth value (0 or 1).
2895 FIXME: one would think we would fold the result, but it causes
2896 problems with the dominator optimizer. */
2897 tree
2899 {
2906 /* If this is a comparison, we can simply invert it, except for
2907 floating-point non-equality comparisons, in which case we just
2908 enclose a TRUTH_NOT_EXPR around what we have. */
2911 {
2914 && flag_trapping_math
2918 else
2919 {
2924 else
2927 }
2928 }
2931 {
2947 /* Here we can invert either operand. We invert the first operand
2948 unless the second operand is a TRUTH_NOT_EXPR in which case our
2949 result is the XOR of the first operand with the inside of the
2950 negation of the second operand. */
2955 else
3009 }
3012 }
3014 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
3015 operands are another bit-wise operation with a common input. If so,
3016 distribute the bit operations to save an operation and possibly two if
3017 constants are involved. For example, convert
3018 (A | B) & (A | C) into A | (B & C)
3019 Further simplification will occur if B and C are constants.
3021 If this optimization cannot be done, 0 will be returned. */
3023 static tree
3025 {
3026 tree common;
3036 {
3040 }
3042 {
3046 }
3048 {
3052 }
3054 {
3058 }
3059 else
3064 }
3065 \f
3066 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
3067 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero. */
3069 static tree
3072 {
3079 }
3081 /* Optimize a bit-field compare.
3083 There are two cases: First is a compare against a constant and the
3084 second is a comparison of two items where the fields are at the same
3085 bit position relative to the start of a chunk (byte, halfword, word)
3086 large enough to contain it. In these cases we can avoid the shift
3087 implicit in bitfield extractions.
3089 For constants, we emit a compare of the shifted constant with the
3090 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
3091 compared. For two fields at the same position, we do the ANDs with the
3092 similar mask and compare the result of the ANDs.
3094 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
3095 COMPARE_TYPE is the type of the comparison, and LHS and RHS
3096 are the left and right operands of the comparison, respectively.
3098 If the optimization described above can be done, we return the resulting
3099 tree. Otherwise we return zero. */
3101 static tree
3104 {
3113 tree mask;
3114 tree offset;
3116 /* Get all the information about the extractions being done. If the bit size
3117 if the same as the size of the underlying object, we aren't doing an
3118 extraction at all and so can do nothing. We also don't want to
3119 do anything if the inner expression is a PLACEHOLDER_EXPR since we
3120 then will no longer be able to replace it. */
3128 {
3129 /* If this is not a constant, we can only do something if bit positions,
3130 sizes, and signedness are the same. */
3138 }
3140 /* See if we can find a mode to refer to this field. We should be able to,
3141 but fail if we can't. */
3150 /* Set signed and unsigned types of the precision of this mode for the
3151 shifts below. */
3155 /* Compute the bit position and size for the new reference and our offset
3156 within it. If the new reference is the same size as the original, we
3157 won't optimize anything, so return zero. */
3167 /* Make the mask to be used against the extracted field. */
3176 /* If not comparing with constant, just rework the comparison
3177 and return. */
3182 mask),
3186 mask));
3188 /* Otherwise, we are handling the constant case. See if the constant is too
3189 big for the field. Warn and return a tree of for 0 (false) if so. We do
3190 this not only for its own sake, but to avoid having to test for this
3191 error case below. If we didn't, we might generate wrong code.
3193 For unsigned fields, the constant shifted right by the field length should
3194 be all zero. For signed fields, the high-order bits should agree with
3195 the sign bit. */
3198 {
3202 {
3206 }
3207 }
3208 else
3209 {
3213 {
3217 }
3218 }
3220 /* Single-bit compares should always be against zero. */
3222 {
3225 }
3227 /* Make a new bitfield reference, shift the constant over the
3228 appropriate number of bits and mask it with the computed mask
3229 (in case this was a signed field). If we changed it, make a new one. */
3232 {
3235 }
3245 rhs);
3246 }
3247 \f
3248 /* Subroutine for fold_truthop: decode a field reference.
3250 If EXP is a comparison reference, we return the innermost reference.
3252 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
3253 set to the starting bit number.
3255 If the innermost field can be completely contained in a mode-sized
3256 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
3258 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
3259 otherwise it is not changed.
3261 *PUNSIGNEDP is set to the signedness of the field.
3263 *PMASK is set to the mask used. This is either contained in a
3264 BIT_AND_EXPR or derived from the width of the field.
3266 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
3268 Return 0 if this is not a component reference or is one that we can't
3269 do anything with. */
3271 static tree
3276 {
3280 tree unsigned_type;
3283 /* All the optimizations using this function assume integer fields.
3284 There are problems with FP fields since the type_for_size call
3285 below can fail for, e.g., XFmode. */
3289 /* We are interested in the bare arrangement of bits, so strip everything
3290 that doesn't affect the machine mode. However, record the type of the
3291 outermost expression if it may matter below. */
3299 {
3305 }
3314 /* If the number of bits in the reference is the same as the bitsize of
3315 the outer type, then the outer type gives the signedness. Otherwise
3316 (in case of a small bitfield) the signedness is unchanged. */
3320 /* Compute the mask to access the bitfield. */
3330 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
3338 }
3340 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
3341 bit positions. */
3343 static int
3345 {
3348 tree tmask;
3353 return
3360 }
3362 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
3363 represents the sign bit of EXP's type. If EXP represents a sign
3364 or zero extension, also test VAL against the unextended type.
3365 The return value is the (sub)expression whose sign bit is VAL,
3366 or NULL_TREE otherwise. */
3368 static tree
3370 {
3374 tree t;
3376 /* Tree EXP must have an integral type. */
3381 /* Tree VAL must be an integer constant. */
3388 {
3395 }
3396 else
3397 {
3404 }
3406 /* We mask off those bits beyond TREE_TYPE (exp) so that we can
3407 treat VAL as if it were unsigned. */
3412 /* Handle extension from a narrower type. */
3418 }
3420 /* Subroutine for fold_truthop: determine if an operand is simple enough
3421 to be evaluated unconditionally. */
3423 static int
3425 {
3426 /* Strip any conversions that don't change the machine mode. */
3435 /* Don't regard global variables as simple. They may be
3436 allocated in ways unknown to the compiler (shared memory,
3437 #pragma weak, etc). */
3440 /* Loading a static variable is unduly expensive, but global
3441 registers aren't expensive. */
3443 }
3444 \f
3445 /* The following functions are subroutines to fold_range_test and allow it to
3446 try to change a logical combination of comparisons into a range test.
3448 For example, both
3449 X == 2 || X == 3 || X == 4 || X == 5
3450 and
3451 X >= 2 && X <= 5
3452 are converted to
3453 (unsigned) (X - 2) <= 3
3455 We describe each set of comparisons as being either inside or outside
3456 a range, using a variable named like IN_P, and then describe the
3457 range with a lower and upper bound. If one of the bounds is omitted,
3458 it represents either the highest or lowest value of the type.
3460 In the comments below, we represent a range by two numbers in brackets
3461 preceded by a "+" to designate being inside that range, or a "-" to
3462 designate being outside that range, so the condition can be inverted by
3463 flipping the prefix. An omitted bound is represented by a "-". For
3464 example, "- [-, 10]" means being outside the range starting at the lowest
3465 possible value and ending at 10, in other words, being greater than 10.
3466 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
3467 always false.
3469 We set up things so that the missing bounds are handled in a consistent
3470 manner so neither a missing bound nor "true" and "false" need to be
3471 handled using a special case. */
3473 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
3474 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
3475 and UPPER1_P are nonzero if the respective argument is an upper bound
3476 and zero for a lower. TYPE, if nonzero, is the type of the result; it
3477 must be specified for a comparison. ARG1 will be converted to ARG0's
3478 type if both are specified. */
3480 static tree
3483 {
3484 tree tem;
3488 /* If neither arg represents infinity, do the normal operation.
3489 Else, if not a comparison, return infinity. Else handle the special
3490 comparison rules. Note that most of the cases below won't occur, but
3491 are handled for consistency. */
3494 {
3499 }
3504 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
3505 for neither. In real maths, we cannot assume open ended ranges are
3506 the same. But, this is computer arithmetic, where numbers are finite.
3507 We can therefore make the transformation of any unbounded range with
3508 the value Z, Z being greater than any representable number. This permits
3509 us to treat unbounded ranges as equal. */
3513 {
3534 }
3537 }
3538 \f
3539 /* Given EXP, a logical expression, set the range it is testing into
3540 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
3541 actually being tested. *PLOW and *PHIGH will be made of the same type
3542 as the returned expression. If EXP is not a comparison, we will most
3543 likely not be returning a useful value and range. */
3545 static tree
3547 {
3554 /* Start with simply saying "EXP != 0" and then look at the code of EXP
3555 and see if we can refine the range. Some of the cases below may not
3556 happen, but it doesn't seem worth worrying about this. We "continue"
3557 the outer loop when we've changed something; otherwise we "break"
3558 the switch, which will "break" the while. */
3564 {
3569 {
3581 }
3584 {
3591 /* We can only do something if the range is testing for zero
3592 and if the second operand is an integer constant. Note that
3593 saying something is "in" the range we make is done by
3594 complementing IN_P since it will set in the initial case of
3595 being not equal to zero; "out" is leaving it alone. */
3602 {
3623 }
3625 /* If this is an unsigned comparison, we also know that EXP is
3626 greater than or equal to zero. We base the range tests we make
3627 on that fact, so we record it here so we can parse existing
3628 range tests. We test arg0_type since often the return type
3629 of, e.g. EQ_EXPR, is boolean. */
3631 {
3635 NULL_TREE))
3640 /* If the high bound is missing, but we have a nonzero low
3641 bound, reverse the range so it goes from zero to the low bound
3642 minus 1. */
3644 {
3649 }
3650 }
3656 /* (-x) IN [a,b] -> x in [-b, -a] */
3668 /* ~ X -> -X - 1 */
3677 /* If EXP is signed, any overflow in the computation is undefined,
3678 so we don't worry about it so long as our computations on
3679 the bounds don't overflow. For unsigned, overflow is defined
3680 and this is exactly the right thing. */
3689 /* Check for an unsigned range which has wrapped around the maximum
3690 value thus making n_high < n_low, and normalize it. */
3692 {
3698 /* If the range is of the form +/- [ x+1, x ], we won't
3699 be able to normalize it. But then, it represents the
3700 whole range or the empty set, so make it
3701 +/- [ -, - ]. */
3705 else
3707 }
3708 else
3732 /* If we're converting arg0 from an unsigned type, to exp,
3733 a signed type, we will be doing the comparison as unsigned.
3734 The tests above have already verified that LOW and HIGH
3735 are both positive.
3737 So we have to ensure that we will handle large unsigned
3738 values the same way that the current signed bounds treat
3739 negative values. */
3742 {
3743 tree high_positive;
3747 /* A range without an upper bound is, naturally, unbounded.
3748 Since convert would have cropped a very large value, use
3749 the max value for the destination type. */
3750 high_positive
3757 high_positive),
3759 integer_one_node)));
3761 /* If the low bound is specified, "and" the range with the
3762 range for which the original unsigned value will be
3763 positive. */
3765 {
3769 integer_zero_node),
3770 high_positive))
3774 }
3775 else
3776 {
3777 /* Otherwise, "or" the range with the range of the input
3778 that will be interpreted as negative. */
3782 integer_zero_node),
3783 high_positive))
3787 }
3788 }
3796 }
3799 }
3801 /* If EXP is a constant, we can evaluate whether this is true or false. */
3803 {
3810 }
3814 }
3815 \f
3816 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
3817 type, TYPE, return an expression to test if EXP is in (or out of, depending
3818 on IN_P) the range. Return 0 if the test couldn't be created. */
3820 static tree
3822 {
3824 tree value;
3827 {
3833 }
3848 {
3850 {
3854 }
3856 }
3858 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
3860 {
3862 HOST_WIDE_INT hi;
3867 {
3870 }
3871 else
3872 {
3875 }
3878 {
3880 {
3883 }
3886 }
3887 }
3891 {
3894 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
3895 for the type in question, as we rely on this here. */
3897 {
3908 {
3914 }
3918 }
3919 }
3925 value);
3928 }
3929 \f
3930 /* Given two ranges, see if we can merge them into one. Return 1 if we
3931 can, 0 if we can't. Set the output range into the specified parameters. */
3933 static int
3936 {
3940 tree tem;
3950 /* Make range 0 be the range that starts first, or ends last if they
3951 start at the same value. Swap them if it isn't. */
3954 || (lowequal
3957 {
3961 }
3963 /* Now flag two cases, whether the ranges are disjoint or whether the
3964 second range is totally subsumed in the first. Note that the tests
3965 below are simplified by the ones above. */
3971 /* We now have four cases, depending on whether we are including or
3972 excluding the two ranges. */
3974 {
3975 /* If they don't overlap, the result is false. If the second range
3976 is a subset it is the result. Otherwise, the range is from the start
3977 of the second to the end of the first. */
3982 else
3984 }
3987 {
3988 /* If they don't overlap, the result is the first range. If they are
3989 equal, the result is false. If the second range is a subset of the
3990 first, and the ranges begin at the same place, we go from just after
3991 the end of the first range to the end of the second. If the second
3992 range is not a subset of the first, or if it is a subset and both
3993 ranges end at the same place, the range starts at the start of the
3994 first range and ends just before the second range.
3995 Otherwise, we can't describe this as a single range. */
4001 {
4005 }
4007 {
4011 }
4012 else
4014 }
4017 {
4018 /* If they don't overlap, the result is the second range. If the second
4019 is a subset of the first, the result is false. Otherwise,
4020 the range starts just after the first range and ends at the
4021 end of the second. */
4026 else
4027 {
4031 }
4032 }
4034 else
4035 {
4036 /* The case where we are excluding both ranges. Here the complex case
4037 is if they don't overlap. In that case, the only time we have a
4038 range is if they are adjacent. If the second is a subset of the
4039 first, the result is the first. Otherwise, the range to exclude
4040 starts at the beginning of the first range and ends at the end of the
4041 second. */
4043 {
4050 else
4051 {
4052 /* Canonicalize - [min, x] into - [-, x]. */
4055 {
4060 /* FALLTHROUGH */
4074 }
4076 /* Canonicalize - [x, max] into - [x, -]. */
4079 {
4084 /* FALLTHROUGH */
4100 }
4102 /* The ranges might be also adjacent between the maximum and
4103 minimum values of the given type. For
4104 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
4105 return + [x + 1, y - 1]. */
4107 {
4116 }
4117 else
4119 }
4120 }
4123 else
4125 }
4129 }
4130 \f
4132 /* Subroutine of fold, looking inside expressions of the form
4133 A op B ? A : C, where ARG0, ARG1 and ARG2 are the three operands
4134 of the COND_EXPR. This function is being used also to optimize
4135 A op B ? C : A, by reversing the comparison first.
4137 Return a folded expression whose code is not a COND_EXPR
4138 anymore, or NULL_TREE if no folding opportunity is found. */
4140 static tree
4142 {
4147 tree tem;
4152 /* If we have A op 0 ? A : -A, consider applying the following
4153 transformations:
4155 A == 0? A : -A same as -A
4156 A != 0? A : -A same as A
4157 A >= 0? A : -A same as abs (A)
4158 A > 0? A : -A same as abs (A)
4159 A <= 0? A : -A same as -abs (A)
4160 A < 0? A : -A same as -abs (A)
4162 None of these transformations work for modes with signed
4163 zeros. If A is +/-0, the first two transformations will
4164 change the sign of the result (from +0 to -0, or vice
4165 versa). The last four will fix the sign of the result,
4166 even though the original expressions could be positive or
4167 negative, depending on the sign of A.
4169 Note that all these transformations are correct if A is
4170 NaN, since the two alternatives (A and -A) are also NaNs. */
4177 {
4189 /* Fall through. */
4211 }
4213 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
4214 A == 0 ? A : 0 is always 0 unless A is -0. Note that
4215 both transformations are correct when A is NaN: A != 0
4216 is then true, and A == 0 is false. */
4219 {
4224 }
4226 /* Try some transformations of A op B ? A : B.
4228 A == B? A : B same as B
4229 A != B? A : B same as A
4230 A >= B? A : B same as max (A, B)
4231 A > B? A : B same as max (B, A)
4232 A <= B? A : B same as min (A, B)
4233 A < B? A : B same as min (B, A)
4235 As above, these transformations don't work in the presence
4236 of signed zeros. For example, if A and B are zeros of
4237 opposite sign, the first two transformations will change
4238 the sign of the result. In the last four, the original
4239 expressions give different results for (A=+0, B=-0) and
4240 (A=-0, B=+0), but the transformed expressions do not.
4242 The first two transformations are correct if either A or B
4243 is a NaN. In the first transformation, the condition will
4244 be false, and B will indeed be chosen. In the case of the
4245 second transformation, the condition A != B will be true,
4246 and A will be chosen.
4248 The conversions to max() and min() are not correct if B is
4249 a number and A is not. The conditions in the original
4250 expressions will be false, so all four give B. The min()
4251 and max() versions would give a NaN instead. */
4253 {
4258 /* Avoid adding NOP_EXPRs in case this is an lvalue. */
4260 {
4264 }
4267 {
4276 /* In C++ a ?: expression can be an lvalue, so put the
4277 operand which will be used if they are equal first
4278 so that we can convert this back to the
4279 corresponding COND_EXPR. */
4281 {
4288 }
4295 {
4302 }
4315 }
4316 }
4318 /* If this is A op C1 ? A : C2 with C1 and C2 constant integers,
4319 we might still be able to simplify this. For example,
4320 if C1 is one less or one more than C2, this might have started
4321 out as a MIN or MAX and been transformed by this function.
4322 Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE. */
4328 {
4330 /* We can replace A with C1 in this case. */
4335 /* If C1 is C2 + 1, this is min(A, C2). */
4337 OEP_ONLY_CONST)
4341 OEP_ONLY_CONST))
4347 /* If C1 is C2 - 1, this is min(A, C2). */
4349 OEP_ONLY_CONST)
4353 OEP_ONLY_CONST))
4359 /* If C1 is C2 - 1, this is max(A, C2). */
4361 OEP_ONLY_CONST)
4365 OEP_ONLY_CONST))
4371 /* If C1 is C2 + 1, this is max(A, C2). */
4373 OEP_ONLY_CONST)
4377 OEP_ONLY_CONST))
4385 }
4388 }
4391 \f
4392 #ifndef LOGICAL_OP_NON_SHORT_CIRCUIT
4393 #define LOGICAL_OP_NON_SHORT_CIRCUIT (BRANCH_COST >= 2)
4394 #endif
4396 /* EXP is some logical combination of boolean tests. See if we can
4397 merge it into some range test. Return the new tree if so. */
4399 static tree
4401 {
4408 tree tem;
4410 /* If this is an OR operation, invert both sides; we will invert
4411 again at the end. */
4415 /* If both expressions are the same, if we can merge the ranges, and we
4416 can build the range test, return it or it inverted. If one of the
4417 ranges is always true or always false, consider it to be the same
4418 expression as the other. */
4428 /* On machines where the branch cost is expensive, if this is a
4429 short-circuited branch and the underlying object on both sides
4430 is the same, make a non-short-circuit operation. */
4436 {
4437 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
4438 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
4439 which cases we can't do this. */
4448 {
4460 }
4461 }
4464 }
4465 \f
4466 /* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
4467 bit value. Arrange things so the extra bits will be set to zero if and
4468 only if C is signed-extended to its full width. If MASK is nonzero,
4469 it is an INTEGER_CST that should be AND'ed with the extra bits. */
4471 static tree
4473 {
4476 tree temp;
4481 /* We work by getting just the sign bit into the low-order bit, then
4482 into the high-order bit, then sign-extend. We then XOR that value
4483 with C. */
4487 /* We must use a signed type in order to get an arithmetic right shift.
4488 However, we must also avoid introducing accidental overflows, so that
4489 a subsequent call to integer_zerop will work. Hence we must
4490 do the type conversion here. At this point, the constant is either
4491 zero or one, and the conversion to a signed type can never overflow.
4492 We could get an overflow if this conversion is done anywhere else. */
4501 /* If necessary, convert the type back to match the type of C. */
4506 }
4507 \f
4508 /* Find ways of folding logical expressions of LHS and RHS:
4509 Try to merge two comparisons to the same innermost item.
4510 Look for range tests like "ch >= '0' && ch <= '9'".
4511 Look for combinations of simple terms on machines with expensive branches
4512 and evaluate the RHS unconditionally.
4514 For example, if we have p->a == 2 && p->b == 4 and we can make an
4515 object large enough to span both A and B, we can do this with a comparison
4516 against the object ANDed with the a mask.
4518 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
4519 operations to do this with one comparison.
4521 We check for both normal comparisons and the BIT_AND_EXPRs made this by
4522 function and the one above.
4524 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
4525 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
4527 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
4528 two operands.
4530 We return the simplified tree or 0 if no optimization is possible. */
4532 static tree
4534 {
4535 /* If this is the "or" of two comparisons, we can do something if
4536 the comparisons are NE_EXPR. If this is the "and", we can do something
4537 if the comparisons are EQ_EXPR. I.e.,
4538 (a->b == 2 && a->c == 4) can become (a->new == NEW).
4540 WANTED_CODE is this operation code. For single bit fields, we can
4541 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
4542 comparison for one-bit fields. */
4562 /* Start by getting the comparison codes. Fail if anything is volatile.
4563 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
4564 it were surrounded with a NE_EXPR. */
4573 {
4577 }
4580 {
4584 }
4595 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
4598 {
4599 tree result;
4602 {
4607 }
4610 {
4616 }
4617 }
4622 /* If the RHS can be evaluated unconditionally and its operands are
4623 simple, it wins to evaluate the RHS unconditionally on machines
4624 with expensive branches. In this case, this isn't a comparison
4625 that can be merged. Avoid doing this if the RHS is a floating-point
4626 comparison since those can trap. */
4632 {
4633 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
4643 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
4655 }
4657 /* See if the comparisons can be merged. Then get all the parameters for
4658 each side. */
4682 /* It must be true that the inner operation on the lhs of each
4683 comparison must be the same if we are to be able to do anything.
4684 Then see if we have constants. If not, the same must be true for
4685 the rhs's. */
4696 else
4699 /* If either comparison code is not correct for our logical operation,
4700 fail. However, we can convert a one-bit comparison against zero into
4701 the opposite comparison against that bit being set in the field. */
4705 {
4707 {
4708 /* Make the left operand unsigned, since we are only interested
4709 in the value of one bit. Otherwise we are doing the wrong
4710 thing below. */
4713 }
4714 else
4716 }
4718 /* This is analogous to the code for l_const above. */
4720 {
4722 {
4725 }
4726 else
4728 }
4730 /* After this point all optimizations will generate bit-field
4731 references, which we might not want. */
4735 /* See if we can find a mode that contains both fields being compared on
4736 the left. If we can't, fail. Otherwise, update all constants and masks
4737 to be relative to a field of that size. */
4742 volatilep);
4752 {
4755 }
4763 {
4771 {
4775 }
4776 }
4778 {
4786 {
4790 }
4791 }
4793 /* If the right sides are not constant, do the same for it. Also,
4794 disallow this optimization if a size or signedness mismatch occurs
4795 between the left and right sides. */
4797 {
4800 /* Make sure the two fields on the right
4801 correspond to the left without being swapped. */
4809 volatilep);
4819 {
4822 }
4829 /* Make a mask that corresponds to both fields being compared.
4830 Do this for both items being compared. If the operands are the
4831 same size and the bits being compared are in the same position
4832 then we can do this by masking both and comparing the masked
4833 results. */
4837 {
4849 }
4851 /* There is still another way we can do something: If both pairs of
4852 fields being compared are adjacent, we may be able to make a wider
4853 field containing them both.
4855 Note that we still must mask the lhs/rhs expressions. Furthermore,
4856 the mask must be shifted to account for the shift done by
4857 make_bit_field_ref. */
4862 {
4863 tree type;
4875 /* Convert to the smaller type before masking out unwanted bits. */
4878 {
4880 {
4884 }
4886 {
4890 }
4891 }
4900 }
4903 }
4905 /* Handle the case of comparisons with constants. If there is something in
4906 common between the masks, those bits of the constants must be the same.
4907 If not, the condition is always false. Test for this to avoid generating
4908 incorrect code below. */
4913 {
4915 {
4918 }
4919 else
4920 {
4923 }
4924 }
4926 /* Construct the expression we will return. First get the component
4927 reference we will make. Unless the mask is all ones the width of
4928 that field, perform the mask operation. Then compare with the
4929 merged constant. */
4939 }
4940 \f
4941 /* Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a
4942 constant. */
4944 static tree
4946 {
4951 tree minmax_const;
4953 tree inner;
4963 /* If something does not permit us to optimize, return the original tree. */
4971 /* Now handle all the various comparison codes. We only handle EQ_EXPR
4972 and GT_EXPR, doing the rest with recursive calls using logical
4973 simplifications. */
4975 {
4977 return
4981 return
4983 optimize_minmax_comparison
4985 optimize_minmax_comparison
4990 /* MAX (X, 0) == 0 -> X <= 0 */
4994 /* MAX (X, 0) == 5 -> X == 5 */
4998 /* MAX (X, 0) == -1 -> false */
5002 /* MIN (X, 0) == 0 -> X >= 0 */
5006 /* MIN (X, 0) == 5 -> false */
5009 else
5010 /* MIN (X, 0) == -1 -> X == -1 */
5015 /* MAX (X, 0) > 0 -> X > 0
5016 MAX (X, 0) > 5 -> X > 5 */
5020 /* MAX (X, 0) > -1 -> true */
5024 /* MIN (X, 0) > 0 -> false
5025 MIN (X, 0) > 5 -> false */
5028 else
5029 /* MIN (X, 0) > -1 -> X > -1 */
5034 }
5035 }
5036 \f
5037 /* T is an integer expression that is being multiplied, divided, or taken a
5038 modulus (CODE says which and what kind of divide or modulus) by a
5039 constant C. See if we can eliminate that operation by folding it with
5040 other operations already in T. WIDE_TYPE, if non-null, is a type that
5041 should be used for the computation if wider than our type.
5043 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
5044 (X * 2) + (Y * 4). We must, however, be assured that either the original
5045 expression would not overflow or that overflow is undefined for the type
5046 in the language in question.
5048 We also canonicalize (X + 7) * 4 into X * 4 + 28 in the hope that either
5049 the machine has a multiply-accumulate insn or that this is part of an
5050 addressing calculation.
5052 If we return a non-null expression, it is an equivalent form of the
5053 original computation, but need not be in the original type. */
5055 static tree
5057 {
5058 /* To avoid exponential search depth, refuse to allow recursion past
5059 three levels. Beyond that (1) it's highly unlikely that we'll find
5060 something interesting and (2) we've probably processed it before
5061 when we built the inner expression. */
5064 tree ret;
5069 depth++;
5071 depth--;
5074 }
5076 static tree
5078 {
5088 /* Don't deal with constants of zero here; they confuse the code below. */
5098 /* Note that we need not handle conditional operations here since fold
5099 already handles those cases. So just do arithmetic here. */
5101 {
5103 /* For a constant, we can always simplify if we are a multiply
5104 or (for divide and modulus) if it is a multiple of our constant. */
5112 /* If op0 is an expression ... */
5117 /* ... and is unsigned, and its type is smaller than ctype,
5118 then we cannot pass through as widening. */
5124 /* ... or this is a truncation (t is narrower than op0),
5125 then we cannot pass through this narrowing. */
5128 /* ... or signedness changes for division or modulus,
5129 then we cannot pass through this conversion. */
5135 /* Pass the constant down and see if we can make a simplification. If
5136 we can, replace this expression with the inner simplification for
5137 possible later conversion to our or some other type. */
5142 code == MULT_EXPR
5153 /* If widening the type changes the signedness, then we can't perform
5154 this optimization as that changes the result. */
5158 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
5161 {
5167 }
5171 /* If the second operand is constant, this is a multiplication
5172 or floor division, by a power of two, so we can treat it that
5173 way unless the multiplier or divisor overflows. Signed
5174 left-shift overflow is implementation-defined rather than
5175 undefined in C90, so do not convert signed left shift into
5176 multiplication. */
5179 /* const_binop may not detect overflow correctly,
5180 so check for it explicitly here. */
5185 size_one_node,
5195 /* See if we can eliminate the operation on both sides. If we can, we
5196 can return a new PLUS or MINUS. If we can't, the only remaining
5197 cases where we can do anything are if the second operand is a
5198 constant. */
5203 /* If not multiplication, we can only do this if both operands
5204 are divisible by c. */
5210 /* If this was a subtraction, negate OP1 and set it to be an addition.
5211 This simplifies the logic below. */
5218 /* If either OP1 or C are negative, this optimization is not safe for
5219 some of the division and remainder types while for others we need
5220 to change the code. */
5222 {
5230 }
5232 /* If it's a multiply or a division/modulus operation of a multiple
5233 of our constant, do the operation and verify it doesn't overflow. */
5236 {
5239 /* We allow the constant to overflow with wrapping semantics. */
5243 }
5244 else
5247 /* If we have an unsigned type is not a sizetype, we cannot widen
5248 the operation since it will change the result if the original
5249 computation overflowed. */
5255 /* If we were able to eliminate our operation from the first side,
5256 apply our operation to the second side and reform the PLUS. */
5260 /* The last case is if we are a multiply. In that case, we can
5261 apply the distributive law to commute the multiply and addition
5262 if the multiplication of the constants doesn't overflow. */
5268 op1));
5273 /* We have a special case here if we are doing something like
5274 (C * 8) % 4 since we know that's zero. */
5281 /* ... fall through ... */
5285 /* If we can extract our operation from the LHS, do so and return a
5286 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
5287 do something only if the second operand is a constant. */
5299 /* If these are the same operation types, we can associate them
5300 assuming no overflow. */
5307 /* If these operations "cancel" each other, we have the main
5308 optimizations of this pass, which occur when either constant is a
5309 multiple of the other, in which case we replace this with either an
5310 operation or CODE or TCODE.
5312 If we have an unsigned type that is not a sizetype, we cannot do
5313 this since it will change the result if the original computation
5314 overflowed. */
5317 && ! flag_wrapv
5322 {
5333 }
5338 }
5341 }
5342 \f
5343 /* Return a node which has the indicated constant VALUE (either 0 or
5344 1), and is of the indicated TYPE. */
5346 tree
5348 {
5356 else
5358 }
5360 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
5361 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
5362 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
5363 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
5364 COND is the first argument to CODE; otherwise (as in the example
5365 given here), it is the second argument. TYPE is the type of the
5366 original expression. Return NULL_TREE if no simplification is
5367 possible. */
5369 static tree
5372 {
5377 /* This transformation is only worthwhile if we don't have to wrap
5378 arg in a SAVE_EXPR, and the operation can be simplified on atleast
5379 one of the branches once its pushed inside the COND_EXPR. */
5384 {
5388 /* If this operand throws an expression, then it does not make
5389 sense to try to perform a logical or arithmetic operation
5390 involving it. */
5395 }
5396 else
5397 {
5402 }
5413 }
5415 \f
5416 /* Subroutine of fold() that checks for the addition of +/- 0.0.
5418 If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type
5419 TYPE, X + ADDEND is the same as X. If NEGATE, return true if X -
5420 ADDEND is the same as X.
5422 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
5423 and finite. The problematic cases are when X is zero, and its mode
5424 has signed zeros. In the case of rounding towards -infinity,
5425 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
5426 modes, X + 0 is not the same as X because -0 + 0 is 0. */
5428 static bool
5430 {
5434 /* Don't allow the fold with -fsignaling-nans. */
5438 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
5442 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
5447 /* The mode has signed zeros, and we have to honor their sign.
5448 In this situation, there is only one case we can return true for.
5449 X - 0 is the same as X unless rounding towards -infinity is
5450 supported. */
5452 }
5454 /* Subroutine of fold() that checks comparisons of built-in math
5455 functions against real constants.
5457 FCODE is the DECL_FUNCTION_CODE of the built-in, CODE is the comparison
5458 operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE
5459 is the type of the result and ARG0 and ARG1 are the operands of the
5460 comparison. ARG1 must be a TREE_REAL_CST.
5462 The function returns the constant folded tree if a simplification
5463 can be made, and NULL_TREE otherwise. */
5465 static tree
5468 {
5469 REAL_VALUE_TYPE c;
5472 {
5478 {
5479 /* sqrt(x) < y is always false, if y is negative. */
5483 /* sqrt(x) > y is always true, if y is negative and we
5484 don't care about NaNs, i.e. negative values of x. */
5488 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
5491 }
5493 {
5494 REAL_VALUE_TYPE c2;
5500 {
5501 /* sqrt(x) > y is x == +Inf, when y is very large. */
5506 /* sqrt(x) > y is always false, when y is very large
5507 and we don't care about infinities. */
5509 }
5511 /* sqrt(x) > c is the same as x > c*c. */
5514 }
5516 {
5517 REAL_VALUE_TYPE c2;
5523 {
5524 /* sqrt(x) < y is always true, when y is a very large
5525 value and we don't care about NaNs or Infinities. */
5529 /* sqrt(x) < y is x != +Inf when y is very large and we
5530 don't care about NaNs. */
5535 /* sqrt(x) < y is x >= 0 when y is very large and we
5536 don't care about Infinities. */
5541 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
5550 dconst0))),
5553 c2)))));
5554 }
5556 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
5561 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
5564 {
5569 dconst0))),
5572 c2)))));
5573 }
5574 }
5575 }
5578 }
5580 /* Subroutine of fold() that optimizes comparisons against Infinities,
5581 either +Inf or -Inf.
5583 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5584 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5585 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5587 The function returns the constant folded tree if a simplification
5588 can be made, and NULL_TREE otherwise. */
5590 static tree
5592 {
5594 REAL_VALUE_TYPE max;
5595 tree temp;
5600 /* For negative infinity swap the sense of the comparison. */
5606 {
5608 /* x > +Inf is always false, if with ignore sNANs. */
5614 /* x <= +Inf is always true, if we don't case about NaNs. */
5618 /* x <= +Inf is the same as x == x, i.e. isfinite(x). */
5621 {
5624 }
5629 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
5635 /* x < +Inf is always equal to x <= DBL_MAX. */
5641 /* x != +Inf is always equal to !(x > DBL_MAX). */
5647 /* The transformation below creates non-gimple code and thus is
5648 not appropriate if we are in gimple form. */
5658 }
5661 }
5663 /* Subroutine of fold() that optimizes comparisons of a division by
5664 a nonzero integer constant against an integer constant, i.e.
5665 X/C1 op C2.
5667 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5668 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5669 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5671 The function returns the constant folded tree if a simplification
5672 can be made, and NULL_TREE otherwise. */
5674 static tree
5676 {
5681 HOST_WIDE_INT hpart;
5684 /* We have to do this the hard way to detect unsigned overflow.
5685 prod = int_const_binop (MULT_EXPR, arg01, arg1, 0); */
5694 {
5698 /* Likewise hi = int_const_binop (PLUS_EXPR, prod, tmp, 0). */
5707 }
5709 {
5712 {
5730 }
5731 }
5732 else
5733 {
5734 /* A negative divisor reverses the relational operators. */
5739 {
5757 }
5758 }
5761 {
5802 }
5805 }
5808 /* If CODE with arguments ARG0 and ARG1 represents a single bit
5809 equality/inequality test, then return a simplified form of
5810 the test using shifts and logical operations. Otherwise return
5811 NULL. TYPE is the desired result type. */
5813 tree
5815 tree result_type)
5816 {
5817 /* If this is a TRUTH_NOT_EXPR, it may have a single bit test inside
5818 operand 0. */
5820 {
5825 /* Extract the arguments of the EQ/NE. */
5829 /* This requires us to invert the code. */
5831 }
5833 /* If this is testing a single bit, we can optimize the test. */
5837 {
5844 tree arg00;
5846 /* If we have (A & C) != 0 where C is the sign bit of A, convert
5847 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
5850 /* This is only a win if casting to a signed type is cheap,
5851 i.e. when arg00's type is not a partial mode. */
5854 {
5859 }
5861 /* Otherwise we have (A & C) != 0 where C is a single bit,
5862 convert that into ((A >> C2) & 1). Where C2 = log2(C).
5863 Similarly for (A & C) == 0. */
5865 /* If INNER is a right shift of a constant and it plus BITNUM does
5866 not overflow, adjust BITNUM and INNER. */
5873 {
5876 }
5878 /* If we are going to be able to omit the AND below, we must do our
5879 operations as unsigned. If we must use the AND, we have a choice.
5880 Normally unsigned is faster, but for some machines signed is. */
5881 #ifdef LOAD_EXTEND_OP
5883 #else
5885 #endif
5900 /* Put the AND last so it can combine with more things. */
5904 /* Make sure to return the proper type. */
5908 }
5910 }
5912 /* Check whether we are allowed to reorder operands arg0 and arg1,
5913 such that the evaluation of arg1 occurs before arg0. */
5915 static bool
5917 {
5924 }
5926 /* Test whether it is preferable two swap two operands, ARG0 and
5927 ARG1, for example because ARG0 is an integer constant and ARG1
5928 isn't. If REORDER is true, only recommend swapping if we can
5929 evaluate the operands in reverse order. */
5931 bool
5933 {
5969 /* It is preferable to swap two SSA_NAME to ensure a canonical form
5970 for commutative and comparison operators. Ensuring a canonical
5971 form allows the optimizers to find additional redundancies without
5972 having to explicitly check for both orderings. */
5979 }
5981 /* Fold comparison ARG0 CODE ARG1 (with result in TYPE), where
5982 ARG0 is extended to a wider type. */
5984 static tree
5986 {
5988 tree arg1_unw;
6001 /* If possible, express the comparison in the shorter mode. */
6013 /* If we are comparing with the integer that does not fit into the range
6014 of the shorter type, the result is known. */
6025 {
6052 }
6055 }
6057 /* Fold comparison ARG0 CODE ARG1 (with result in TYPE), where for
6058 ARG0 just the signedness is changed. */
6060 static tree
6063 {
6088 {
6095 }
6096 else
6100 }
6102 /* Tries to replace &a[idx] CODE s * delta with &a[idx CODE delta], if s is
6103 step of the array. TYPE is the type of the expression. ADDR is the address.
6104 MULT is the multiplicative expression. If the function succeeds, the new
6105 address expression is returned. Otherwise NULL_TREE is returned. */
6107 static tree
6109 {
6114 tree itype;
6120 {
6123 }
6125 {
6128 }
6129 else
6133 {
6135 {
6143 /* If the type sizes do not match, we might run into problems
6144 when one of them would overflow. */
6153 }
6157 }
6159 /* We found the suitable array reference. So copy everything up to it,
6160 and replace the index. */
6167 {
6171 }
6175 delta));
6178 }
6181 /* Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y
6182 means A >= Y && A != MAX, but in this case we know that
6183 A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X. */
6185 static tree
6187 {
6194 else
6203 {
6206 }
6208 {
6211 }
6212 else
6223 }
6225 /* Perform constant folding and related simplification of EXPR.
6226 The related simplifications include x*1 => x, x*0 => 0, etc.,
6227 and application of the associative law.
6228 NOP_EXPR conversions may be removed freely (as long as we
6229 are careful not to change the type of the overall expression).
6230 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
6231 but we can constant-fold them if they have constant operands. */
6233 #ifdef ENABLE_FOLD_CHECKING
6234 # define fold(x) fold_1 (x)
6236 static
6237 #endif
6238 tree
6240 {
6244 tree tem;
6249 /* WINS will be nonzero when the switch is done
6250 if all operands are constant. */
6253 /* Return right away if a constant. */
6258 {
6259 tree subop;
6261 /* Special case for conversion ops that can have fixed point args. */
6264 /* Don't use STRIP_NOPS, because signedness of argument type matters. */
6270 else
6275 /* Note that TREE_CONSTANT isn't enough:
6276 static var addresses are constant but we can't
6277 do arithmetic on them. */
6279 }
6281 {
6285 {
6287 tree subop;
6292 /* Strip any conversions that don't change the mode. This is
6293 safe for every expression, except for a comparison expression
6294 because its signedness is derived from its operands. So, in
6295 the latter case, only strip conversions that don't change the
6296 signedness.
6298 Note that this is done as an internal manipulation within the
6299 constant folder, in order to find the simplest representation
6300 of the arguments so that their form can be studied. In any
6301 cases, the appropriate type conversions should be put back in
6302 the tree that will get out of the constant folder. */
6305 else
6310 else
6315 /* Note that TREE_CONSTANT isn't enough:
6316 static var addresses are constant but we can't
6317 do arithmetic on them. */
6324 }
6325 }
6327 /* If this is a commutative operation, and ARG0 is a constant, move it
6328 to ARG1 to reduce the number of tests below. */
6334 /* Now WINS is set as described above,
6335 ARG0 is the first operand of EXPR,
6336 and ARG1 is the second operand (if it has more than one operand).
6338 First check for cases where an arithmetic operation is applied to a
6339 compound, conditional, or comparison operation. Push the arithmetic
6340 operation inside the compound or conditional to see if any folding
6341 can then be done. Convert comparison to conditional for this purpose.
6342 The also optimizes non-constant cases that used to be done in
6343 expand_expr.
6345 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
6346 one of the operands is a comparison and the other is a comparison, a
6347 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
6348 code below would make the expression more complex. Change it to a
6349 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
6350 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
6362 {
6373 }
6376 {
6381 {
6391 /* If this was a conversion, and all we did was to move into
6392 inside the COND_EXPR, bring it back out. But leave it if
6393 it is a conversion from integer to integer and the
6394 result precision is no wider than a word since such a
6395 conversion is cheap and may be optimized away by combine,
6396 while it couldn't if it were outside the COND_EXPR. Then return
6397 so we don't get into an infinite recursion loop taking the
6398 conversion out and then back in. */
6410 && (INTEGRAL_TYPE_P
6421 }
6423 {
6425 {
6429 }
6433 integer_one_node)),
6435 integer_zero_node))));
6436 }
6437 }
6448 {
6452 arg1)));
6460 {
6465 }
6468 {
6473 }
6474 }
6477 {
6491 /* Handle cases of two conversions in a row. */
6494 {
6513 /* In addition to the cases of two conversions in a row
6514 handled below, if we are converting something to its own
6515 type via an object of identical or wider precision, neither
6516 conversion is needed. */
6523 /* Likewise, if the intermediate and final types are either both
6524 float or both integer, we don't need the middle conversion if
6525 it is wider than the final type and doesn't change the signedness
6526 (for integers). Avoid this if the final type is a pointer
6527 since then we sometimes need the inner conversion. Likewise if
6528 the outer has a precision not equal to the size of its mode. */
6539 /* If we have a sign-extension of a zero-extended value, we can
6540 replace that by a single zero-extension. */
6547 /* Two conversions in a row are not needed unless:
6548 - some conversion is floating-point (overstrict for now), or
6549 - the intermediate type is narrower than both initial and
6550 final, or
6551 - the intermediate type and innermost type differ in signedness,
6552 and the outermost type is wider than the intermediate, or
6553 - the initial type is a pointer type and the precisions of the
6554 intermediate and final types differ, or
6555 - the final type is a pointer type and the precisions of the
6556 initial and intermediate types differ. */
6571 }
6575 /* Detect assigning a bitfield. */
6578 {
6579 /* Don't leave an assignment inside a conversion
6580 unless assigning a bitfield. */
6584 /* First do the assignment, then return converted constant. */
6589 }
6591 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
6592 constants (if x has signed type, the sign bit cannot be set
6593 in c). This folds extension into the BIT_AND_EXPR. */
6598 {
6608 <= HOST_BITS_PER_WIDE_INT
6610 {
6617 #ifdef LOAD_EXTEND_OP
6621 {
6625 }
6626 #endif
6627 }
6632 }
6634 /* Convert (T1)((T2)X op Y) into (T1)X op Y, for pointer types T1 and
6635 T2 being pointers to types of the same size. */
6640 {
6652 }
6666 {
6670 }
6675 {
6680 }
6693 /* Convert fabs((double)float) into (double)fabsf(float). */
6696 {
6701 targ0)));
6702 }
6735 /* A + (-B) -> A - B */
6738 /* (-A) + B -> B - A */
6743 {
6747 /* If we are adding two BIT_AND_EXPR's, both of which are and'ing
6748 with a constant, and the two constants have no bits in common,
6749 we should treat this as a BIT_IOR_EXPR since this may produce more
6750 simplifications. */
6758 {
6761 }
6763 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
6764 (plus (plus (mult) (mult)) (foo)) so that we can
6765 take advantage of the factoring cases below. */
6772 {
6778 else
6800 parg1)))));
6801 }
6804 {
6808 /* (A * C) + (B * C) -> (A+B) * C.
6809 We are most concerned about the case where C is a constant,
6810 but other combinations show up during loop reduction. Since
6811 it is not difficult, try all four possibilities. */
6828 /* No identical multiplicands; see if we can find a common
6829 power-of-two factor in non-power-of-two multiplies. This
6830 can help in multi-dimensional array access. */
6835 {
6840 /* Move min of absolute values to int11. */
6843 {
6847 }
6850 {
6856 }
6857 }
6864 same));
6865 }
6867 /* Try replacing &a[i1] + c * i2 with &a[i1 + i2], if c is step
6868 of the array. Loop optimizer sometimes produce this type of
6869 expressions. */
6872 {
6876 }
6879 {
6883 }
6884 }
6885 else
6886 {
6887 /* See if ARG1 is zero and X + ARG1 reduces to X. */
6891 /* Likewise if the operands are reversed. */
6895 /* Convert X + -C into X - C. */
6898 {
6904 }
6906 /* Convert x+x into x*2.0. */
6912 /* Convert x*c+x into x*(c+1). */
6918 {
6919 REAL_VALUE_TYPE c;
6925 }
6927 /* Convert x+x*c into x*(c+1). */
6933 {
6934 REAL_VALUE_TYPE c;
6940 }
6942 /* Convert x*c1+x*c2 into x*(c1+c2). */
6952 {
6961 }
6962 /* Convert a + (b*c + d*e) into (a + b*c) + d*e. */
6966 {
6971 {
6972 tree tree0;
6975 }
6976 }
6977 /* Convert (b*c + d*e) + a into b*c + (d*e +a). */
6981 {
6986 {
6987 tree tree0;
6990 }
6991 }
6992 }
6994 bit_rotate:
6995 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
6996 is a rotate of A by C1 bits. */
6997 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
6998 is a rotate of A by B bits. */
6999 {
7008 {
7027 {
7035 TYPE_PRECISION
7040 ? LROTATE_EXPR
7043 }
7045 {
7053 TYPE_PRECISION
7058 ? LROTATE_EXPR
7061 }
7062 }
7063 }
7065 associate:
7066 /* In most languages, can't associate operations on floats through
7067 parentheses. Rather than remember where the parentheses were, we
7068 don't associate floats at all, unless the user has specified
7069 -funsafe-math-optimizations. */
7073 {
7077 /* Split both trees into variables, constants, and literals. Then
7078 associate each group together, the constants with literals,
7079 then the result with variables. This increases the chances of
7080 literals being recombined later and of generating relocatable
7081 expressions for the sum of a constant and literal. */
7086 /* Only do something if we found more than two objects. Otherwise,
7087 nothing has changed and we risk infinite recursion. */
7092 {
7093 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
7102 /* Preserve the MINUS_EXPR if the negative part of the literal is
7103 greater than the positive part. Otherwise, the multiplicative
7104 folding code (i.e extract_muldiv) may be fooled in case
7105 unsigned constants are subtracted, like in the following
7106 example: ((X*2 + 4) - 8U)/2. */
7108 {
7112 {
7116 }
7117 else
7118 {
7122 }
7123 }
7125 {
7130 else
7131 {
7137 }
7138 }
7143 }
7144 }
7146 binary:
7150 {
7151 /* The return value should always have
7152 the same type as the original expression. */
7157 }
7161 /* A - (-B) -> A + B */
7164 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
7174 {
7180 /* Fold A - (A & B) into ~B & A. */
7183 {
7188 arg0));
7193 arg0));
7194 }
7196 /* Fold (A & ~B) - (A & B) into (A ^ B) - B, where B is
7197 any power of 2 minus 1. */
7202 {
7208 {
7212 }
7213 }
7214 }
7216 /* See if ARG1 is zero and X - ARG1 reduces to X. */
7220 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
7221 ARG0 is zero and X + ARG0 reduces to X, since that would mean
7222 (-ARG1 + ARG0) reduces to -ARG1. */
7226 /* Fold &x - &x. This can happen from &x.foo - &x.
7227 This is unsafe for certain floats even in non-IEEE formats.
7228 In IEEE, it is unsafe because it does wrong for NaNs.
7229 Also note that operand_equal_p is always false if an operand
7230 is volatile. */
7236 /* A - B -> A + (-B) if B is easily negatable. */
7239 /* Avoid this transformation if B is a positive REAL_CST. */
7245 /* Try folding difference of addresses. */
7246 {
7247 HOST_WIDE_INT diff;
7253 }
7255 /* Try replacing &a[i1] - c * i2 with &a[i1 - i2], if c is step
7256 of the array. Loop optimizer sometimes produce this type of
7257 expressions. */
7260 {
7264 }
7269 {
7270 /* (A * C) - (B * C) -> (A-B) * C. */
7278 /* (A * C1) - (A * C2) -> A * (C1-C2). */
7286 }
7291 /* (-A) * (-B) -> A * B */
7302 {
7308 /* (a * (1 << b)) is (a << b) */
7324 }
7325 else
7326 {
7327 /* Maybe fold x * 0 to 0. The expressions aren't the same
7328 when x is NaN, since x * 0 is also NaN. Nor are they the
7329 same in modes with signed zeros, since multiplying a
7330 negative value by 0 gives -0, not +0. */
7335 /* In IEEE floating point, x*1 is not equivalent to x for snans. */
7340 /* Transform x * -1.0 into -x. */
7345 /* Convert (C1/X)*C2 into (C1*C2)/X. */
7350 {
7356 }
7359 {
7363 /* Optimizations of root(...)*root(...). */
7365 {
7370 /* Optimize sqrt(x)*sqrt(x) as x. */
7376 /* Optimize root(x)*root(y) as root(x*y). */
7381 }
7383 /* Optimize expN(x)*expN(y) as expN(x+y). */
7385 {
7392 }
7394 /* Optimizations of pow(...)*pow(...). */
7398 {
7406 /* Optimize pow(x,y)*pow(z,y) as pow(x*z,y). */
7408 {
7413 arg01));
7415 }
7417 /* Optimize pow(x,y)*pow(x,z) as pow(x,y+z). */
7419 {
7424 arg));
7426 }
7427 }
7429 /* Optimize tan(x)*cos(x) as sin(x). */
7438 {
7444 }
7446 /* Optimize x*pow(x,c) as pow(x,c+1). */
7450 {
7457 {
7459 REAL_VALUE_TYPE c;
7468 }
7469 }
7471 /* Optimize pow(x,c)*x as pow(x,c+1). */
7475 {
7482 {
7484 REAL_VALUE_TYPE c;
7493 }
7494 }
7496 /* Optimize x*x as pow(x,2.0), which is expanded as x*x. */
7499 {
7503 {
7508 }
7509 }
7510 }
7511 }
7515 bit_ior:
7523 /* ~X | X is -1. */
7526 {
7530 }
7532 /* X | ~X is -1. */
7535 {
7539 }
7545 /* Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
7547 This results in more efficient code for machines without a NAND
7548 instruction. Combine will canonicalize to the first form
7549 which will allow use of NAND instructions provided by the
7550 backend if they exist. */
7553 {
7558 }
7560 /* See if this can be simplified into a rotate first. If that
7561 is unsuccessful continue in the association code. */
7572 /* ~X ^ X is -1. */
7575 {
7579 }
7581 /* X ^ ~X is -1. */
7584 {
7588 }
7590 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
7591 with a constant, and the two constants have no bits in common,
7592 we should treat this as a BIT_IOR_EXPR since this may produce more
7593 simplifications. */
7601 {
7604 }
7606 /* See if this can be simplified into a rotate first. If that
7607 is unsuccessful continue in the association code. */
7618 /* ~X & X is always zero. */
7623 /* X & ~X is always zero. */
7631 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
7634 {
7635 unsigned int prec
7642 }
7644 /* Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
7646 This results in more efficient code for machines without a NOR
7647 instruction. Combine will canonicalize to the first form
7648 which will allow use of NOR instructions provided by the
7649 backend if they exist. */
7652 {
7657 }
7662 /* Don't touch a floating-point divide by zero unless the mode
7663 of the constant can represent infinity. */
7669 /* (-A) / (-B) -> A / B */
7679 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
7684 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
7689 /* If ARG1 is a constant, we can convert this to a multiply by the
7690 reciprocal. This does not have the same rounding properties,
7691 so only do this if -funsafe-math-optimizations. We can actually
7692 always safely do it if ARG1 is a power of two, but it's hard to
7693 tell if it is or not in a portable manner. */
7695 {
7700 /* Find the reciprocal if optimizing and the result is exact. */
7702 {
7703 REAL_VALUE_TYPE r;
7706 {
7709 }
7710 }
7711 }
7712 /* Convert A/B/C to A/(B*C). */
7719 /* Convert A/(B/C) to (A/B)*C. */
7727 /* Convert C1/(X*C2) into (C1/C2)/X. */
7732 {
7738 }
7741 {
7743 /* Optimize x/expN(y) into x*expN(-y). */
7745 {
7752 }
7754 /* Optimize x/pow(y,z) into x*pow(y,-z). */
7758 {
7767 }
7768 }
7771 {
7775 /* Optimize sin(x)/cos(x) as tan(x). */
7781 {
7787 }
7789 /* Optimize cos(x)/sin(x) as 1.0/tan(x). */
7795 {
7799 {
7804 }
7805 }
7807 /* Optimize pow(x,c)/x as pow(x,c-1). */
7811 {
7817 {
7819 REAL_VALUE_TYPE c;
7828 }
7829 }
7830 }
7842 /* X / -1 is -X. */
7849 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
7850 operation, EXACT_DIV_EXPR.
7852 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
7853 At one time others generated faster code, it's not clear if they do
7854 after the last round to changes to the DIV code in expmed.c. */
7875 /* X % -1 is zero. */
7882 /* Optimize unsigned TRUNC_MOD_EXPR by a power of two into a
7883 BIT_AND_EXPR, i.e. "X % C" into "X & C2". */
7887 {
7889 tree mask;
7894 {
7898 }
7899 else
7900 {
7903 }
7908 }
7910 /* X % -C is the same as X % C. */
7915 && !flag_trapv
7916 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
7921 /* X % -Y is the same as X % Y. */
7943 /* Optimize -1 >> x for arithmetic right shifts. */
7946 /* ... fall through ... */
7949 shift:
7955 /* Since negative shift count is not well-defined,
7956 don't try to compute it in the compiler. */
7959 /* Rewrite an LROTATE_EXPR by a constant into an
7960 RROTATE_EXPR by a new constant. */
7962 {
7968 }
7970 /* If we have a rotate of a bit operation with the rotate count and
7971 the second operand of the bit operation both constant,
7972 permute the two operations. */
7984 /* Two consecutive rotates adding up to the width of the mode can
7985 be ignored. */
8016 /* The argument to invert_truthvalue must have Boolean type. */
8020 /* Note that the operand of this must be an int
8021 and its values must be 0 or 1.
8022 ("true" is a fixed value perhaps depending on the language,
8023 but we don't handle values other than 1 correctly yet.) */
8025 /* Avoid infinite recursion. */
8027 {
8032 }
8036 /* Note that the operands of this must be ints
8037 and their values must be 0 or 1.
8038 ("true" is a fixed value perhaps depending on the language.) */
8039 /* If first arg is constant zero, return it. */
8043 /* If either arg is constant true, drop it. */
8047 /* Preserve sequence points. */
8050 /* If second arg is constant zero, result is zero, but first arg
8051 must be evaluated. */
8054 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
8055 case will be handled here. */
8059 /* !X && X is always false. */
8063 /* X && !X is always false. */
8068 /* A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
8069 means A >= Y && A != MAX, but in this case we know that
8070 A < X <= MAX. */
8074 {
8082 }
8084 truth_andor:
8085 /* We only do these simplifications if we are optimizing. */
8089 /* Check for things like (A || B) && (A || C). We can convert this
8090 to A || (B && C). Note that either operator can be any of the four
8091 truth and/or operations and the transformation will still be
8092 valid. Also note that we only care about order for the
8093 ANDIF and ORIF operators. If B contains side effects, this
8094 might change the truth-value of A. */
8101 {
8121 /* This case if tricky because we must either have commutative
8122 operators or else A10 must not have side-effects. */
8128 a01));
8129 }
8131 /* See if we can build a range comparison. */
8135 /* Check for the possibility of merging component references. If our
8136 lhs is another similar operation, try to merge its rhs with our
8137 rhs. Then try to merge our lhs and rhs. */
8149 /* Note that the operands of this must be ints
8150 and their values must be 0 or true.
8151 ("true" is a fixed value perhaps depending on the language.) */
8152 /* If first arg is constant true, return it. */
8156 /* If either arg is constant zero, drop it. */
8160 /* Preserve sequence points. */
8163 /* If second arg is constant true, result is true, but we must
8164 evaluate first arg. */
8167 /* Likewise for first arg, but note this only occurs here for
8168 TRUTH_OR_EXPR. */
8172 /* !X || X is always true. */
8176 /* X || !X is always true. */
8184 /* If the second arg is constant zero, drop it. */
8187 /* If the second arg is constant true, this is a logical inversion. */
8190 /* Identical arguments cancel to zero. */
8194 /* !X ^ X is always true. */
8199 /* X ^ !X is always true. */
8212 /* If one arg is a real or integer constant, put it last. */
8216 /* If this is an equality comparison of the address of a non-weak
8217 object against zero, then we know the result. */
8225 /* If this is an equality comparison of the address of two non-weak,
8226 unaliased symbols neither of which are extern (since we do not
8227 have access to attributes for externs), then we know the result. */
8243 type);
8246 {
8254 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
8259 /* (-a) CMP (-b) -> b CMP a */
8266 {
8267 REAL_VALUE_TYPE cst;
8270 /* (-a) CMP CST -> a swap(CMP) (-CST) */
8272 return
8278 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
8279 /* a CMP (-0) -> a CMP 0 */
8284 /* x != NaN is always true, other ops are always false. */
8287 {
8290 }
8292 /* Fold comparisons against infinity. */
8294 {
8298 }
8299 }
8301 /* If this is a comparison of a real constant with a PLUS_EXPR
8302 or a MINUS_EXPR of a real constant, we can convert it into a
8303 comparison with a revised real constant as long as no overflow
8304 occurs when unsafe_math_optimizations are enabled. */
8316 /* Likewise, we can simplify a comparison of a real constant with
8317 a MINUS_EXPR whose first operand is also a real constant, i.e.
8318 (c1 - x) < c2 becomes x > c1-c2. */
8329 /* Fold comparisons against built-in math functions. */
8331 && flag_unsafe_math_optimizations
8333 {
8337 {
8341 }
8342 }
8343 }
8345 /* Convert foo++ == CONST into ++foo == CONST + INCR. */
8349 /* This optimization is invalid for ordered comparisons
8350 if CONST+INCR overflows or if foo+incr might overflow.
8351 This optimization is invalid for floating point due to rounding.
8352 For pointer types we assume overflow doesn't happen. */
8356 {
8360 {
8366 }
8367 else
8368 {
8374 }
8377 /* If VAROP is a reference to a bitfield, we must mask
8378 the constant by the width of the field. */
8383 {
8388 /* First check whether the comparison would come out
8389 always the same. If we don't do that we would
8390 change the meaning with the masking. */
8404 }
8407 }
8409 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
8410 This transformation affects the cases which are handled in later
8411 optimizations involving comparisons with non-negative constants. */
8415 {
8417 {
8428 }
8429 }
8431 /* Comparisons with the highest or lowest possible integer of
8432 the specified size will have known values.
8434 This is quite similar to fold_relational_hi_lo; however, my
8435 attempts to share the code have been nothing but trouble.
8436 I give up for now. */
8437 {
8445 {
8452 {
8455 }
8456 else
8457 {
8460 }
8465 {
8478 /* The GE_EXPR and LT_EXPR cases above are not normally
8479 reached because of previous transformations. */
8483 }
8487 {
8496 }
8500 {
8515 }
8519 {
8528 }
8534 /* signed_type does not work on pointer types. */
8536 {
8537 /* The following case also applies to X < signed_max+1
8538 and X >= signed_max+1 because previous transformations. */
8540 {
8544 return fold
8548 }
8549 }
8550 }
8551 }
8553 /* If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
8554 a MINUS_EXPR of a constant, we can convert it into a comparison with
8555 a revised constant as long as no overflow occurs. */
8567 /* Similarly for a NEGATE_EXPR. */
8576 /* If we have X - Y == 0, we can convert that to X == Y and similarly
8577 for !=. Don't do this for ordered comparisons due to overflow. */
8585 {
8586 /* If we are widening one operand of an integer comparison,
8587 see if the other operand is similarly being widened. Perhaps we
8588 can do the comparison in the narrower type. */
8593 /* Or if we are changing signedness. */
8597 }
8599 /* If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a
8600 constant, we can simplify it. */
8607 /* If we are comparing an ABS_EXPR with a constant, we can
8608 convert all the cases into explicit comparisons, but they may
8609 well not be faster than doing the ABS and one comparison.
8610 But ABS (X) <= C is a range comparison, which becomes a subtraction
8611 and a comparison, and is probably faster. */
8624 /* If this is an EQ or NE comparison with zero and ARG0 is
8625 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
8626 two operations, but the latter can be done in one less insn
8627 on machines that have only two-operand insns or on which a
8628 constant cannot be the first operand. */
8631 {
8636 return
8642 integer_one_node)),
8643 arg1));
8646 return
8652 integer_one_node)),
8653 arg1));
8654 }
8656 /* If this is an NE or EQ comparison of zero against the result of a
8657 signed MOD operation whose second operand is a power of 2, make
8658 the MOD operation unsigned since it is simpler and equivalent. */
8667 {
8677 }
8679 /* If this is an NE comparison of zero with an AND of one, remove the
8680 comparison since the AND will give the correct value. */
8686 /* If we have (A & C) == C where C is a power of 2, convert this into
8687 (A & C) != 0. Similarly for NE_EXPR. */
8694 integer_zero_node)));
8696 /* If we have (A & C) != 0 or (A & C) == 0 and C is a power of
8697 2, then fold the expression into shifts and logical operations. */
8702 /* If we have (A & C) == D where D & ~C != 0, convert this into 0.
8703 Similarly for NE_EXPR. */
8708 {
8717 }
8719 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
8720 Similarly for NE_EXPR. */
8725 {
8732 }
8734 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
8735 and similarly for >= into !=. */
8751 return
8759 /* Simplify comparison of something with itself. (For IEEE
8760 floating-point, we can only do some of these simplifications.) */
8762 {
8764 {
8779 /* For NE, we can only do this simplification if integer
8780 or we don't honor IEEE floating point NaNs. */
8784 /* ... fall through ... */
8790 }
8791 }
8793 /* If we are comparing an expression that just has comparisons
8794 of two integer values, arithmetic expressions of those comparisons,
8795 and constants, we can simplify it. There are only three cases
8796 to check: the two values can either be equal, the first can be
8797 greater, or the second can be greater. Fold the expression for
8798 those three values. Since each value must be 0 or 1, we have
8799 eight possibilities, each of which corresponds to the constant 0
8800 or 1 or one of the six possible comparisons.
8802 This handles common cases like (a > b) == 0 but also handles
8803 expressions like ((x > y) - (y > x)) > 0, which supposedly
8804 occur in macroized code. */
8807 {
8812 /* Don't handle degenerate cases here; they should already
8813 have been handled anyway. */
8822 {
8826 /* We can't just pass T to eval_subst in case cval1 or cval2
8827 was the same as ARG1. */
8829 tree high_result
8833 arg1));
8834 tree equal_result
8838 arg1));
8839 tree low_result
8843 arg1));
8845 /* All three of these results should be 0 or 1. Confirm they
8846 are. Then use those values to select the proper code
8847 to use. */
8855 {
8856 /* Make a 3-bit mask with the high-order bit being the
8857 value for `>', the next for '=', and the low for '<'. */
8861 {
8863 /* Always false. */
8884 /* Always true. */
8886 }
8891 else
8893 }
8894 }
8895 }
8897 /* If this is a comparison of a field, we may be able to simplify it. */
8902 /* Handle the constant case even without -O
8903 to make sure the warnings are given. */
8905 {
8909 }
8911 /* If this is a comparison of complex values and either or both sides
8912 are a COMPLEX_EXPR or COMPLEX_CST, it is best to split up the
8913 comparisons and join them with a TRUTH_ANDIF_EXPR or TRUTH_ORIF_EXPR.
8914 This may prevent needless evaluations. */
8921 {
8934 type,
8937 }
8939 /* Optimize comparisons of strlen vs zero to a compare of the
8940 first character of the string vs zero. To wit,
8941 strlen(ptr) == 0 => *ptr == 0
8942 strlen(ptr) != 0 => *ptr != 0
8943 Other cases should reduce to one of these two (or a constant)
8944 due to the return value of strlen being unsigned. */
8948 {
8950 tree arglist;
8963 integer_zero_node)));
8964 }
8966 /* We can fold X/C1 op C2 where C1 and C2 are integer constants
8967 into a single range test. */
8974 {
8978 }
8998 {
9002 }
9004 /* If the first operand is NaN, the result is constant. */
9008 {
9010 ? integer_zero_node
9013 }
9015 /* If the second operand is NaN, the result is constant. */
9019 {
9021 ? integer_zero_node
9024 }
9026 /* Simplify unordered comparison of something with itself. */
9032 && !flag_trapping_math
9036 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
9037 {
9048 }
9053 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
9054 so all simple results must be passed through pedantic_non_lvalue. */
9056 {
9058 /* Only optimize constant conditions when the selected branch
9059 has the same type as the COND_EXPR. This avoids optimizing
9060 away "c ? x : throw", where the throw has a void type. */
9065 }
9069 /* If we have A op B ? A : C, we may be able to convert this to a
9070 simpler expression, depending on the operation and the values
9071 of B and C. Signed zeros prevent all of these transformations,
9072 for reasons given above each one.
9074 Also try swapping the arguments and inverting the conditional. */
9079 {
9085 }
9092 {
9095 {
9101 }
9102 }
9104 /* If the second operand is simpler than the third, swap them
9105 since that produces better jump optimization results. */
9108 {
9109 /* See if this can be inverted. If it can't, possibly because
9110 it was a floating-point inequality comparison, don't do
9111 anything. */
9117 }
9119 /* Convert A ? 1 : 0 to simply A. */
9122 /* If we try to convert TREE_OPERAND (t, 0) to our type, the
9123 call to fold will try to move the conversion inside
9124 a COND, which will recurse. In that case, the COND_EXPR
9125 is probably the best choice, so leave it alone. */
9129 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
9130 over COND_EXPR in cases such as floating point comparisons. */
9137 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
9145 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
9146 already handled above. */
9151 {
9160 }
9162 /* A & N ? N : 0 is simply A & N if N is a power of two. This
9163 is probably obsolete because the first operand should be a
9164 truth value (that's why we have the two cases above), but let's
9165 leave it in until we can confirm this for all front-ends. */
9176 /* Convert A ? B : 0 into A && B if A and B are truth values. */
9182 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
9186 {
9187 /* Only perform transformation if ARG0 is easily inverted. */
9191 }
9193 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
9197 {
9198 /* Only perform transformation if ARG0 is easily inverted. */
9203 }
9205 /* Convert A ? 1 : B into A || B if A and B are truth values. */
9215 /* When pedantic, a compound expression can be neither an lvalue
9216 nor an integer constant expression. */
9219 /* Don't let (0, 0) be null pointer constant. */
9262 /* Check for a built-in function. */
9267 {
9271 }
9277 }
9279 #ifdef ENABLE_FOLD_CHECKING
9280 #undef fold
9286 /* When --enable-checking=fold, compute a digest of expr before
9287 and after actual fold call to see if fold did not accidentally
9288 change original expr. */
9290 tree
9292 {
9293 tree ret;
9296 htab_t ht;
9315 }
9317 void
9319 {
9322 htab_t ht;
9332 }
9334 static void
9336 {
9338 }
9340 static void
9342 {
9360 {
9361 /* Allow DECL_ASSEMBLER_NAME to be modified. */
9365 }
9369 {
9370 /* Allow these fields to be modified. */
9377 }
9384 {
9387 {
9401 }
9405 {
9416 }
9450 {
9453 }
9463 }
9464 }
9466 #endif
9468 /* Perform constant folding and related simplification of initializer
9469 expression EXPR. This behaves identically to "fold" but ignores
9470 potential run-time traps and exceptions that fold must preserve. */
9472 tree
9474 {
9478 tree result;
9491 }
9493 /* Determine if first argument is a multiple of second argument. Return 0 if
9494 it is not, or we cannot easily determined it to be.
9496 An example of the sort of thing we care about (at this point; this routine
9497 could surely be made more general, and expanded to do what the *_DIV_EXPR's
9498 fold cases do now) is discovering that
9500 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9502 is a multiple of
9504 SAVE_EXPR (J * 8)
9506 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
9508 This code also handles discovering that
9510 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9512 is a multiple of 8 so we don't have to worry about dealing with a
9513 possible remainder.
9515 Note that we *look* inside a SAVE_EXPR only to determine how it was
9516 calculated; it is not safe for fold to do much of anything else with the
9517 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
9518 at run time. For example, the latter example above *cannot* be implemented
9519 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
9520 evaluation time of the original SAVE_EXPR is not necessarily the same at
9521 the time the new expression is evaluated. The only optimization of this
9522 sort that would be valid is changing
9524 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
9526 divided by 8 to
9528 SAVE_EXPR (I) * SAVE_EXPR (J)
9530 (where the same SAVE_EXPR (J) is used in the original and the
9531 transformed version). */
9533 static int
9535 {
9543 {
9555 {
9559 /* const_binop may not detect overflow correctly,
9560 so check for it explicitly here. */
9566 size_one_node,
9570 }
9574 /* Can't handle conversions from non-integral or wider integral type. */
9580 /* .. fall through ... */
9596 }
9597 }
9599 /* Return true if `t' is known to be non-negative. */
9601 int
9603 {
9605 {
9620 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
9621 both unsigned and at least 2 bits shorter than the result. */
9625 {
9630 {
9634 }
9635 }
9640 {
9641 /* x * x for floating point x is always non-negative. */
9646 }
9648 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
9649 both unsigned and their total bits is shorter than the result. */
9653 {
9660 }
9689 {
9694 {
9698 {
9702 }
9703 }
9705 {
9711 }
9712 }
9738 {
9742 /* If the initializer is non-void, then it's a normal expression
9743 that will be assigned to the slot. */
9747 /* Otherwise, the initializer sets the slot in some way. One common
9748 way is an assignment statement at the end of the initializer. */
9750 {
9758 else
9760 }
9766 }
9769 {
9776 {
9777 #define CASE_BUILTIN_F(BUILT_IN_FN) \
9778 case BUILT_IN_FN: case BUILT_IN_FN##F: case BUILT_IN_FN##L:
9779 #define CASE_BUILTIN_I(BUILT_IN_FN) \
9780 case BUILT_IN_FN: case BUILT_IN_FN##L: case BUILT_IN_FN##LL:
9798 /* Always true. */
9802 /* sqrt(-0.0) is -0.0. */
9830 /* True if the 1st argument is nonnegative. */
9834 /* True if the 1st OR 2nd arguments are nonnegative. */
9839 /* True if the 1st AND 2nd arguments are nonnegative. */
9844 /* True if the 2nd argument is nonnegative. */
9849 #undef CASE_BUILTIN_F
9850 #undef CASE_BUILTIN_I
9851 }
9852 }
9854 /* ... fall through ... */
9858 /* Truth values evaluate to 0 or 1, which is nonnegative. */
9860 }
9862 /* We don't know sign of `t', so be conservative and return false. */
9864 }
9866 /* Return true when T is an address and is known to be nonzero.
9867 For floating point we further ensure that T is not denormal.
9868 Similar logic is present in nonzero_address in rtlanal.h. */
9870 static bool
9872 {
9875 /* Doing something useful for floating point would need more work. */
9880 {
9886 /* We used to test for !integer_zerop here. This does not work correctly
9887 if TREE_CONSTANT_OVERFLOW (t). */
9893 {
9894 /* With the presence of negative values it is hard
9895 to say something. */
9899 /* One of operands must be positive and the other non-negative. */
9902 }
9907 {
9910 }
9914 {
9920 }
9924 {
9930 /* Weak declarations may link to NULL. */
9934 /* Constants are never weak. */
9939 }
9951 {
9952 /* When both operands are nonzero, then MAX must be too. */
9956 /* MAX where operand 0 is positive is positive. */
9958 }
9959 /* MAX where operand 1 is positive is positive. */
9980 }
9982 }
9984 /* See if we are applying CODE, a relational to the highest or lowest
9985 possible integer of TYPE. If so, then the result is a compile
9986 time constant. */
9988 static tree
9991 {
10002 {
10009 {
10012 }
10013 else
10014 {
10017 }
10022 {
10036 /* The GE_EXPR and LT_EXPR cases above are not normally
10037 reached because of previous transformations. */
10041 }
10045 {
10056 }
10060 {
10077 }
10081 {
10092 }
10097 /* signed_type does not work on pointer types. */
10099 {
10100 /* The following case also applies to X < signed_max+1
10101 and X >= signed_max+1 because previous transformations. */
10103 {
10109 type,
10118 /* If we are in gimple form, then returning EXP would create
10119 non-gimple expressions. Clearing it is safe and insures
10120 we do not allow a non-gimple expression to escape. */
10125 }
10126 }
10127 }
10130 }
10133 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
10134 attempt to fold the expression to a constant without modifying TYPE,
10135 OP0 or OP1.
10137 If the expression could be simplified to a constant, then return
10138 the constant. If the expression would not be simplified to a
10139 constant, then return NULL_TREE.
10141 Note this is primarily designed to be called after gimplification
10142 of the tree structures and when at least one operand is a constant.
10143 As a result of those simplifying assumptions this routine is far
10144 simpler than the generic fold routine. */
10146 tree
10148 {
10150 tree subop0;
10151 tree subop1;
10152 tree tem;
10154 /* If this is a commutative operation, and ARG0 is a constant, move it
10155 to ARG1 to reduce the number of tests below. */
10158 {
10162 }
10164 /* If either operand is a complex type, extract its real component. */
10167 else
10172 else
10175 /* Note if either argument is not a real or integer constant.
10176 With a few exceptions, simplification is limited to cases
10177 where both arguments are constants. */
10185 {
10187 /* (plus (address) (const_int)) is a constant. */
10195 {
10199 }
10202 binary:
10206 /* Both arguments are constants. Simplify. */
10209 {
10210 /* The return value should always have the same type as
10211 the original expression. */
10216 }
10220 /* Fold &x - &x. This can happen from &x.foo - &x.
10221 This is unsafe for certain floats even in non-IEEE formats.
10222 In IEEE, it is unsafe because it does wrong for NaNs.
10223 Also note that operand_equal_p is always false if an
10224 operand is volatile. */
10232 /* Special case multiplication or bitwise AND where one argument
10233 is zero. */
10236 else
10245 /* Special case when we know the result will be all ones. */
10261 /* Division by zero is undefined. */
10288 /* Optimize -1 >> x for arithmetic right shifts. */
10291 /* ... fall through ... */
10297 /* Since negative shift count is not well-defined, don't
10298 try to compute it in the compiler. */
10306 /* -1 rotated either direction by any amount is still -1. */
10310 /* 0 rotated either direction by any amount is still zero. */
10327 /* If one arg is a real or integer constant, put it last. */
10332 {
10333 tree temp;
10339 }
10341 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
10342 This transformation affects the cases which are handled in later
10343 optimizations involving comparisons with non-negative constants. */
10347 {
10349 {
10362 }
10363 }
10369 /* Fall through. */
10385 /* This could probably be handled. */
10389 /* If second arg is constant zero, result is zero, but first arg
10390 must be evaluated. */
10393 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
10394 case will be handled here. */
10402 /* If second arg is constant true, result is true, but we must
10403 evaluate first arg. */
10406 /* Likewise for first arg, but note this only occurs here for
10407 TRUTH_OR_EXPR. */
10416 {
10419 }
10424 }
10425 }
10427 /* Given the components of a unary expression CODE, TYPE and OP0,
10428 attempt to fold the expression to a constant without modifying
10429 TYPE or OP0.
10431 If the expression could be simplified to a constant, then return
10432 the constant. If the expression would not be simplified to a
10433 constant, then return NULL_TREE.
10435 Note this is primarily designed to be called after gimplification
10436 of the tree structures and when op0 is a constant. As a result
10437 of those simplifying assumptions this routine is far simpler than
10438 the generic fold routine. */
10440 tree
10442 {
10443 /* Make sure we have a suitable constant argument. */
10445 {
10446 tree subop;
10450 else
10455 }
10458 {
10470 else
10476 else
10482 else
10488 else
10494 else
10506 }
10507 }
10509 /* If EXP represents referencing an element in a constant string
10510 (either via pointer arithmetic or array indexing), return the
10511 tree representing the value accessed, otherwise return NULL. */
10513 tree
10515 {
10517 {
10519 tree index;
10520 tree string;
10524 else
10525 {
10529 /* Optimize the special-case of a zero lower bound.
10531 We convert the low_bound to sizetype to avoid some problems
10532 with constant folding. (E.g. suppose the lower bound is 1,
10533 and its mode is QI. Without the conversion,l (ARRAY
10534 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
10535 +INDEX), which becomes (ARRAY+255+INDEX). Opps!) */
10540 }
10554 }
10556 }
10558 /* Return the tree for neg (ARG0) when ARG0 is known to be either
10559 an integer constant or real constant.
10561 TYPE is the type of the result. */
10563 static tree
10565 {
10569 {
10571 {
10573 HOST_WIDE_INT high;
10583 }
10591 }
10594 }
10596 /* Return the tree for abs (ARG0) when ARG0 is known to be either
10597 an integer constant or real constant.
10599 TYPE is the type of the result. */
10601 tree
10603 {
10607 {
10609 /* If the value is unsigned, then the absolute value is
10610 the same as the ordinary value. */
10613 /* Similarly, if the value is non-negative. */
10616 /* If the value is negative, then the absolute value is
10617 its negation. */
10618 else
10619 {
10621 HOST_WIDE_INT high;
10628 }
10634 else
10640 }
10643 }
10645 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
10646 constant. TYPE is the type of the result. */
10648 static tree
10650 {
10662 }
10664 /* Given CODE, a relational operator, the target type, TYPE and two
10665 constant operands OP0 and OP1, return the result of the
10666 relational operation. If the result is not a compile time
10667 constant, then return NULL_TREE. */
10669 static tree
10671 {
10674 /* From here on, the only cases we handle are when the result is
10675 known to be a constant. */
10678 {
10682 /* Handle the cases where either operand is a NaN. */
10684 {
10686 {
10714 }
10717 }
10720 }
10722 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
10724 To compute GT, swap the arguments and do LT.
10725 To compute GE, do LT and invert the result.
10726 To compute LE, swap the arguments, do LT and invert the result.
10727 To compute NE, do EQ and invert the result.
10729 Therefore, the code below must handle only EQ and LT. */
10732 {
10737 }
10739 /* Note that it is safe to invert for real values here because we
10740 have already handled the one case that it matters. */
10744 {
10747 }
10749 /* Compute a result for LT or EQ if args permit;
10750 Otherwise return T. */
10752 {
10757 else
10759 }
10760 else
10766 }
10768 /* Build an expression for the a clean point containing EXPR with type TYPE.
10769 Don't build a cleanup point expression for EXPR which don't have side
10770 effects. */
10772 tree
10774 {
10775 /* If the expression does not have side effects then we don't have to wrap
10776 it with a cleanup point expression. */
10781 }
10783 /* Build an expression for the address of T. Folds away INDIRECT_REF to
10784 avoid confusing the gimplify process. */
10786 tree
10788 {
10789 /* The size of the object is not relevant when talking about its address. */
10793 /* Note: doesn't apply to ALIGN_INDIRECT_REF */
10796 {
10800 }
10801 else
10802 {
10813 }
10816 }
10818 tree
10820 {
10822 }
10824 /* Builds an expression for an indirection through T, simplifying some
10825 cases. */
10827 tree
10829 {
10832 tree subtype;
10836 {
10839 /* *&p => p */
10842 /* *(foo *)&fooarray => fooarray[0] */
10846 }
10848 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
10852 {
10855 }
10858 }
10860 /* Strip non-trapping, non-side-effecting tree nodes from an expression
10861 whose result is ignored. The type of the returned tree need not be
10862 the same as the original expression. */
10864 tree
10866 {
10872 {
10883 else
10889 {
10905 }
10910 }
10911 }
10913 /* Return the value of VALUE, rounded up to a multiple of DIVISOR.
10914 This can only be applied to objects of a sizetype. */
10916 tree
10918 {
10925 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10926 have to do anything. Only do this when we are not given a const,
10927 because in that case, this check is more expensive than just
10928 doing it. */
10930 {
10935 }
10937 /* If divisor is a power of two, simplify this to bit manipulation. */
10939 {
10940 tree t;
10946 }
10947 else
10948 {
10953 }
10956 }
10958 /* Likewise, but round down. */
10960 tree
10962 {
10969 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10970 have to do anything. Only do this when we are not given a const,
10971 because in that case, this check is more expensive than just
10972 doing it. */
10974 {
10979 }
10981 /* If divisor is a power of two, simplify this to bit manipulation. */
10983 {
10984 tree t;
10988 }
10989 else
10990 {
10995 }
10998 }
11000 /* Returns the pointer to the base of the object addressed by EXP and
11001 extracts the information about the offset of the access, storing it
11002 to PBITPOS and POFFSET. */
11004 static tree
11007 {
11008 tree core;
11011 HOST_WIDE_INT bitsize;
11014 {
11020 }
11021 else
11022 {
11026 }
11029 }
11031 /* Returns true if addresses of E1 and E2 differ by a constant, false
11032 otherwise. If they do, E1 - E2 is stored in *DIFF. */
11034 bool
11036 {
11050 {
11060 }
11062 {
11063 /* If only one of the offsets is non-constant, the difference cannot
11064 be a constant. */
11066 }
11067 else
11072 }