| blob | 25c84a459149aa0aa73ffb8b267f9aaf07a79fa2 |
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 {
1270 }
1273 }
1277 }
1278 \f
1279 /* Combine two integer constants ARG1 and ARG2 under operation CODE
1280 to produce a new constant.
1282 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1284 tree
1286 {
1290 HOST_WIDE_INT hi;
1292 HOST_WIDE_INT garbageh;
1293 tree t;
1296 int is_sizetype
1307 {
1323 /* It's unclear from the C standard whether shifts can overflow.
1324 The following code ignores overflow; perhaps a C standard
1325 interpretation ruling is needed. */
1355 /* This is a shortcut for a common special case. */
1360 {
1366 }
1368 /* ... fall through ... */
1372 {
1375 }
1378 {
1381 }
1388 /* This is a shortcut for a common special case. */
1393 {
1398 }
1400 /* ... fall through ... */
1416 else
1422 else
1428 }
1433 {
1434 /* Propagate overflow flags ourselves. */
1437 {
1441 }
1443 {
1446 }
1447 }
1448 else
1456 }
1458 /* Combine two constants ARG1 and ARG2 under operation CODE to produce a new
1459 constant. We assume ARG1 and ARG2 have the same data type, or at least
1460 are the same kind of constant and the same machine mode.
1462 If NOTRUNC is nonzero, do not truncate the result to fit the data type. */
1464 static tree
1466 {
1474 {
1476 REAL_VALUE_TYPE d1;
1477 REAL_VALUE_TYPE d2;
1478 REAL_VALUE_TYPE value;
1487 /* Don't perform operation if we honor signaling NaNs and
1488 either operand is a NaN. */
1493 /* Don't perform operation if it would raise a division
1494 by zero exception. */
1500 /* If either operand is a NaN, just return it. Otherwise, set up
1501 for floating-point trap; we return an overflow. */
1517 }
1519 {
1525 tree t;
1528 {
1548 notrunc),
1554 notrunc));
1558 {
1559 tree magsquared
1563 notrunc);
1566 const_binop
1571 notrunc),
1573 notrunc),
1574 notrunc),
1576 const_binop
1581 notrunc),
1583 notrunc),
1584 notrunc),
1586 }
1591 }
1593 }
1595 }
1597 /* Create a size type INT_CST node with NUMBER sign extended. KIND
1598 indicates which particular sizetype to create. */
1600 tree
1602 {
1604 }
1605 \f
1606 /* Combine operands OP1 and OP2 with arithmetic operation CODE. CODE
1607 is a tree code. The type of the result is taken from the operands.
1608 Both must be the same type integer type and it must be a size type.
1609 If the operands are constant, so is the result. */
1611 tree
1613 {
1619 /* Handle the special case of two integer constants faster. */
1621 {
1622 /* And some specific cases even faster than that. */
1631 /* Handle general case of two integer constants. */
1633 }
1639 }
1641 /* Given two values, either both of sizetype or both of bitsizetype,
1642 compute the difference between the two values. Return the value
1643 in signed type corresponding to the type of the operands. */
1645 tree
1647 {
1649 tree ctype;
1654 /* If the type is already signed, just do the simple thing. */
1660 /* If either operand is not a constant, do the conversions to the signed
1661 type and subtract. The hardware will do the right thing with any
1662 overflow in the subtraction. */
1667 /* If ARG0 is larger than ARG1, subtract and return the result in CTYPE.
1668 Otherwise, subtract the other way, convert to CTYPE (we know that can't
1669 overflow) and negate (which can't either). Special-case a result
1670 of zero while we're here. */
1675 else
1679 }
1680 \f
1681 /* Construct a vector of zero elements of vector type TYPE. */
1683 static tree
1685 {
1696 }
1699 /* Attempt to fold type conversion operation CODE of expression ARG1 to
1700 type TYPE. If no simplification can be done return NULL_TREE. */
1702 static tree
1704 {
1706 tree t;
1712 {
1714 {
1715 /* If we would build a constant wider than GCC supports,
1716 leave the conversion unfolded. */
1720 /* Given an integer constant, make new constant with new type,
1721 appropriately sign-extended or truncated. */
1726 /* Don't set the overflow when
1727 converting a pointer */
1735 }
1737 {
1738 /* The following code implements the floating point to integer
1739 conversion rules required by the Java Language Specification,
1740 that IEEE NaNs are mapped to zero and values that overflow
1741 the target precision saturate, i.e. values greater than
1742 INT_MAX are mapped to INT_MAX, and values less than INT_MIN
1743 are mapped to INT_MIN. These semantics are allowed by the
1744 C and C++ standards that simply state that the behavior of
1745 FP-to-integer conversion is unspecified upon overflow. */
1748 REAL_VALUE_TYPE r;
1752 {
1771 }
1773 /* If R is NaN, return zero and show we have an overflow. */
1775 {
1779 }
1781 /* See if R is less than the lower bound or greater than the
1782 upper bound. */
1785 {
1789 {
1793 }
1794 }
1797 {
1800 {
1803 {
1807 }
1808 }
1809 }
1819 }
1820 }
1822 {
1826 {
1828 {
1829 /* We make a copy of ARG1 so that we don't modify an
1830 existing constant tree. */
1834 }
1844 }
1845 }
1847 }
1849 /* Convert expression ARG to type TYPE. Used by the middle-end for
1850 simple conversions in preference to calling the front-end's convert. */
1852 tree
1854 {
1856 tree tem;
1872 {
1877 {
1881 }
1886 {
1889 }
1896 {
1900 }
1902 {
1906 }
1909 {
1925 }
1929 {
1938 {
1942 {
1946 }
1954 }
1958 }
1973 }
1974 }
1975 \f
1976 /* Return an expr equal to X but certainly not valid as an lvalue. */
1978 tree
1980 {
1981 /* We only need to wrap lvalue tree codes. */
1983 {
2015 /* Assume the worst for front-end tree codes. */
2019 }
2021 }
2023 /* Nonzero means lvalues are limited to those valid in pedantic ANSI C.
2024 Zero means allow extended lvalues. */
2028 /* When pedantic, return an expr equal to X but certainly not valid as a
2029 pedantic lvalue. Otherwise, return X. */
2031 tree
2033 {
2036 else
2038 }
2039 \f
2040 /* Given a tree comparison code, return the code that is the logical inverse
2041 of the given code. It is not safe to do this for floating-point
2042 comparisons, except for NE_EXPR and EQ_EXPR, so we receive a machine mode
2043 as well: if reversing the comparison is unsafe, return ERROR_MARK. */
2045 static enum tree_code
2047 {
2052 {
2083 }
2084 }
2086 /* Similar, but return the comparison that results if the operands are
2087 swapped. This is safe for floating-point. */
2089 enum tree_code
2091 {
2093 {
2107 }
2108 }
2111 /* Convert a comparison tree code from an enum tree_code representation
2112 into a compcode bit-based encoding. This function is the inverse of
2113 compcode_to_comparison. */
2115 static enum comparison_code
2117 {
2119 {
2150 }
2151 }
2153 /* Convert a compcode bit-based encoding of a comparison operator back
2154 to GCC's enum tree_code representation. This function is the
2155 inverse of comparison_to_compcode. */
2157 static enum tree_code
2159 {
2161 {
2192 }
2193 }
2195 /* Return a tree for the comparison which is the combination of
2196 doing the AND or OR (depending on CODE) of the two operations LCODE
2197 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2198 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2199 if this makes the transformation invalid. */
2201 tree
2205 {
2212 {
2223 }
2226 {
2227 /* Eliminate unordered comparisons, as well as LTGT and ORD
2228 which are not used unless the mode has NaNs. */
2234 }
2236 {
2237 /* Check that the original operation and the optimized ones will trap
2238 under the same condition. */
2249 /* In a short-circuited boolean expression the LHS might be
2250 such that the RHS, if evaluated, will never trap. For
2251 example, in ORD (x, y) && (x < y), we evaluate the RHS only
2252 if neither x nor y is NaN. (This is a mixed blessing: for
2253 example, the expression above will never trap, hence
2254 optimizing it to x < y would be invalid). */
2259 /* If the comparison was short-circuited, and only the RHS
2260 trapped, we may now generate a spurious trap. */
2265 /* If we changed the conditions that cause a trap, we lose. */
2268 }
2274 else
2277 }
2279 /* Return nonzero if CODE is a tree code that represents a truth value. */
2281 static int
2283 {
2288 }
2289 \f
2290 /* Return nonzero if two operands (typically of the same tree node)
2291 are necessarily equal. If either argument has side-effects this
2292 function returns zero. FLAGS modifies behavior as follows:
2294 If OEP_ONLY_CONST is set, only return nonzero for constants.
2295 This function tests whether the operands are indistinguishable;
2296 it does not test whether they are equal using C's == operation.
2297 The distinction is important for IEEE floating point, because
2298 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
2299 (2) two NaNs may be indistinguishable, but NaN!=NaN.
2301 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
2302 even though it may hold multiple values during a function.
2303 This is because a GCC tree node guarantees that nothing else is
2304 executed between the evaluation of its "operands" (which may often
2305 be evaluated in arbitrary order). Hence if the operands themselves
2306 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
2307 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
2308 unset means assuming isochronic (or instantaneous) tree equivalence.
2309 Unless comparing arbitrary expression trees, such as from different
2310 statements, this flag can usually be left unset.
2312 If OEP_PURE_SAME is set, then pure functions with identical arguments
2313 are considered the same. It is used when the caller has other ways
2314 to ensure that global memory is unchanged in between. */
2316 int
2318 {
2319 /* If one is specified and the other isn't, they aren't equal and if
2320 neither is specified, they are.
2322 ??? This is temporary and is meant only to handle the cases of the
2323 optional operands for COMPONENT_REF and ARRAY_REF. */
2328 /* If either is ERROR_MARK, they aren't equal. */
2332 /* If both types don't have the same signedness, then we can't consider
2333 them equal. We must check this before the STRIP_NOPS calls
2334 because they may change the signedness of the arguments. */
2342 /* This is needed for conversions and for COMPONENT_REF.
2343 Might as well play it safe and always test this. */
2349 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
2350 We don't care about side effects in that case because the SAVE_EXPR
2351 takes care of that for us. In all other cases, two expressions are
2352 equal if they have no side effects. If we have two identical
2353 expressions with side effects that should be treated the same due
2354 to the only side effects being identical SAVE_EXPR's, that will
2355 be detected in the recursive calls below. */
2361 /* Next handle constant cases, those for which we can return 1 even
2362 if ONLY_CONST is set. */
2365 {
2378 {
2388 {
2390 flags))
2394 }
2397 }
2401 flags)
2403 flags));
2416 }
2422 {
2424 /* Two conversions are equal only if signedness and modes match. */
2426 {
2439 }
2452 /* For commutative ops, allow the other order. */
2460 /* If either of the pointer (or reference) expressions we are
2461 dereferencing contain a side effect, these cannot be equal. */
2467 {
2504 }
2508 {
2534 /* If the CALL_EXPRs call different functions, then they
2535 clearly can not be equal. */
2540 {
2544 else
2548 }
2550 /* Now see if all the arguments are the same. operand_equal_p
2551 does not handle TREE_LIST, so we walk the operands here
2552 feeding them to operand_equal_p. */
2556 {
2558 flags))
2563 }
2565 /* If we get here and both argument lists are exhausted
2566 then the CALL_EXPRs are equal. */
2571 }
2574 /* Consider __builtin_sqrt equal to sqrt. */
2582 }
2583 }
2584 \f
2585 /* Similar to operand_equal_p, but see if ARG0 might have been made by
2586 shorten_compare from ARG1 when ARG1 was being compared with OTHER.
2588 When in doubt, return 0. */
2590 static int
2592 {
2604 /* Discard any conversions that don't change the modes of ARG0 and ARG1
2605 and see if the inner values are the same. This removes any
2606 signedness comparison, which doesn't matter here. */
2613 /* Duplicate what shorten_compare does to ARG1 and see if that gives the
2614 actual comparison operand, ARG0.
2616 First throw away any conversions to wider types
2617 already present in the operands. */
2626 {
2629 /* Make sure shorter operand is extended the right way
2630 to match the longer operand. */
2636 }
2639 }
2640 \f
2641 /* See if ARG is an expression that is either a comparison or is performing
2642 arithmetic on comparisons. The comparisons must only be comparing
2643 two different values, which will be stored in *CVAL1 and *CVAL2; if
2644 they are nonzero it means that some operands have already been found.
2645 No variables may be used anywhere else in the expression except in the
2646 comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around
2647 the expression and save_expr needs to be called with CVAL1 and CVAL2.
2649 If this is true, return 1. Otherwise, return zero. */
2651 static int
2653 {
2657 /* We can handle some of the tcc_expression cases here. */
2667 {
2668 /* If we've already found a CVAL1 or CVAL2, this expression is
2669 two complex to handle. */
2675 }
2678 {
2701 /* First see if we can handle the first operand, then the second. For
2702 the second operand, we know *CVAL1 can't be zero. It must be that
2703 one side of the comparison is each of the values; test for the
2704 case where this isn't true by failing if the two operands
2705 are the same. */
2714 ;
2718 ;
2719 else
2723 ;
2727 ;
2728 else
2735 }
2736 }
2737 \f
2738 /* ARG is a tree that is known to contain just arithmetic operations and
2739 comparisons. Evaluate the operations in the tree substituting NEW0 for
2740 any occurrence of OLD0 as an operand of a comparison and likewise for
2741 NEW1 and OLD1. */
2743 static tree
2745 {
2750 /* We can handle some of the tcc_expression cases here. */
2758 {
2773 {
2790 }
2791 /* Fall through - ??? */
2794 {
2798 /* We need to check both for exact equality and tree equality. The
2799 former will be true if the operand has a side-effect. In that
2800 case, we know the operand occurred exactly once. */
2813 }
2817 }
2818 }
2819 \f
2820 /* Return a tree for the case when the result of an expression is RESULT
2821 converted to TYPE and OMITTED was previously an operand of the expression
2822 but is now not needed (e.g., we folded OMITTED * 0).
2824 If OMITTED has side effects, we must evaluate it. Otherwise, just do
2825 the conversion of RESULT to TYPE. */
2827 tree
2829 {
2836 }
2838 /* Similar, but call pedantic_non_lvalue instead of non_lvalue. */
2840 static tree
2842 {
2849 }
2851 /* Return a tree for the case when the result of an expression is RESULT
2852 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
2853 of the expression but are now not needed.
2855 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
2856 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
2857 evaluated before OMITTED2. Otherwise, if neither has side effects,
2858 just do the conversion of RESULT to TYPE. */
2860 tree
2862 {
2871 }
2873 \f
2874 /* Return a simplified tree node for the truth-negation of ARG. This
2875 never alters ARG itself. We assume that ARG is an operation that
2876 returns a truth value (0 or 1).
2878 FIXME: one would think we would fold the result, but it causes
2879 problems with the dominator optimizer. */
2880 tree
2882 {
2889 /* If this is a comparison, we can simply invert it, except for
2890 floating-point non-equality comparisons, in which case we just
2891 enclose a TRUTH_NOT_EXPR around what we have. */
2894 {
2897 && flag_trapping_math
2901 else
2902 {
2907 else
2910 }
2911 }
2914 {
2930 /* Here we can invert either operand. We invert the first operand
2931 unless the second operand is a TRUTH_NOT_EXPR in which case our
2932 result is the XOR of the first operand with the inside of the
2933 negation of the second operand. */
2938 else
2992 }
2995 }
2997 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
2998 operands are another bit-wise operation with a common input. If so,
2999 distribute the bit operations to save an operation and possibly two if
3000 constants are involved. For example, convert
3001 (A | B) & (A | C) into A | (B & C)
3002 Further simplification will occur if B and C are constants.
3004 If this optimization cannot be done, 0 will be returned. */
3006 static tree
3008 {
3009 tree common;
3019 {
3023 }
3025 {
3029 }
3031 {
3035 }
3037 {
3041 }
3042 else
3047 }
3048 \f
3049 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
3050 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero. */
3052 static tree
3055 {
3062 }
3064 /* Optimize a bit-field compare.
3066 There are two cases: First is a compare against a constant and the
3067 second is a comparison of two items where the fields are at the same
3068 bit position relative to the start of a chunk (byte, halfword, word)
3069 large enough to contain it. In these cases we can avoid the shift
3070 implicit in bitfield extractions.
3072 For constants, we emit a compare of the shifted constant with the
3073 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
3074 compared. For two fields at the same position, we do the ANDs with the
3075 similar mask and compare the result of the ANDs.
3077 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
3078 COMPARE_TYPE is the type of the comparison, and LHS and RHS
3079 are the left and right operands of the comparison, respectively.
3081 If the optimization described above can be done, we return the resulting
3082 tree. Otherwise we return zero. */
3084 static tree
3087 {
3096 tree mask;
3097 tree offset;
3099 /* Get all the information about the extractions being done. If the bit size
3100 if the same as the size of the underlying object, we aren't doing an
3101 extraction at all and so can do nothing. We also don't want to
3102 do anything if the inner expression is a PLACEHOLDER_EXPR since we
3103 then will no longer be able to replace it. */
3111 {
3112 /* If this is not a constant, we can only do something if bit positions,
3113 sizes, and signedness are the same. */
3121 }
3123 /* See if we can find a mode to refer to this field. We should be able to,
3124 but fail if we can't. */
3133 /* Set signed and unsigned types of the precision of this mode for the
3134 shifts below. */
3138 /* Compute the bit position and size for the new reference and our offset
3139 within it. If the new reference is the same size as the original, we
3140 won't optimize anything, so return zero. */
3150 /* Make the mask to be used against the extracted field. */
3159 /* If not comparing with constant, just rework the comparison
3160 and return. */
3165 mask),
3169 mask));
3171 /* Otherwise, we are handling the constant case. See if the constant is too
3172 big for the field. Warn and return a tree of for 0 (false) if so. We do
3173 this not only for its own sake, but to avoid having to test for this
3174 error case below. If we didn't, we might generate wrong code.
3176 For unsigned fields, the constant shifted right by the field length should
3177 be all zero. For signed fields, the high-order bits should agree with
3178 the sign bit. */
3181 {
3185 {
3189 }
3190 }
3191 else
3192 {
3196 {
3200 }
3201 }
3203 /* Single-bit compares should always be against zero. */
3205 {
3208 }
3210 /* Make a new bitfield reference, shift the constant over the
3211 appropriate number of bits and mask it with the computed mask
3212 (in case this was a signed field). If we changed it, make a new one. */
3215 {
3218 }
3228 rhs);
3229 }
3230 \f
3231 /* Subroutine for fold_truthop: decode a field reference.
3233 If EXP is a comparison reference, we return the innermost reference.
3235 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
3236 set to the starting bit number.
3238 If the innermost field can be completely contained in a mode-sized
3239 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
3241 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
3242 otherwise it is not changed.
3244 *PUNSIGNEDP is set to the signedness of the field.
3246 *PMASK is set to the mask used. This is either contained in a
3247 BIT_AND_EXPR or derived from the width of the field.
3249 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
3251 Return 0 if this is not a component reference or is one that we can't
3252 do anything with. */
3254 static tree
3259 {
3263 tree unsigned_type;
3266 /* All the optimizations using this function assume integer fields.
3267 There are problems with FP fields since the type_for_size call
3268 below can fail for, e.g., XFmode. */
3272 /* We are interested in the bare arrangement of bits, so strip everything
3273 that doesn't affect the machine mode. However, record the type of the
3274 outermost expression if it may matter below. */
3282 {
3288 }
3297 /* If the number of bits in the reference is the same as the bitsize of
3298 the outer type, then the outer type gives the signedness. Otherwise
3299 (in case of a small bitfield) the signedness is unchanged. */
3303 /* Compute the mask to access the bitfield. */
3313 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
3321 }
3323 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
3324 bit positions. */
3326 static int
3328 {
3331 tree tmask;
3336 return
3343 }
3345 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
3346 represents the sign bit of EXP's type. If EXP represents a sign
3347 or zero extension, also test VAL against the unextended type.
3348 The return value is the (sub)expression whose sign bit is VAL,
3349 or NULL_TREE otherwise. */
3351 static tree
3353 {
3357 tree t;
3359 /* Tree EXP must have an integral type. */
3364 /* Tree VAL must be an integer constant. */
3371 {
3378 }
3379 else
3380 {
3387 }
3389 /* We mask off those bits beyond TREE_TYPE (exp) so that we can
3390 treat VAL as if it were unsigned. */
3395 /* Handle extension from a narrower type. */
3401 }
3403 /* Subroutine for fold_truthop: determine if an operand is simple enough
3404 to be evaluated unconditionally. */
3406 static int
3408 {
3409 /* Strip any conversions that don't change the machine mode. */
3421 /* Don't regard global variables as simple. They may be
3422 allocated in ways unknown to the compiler (shared memory,
3423 #pragma weak, etc). */
3426 /* Loading a static variable is unduly expensive, but global
3427 registers aren't expensive. */
3429 }
3430 \f
3431 /* The following functions are subroutines to fold_range_test and allow it to
3432 try to change a logical combination of comparisons into a range test.
3434 For example, both
3435 X == 2 || X == 3 || X == 4 || X == 5
3436 and
3437 X >= 2 && X <= 5
3438 are converted to
3439 (unsigned) (X - 2) <= 3
3441 We describe each set of comparisons as being either inside or outside
3442 a range, using a variable named like IN_P, and then describe the
3443 range with a lower and upper bound. If one of the bounds is omitted,
3444 it represents either the highest or lowest value of the type.
3446 In the comments below, we represent a range by two numbers in brackets
3447 preceded by a "+" to designate being inside that range, or a "-" to
3448 designate being outside that range, so the condition can be inverted by
3449 flipping the prefix. An omitted bound is represented by a "-". For
3450 example, "- [-, 10]" means being outside the range starting at the lowest
3451 possible value and ending at 10, in other words, being greater than 10.
3452 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
3453 always false.
3455 We set up things so that the missing bounds are handled in a consistent
3456 manner so neither a missing bound nor "true" and "false" need to be
3457 handled using a special case. */
3459 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
3460 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
3461 and UPPER1_P are nonzero if the respective argument is an upper bound
3462 and zero for a lower. TYPE, if nonzero, is the type of the result; it
3463 must be specified for a comparison. ARG1 will be converted to ARG0's
3464 type if both are specified. */
3466 static tree
3469 {
3470 tree tem;
3474 /* If neither arg represents infinity, do the normal operation.
3475 Else, if not a comparison, return infinity. Else handle the special
3476 comparison rules. Note that most of the cases below won't occur, but
3477 are handled for consistency. */
3480 {
3485 }
3490 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
3491 for neither. In real maths, we cannot assume open ended ranges are
3492 the same. But, this is computer arithmetic, where numbers are finite.
3493 We can therefore make the transformation of any unbounded range with
3494 the value Z, Z being greater than any representable number. This permits
3495 us to treat unbounded ranges as equal. */
3499 {
3520 }
3523 }
3524 \f
3525 /* Given EXP, a logical expression, set the range it is testing into
3526 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
3527 actually being tested. *PLOW and *PHIGH will be made of the same type
3528 as the returned expression. If EXP is not a comparison, we will most
3529 likely not be returning a useful value and range. */
3531 static tree
3533 {
3540 /* Start with simply saying "EXP != 0" and then look at the code of EXP
3541 and see if we can refine the range. Some of the cases below may not
3542 happen, but it doesn't seem worth worrying about this. We "continue"
3543 the outer loop when we've changed something; otherwise we "break"
3544 the switch, which will "break" the while. */
3550 {
3555 {
3567 }
3570 {
3577 /* We can only do something if the range is testing for zero
3578 and if the second operand is an integer constant. Note that
3579 saying something is "in" the range we make is done by
3580 complementing IN_P since it will set in the initial case of
3581 being not equal to zero; "out" is leaving it alone. */
3588 {
3609 }
3611 /* If this is an unsigned comparison, we also know that EXP is
3612 greater than or equal to zero. We base the range tests we make
3613 on that fact, so we record it here so we can parse existing
3614 range tests. We test arg0_type since often the return type
3615 of, e.g. EQ_EXPR, is boolean. */
3617 {
3621 NULL_TREE))
3626 /* If the high bound is missing, but we have a nonzero low
3627 bound, reverse the range so it goes from zero to the low bound
3628 minus 1. */
3630 {
3635 }
3636 }
3642 /* (-x) IN [a,b] -> x in [-b, -a] */
3654 /* ~ X -> -X - 1 */
3663 /* If EXP is signed, any overflow in the computation is undefined,
3664 so we don't worry about it so long as our computations on
3665 the bounds don't overflow. For unsigned, overflow is defined
3666 and this is exactly the right thing. */
3675 /* Check for an unsigned range which has wrapped around the maximum
3676 value thus making n_high < n_low, and normalize it. */
3678 {
3684 /* If the range is of the form +/- [ x+1, x ], we won't
3685 be able to normalize it. But then, it represents the
3686 whole range or the empty set, so make it
3687 +/- [ -, - ]. */
3691 else
3693 }
3694 else
3718 /* If we're converting arg0 from an unsigned type, to exp,
3719 a signed type, we will be doing the comparison as unsigned.
3720 The tests above have already verified that LOW and HIGH
3721 are both positive.
3723 So we have to ensure that we will handle large unsigned
3724 values the same way that the current signed bounds treat
3725 negative values. */
3728 {
3729 tree high_positive;
3733 /* A range without an upper bound is, naturally, unbounded.
3734 Since convert would have cropped a very large value, use
3735 the max value for the destination type. */
3736 high_positive
3743 high_positive),
3745 integer_one_node)));
3747 /* If the low bound is specified, "and" the range with the
3748 range for which the original unsigned value will be
3749 positive. */
3751 {
3755 integer_zero_node),
3756 high_positive))
3760 }
3761 else
3762 {
3763 /* Otherwise, "or" the range with the range of the input
3764 that will be interpreted as negative. */
3768 integer_zero_node),
3769 high_positive))
3773 }
3774 }
3782 }
3785 }
3787 /* If EXP is a constant, we can evaluate whether this is true or false. */
3789 {
3796 }
3800 }
3801 \f
3802 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
3803 type, TYPE, return an expression to test if EXP is in (or out of, depending
3804 on IN_P) the range. Return 0 if the test couldn't be created. */
3806 static tree
3808 {
3810 tree value;
3813 {
3819 }
3834 {
3836 {
3840 }
3842 }
3844 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
3846 {
3848 HOST_WIDE_INT hi;
3853 {
3856 }
3857 else
3858 {
3861 }
3864 {
3866 {
3869 }
3872 }
3873 }
3877 {
3880 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
3881 for the type in question, as we rely on this here. */
3883 {
3894 {
3900 }
3904 }
3905 }
3911 value);
3914 }
3915 \f
3916 /* Given two ranges, see if we can merge them into one. Return 1 if we
3917 can, 0 if we can't. Set the output range into the specified parameters. */
3919 static int
3922 {
3926 tree tem;
3936 /* Make range 0 be the range that starts first, or ends last if they
3937 start at the same value. Swap them if it isn't. */
3940 || (lowequal
3943 {
3947 }
3949 /* Now flag two cases, whether the ranges are disjoint or whether the
3950 second range is totally subsumed in the first. Note that the tests
3951 below are simplified by the ones above. */
3957 /* We now have four cases, depending on whether we are including or
3958 excluding the two ranges. */
3960 {
3961 /* If they don't overlap, the result is false. If the second range
3962 is a subset it is the result. Otherwise, the range is from the start
3963 of the second to the end of the first. */
3968 else
3970 }
3973 {
3974 /* If they don't overlap, the result is the first range. If they are
3975 equal, the result is false. If the second range is a subset of the
3976 first, and the ranges begin at the same place, we go from just after
3977 the end of the first range to the end of the second. If the second
3978 range is not a subset of the first, or if it is a subset and both
3979 ranges end at the same place, the range starts at the start of the
3980 first range and ends just before the second range.
3981 Otherwise, we can't describe this as a single range. */
3987 {
3991 }
3993 {
3997 }
3998 else
4000 }
4003 {
4004 /* If they don't overlap, the result is the second range. If the second
4005 is a subset of the first, the result is false. Otherwise,
4006 the range starts just after the first range and ends at the
4007 end of the second. */
4012 else
4013 {
4017 }
4018 }
4020 else
4021 {
4022 /* The case where we are excluding both ranges. Here the complex case
4023 is if they don't overlap. In that case, the only time we have a
4024 range is if they are adjacent. If the second is a subset of the
4025 first, the result is the first. Otherwise, the range to exclude
4026 starts at the beginning of the first range and ends at the end of the
4027 second. */
4029 {
4036 else
4037 {
4038 /* Canonicalize - [min, x] into - [-, x]. */
4041 {
4046 /* FALLTHROUGH */
4060 }
4062 /* Canonicalize - [x, max] into - [x, -]. */
4065 {
4070 /* FALLTHROUGH */
4086 }
4088 /* The ranges might be also adjacent between the maximum and
4089 minimum values of the given type. For
4090 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
4091 return + [x + 1, y - 1]. */
4093 {
4102 }
4103 else
4105 }
4106 }
4109 else
4111 }
4115 }
4116 \f
4118 /* Subroutine of fold, looking inside expressions of the form
4119 A op B ? A : C, where ARG0, ARG1 and ARG2 are the three operands
4120 of the COND_EXPR. This function is being used also to optimize
4121 A op B ? C : A, by reversing the comparison first.
4123 Return a folded expression whose code is not a COND_EXPR
4124 anymore, or NULL_TREE if no folding opportunity is found. */
4126 static tree
4128 {
4133 tree tem;
4138 /* If we have A op 0 ? A : -A, consider applying the following
4139 transformations:
4141 A == 0? A : -A same as -A
4142 A != 0? A : -A same as A
4143 A >= 0? A : -A same as abs (A)
4144 A > 0? A : -A same as abs (A)
4145 A <= 0? A : -A same as -abs (A)
4146 A < 0? A : -A same as -abs (A)
4148 None of these transformations work for modes with signed
4149 zeros. If A is +/-0, the first two transformations will
4150 change the sign of the result (from +0 to -0, or vice
4151 versa). The last four will fix the sign of the result,
4152 even though the original expressions could be positive or
4153 negative, depending on the sign of A.
4155 Note that all these transformations are correct if A is
4156 NaN, since the two alternatives (A and -A) are also NaNs. */
4163 {
4175 /* Fall through. */
4197 }
4199 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
4200 A == 0 ? A : 0 is always 0 unless A is -0. Note that
4201 both transformations are correct when A is NaN: A != 0
4202 is then true, and A == 0 is false. */
4205 {
4210 }
4212 /* Try some transformations of A op B ? A : B.
4214 A == B? A : B same as B
4215 A != B? A : B same as A
4216 A >= B? A : B same as max (A, B)
4217 A > B? A : B same as max (B, A)
4218 A <= B? A : B same as min (A, B)
4219 A < B? A : B same as min (B, A)
4221 As above, these transformations don't work in the presence
4222 of signed zeros. For example, if A and B are zeros of
4223 opposite sign, the first two transformations will change
4224 the sign of the result. In the last four, the original
4225 expressions give different results for (A=+0, B=-0) and
4226 (A=-0, B=+0), but the transformed expressions do not.
4228 The first two transformations are correct if either A or B
4229 is a NaN. In the first transformation, the condition will
4230 be false, and B will indeed be chosen. In the case of the
4231 second transformation, the condition A != B will be true,
4232 and A will be chosen.
4234 The conversions to max() and min() are not correct if B is
4235 a number and A is not. The conditions in the original
4236 expressions will be false, so all four give B. The min()
4237 and max() versions would give a NaN instead. */
4239 {
4244 /* Avoid adding NOP_EXPRs in case this is an lvalue. */
4246 {
4250 }
4253 {
4262 /* In C++ a ?: expression can be an lvalue, so put the
4263 operand which will be used if they are equal first
4264 so that we can convert this back to the
4265 corresponding COND_EXPR. */
4267 {
4274 }
4281 {
4288 }
4301 }
4302 }
4304 /* If this is A op C1 ? A : C2 with C1 and C2 constant integers,
4305 we might still be able to simplify this. For example,
4306 if C1 is one less or one more than C2, this might have started
4307 out as a MIN or MAX and been transformed by this function.
4308 Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE. */
4314 {
4316 /* We can replace A with C1 in this case. */
4321 /* If C1 is C2 + 1, this is min(A, C2). */
4323 OEP_ONLY_CONST)
4327 OEP_ONLY_CONST))
4333 /* If C1 is C2 - 1, this is min(A, C2). */
4335 OEP_ONLY_CONST)
4339 OEP_ONLY_CONST))
4345 /* If C1 is C2 - 1, this is max(A, C2). */
4347 OEP_ONLY_CONST)
4351 OEP_ONLY_CONST))
4357 /* If C1 is C2 + 1, this is max(A, C2). */
4359 OEP_ONLY_CONST)
4363 OEP_ONLY_CONST))
4371 }
4374 }
4377 \f
4378 #ifndef RANGE_TEST_NON_SHORT_CIRCUIT
4379 #define RANGE_TEST_NON_SHORT_CIRCUIT (BRANCH_COST >= 2)
4380 #endif
4382 /* EXP is some logical combination of boolean tests. See if we can
4383 merge it into some range test. Return the new tree if so. */
4385 static tree
4387 {
4394 tree tem;
4396 /* If this is an OR operation, invert both sides; we will invert
4397 again at the end. */
4401 /* If both expressions are the same, if we can merge the ranges, and we
4402 can build the range test, return it or it inverted. If one of the
4403 ranges is always true or always false, consider it to be the same
4404 expression as the other. */
4414 /* On machines where the branch cost is expensive, if this is a
4415 short-circuited branch and the underlying object on both sides
4416 is the same, make a non-short-circuit operation. */
4422 {
4423 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
4424 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
4425 which cases we can't do this. */
4434 {
4446 }
4447 }
4450 }
4451 \f
4452 /* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
4453 bit value. Arrange things so the extra bits will be set to zero if and
4454 only if C is signed-extended to its full width. If MASK is nonzero,
4455 it is an INTEGER_CST that should be AND'ed with the extra bits. */
4457 static tree
4459 {
4462 tree temp;
4467 /* We work by getting just the sign bit into the low-order bit, then
4468 into the high-order bit, then sign-extend. We then XOR that value
4469 with C. */
4473 /* We must use a signed type in order to get an arithmetic right shift.
4474 However, we must also avoid introducing accidental overflows, so that
4475 a subsequent call to integer_zerop will work. Hence we must
4476 do the type conversion here. At this point, the constant is either
4477 zero or one, and the conversion to a signed type can never overflow.
4478 We could get an overflow if this conversion is done anywhere else. */
4487 /* If necessary, convert the type back to match the type of C. */
4492 }
4493 \f
4494 /* Find ways of folding logical expressions of LHS and RHS:
4495 Try to merge two comparisons to the same innermost item.
4496 Look for range tests like "ch >= '0' && ch <= '9'".
4497 Look for combinations of simple terms on machines with expensive branches
4498 and evaluate the RHS unconditionally.
4500 For example, if we have p->a == 2 && p->b == 4 and we can make an
4501 object large enough to span both A and B, we can do this with a comparison
4502 against the object ANDed with the a mask.
4504 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
4505 operations to do this with one comparison.
4507 We check for both normal comparisons and the BIT_AND_EXPRs made this by
4508 function and the one above.
4510 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
4511 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
4513 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
4514 two operands.
4516 We return the simplified tree or 0 if no optimization is possible. */
4518 static tree
4520 {
4521 /* If this is the "or" of two comparisons, we can do something if
4522 the comparisons are NE_EXPR. If this is the "and", we can do something
4523 if the comparisons are EQ_EXPR. I.e.,
4524 (a->b == 2 && a->c == 4) can become (a->new == NEW).
4526 WANTED_CODE is this operation code. For single bit fields, we can
4527 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
4528 comparison for one-bit fields. */
4548 /* Start by getting the comparison codes. Fail if anything is volatile.
4549 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
4550 it were surrounded with a NE_EXPR. */
4559 {
4563 }
4566 {
4570 }
4581 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
4584 {
4585 tree result;
4588 {
4593 }
4596 {
4602 }
4603 }
4608 /* If the RHS can be evaluated unconditionally and its operands are
4609 simple, it wins to evaluate the RHS unconditionally on machines
4610 with expensive branches. In this case, this isn't a comparison
4611 that can be merged. Avoid doing this if the RHS is a floating-point
4612 comparison since those can trap. */
4618 {
4619 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
4629 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
4640 }
4642 /* See if the comparisons can be merged. Then get all the parameters for
4643 each side. */
4667 /* It must be true that the inner operation on the lhs of each
4668 comparison must be the same if we are to be able to do anything.
4669 Then see if we have constants. If not, the same must be true for
4670 the rhs's. */
4681 else
4684 /* If either comparison code is not correct for our logical operation,
4685 fail. However, we can convert a one-bit comparison against zero into
4686 the opposite comparison against that bit being set in the field. */
4690 {
4692 {
4693 /* Make the left operand unsigned, since we are only interested
4694 in the value of one bit. Otherwise we are doing the wrong
4695 thing below. */
4698 }
4699 else
4701 }
4703 /* This is analogous to the code for l_const above. */
4705 {
4707 {
4710 }
4711 else
4713 }
4715 /* After this point all optimizations will generate bit-field
4716 references, which we might not want. */
4720 /* See if we can find a mode that contains both fields being compared on
4721 the left. If we can't, fail. Otherwise, update all constants and masks
4722 to be relative to a field of that size. */
4727 volatilep);
4737 {
4740 }
4748 {
4756 {
4760 }
4761 }
4763 {
4771 {
4775 }
4776 }
4778 /* If the right sides are not constant, do the same for it. Also,
4779 disallow this optimization if a size or signedness mismatch occurs
4780 between the left and right sides. */
4782 {
4785 /* Make sure the two fields on the right
4786 correspond to the left without being swapped. */
4794 volatilep);
4804 {
4807 }
4814 /* Make a mask that corresponds to both fields being compared.
4815 Do this for both items being compared. If the operands are the
4816 same size and the bits being compared are in the same position
4817 then we can do this by masking both and comparing the masked
4818 results. */
4822 {
4834 }
4836 /* There is still another way we can do something: If both pairs of
4837 fields being compared are adjacent, we may be able to make a wider
4838 field containing them both.
4840 Note that we still must mask the lhs/rhs expressions. Furthermore,
4841 the mask must be shifted to account for the shift done by
4842 make_bit_field_ref. */
4847 {
4848 tree type;
4860 /* Convert to the smaller type before masking out unwanted bits. */
4863 {
4865 {
4869 }
4871 {
4875 }
4876 }
4885 }
4888 }
4890 /* Handle the case of comparisons with constants. If there is something in
4891 common between the masks, those bits of the constants must be the same.
4892 If not, the condition is always false. Test for this to avoid generating
4893 incorrect code below. */
4898 {
4900 {
4903 }
4904 else
4905 {
4908 }
4909 }
4911 /* Construct the expression we will return. First get the component
4912 reference we will make. Unless the mask is all ones the width of
4913 that field, perform the mask operation. Then compare with the
4914 merged constant. */
4924 }
4925 \f
4926 /* Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a
4927 constant. */
4929 static tree
4931 {
4936 tree minmax_const;
4938 tree inner;
4948 /* If something does not permit us to optimize, return the original tree. */
4956 /* Now handle all the various comparison codes. We only handle EQ_EXPR
4957 and GT_EXPR, doing the rest with recursive calls using logical
4958 simplifications. */
4960 {
4962 return
4966 return
4968 optimize_minmax_comparison
4970 optimize_minmax_comparison
4975 /* MAX (X, 0) == 0 -> X <= 0 */
4979 /* MAX (X, 0) == 5 -> X == 5 */
4983 /* MAX (X, 0) == -1 -> false */
4987 /* MIN (X, 0) == 0 -> X >= 0 */
4991 /* MIN (X, 0) == 5 -> false */
4994 else
4995 /* MIN (X, 0) == -1 -> X == -1 */
5000 /* MAX (X, 0) > 0 -> X > 0
5001 MAX (X, 0) > 5 -> X > 5 */
5005 /* MAX (X, 0) > -1 -> true */
5009 /* MIN (X, 0) > 0 -> false
5010 MIN (X, 0) > 5 -> false */
5013 else
5014 /* MIN (X, 0) > -1 -> X > -1 */
5019 }
5020 }
5021 \f
5022 /* T is an integer expression that is being multiplied, divided, or taken a
5023 modulus (CODE says which and what kind of divide or modulus) by a
5024 constant C. See if we can eliminate that operation by folding it with
5025 other operations already in T. WIDE_TYPE, if non-null, is a type that
5026 should be used for the computation if wider than our type.
5028 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
5029 (X * 2) + (Y * 4). We must, however, be assured that either the original
5030 expression would not overflow or that overflow is undefined for the type
5031 in the language in question.
5033 We also canonicalize (X + 7) * 4 into X * 4 + 28 in the hope that either
5034 the machine has a multiply-accumulate insn or that this is part of an
5035 addressing calculation.
5037 If we return a non-null expression, it is an equivalent form of the
5038 original computation, but need not be in the original type. */
5040 static tree
5042 {
5043 /* To avoid exponential search depth, refuse to allow recursion past
5044 three levels. Beyond that (1) it's highly unlikely that we'll find
5045 something interesting and (2) we've probably processed it before
5046 when we built the inner expression. */
5049 tree ret;
5054 depth++;
5056 depth--;
5059 }
5061 static tree
5063 {
5073 /* Don't deal with constants of zero here; they confuse the code below. */
5083 /* Note that we need not handle conditional operations here since fold
5084 already handles those cases. So just do arithmetic here. */
5086 {
5088 /* For a constant, we can always simplify if we are a multiply
5089 or (for divide and modulus) if it is a multiple of our constant. */
5097 /* If op0 is an expression ... */
5102 /* ... and is unsigned, and its type is smaller than ctype,
5103 then we cannot pass through as widening. */
5109 /* ... or this is a truncation (t is narrower than op0),
5110 then we cannot pass through this narrowing. */
5113 /* ... or signedness changes for division or modulus,
5114 then we cannot pass through this conversion. */
5120 /* Pass the constant down and see if we can make a simplification. If
5121 we can, replace this expression with the inner simplification for
5122 possible later conversion to our or some other type. */
5127 code == MULT_EXPR
5138 /* If widening the type changes the signedness, then we can't perform
5139 this optimization as that changes the result. */
5143 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
5146 {
5152 }
5156 /* If the second operand is constant, this is a multiplication
5157 or floor division, by a power of two, so we can treat it that
5158 way unless the multiplier or divisor overflows. Signed
5159 left-shift overflow is implementation-defined rather than
5160 undefined in C90, so do not convert signed left shift into
5161 multiplication. */
5164 /* const_binop may not detect overflow correctly,
5165 so check for it explicitly here. */
5170 size_one_node,
5180 /* See if we can eliminate the operation on both sides. If we can, we
5181 can return a new PLUS or MINUS. If we can't, the only remaining
5182 cases where we can do anything are if the second operand is a
5183 constant. */
5188 /* If not multiplication, we can only do this if both operands
5189 are divisible by c. */
5195 /* If this was a subtraction, negate OP1 and set it to be an addition.
5196 This simplifies the logic below. */
5203 /* If either OP1 or C are negative, this optimization is not safe for
5204 some of the division and remainder types while for others we need
5205 to change the code. */
5207 {
5215 }
5217 /* If it's a multiply or a division/modulus operation of a multiple
5218 of our constant, do the operation and verify it doesn't overflow. */
5221 {
5224 /* We allow the constant to overflow with wrapping semantics. */
5228 }
5229 else
5232 /* If we have an unsigned type is not a sizetype, we cannot widen
5233 the operation since it will change the result if the original
5234 computation overflowed. */
5240 /* If we were able to eliminate our operation from the first side,
5241 apply our operation to the second side and reform the PLUS. */
5245 /* The last case is if we are a multiply. In that case, we can
5246 apply the distributive law to commute the multiply and addition
5247 if the multiplication of the constants doesn't overflow. */
5253 op1));
5258 /* We have a special case here if we are doing something like
5259 (C * 8) % 4 since we know that's zero. */
5266 /* ... fall through ... */
5270 /* If we can extract our operation from the LHS, do so and return a
5271 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
5272 do something only if the second operand is a constant. */
5284 /* If these are the same operation types, we can associate them
5285 assuming no overflow. */
5292 /* If these operations "cancel" each other, we have the main
5293 optimizations of this pass, which occur when either constant is a
5294 multiple of the other, in which case we replace this with either an
5295 operation or CODE or TCODE.
5297 If we have an unsigned type that is not a sizetype, we cannot do
5298 this since it will change the result if the original computation
5299 overflowed. */
5302 && ! flag_wrapv
5307 {
5318 }
5323 }
5326 }
5327 \f
5328 /* Return a node which has the indicated constant VALUE (either 0 or
5329 1), and is of the indicated TYPE. */
5331 tree
5333 {
5341 else
5343 }
5345 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
5346 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
5347 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
5348 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
5349 COND is the first argument to CODE; otherwise (as in the example
5350 given here), it is the second argument. TYPE is the type of the
5351 original expression. Return NULL_TREE if no simplification is
5352 possible. */
5354 static tree
5357 {
5362 /* This transformation is only worthwhile if we don't have to wrap
5363 arg in a SAVE_EXPR, and the operation can be simplified on atleast
5364 one of the branches once its pushed inside the COND_EXPR. */
5369 {
5373 /* If this operand throws an expression, then it does not make
5374 sense to try to perform a logical or arithmetic operation
5375 involving it. */
5380 }
5381 else
5382 {
5387 }
5398 }
5400 \f
5401 /* Subroutine of fold() that checks for the addition of +/- 0.0.
5403 If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type
5404 TYPE, X + ADDEND is the same as X. If NEGATE, return true if X -
5405 ADDEND is the same as X.
5407 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
5408 and finite. The problematic cases are when X is zero, and its mode
5409 has signed zeros. In the case of rounding towards -infinity,
5410 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
5411 modes, X + 0 is not the same as X because -0 + 0 is 0. */
5413 static bool
5415 {
5419 /* Don't allow the fold with -fsignaling-nans. */
5423 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
5427 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
5432 /* The mode has signed zeros, and we have to honor their sign.
5433 In this situation, there is only one case we can return true for.
5434 X - 0 is the same as X unless rounding towards -infinity is
5435 supported. */
5437 }
5439 /* Subroutine of fold() that checks comparisons of built-in math
5440 functions against real constants.
5442 FCODE is the DECL_FUNCTION_CODE of the built-in, CODE is the comparison
5443 operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE
5444 is the type of the result and ARG0 and ARG1 are the operands of the
5445 comparison. ARG1 must be a TREE_REAL_CST.
5447 The function returns the constant folded tree if a simplification
5448 can be made, and NULL_TREE otherwise. */
5450 static tree
5453 {
5454 REAL_VALUE_TYPE c;
5457 {
5463 {
5464 /* sqrt(x) < y is always false, if y is negative. */
5468 /* sqrt(x) > y is always true, if y is negative and we
5469 don't care about NaNs, i.e. negative values of x. */
5473 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
5476 }
5478 {
5479 REAL_VALUE_TYPE c2;
5485 {
5486 /* sqrt(x) > y is x == +Inf, when y is very large. */
5491 /* sqrt(x) > y is always false, when y is very large
5492 and we don't care about infinities. */
5494 }
5496 /* sqrt(x) > c is the same as x > c*c. */
5499 }
5501 {
5502 REAL_VALUE_TYPE c2;
5508 {
5509 /* sqrt(x) < y is always true, when y is a very large
5510 value and we don't care about NaNs or Infinities. */
5514 /* sqrt(x) < y is x != +Inf when y is very large and we
5515 don't care about NaNs. */
5520 /* sqrt(x) < y is x >= 0 when y is very large and we
5521 don't care about Infinities. */
5526 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
5535 dconst0))),
5538 c2)))));
5539 }
5541 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
5546 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
5549 {
5554 dconst0))),
5557 c2)))));
5558 }
5559 }
5560 }
5563 }
5565 /* Subroutine of fold() that optimizes comparisons against Infinities,
5566 either +Inf or -Inf.
5568 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5569 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5570 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5572 The function returns the constant folded tree if a simplification
5573 can be made, and NULL_TREE otherwise. */
5575 static tree
5577 {
5579 REAL_VALUE_TYPE max;
5580 tree temp;
5585 /* For negative infinity swap the sense of the comparison. */
5591 {
5593 /* x > +Inf is always false, if with ignore sNANs. */
5599 /* x <= +Inf is always true, if we don't case about NaNs. */
5603 /* x <= +Inf is the same as x == x, i.e. isfinite(x). */
5606 {
5609 }
5614 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
5620 /* x < +Inf is always equal to x <= DBL_MAX. */
5626 /* x != +Inf is always equal to !(x > DBL_MAX). */
5632 /* The transformation below creates non-gimple code and thus is
5633 not appropriate if we are in gimple form. */
5643 }
5646 }
5648 /* Subroutine of fold() that optimizes comparisons of a division by
5649 a nonzero integer constant against an integer constant, i.e.
5650 X/C1 op C2.
5652 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5653 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5654 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5656 The function returns the constant folded tree if a simplification
5657 can be made, and NULL_TREE otherwise. */
5659 static tree
5661 {
5666 HOST_WIDE_INT hpart;
5669 /* We have to do this the hard way to detect unsigned overflow.
5670 prod = int_const_binop (MULT_EXPR, arg01, arg1, 0); */
5679 {
5683 /* Likewise hi = int_const_binop (PLUS_EXPR, prod, tmp, 0). */
5692 }
5694 {
5697 {
5715 }
5716 }
5717 else
5718 {
5721 {
5739 }
5740 }
5743 {
5784 }
5787 }
5790 /* If CODE with arguments ARG0 and ARG1 represents a single bit
5791 equality/inequality test, then return a simplified form of
5792 the test using shifts and logical operations. Otherwise return
5793 NULL. TYPE is the desired result type. */
5795 tree
5797 tree result_type)
5798 {
5799 /* If this is a TRUTH_NOT_EXPR, it may have a single bit test inside
5800 operand 0. */
5802 {
5807 /* Extract the arguments of the EQ/NE. */
5811 /* This requires us to invert the code. */
5813 }
5815 /* If this is testing a single bit, we can optimize the test. */
5819 {
5826 tree arg00;
5828 /* If we have (A & C) != 0 where C is the sign bit of A, convert
5829 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
5832 /* This is only a win if casting to a signed type is cheap,
5833 i.e. when arg00's type is not a partial mode. */
5836 {
5841 }
5843 /* Otherwise we have (A & C) != 0 where C is a single bit,
5844 convert that into ((A >> C2) & 1). Where C2 = log2(C).
5845 Similarly for (A & C) == 0. */
5847 /* If INNER is a right shift of a constant and it plus BITNUM does
5848 not overflow, adjust BITNUM and INNER. */
5855 {
5858 }
5860 /* If we are going to be able to omit the AND below, we must do our
5861 operations as unsigned. If we must use the AND, we have a choice.
5862 Normally unsigned is faster, but for some machines signed is. */
5863 #ifdef LOAD_EXTEND_OP
5865 #else
5867 #endif
5882 /* Put the AND last so it can combine with more things. */
5886 /* Make sure to return the proper type. */
5890 }
5892 }
5894 /* Check whether we are allowed to reorder operands arg0 and arg1,
5895 such that the evaluation of arg1 occurs before arg0. */
5897 static bool
5899 {
5906 }
5908 /* Test whether it is preferable two swap two operands, ARG0 and
5909 ARG1, for example because ARG0 is an integer constant and ARG1
5910 isn't. If REORDER is true, only recommend swapping if we can
5911 evaluate the operands in reverse order. */
5913 bool
5915 {
5960 /* It is preferable to swap two SSA_NAME to ensure a canonical form
5961 for commutative and comparison operators. Ensuring a canonical
5962 form allows the optimizers to find additional redundancies without
5963 having to explicitly check for both orderings. */
5970 }
5972 /* Perform constant folding and related simplification of EXPR.
5973 The related simplifications include x*1 => x, x*0 => 0, etc.,
5974 and application of the associative law.
5975 NOP_EXPR conversions may be removed freely (as long as we
5976 are careful not to change the type of the overall expression).
5977 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
5978 but we can constant-fold them if they have constant operands. */
5980 #ifdef ENABLE_FOLD_CHECKING
5981 # define fold(x) fold_1 (x)
5983 static
5984 #endif
5985 tree
5987 {
5991 tree tem;
5996 /* WINS will be nonzero when the switch is done
5997 if all operands are constant. */
6000 /* Return right away if a constant. */
6005 {
6006 tree subop;
6008 /* Special case for conversion ops that can have fixed point args. */
6011 /* Don't use STRIP_NOPS, because signedness of argument type matters. */
6017 else
6022 /* Note that TREE_CONSTANT isn't enough:
6023 static var addresses are constant but we can't
6024 do arithmetic on them. */
6026 }
6028 {
6032 {
6034 tree subop;
6039 /* Strip any conversions that don't change the mode. This is
6040 safe for every expression, except for a comparison expression
6041 because its signedness is derived from its operands. So, in
6042 the latter case, only strip conversions that don't change the
6043 signedness.
6045 Note that this is done as an internal manipulation within the
6046 constant folder, in order to find the simplest representation
6047 of the arguments so that their form can be studied. In any
6048 cases, the appropriate type conversions should be put back in
6049 the tree that will get out of the constant folder. */
6052 else
6057 else
6062 /* Note that TREE_CONSTANT isn't enough:
6063 static var addresses are constant but we can't
6064 do arithmetic on them. */
6071 }
6072 }
6074 /* If this is a commutative operation, and ARG0 is a constant, move it
6075 to ARG1 to reduce the number of tests below. */
6081 /* Now WINS is set as described above,
6082 ARG0 is the first operand of EXPR,
6083 and ARG1 is the second operand (if it has more than one operand).
6085 First check for cases where an arithmetic operation is applied to a
6086 compound, conditional, or comparison operation. Push the arithmetic
6087 operation inside the compound or conditional to see if any folding
6088 can then be done. Convert comparison to conditional for this purpose.
6089 The also optimizes non-constant cases that used to be done in
6090 expand_expr.
6092 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
6093 one of the operands is a comparison and the other is a comparison, a
6094 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
6095 code below would make the expression more complex. Change it to a
6096 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
6097 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
6109 {
6120 }
6123 {
6128 {
6138 /* If this was a conversion, and all we did was to move into
6139 inside the COND_EXPR, bring it back out. But leave it if
6140 it is a conversion from integer to integer and the
6141 result precision is no wider than a word since such a
6142 conversion is cheap and may be optimized away by combine,
6143 while it couldn't if it were outside the COND_EXPR. Then return
6144 so we don't get into an infinite recursion loop taking the
6145 conversion out and then back in. */
6157 && (INTEGRAL_TYPE_P
6168 }
6170 {
6172 {
6176 }
6180 integer_one_node)),
6182 integer_zero_node))));
6183 }
6184 }
6195 {
6199 arg1)));
6207 {
6212 }
6215 {
6220 }
6221 }
6224 {
6238 /* Handle cases of two conversions in a row. */
6241 {
6260 /* In addition to the cases of two conversions in a row
6261 handled below, if we are converting something to its own
6262 type via an object of identical or wider precision, neither
6263 conversion is needed. */
6270 /* Likewise, if the intermediate and final types are either both
6271 float or both integer, we don't need the middle conversion if
6272 it is wider than the final type and doesn't change the signedness
6273 (for integers). Avoid this if the final type is a pointer
6274 since then we sometimes need the inner conversion. Likewise if
6275 the outer has a precision not equal to the size of its mode. */
6286 /* If we have a sign-extension of a zero-extended value, we can
6287 replace that by a single zero-extension. */
6294 /* Two conversions in a row are not needed unless:
6295 - some conversion is floating-point (overstrict for now), or
6296 - the intermediate type is narrower than both initial and
6297 final, or
6298 - the intermediate type and innermost type differ in signedness,
6299 and the outermost type is wider than the intermediate, or
6300 - the initial type is a pointer type and the precisions of the
6301 intermediate and final types differ, or
6302 - the final type is a pointer type and the precisions of the
6303 initial and intermediate types differ. */
6318 }
6322 /* Detect assigning a bitfield. */
6325 {
6326 /* Don't leave an assignment inside a conversion
6327 unless assigning a bitfield. */
6331 /* First do the assignment, then return converted constant. */
6336 }
6338 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
6339 constants (if x has signed type, the sign bit cannot be set
6340 in c). This folds extension into the BIT_AND_EXPR. */
6345 {
6355 <= HOST_BITS_PER_WIDE_INT
6357 {
6364 #ifdef LOAD_EXTEND_OP
6368 {
6372 }
6373 #endif
6374 }
6379 }
6381 /* Convert (T1)((T2)X op Y) into (T1)X op Y, for pointer types T1 and
6382 T2 being pointers to types of the same size. */
6387 {
6399 }
6413 {
6417 }
6422 {
6427 }
6440 /* Convert fabs((double)float) into (double)fabsf(float). */
6443 {
6448 targ0)));
6449 }
6482 /* A + (-B) -> A - B */
6485 /* (-A) + B -> B - A */
6490 {
6494 /* If we are adding two BIT_AND_EXPR's, both of which are and'ing
6495 with a constant, and the two constants have no bits in common,
6496 we should treat this as a BIT_IOR_EXPR since this may produce more
6497 simplifications. */
6505 {
6508 }
6510 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
6511 (plus (plus (mult) (mult)) (foo)) so that we can
6512 take advantage of the factoring cases below. */
6517 {
6522 else
6543 }
6546 {
6550 /* (A * C) + (B * C) -> (A+B) * C.
6551 We are most concerned about the case where C is a constant,
6552 but other combinations show up during loop reduction. Since
6553 it is not difficult, try all four possibilities. */
6570 /* No identical multiplicands; see if we can find a common
6571 power-of-two factor in non-power-of-two multiplies. This
6572 can help in multi-dimensional array access. */
6577 {
6582 /* Move min of absolute values to int11. */
6585 {
6589 }
6592 {
6598 }
6599 }
6605 same));
6606 }
6607 }
6608 else
6609 {
6610 /* See if ARG1 is zero and X + ARG1 reduces to X. */
6614 /* Likewise if the operands are reversed. */
6618 /* Convert X + -C into X - C. */
6621 {
6627 }
6629 /* Convert x+x into x*2.0. */
6635 /* Convert x*c+x into x*(c+1). */
6641 {
6642 REAL_VALUE_TYPE c;
6648 }
6650 /* Convert x+x*c into x*(c+1). */
6656 {
6657 REAL_VALUE_TYPE c;
6663 }
6665 /* Convert x*c1+x*c2 into x*(c1+c2). */
6675 {
6684 }
6685 /* Convert a + (b*c + d*e) into (a + b*c) + d*e */
6689 {
6694 {
6695 tree tree0;
6698 }
6699 }
6700 /* Convert (b*c + d*e) + a into b*c + (d*e +a) */
6704 {
6709 {
6710 tree tree0;
6713 }
6714 }
6715 }
6717 bit_rotate:
6718 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
6719 is a rotate of A by C1 bits. */
6720 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
6721 is a rotate of A by B bits. */
6722 {
6731 {
6750 {
6758 TYPE_PRECISION
6763 ? LROTATE_EXPR
6766 }
6768 {
6776 TYPE_PRECISION
6781 ? LROTATE_EXPR
6784 }
6785 }
6786 }
6788 associate:
6789 /* In most languages, can't associate operations on floats through
6790 parentheses. Rather than remember where the parentheses were, we
6791 don't associate floats at all, unless the user has specified
6792 -funsafe-math-optimizations. */
6796 {
6800 /* Split both trees into variables, constants, and literals. Then
6801 associate each group together, the constants with literals,
6802 then the result with variables. This increases the chances of
6803 literals being recombined later and of generating relocatable
6804 expressions for the sum of a constant and literal. */
6809 /* Only do something if we found more than two objects. Otherwise,
6810 nothing has changed and we risk infinite recursion. */
6815 {
6816 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
6825 /* Preserve the MINUS_EXPR if the negative part of the literal is
6826 greater than the positive part. Otherwise, the multiplicative
6827 folding code (i.e extract_muldiv) may be fooled in case
6828 unsigned constants are subtracted, like in the following
6829 example: ((X*2 + 4) - 8U)/2. */
6831 {
6835 {
6839 }
6840 else
6841 {
6845 }
6846 }
6848 {
6853 else
6854 {
6860 }
6861 }
6866 }
6867 }
6869 binary:
6873 {
6874 /* The return value should always have
6875 the same type as the original expression. */
6880 }
6884 /* A - (-B) -> A + B */
6887 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
6897 {
6903 /* Fold A - (A & B) into ~B & A. */
6906 {
6911 arg0));
6916 arg0));
6917 }
6919 /* Fold (A & ~B) - (A & B) into (A ^ B) - B, where B is
6920 any power of 2 minus 1. */
6925 {
6931 {
6935 }
6936 }
6937 }
6939 /* See if ARG1 is zero and X - ARG1 reduces to X. */
6943 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
6944 ARG0 is zero and X + ARG0 reduces to X, since that would mean
6945 (-ARG1 + ARG0) reduces to -ARG1. */
6949 /* Fold &x - &x. This can happen from &x.foo - &x.
6950 This is unsafe for certain floats even in non-IEEE formats.
6951 In IEEE, it is unsafe because it does wrong for NaNs.
6952 Also note that operand_equal_p is always false if an operand
6953 is volatile. */
6959 /* A - B -> A + (-B) if B is easily negatable. */
6962 /* Avoid this transformation if B is a positive REAL_CST. */
6968 /* Try folding difference of addresses. */
6969 {
6970 HOST_WIDE_INT diff;
6978 }
6983 {
6984 /* (A * C) - (B * C) -> (A-B) * C. */
6992 /* (A * C1) - (A * C2) -> A * (C1-C2). */
7000 }
7005 /* (-A) * (-B) -> A * B */
7016 {
7022 /* (a * (1 << b)) is (a << b) */
7038 }
7039 else
7040 {
7041 /* Maybe fold x * 0 to 0. The expressions aren't the same
7042 when x is NaN, since x * 0 is also NaN. Nor are they the
7043 same in modes with signed zeros, since multiplying a
7044 negative value by 0 gives -0, not +0. */
7049 /* In IEEE floating point, x*1 is not equivalent to x for snans. */
7054 /* Transform x * -1.0 into -x. */
7059 /* Convert (C1/X)*C2 into (C1*C2)/X. */
7064 {
7070 }
7073 {
7077 /* Optimizations of root(...)*root(...). */
7079 {
7084 /* Optimize sqrt(x)*sqrt(x) as x. */
7090 /* Optimize root(x)*root(y) as root(x*y). */
7095 }
7097 /* Optimize expN(x)*expN(y) as expN(x+y). */
7099 {
7106 }
7108 /* Optimizations of pow(...)*pow(...). */
7112 {
7120 /* Optimize pow(x,y)*pow(z,y) as pow(x*z,y). */
7122 {
7127 arg01));
7129 }
7131 /* Optimize pow(x,y)*pow(x,z) as pow(x,y+z). */
7133 {
7138 arg));
7140 }
7141 }
7143 /* Optimize tan(x)*cos(x) as sin(x). */
7152 {
7158 }
7160 /* Optimize x*pow(x,c) as pow(x,c+1). */
7164 {
7171 {
7173 REAL_VALUE_TYPE c;
7182 }
7183 }
7185 /* Optimize pow(x,c)*x as pow(x,c+1). */
7189 {
7196 {
7198 REAL_VALUE_TYPE c;
7207 }
7208 }
7210 /* Optimize x*x as pow(x,2.0), which is expanded as x*x. */
7213 {
7217 {
7222 }
7223 }
7224 }
7225 }
7229 bit_ior:
7237 /* ~X | X is -1. */
7240 {
7244 }
7246 /* X | ~X is -1. */
7249 {
7253 }
7259 /* Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
7261 This results in more efficient code for machines without a NAND
7262 instruction. Combine will canonicalize to the first form
7263 which will allow use of NAND instructions provided by the
7264 backend if they exist. */
7267 {
7272 }
7274 /* See if this can be simplified into a rotate first. If that
7275 is unsuccessful continue in the association code. */
7286 /* ~X ^ X is -1. */
7289 {
7293 }
7295 /* X ^ ~X is -1. */
7298 {
7302 }
7304 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
7305 with a constant, and the two constants have no bits in common,
7306 we should treat this as a BIT_IOR_EXPR since this may produce more
7307 simplifications. */
7315 {
7318 }
7320 /* See if this can be simplified into a rotate first. If that
7321 is unsuccessful continue in the association code. */
7332 /* ~X & X is always zero. */
7337 /* X & ~X is always zero. */
7345 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
7348 {
7349 unsigned int prec
7356 }
7358 /* Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
7360 This results in more efficient code for machines without a NOR
7361 instruction. Combine will canonicalize to the first form
7362 which will allow use of NOR instructions provided by the
7363 backend if they exist. */
7366 {
7371 }
7376 /* Don't touch a floating-point divide by zero unless the mode
7377 of the constant can represent infinity. */
7383 /* (-A) / (-B) -> A / B */
7393 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
7398 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
7403 /* If ARG1 is a constant, we can convert this to a multiply by the
7404 reciprocal. This does not have the same rounding properties,
7405 so only do this if -funsafe-math-optimizations. We can actually
7406 always safely do it if ARG1 is a power of two, but it's hard to
7407 tell if it is or not in a portable manner. */
7409 {
7414 /* Find the reciprocal if optimizing and the result is exact. */
7416 {
7417 REAL_VALUE_TYPE r;
7420 {
7423 }
7424 }
7425 }
7426 /* Convert A/B/C to A/(B*C). */
7433 /* Convert A/(B/C) to (A/B)*C. */
7441 /* Convert C1/(X*C2) into (C1/C2)/X. */
7446 {
7452 }
7455 {
7457 /* Optimize x/expN(y) into x*expN(-y). */
7459 {
7466 }
7468 /* Optimize x/pow(y,z) into x*pow(y,-z). */
7472 {
7481 }
7482 }
7485 {
7489 /* Optimize sin(x)/cos(x) as tan(x). */
7495 {
7501 }
7503 /* Optimize cos(x)/sin(x) as 1.0/tan(x). */
7509 {
7513 {
7518 }
7519 }
7521 /* Optimize pow(x,c)/x as pow(x,c-1). */
7525 {
7531 {
7533 REAL_VALUE_TYPE c;
7542 }
7543 }
7544 }
7556 /* X / -1 is -X. */
7563 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
7564 operation, EXACT_DIV_EXPR.
7566 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
7567 At one time others generated faster code, it's not clear if they do
7568 after the last round to changes to the DIV code in expmed.c. */
7589 /* X % -1 is zero. */
7596 /* Optimize unsigned TRUNC_MOD_EXPR by a power of two into a
7597 BIT_AND_EXPR, i.e. "X % C" into "X & C2". */
7601 {
7603 tree mask;
7608 {
7612 }
7613 else
7614 {
7617 }
7622 }
7624 /* X % -C is the same as X % C. */
7629 && !flag_trapv
7630 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
7635 /* X % -Y is the same as X % Y. */
7657 /* Optimize -1 >> x for arithmetic right shifts. */
7660 /* ... fall through ... */
7663 shift:
7669 /* Since negative shift count is not well-defined,
7670 don't try to compute it in the compiler. */
7673 /* Rewrite an LROTATE_EXPR by a constant into an
7674 RROTATE_EXPR by a new constant. */
7676 {
7682 }
7684 /* If we have a rotate of a bit operation with the rotate count and
7685 the second operand of the bit operation both constant,
7686 permute the two operations. */
7698 /* Two consecutive rotates adding up to the width of the mode can
7699 be ignored. */
7730 /* The argument to invert_truthvalue must have Boolean type. */
7734 /* Note that the operand of this must be an int
7735 and its values must be 0 or 1.
7736 ("true" is a fixed value perhaps depending on the language,
7737 but we don't handle values other than 1 correctly yet.) */
7739 /* Avoid infinite recursion. */
7741 {
7746 }
7750 /* Note that the operands of this must be ints
7751 and their values must be 0 or 1.
7752 ("true" is a fixed value perhaps depending on the language.) */
7753 /* If first arg is constant zero, return it. */
7757 /* If either arg is constant true, drop it. */
7761 /* Preserve sequence points. */
7764 /* If second arg is constant zero, result is zero, but first arg
7765 must be evaluated. */
7768 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
7769 case will be handled here. */
7773 /* !X && X is always false. */
7777 /* X && !X is always false. */
7782 truth_andor:
7783 /* We only do these simplifications if we are optimizing. */
7787 /* Check for things like (A || B) && (A || C). We can convert this
7788 to A || (B && C). Note that either operator can be any of the four
7789 truth and/or operations and the transformation will still be
7790 valid. Also note that we only care about order for the
7791 ANDIF and ORIF operators. If B contains side effects, this
7792 might change the truth-value of A. */
7799 {
7819 /* This case if tricky because we must either have commutative
7820 operators or else A10 must not have side-effects. */
7826 a01));
7827 }
7829 /* See if we can build a range comparison. */
7833 /* Check for the possibility of merging component references. If our
7834 lhs is another similar operation, try to merge its rhs with our
7835 rhs. Then try to merge our lhs and rhs. */
7847 /* Note that the operands of this must be ints
7848 and their values must be 0 or true.
7849 ("true" is a fixed value perhaps depending on the language.) */
7850 /* If first arg is constant true, return it. */
7854 /* If either arg is constant zero, drop it. */
7858 /* Preserve sequence points. */
7861 /* If second arg is constant true, result is true, but we must
7862 evaluate first arg. */
7865 /* Likewise for first arg, but note this only occurs here for
7866 TRUTH_OR_EXPR. */
7870 /* !X || X is always true. */
7874 /* X || !X is always true. */
7882 /* If the second arg is constant zero, drop it. */
7885 /* If the second arg is constant true, this is a logical inversion. */
7888 /* Identical arguments cancel to zero. */
7892 /* !X ^ X is always true. */
7897 /* X ^ !X is always true. */
7910 /* If one arg is a real or integer constant, put it last. */
7914 /* If this is an equality comparison of the address of a non-weak
7915 object against zero, then we know the result. */
7923 /* If this is an equality comparison of the address of two non-weak,
7924 unaliased symbols neither of which are extern (since we do not
7925 have access to attributes for externs), then we know the result. */
7941 type);
7944 {
7952 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
7957 /* (-a) CMP (-b) -> b CMP a */
7964 {
7965 REAL_VALUE_TYPE cst;
7968 /* (-a) CMP CST -> a swap(CMP) (-CST) */
7970 return
7976 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
7977 /* a CMP (-0) -> a CMP 0 */
7982 /* x != NaN is always true, other ops are always false. */
7985 {
7988 }
7990 /* Fold comparisons against infinity. */
7992 {
7996 }
7997 }
7999 /* If this is a comparison of a real constant with a PLUS_EXPR
8000 or a MINUS_EXPR of a real constant, we can convert it into a
8001 comparison with a revised real constant as long as no overflow
8002 occurs when unsafe_math_optimizations are enabled. */
8014 /* Likewise, we can simplify a comparison of a real constant with
8015 a MINUS_EXPR whose first operand is also a real constant, i.e.
8016 (c1 - x) < c2 becomes x > c1-c2. */
8027 /* Fold comparisons against built-in math functions. */
8029 && flag_unsafe_math_optimizations
8031 {
8035 {
8039 }
8040 }
8041 }
8043 /* Convert foo++ == CONST into ++foo == CONST + INCR. */
8047 /* This optimization is invalid for ordered comparisons
8048 if CONST+INCR overflows or if foo+incr might overflow.
8049 This optimization is invalid for floating point due to rounding.
8050 For pointer types we assume overflow doesn't happen. */
8054 {
8058 {
8064 }
8065 else
8066 {
8072 }
8075 /* If VAROP is a reference to a bitfield, we must mask
8076 the constant by the width of the field. */
8081 {
8086 /* First check whether the comparison would come out
8087 always the same. If we don't do that we would
8088 change the meaning with the masking. */
8102 }
8105 }
8107 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
8108 This transformation affects the cases which are handled in later
8109 optimizations involving comparisons with non-negative constants. */
8113 {
8115 {
8126 }
8127 }
8129 /* Comparisons with the highest or lowest possible integer of
8130 the specified size will have known values.
8132 This is quite similar to fold_relational_hi_lo; however, my
8133 attempts to share the code have been nothing but trouble.
8134 I give up for now. */
8135 {
8143 {
8150 {
8153 }
8154 else
8155 {
8158 }
8163 {
8176 /* The GE_EXPR and LT_EXPR cases above are not normally
8177 reached because of previous transformations. */
8181 }
8185 {
8194 }
8198 {
8213 }
8217 {
8226 }
8232 /* signed_type does not work on pointer types. */
8234 {
8235 /* The following case also applies to X < signed_max+1
8236 and X >= signed_max+1 because previous transformations. */
8238 {
8242 return fold
8246 }
8247 }
8248 }
8249 }
8251 /* If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
8252 a MINUS_EXPR of a constant, we can convert it into a comparison with
8253 a revised constant as long as no overflow occurs. */
8265 /* Similarly for a NEGATE_EXPR. */
8274 /* If we have X - Y == 0, we can convert that to X == Y and similarly
8275 for !=. Don't do this for ordered comparisons due to overflow. */
8281 /* If we are widening one operand of an integer comparison,
8282 see if the other operand is similarly being widened. Perhaps we
8283 can do the comparison in the narrower type. */
8297 /* If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a
8298 constant, we can simplify it. */
8305 /* If we are comparing an ABS_EXPR with a constant, we can
8306 convert all the cases into explicit comparisons, but they may
8307 well not be faster than doing the ABS and one comparison.
8308 But ABS (X) <= C is a range comparison, which becomes a subtraction
8309 and a comparison, and is probably faster. */
8322 /* If this is an EQ or NE comparison with zero and ARG0 is
8323 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
8324 two operations, but the latter can be done in one less insn
8325 on machines that have only two-operand insns or on which a
8326 constant cannot be the first operand. */
8329 {
8334 return
8340 integer_one_node)),
8341 arg1));
8344 return
8350 integer_one_node)),
8351 arg1));
8352 }
8354 /* If this is an NE or EQ comparison of zero against the result of a
8355 signed MOD operation whose second operand is a power of 2, make
8356 the MOD operation unsigned since it is simpler and equivalent. */
8365 {
8375 }
8377 /* If this is an NE comparison of zero with an AND of one, remove the
8378 comparison since the AND will give the correct value. */
8384 /* If we have (A & C) == C where C is a power of 2, convert this into
8385 (A & C) != 0. Similarly for NE_EXPR. */
8392 integer_zero_node)));
8394 /* If we have (A & C) != 0 or (A & C) == 0 and C is a power of
8395 2, then fold the expression into shifts and logical operations. */
8400 /* If we have (A & C) == D where D & ~C != 0, convert this into 0.
8401 Similarly for NE_EXPR. */
8406 {
8407 tree dandnotc
8415 }
8417 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
8418 Similarly for NE_EXPR. */
8423 {
8424 tree candnotd
8431 }
8433 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
8434 and similarly for >= into !=. */
8450 return
8458 /* Simplify comparison of something with itself. (For IEEE
8459 floating-point, we can only do some of these simplifications.) */
8461 {
8463 {
8478 /* For NE, we can only do this simplification if integer
8479 or we don't honor IEEE floating point NaNs. */
8483 /* ... fall through ... */
8489 }
8490 }
8492 /* If we are comparing an expression that just has comparisons
8493 of two integer values, arithmetic expressions of those comparisons,
8494 and constants, we can simplify it. There are only three cases
8495 to check: the two values can either be equal, the first can be
8496 greater, or the second can be greater. Fold the expression for
8497 those three values. Since each value must be 0 or 1, we have
8498 eight possibilities, each of which corresponds to the constant 0
8499 or 1 or one of the six possible comparisons.
8501 This handles common cases like (a > b) == 0 but also handles
8502 expressions like ((x > y) - (y > x)) > 0, which supposedly
8503 occur in macroized code. */
8506 {
8511 /* Don't handle degenerate cases here; they should already
8512 have been handled anyway. */
8521 {
8525 /* We can't just pass T to eval_subst in case cval1 or cval2
8526 was the same as ARG1. */
8528 tree high_result
8532 arg1));
8533 tree equal_result
8537 arg1));
8538 tree low_result
8542 arg1));
8544 /* All three of these results should be 0 or 1. Confirm they
8545 are. Then use those values to select the proper code
8546 to use. */
8554 {
8555 /* Make a 3-bit mask with the high-order bit being the
8556 value for `>', the next for '=', and the low for '<'. */
8560 {
8562 /* Always false. */
8583 /* Always true. */
8585 }
8590 else
8592 }
8593 }
8594 }
8596 /* If this is a comparison of a field, we may be able to simplify it. */
8601 /* Handle the constant case even without -O
8602 to make sure the warnings are given. */
8604 {
8608 }
8610 /* If this is a comparison of complex values and either or both sides
8611 are a COMPLEX_EXPR or COMPLEX_CST, it is best to split up the
8612 comparisons and join them with a TRUTH_ANDIF_EXPR or TRUTH_ORIF_EXPR.
8613 This may prevent needless evaluations. */
8620 {
8633 type,
8636 }
8638 /* Optimize comparisons of strlen vs zero to a compare of the
8639 first character of the string vs zero. To wit,
8640 strlen(ptr) == 0 => *ptr == 0
8641 strlen(ptr) != 0 => *ptr != 0
8642 Other cases should reduce to one of these two (or a constant)
8643 due to the return value of strlen being unsigned. */
8647 {
8649 tree arglist;
8662 integer_zero_node)));
8663 }
8665 /* We can fold X/C1 op C2 where C1 and C2 are integer constants
8666 into a single range test. */
8673 {
8677 }
8697 {
8701 }
8703 /* If the first operand is NaN, the result is constant. */
8707 {
8709 ? integer_zero_node
8712 }
8714 /* If the second operand is NaN, the result is constant. */
8718 {
8720 ? integer_zero_node
8723 }
8725 /* Simplify unordered comparison of something with itself. */
8731 && !flag_trapping_math
8735 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
8736 {
8747 }
8752 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
8753 so all simple results must be passed through pedantic_non_lvalue. */
8755 {
8757 /* Only optimize constant conditions when the selected branch
8758 has the same type as the COND_EXPR. This avoids optimizing
8759 away "c ? x : throw", where the throw has a void type. */
8764 }
8768 /* If we have A op B ? A : C, we may be able to convert this to a
8769 simpler expression, depending on the operation and the values
8770 of B and C. Signed zeros prevent all of these transformations,
8771 for reasons given above each one.
8773 Also try swapping the arguments and inverting the conditional. */
8778 {
8784 }
8791 {
8794 {
8800 }
8801 }
8803 /* If the second operand is simpler than the third, swap them
8804 since that produces better jump optimization results. */
8807 {
8808 /* See if this can be inverted. If it can't, possibly because
8809 it was a floating-point inequality comparison, don't do
8810 anything. */
8816 }
8818 /* Convert A ? 1 : 0 to simply A. */
8821 /* If we try to convert TREE_OPERAND (t, 0) to our type, the
8822 call to fold will try to move the conversion inside
8823 a COND, which will recurse. In that case, the COND_EXPR
8824 is probably the best choice, so leave it alone. */
8828 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
8829 over COND_EXPR in cases such as floating point comparisons. */
8836 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
8844 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
8845 already handled above. */
8850 {
8859 }
8861 /* A & N ? N : 0 is simply A & N if N is a power of two. This
8862 is probably obsolete because the first operand should be a
8863 truth value (that's why we have the two cases above), but let's
8864 leave it in until we can confirm this for all front-ends. */
8875 /* Convert A ? B : 0 into A && B if A and B are truth values. */
8881 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
8885 {
8886 /* Only perform transformation if ARG0 is easily inverted. */
8890 }
8892 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
8896 {
8897 /* Only perform transformation if ARG0 is easily inverted. */
8902 }
8904 /* Convert A ? 1 : B into A || B if A and B are truth values. */
8914 /* When pedantic, a compound expression can be neither an lvalue
8915 nor an integer constant expression. */
8918 /* Don't let (0, 0) be null pointer constant. */
8961 /* Check for a built-in function. */
8966 {
8970 }
8976 }
8978 #ifdef ENABLE_FOLD_CHECKING
8979 #undef fold
8985 /* When --enable-checking=fold, compute a digest of expr before
8986 and after actual fold call to see if fold did not accidentally
8987 change original expr. */
8989 tree
8991 {
8992 tree ret;
8995 htab_t ht;
9014 }
9016 void
9018 {
9021 htab_t ht;
9031 }
9033 static void
9035 {
9037 }
9039 static void
9041 {
9059 {
9060 /* Allow DECL_ASSEMBLER_NAME to be modified. */
9064 }
9068 {
9069 /* Allow these fields to be modified. */
9076 }
9083 {
9086 {
9100 }
9104 {
9115 }
9149 {
9152 }
9162 }
9163 }
9165 #endif
9167 /* Perform constant folding and related simplification of initializer
9168 expression EXPR. This behaves identically to "fold" but ignores
9169 potential run-time traps and exceptions that fold must preserve. */
9171 tree
9173 {
9177 tree result;
9190 }
9192 /* Determine if first argument is a multiple of second argument. Return 0 if
9193 it is not, or we cannot easily determined it to be.
9195 An example of the sort of thing we care about (at this point; this routine
9196 could surely be made more general, and expanded to do what the *_DIV_EXPR's
9197 fold cases do now) is discovering that
9199 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9201 is a multiple of
9203 SAVE_EXPR (J * 8)
9205 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
9207 This code also handles discovering that
9209 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9211 is a multiple of 8 so we don't have to worry about dealing with a
9212 possible remainder.
9214 Note that we *look* inside a SAVE_EXPR only to determine how it was
9215 calculated; it is not safe for fold to do much of anything else with the
9216 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
9217 at run time. For example, the latter example above *cannot* be implemented
9218 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
9219 evaluation time of the original SAVE_EXPR is not necessarily the same at
9220 the time the new expression is evaluated. The only optimization of this
9221 sort that would be valid is changing
9223 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
9225 divided by 8 to
9227 SAVE_EXPR (I) * SAVE_EXPR (J)
9229 (where the same SAVE_EXPR (J) is used in the original and the
9230 transformed version). */
9232 static int
9234 {
9242 {
9254 {
9258 /* const_binop may not detect overflow correctly,
9259 so check for it explicitly here. */
9265 size_one_node,
9269 }
9273 /* Can't handle conversions from non-integral or wider integral type. */
9279 /* .. fall through ... */
9295 }
9296 }
9298 /* Return true if `t' is known to be non-negative. */
9300 int
9302 {
9304 {
9319 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
9320 both unsigned and at least 2 bits shorter than the result. */
9324 {
9329 {
9333 }
9334 }
9339 {
9340 /* x * x for floating point x is always non-negative. */
9345 }
9347 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
9348 both unsigned and their total bits is shorter than the result. */
9352 {
9359 }
9388 {
9393 {
9397 {
9401 }
9402 }
9404 {
9410 }
9411 }
9437 {
9441 /* If the initializer is non-void, then it's a normal expression
9442 that will be assigned to the slot. */
9446 /* Otherwise, the initializer sets the slot in some way. One common
9447 way is an assignment statement at the end of the initializer. */
9449 {
9457 else
9459 }
9465 }
9468 {
9475 {
9476 #define CASE_BUILTIN_F(BUILT_IN_FN) \
9477 case BUILT_IN_FN: case BUILT_IN_FN##F: case BUILT_IN_FN##L:
9478 #define CASE_BUILTIN_I(BUILT_IN_FN) \
9479 case BUILT_IN_FN: case BUILT_IN_FN##L: case BUILT_IN_FN##LL:
9497 /* Always true. */
9501 /* sqrt(-0.0) is -0.0. */
9529 /* True if the 1st argument is nonnegative. */
9533 /* True if the 1st OR 2nd arguments are nonnegative. */
9538 /* True if the 1st AND 2nd arguments are nonnegative. */
9543 /* True if the 2nd argument is nonnegative. */
9548 #undef CASE_BUILTIN_F
9549 #undef CASE_BUILTIN_I
9550 }
9551 }
9553 /* ... fall through ... */
9557 /* Truth values evaluate to 0 or 1, which is nonnegative. */
9559 }
9561 /* We don't know sign of `t', so be conservative and return false. */
9563 }
9565 /* Return true when T is an address and is known to be nonzero.
9566 For floating point we further ensure that T is not denormal.
9567 Similar logic is present in nonzero_address in rtlanal.h */
9569 static bool
9571 {
9574 /* Doing something useful for floating point would need more work. */
9579 {
9585 /* We used to test for !integer_zerop here. This does not work correctly
9586 if TREE_CONSTANT_OVERFLOW (t). */
9592 {
9593 /* With the presence of negative values it is hard
9594 to say something. */
9598 /* One of operands must be positive and the other non-negative. */
9601 }
9606 {
9609 }
9613 {
9619 }
9623 {
9629 /* Weak declarations may link to NULL. */
9633 /* Constants are never weak. */
9638 }
9650 {
9651 /* When both operands are nonzero, then MAX must be too. */
9655 /* MAX where operand 0 is positive is positive. */
9657 }
9658 /* MAX where operand 1 is positive is positive. */
9679 }
9681 }
9683 /* See if we are applying CODE, a relational to the highest or lowest
9684 possible integer of TYPE. If so, then the result is a compile
9685 time constant. */
9687 static tree
9690 {
9701 {
9708 {
9711 }
9712 else
9713 {
9716 }
9721 {
9735 /* The GE_EXPR and LT_EXPR cases above are not normally
9736 reached because of previous transformations. */
9740 }
9744 {
9755 }
9759 {
9776 }
9780 {
9791 }
9796 /* signed_type does not work on pointer types. */
9798 {
9799 /* The following case also applies to X < signed_max+1
9800 and X >= signed_max+1 because previous transformations. */
9802 {
9808 type,
9812 retval
9818 /* If we are in gimple form, then returning EXP would create
9819 non-gimple expressions. Clearing it is safe and insures
9820 we do not allow a non-gimple expression to escape. */
9825 }
9826 }
9827 }
9830 }
9833 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
9834 attempt to fold the expression to a constant without modifying TYPE,
9835 OP0 or OP1.
9837 If the expression could be simplified to a constant, then return
9838 the constant. If the expression would not be simplified to a
9839 constant, then return NULL_TREE.
9841 Note this is primarily designed to be called after gimplification
9842 of the tree structures and when at least one operand is a constant.
9843 As a result of those simplifying assumptions this routine is far
9844 simpler than the generic fold routine. */
9846 tree
9849 {
9851 tree subop0;
9852 tree subop1;
9853 tree tem;
9855 /* If this is a commutative operation, and ARG0 is a constant, move it
9856 to ARG1 to reduce the number of tests below. */
9859 {
9863 }
9865 /* If either operand is a complex type, extract its real component. */
9868 else
9873 else
9876 /* Note if either argument is not a real or integer constant.
9877 With a few exceptions, simplification is limited to cases
9878 where both arguments are constants. */
9886 {
9888 /* (plus (address) (const_int)) is a constant. */
9896 {
9900 }
9903 binary:
9907 /* Both arguments are constants. Simplify. */
9910 {
9911 /* The return value should always have the same type as
9912 the original expression. */
9917 }
9921 /* Fold &x - &x. This can happen from &x.foo - &x.
9922 This is unsafe for certain floats even in non-IEEE formats.
9923 In IEEE, it is unsafe because it does wrong for NaNs.
9924 Also note that operand_equal_p is always false if an
9925 operand is volatile. */
9933 /* Special case multiplication or bitwise AND where one argument
9934 is zero. */
9937 else
9946 /* Special case when we know the result will be all ones. */
9962 /* Division by zero is undefined. */
9989 /* Optimize -1 >> x for arithmetic right shifts. */
9992 /* ... fall through ... */
9998 /* Since negative shift count is not well-defined, don't
9999 try to compute it in the compiler. */
10007 /* -1 rotated either direction by any amount is still -1. */
10011 /* 0 rotated either direction by any amount is still zero. */
10028 /* If one arg is a real or integer constant, put it last. */
10033 {
10034 tree temp;
10040 }
10042 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
10043 This transformation affects the cases which are handled in later
10044 optimizations involving comparisons with non-negative constants. */
10048 {
10050 {
10063 }
10064 }
10070 /* Fall through. */
10086 /* This could probably be handled. */
10090 /* If second arg is constant zero, result is zero, but first arg
10091 must be evaluated. */
10094 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
10095 case will be handled here. */
10103 /* If second arg is constant true, result is true, but we must
10104 evaluate first arg. */
10107 /* Likewise for first arg, but note this only occurs here for
10108 TRUTH_OR_EXPR. */
10117 {
10120 }
10125 }
10126 }
10128 /* Given the components of a unary expression CODE, TYPE and OP0,
10129 attempt to fold the expression to a constant without modifying
10130 TYPE or OP0.
10132 If the expression could be simplified to a constant, then return
10133 the constant. If the expression would not be simplified to a
10134 constant, then return NULL_TREE.
10136 Note this is primarily designed to be called after gimplification
10137 of the tree structures and when op0 is a constant. As a result
10138 of those simplifying assumptions this routine is far simpler than
10139 the generic fold routine. */
10141 tree
10143 tree op0)
10144 {
10145 /* Make sure we have a suitable constant argument. */
10147 {
10148 tree subop;
10152 else
10157 }
10160 {
10172 else
10178 else
10184 else
10190 else
10196 else
10208 }
10209 }
10211 /* If EXP represents referencing an element in a constant string
10212 (either via pointer arithmetic or array indexing), return the
10213 tree representing the value accessed, otherwise return NULL. */
10215 tree
10217 {
10219 {
10221 tree index;
10222 tree string;
10226 else
10227 {
10231 /* Optimize the special-case of a zero lower bound.
10233 We convert the low_bound to sizetype to avoid some problems
10234 with constant folding. (E.g. suppose the lower bound is 1,
10235 and its mode is QI. Without the conversion,l (ARRAY
10236 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
10237 +INDEX), which becomes (ARRAY+255+INDEX). Opps!) */
10242 }
10256 }
10258 }
10260 /* Return the tree for neg (ARG0) when ARG0 is known to be either
10261 an integer constant or real constant.
10263 TYPE is the type of the result. */
10265 static tree
10267 {
10271 {
10273 {
10275 HOST_WIDE_INT high;
10285 }
10293 }
10296 }
10298 /* Return the tree for abs (ARG0) when ARG0 is known to be either
10299 an integer constant or real constant.
10301 TYPE is the type of the result. */
10303 tree
10305 {
10309 {
10311 /* If the value is unsigned, then the absolute value is
10312 the same as the ordinary value. */
10315 /* Similarly, if the value is non-negative. */
10318 /* If the value is negative, then the absolute value is
10319 its negation. */
10320 else
10321 {
10323 HOST_WIDE_INT high;
10330 }
10336 else
10342 }
10345 }
10347 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
10348 constant. TYPE is the type of the result. */
10350 static tree
10352 {
10364 }
10366 /* Given CODE, a relational operator, the target type, TYPE and two
10367 constant operands OP0 and OP1, return the result of the
10368 relational operation. If the result is not a compile time
10369 constant, then return NULL_TREE. */
10371 static tree
10373 {
10376 /* From here on, the only cases we handle are when the result is
10377 known to be a constant. */
10380 {
10384 /* Handle the cases where either operand is a NaN. */
10386 {
10388 {
10416 }
10419 }
10422 }
10424 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
10426 To compute GT, swap the arguments and do LT.
10427 To compute GE, do LT and invert the result.
10428 To compute LE, swap the arguments, do LT and invert the result.
10429 To compute NE, do EQ and invert the result.
10431 Therefore, the code below must handle only EQ and LT. */
10434 {
10439 }
10441 /* Note that it is safe to invert for real values here because we
10442 have already handled the one case that it matters. */
10446 {
10449 }
10451 /* Compute a result for LT or EQ if args permit;
10452 Otherwise return T. */
10454 {
10459 else
10461 }
10462 else
10468 }
10470 /* Build an expression for the address of T. Folds away INDIRECT_REF to
10471 avoid confusing the gimplify process. */
10473 tree
10475 {
10476 /* The size of the object is not relevant when talking about its address. */
10481 {
10485 }
10486 else
10487 {
10498 }
10501 }
10503 tree
10505 {
10507 }
10509 /* Builds an expression for an indirection through T, simplifying some
10510 cases. */
10512 tree
10514 {
10517 tree subtype;
10521 {
10524 /* *&p => p */
10527 /* *(foo *)&fooarray => fooarray[0] */
10531 }
10533 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
10537 {
10540 }
10543 }
10545 /* Strip non-trapping, non-side-effecting tree nodes from an expression
10546 whose result is ignored. The type of the returned tree need not be
10547 the same as the original expression. */
10549 tree
10551 {
10557 {
10568 else
10574 {
10590 }
10595 }
10596 }
10598 /* Return the value of VALUE, rounded up to a multiple of DIVISOR.
10599 This can only be applied to objects of a sizetype. */
10601 tree
10603 {
10610 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10611 have to do anything. Only do this when we are not given a const,
10612 because in that case, this check is more expensive than just
10613 doing it. */
10615 {
10620 }
10622 /* If divisor is a power of two, simplify this to bit manipulation. */
10624 {
10625 tree t;
10631 }
10632 else
10633 {
10638 }
10641 }
10643 /* Likewise, but round down. */
10645 tree
10647 {
10654 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10655 have to do anything. Only do this when we are not given a const,
10656 because in that case, this check is more expensive than just
10657 doing it. */
10659 {
10664 }
10666 /* If divisor is a power of two, simplify this to bit manipulation. */
10668 {
10669 tree t;
10673 }
10674 else
10675 {
10680 }
10683 }
10685 /* Returns true if addresses of E1 and E2 differ by a constant, false
10686 otherwise. If they do, &E1 - &E2 is stored in *DIFF. */
10688 bool
10690 {
10709 {
10719 }
10721 {
10722 /* If only one of the offsets is non-constant, the difference cannot
10723 be a constant. */
10725 }
10726 else
10731 }