| blob | 7de102f98704a062f6c93cc72b42ee256ca91f9e |
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 {
2017 /* Assume the worst for front-end tree codes. */
2021 }
2023 }
2025 /* Nonzero means lvalues are limited to those valid in pedantic ANSI C.
2026 Zero means allow extended lvalues. */
2030 /* When pedantic, return an expr equal to X but certainly not valid as a
2031 pedantic lvalue. Otherwise, return X. */
2033 tree
2035 {
2038 else
2040 }
2041 \f
2042 /* Given a tree comparison code, return the code that is the logical inverse
2043 of the given code. It is not safe to do this for floating-point
2044 comparisons, except for NE_EXPR and EQ_EXPR, so we receive a machine mode
2045 as well: if reversing the comparison is unsafe, return ERROR_MARK. */
2047 static enum tree_code
2049 {
2054 {
2085 }
2086 }
2088 /* Similar, but return the comparison that results if the operands are
2089 swapped. This is safe for floating-point. */
2091 enum tree_code
2093 {
2095 {
2109 }
2110 }
2113 /* Convert a comparison tree code from an enum tree_code representation
2114 into a compcode bit-based encoding. This function is the inverse of
2115 compcode_to_comparison. */
2117 static enum comparison_code
2119 {
2121 {
2152 }
2153 }
2155 /* Convert a compcode bit-based encoding of a comparison operator back
2156 to GCC's enum tree_code representation. This function is the
2157 inverse of comparison_to_compcode. */
2159 static enum tree_code
2161 {
2163 {
2194 }
2195 }
2197 /* Return a tree for the comparison which is the combination of
2198 doing the AND or OR (depending on CODE) of the two operations LCODE
2199 and RCODE on the identical operands LL_ARG and LR_ARG. Take into account
2200 the possibility of trapping if the mode has NaNs, and return NULL_TREE
2201 if this makes the transformation invalid. */
2203 tree
2207 {
2214 {
2225 }
2228 {
2229 /* Eliminate unordered comparisons, as well as LTGT and ORD
2230 which are not used unless the mode has NaNs. */
2236 }
2238 {
2239 /* Check that the original operation and the optimized ones will trap
2240 under the same condition. */
2251 /* In a short-circuited boolean expression the LHS might be
2252 such that the RHS, if evaluated, will never trap. For
2253 example, in ORD (x, y) && (x < y), we evaluate the RHS only
2254 if neither x nor y is NaN. (This is a mixed blessing: for
2255 example, the expression above will never trap, hence
2256 optimizing it to x < y would be invalid). */
2261 /* If the comparison was short-circuited, and only the RHS
2262 trapped, we may now generate a spurious trap. */
2267 /* If we changed the conditions that cause a trap, we lose. */
2270 }
2276 else
2279 }
2281 /* Return nonzero if CODE is a tree code that represents a truth value. */
2283 static int
2285 {
2290 }
2291 \f
2292 /* Return nonzero if two operands (typically of the same tree node)
2293 are necessarily equal. If either argument has side-effects this
2294 function returns zero. FLAGS modifies behavior as follows:
2296 If OEP_ONLY_CONST is set, only return nonzero for constants.
2297 This function tests whether the operands are indistinguishable;
2298 it does not test whether they are equal using C's == operation.
2299 The distinction is important for IEEE floating point, because
2300 (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and
2301 (2) two NaNs may be indistinguishable, but NaN!=NaN.
2303 If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself
2304 even though it may hold multiple values during a function.
2305 This is because a GCC tree node guarantees that nothing else is
2306 executed between the evaluation of its "operands" (which may often
2307 be evaluated in arbitrary order). Hence if the operands themselves
2308 don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the
2309 same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST
2310 unset means assuming isochronic (or instantaneous) tree equivalence.
2311 Unless comparing arbitrary expression trees, such as from different
2312 statements, this flag can usually be left unset.
2314 If OEP_PURE_SAME is set, then pure functions with identical arguments
2315 are considered the same. It is used when the caller has other ways
2316 to ensure that global memory is unchanged in between. */
2318 int
2320 {
2321 /* If one is specified and the other isn't, they aren't equal and if
2322 neither is specified, they are.
2324 ??? This is temporary and is meant only to handle the cases of the
2325 optional operands for COMPONENT_REF and ARRAY_REF. */
2330 /* If either is ERROR_MARK, they aren't equal. */
2334 /* If both types don't have the same signedness, then we can't consider
2335 them equal. We must check this before the STRIP_NOPS calls
2336 because they may change the signedness of the arguments. */
2344 /* This is needed for conversions and for COMPONENT_REF.
2345 Might as well play it safe and always test this. */
2351 /* If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal.
2352 We don't care about side effects in that case because the SAVE_EXPR
2353 takes care of that for us. In all other cases, two expressions are
2354 equal if they have no side effects. If we have two identical
2355 expressions with side effects that should be treated the same due
2356 to the only side effects being identical SAVE_EXPR's, that will
2357 be detected in the recursive calls below. */
2363 /* Next handle constant cases, those for which we can return 1 even
2364 if ONLY_CONST is set. */
2367 {
2380 {
2390 {
2392 flags))
2396 }
2399 }
2403 flags)
2405 flags));
2418 }
2424 {
2426 /* Two conversions are equal only if signedness and modes match. */
2428 {
2441 }
2454 /* For commutative ops, allow the other order. */
2462 /* If either of the pointer (or reference) expressions we are
2463 dereferencing contain a side effect, these cannot be equal. */
2469 {
2508 }
2512 {
2538 /* If the CALL_EXPRs call different functions, then they
2539 clearly can not be equal. */
2544 {
2548 else
2552 }
2554 /* Now see if all the arguments are the same. operand_equal_p
2555 does not handle TREE_LIST, so we walk the operands here
2556 feeding them to operand_equal_p. */
2560 {
2562 flags))
2567 }
2569 /* If we get here and both argument lists are exhausted
2570 then the CALL_EXPRs are equal. */
2575 }
2578 /* Consider __builtin_sqrt equal to sqrt. */
2586 }
2587 }
2588 \f
2589 /* Similar to operand_equal_p, but see if ARG0 might have been made by
2590 shorten_compare from ARG1 when ARG1 was being compared with OTHER.
2592 When in doubt, return 0. */
2594 static int
2596 {
2608 /* Discard any conversions that don't change the modes of ARG0 and ARG1
2609 and see if the inner values are the same. This removes any
2610 signedness comparison, which doesn't matter here. */
2617 /* Duplicate what shorten_compare does to ARG1 and see if that gives the
2618 actual comparison operand, ARG0.
2620 First throw away any conversions to wider types
2621 already present in the operands. */
2630 {
2633 /* Make sure shorter operand is extended the right way
2634 to match the longer operand. */
2640 }
2643 }
2644 \f
2645 /* See if ARG is an expression that is either a comparison or is performing
2646 arithmetic on comparisons. The comparisons must only be comparing
2647 two different values, which will be stored in *CVAL1 and *CVAL2; if
2648 they are nonzero it means that some operands have already been found.
2649 No variables may be used anywhere else in the expression except in the
2650 comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around
2651 the expression and save_expr needs to be called with CVAL1 and CVAL2.
2653 If this is true, return 1. Otherwise, return zero. */
2655 static int
2657 {
2661 /* We can handle some of the tcc_expression cases here. */
2671 {
2672 /* If we've already found a CVAL1 or CVAL2, this expression is
2673 two complex to handle. */
2679 }
2682 {
2705 /* First see if we can handle the first operand, then the second. For
2706 the second operand, we know *CVAL1 can't be zero. It must be that
2707 one side of the comparison is each of the values; test for the
2708 case where this isn't true by failing if the two operands
2709 are the same. */
2718 ;
2722 ;
2723 else
2727 ;
2731 ;
2732 else
2739 }
2740 }
2741 \f
2742 /* ARG is a tree that is known to contain just arithmetic operations and
2743 comparisons. Evaluate the operations in the tree substituting NEW0 for
2744 any occurrence of OLD0 as an operand of a comparison and likewise for
2745 NEW1 and OLD1. */
2747 static tree
2749 {
2754 /* We can handle some of the tcc_expression cases here. */
2762 {
2777 {
2794 }
2795 /* Fall through - ??? */
2798 {
2802 /* We need to check both for exact equality and tree equality. The
2803 former will be true if the operand has a side-effect. In that
2804 case, we know the operand occurred exactly once. */
2817 }
2821 }
2822 }
2823 \f
2824 /* Return a tree for the case when the result of an expression is RESULT
2825 converted to TYPE and OMITTED was previously an operand of the expression
2826 but is now not needed (e.g., we folded OMITTED * 0).
2828 If OMITTED has side effects, we must evaluate it. Otherwise, just do
2829 the conversion of RESULT to TYPE. */
2831 tree
2833 {
2840 }
2842 /* Similar, but call pedantic_non_lvalue instead of non_lvalue. */
2844 static tree
2846 {
2853 }
2855 /* Return a tree for the case when the result of an expression is RESULT
2856 converted to TYPE and OMITTED1 and OMITTED2 were previously operands
2857 of the expression but are now not needed.
2859 If OMITTED1 or OMITTED2 has side effects, they must be evaluated.
2860 If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is
2861 evaluated before OMITTED2. Otherwise, if neither has side effects,
2862 just do the conversion of RESULT to TYPE. */
2864 tree
2866 {
2875 }
2877 \f
2878 /* Return a simplified tree node for the truth-negation of ARG. This
2879 never alters ARG itself. We assume that ARG is an operation that
2880 returns a truth value (0 or 1).
2882 FIXME: one would think we would fold the result, but it causes
2883 problems with the dominator optimizer. */
2884 tree
2886 {
2893 /* If this is a comparison, we can simply invert it, except for
2894 floating-point non-equality comparisons, in which case we just
2895 enclose a TRUTH_NOT_EXPR around what we have. */
2898 {
2901 && flag_trapping_math
2905 else
2906 {
2911 else
2914 }
2915 }
2918 {
2934 /* Here we can invert either operand. We invert the first operand
2935 unless the second operand is a TRUTH_NOT_EXPR in which case our
2936 result is the XOR of the first operand with the inside of the
2937 negation of the second operand. */
2942 else
2996 }
2999 }
3001 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
3002 operands are another bit-wise operation with a common input. If so,
3003 distribute the bit operations to save an operation and possibly two if
3004 constants are involved. For example, convert
3005 (A | B) & (A | C) into A | (B & C)
3006 Further simplification will occur if B and C are constants.
3008 If this optimization cannot be done, 0 will be returned. */
3010 static tree
3012 {
3013 tree common;
3023 {
3027 }
3029 {
3033 }
3035 {
3039 }
3041 {
3045 }
3046 else
3051 }
3052 \f
3053 /* Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER
3054 starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero. */
3056 static tree
3059 {
3066 }
3068 /* Optimize a bit-field compare.
3070 There are two cases: First is a compare against a constant and the
3071 second is a comparison of two items where the fields are at the same
3072 bit position relative to the start of a chunk (byte, halfword, word)
3073 large enough to contain it. In these cases we can avoid the shift
3074 implicit in bitfield extractions.
3076 For constants, we emit a compare of the shifted constant with the
3077 BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being
3078 compared. For two fields at the same position, we do the ANDs with the
3079 similar mask and compare the result of the ANDs.
3081 CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR.
3082 COMPARE_TYPE is the type of the comparison, and LHS and RHS
3083 are the left and right operands of the comparison, respectively.
3085 If the optimization described above can be done, we return the resulting
3086 tree. Otherwise we return zero. */
3088 static tree
3091 {
3100 tree mask;
3101 tree offset;
3103 /* Get all the information about the extractions being done. If the bit size
3104 if the same as the size of the underlying object, we aren't doing an
3105 extraction at all and so can do nothing. We also don't want to
3106 do anything if the inner expression is a PLACEHOLDER_EXPR since we
3107 then will no longer be able to replace it. */
3115 {
3116 /* If this is not a constant, we can only do something if bit positions,
3117 sizes, and signedness are the same. */
3125 }
3127 /* See if we can find a mode to refer to this field. We should be able to,
3128 but fail if we can't. */
3137 /* Set signed and unsigned types of the precision of this mode for the
3138 shifts below. */
3142 /* Compute the bit position and size for the new reference and our offset
3143 within it. If the new reference is the same size as the original, we
3144 won't optimize anything, so return zero. */
3154 /* Make the mask to be used against the extracted field. */
3163 /* If not comparing with constant, just rework the comparison
3164 and return. */
3169 mask),
3173 mask));
3175 /* Otherwise, we are handling the constant case. See if the constant is too
3176 big for the field. Warn and return a tree of for 0 (false) if so. We do
3177 this not only for its own sake, but to avoid having to test for this
3178 error case below. If we didn't, we might generate wrong code.
3180 For unsigned fields, the constant shifted right by the field length should
3181 be all zero. For signed fields, the high-order bits should agree with
3182 the sign bit. */
3185 {
3189 {
3193 }
3194 }
3195 else
3196 {
3200 {
3204 }
3205 }
3207 /* Single-bit compares should always be against zero. */
3209 {
3212 }
3214 /* Make a new bitfield reference, shift the constant over the
3215 appropriate number of bits and mask it with the computed mask
3216 (in case this was a signed field). If we changed it, make a new one. */
3219 {
3222 }
3232 rhs);
3233 }
3234 \f
3235 /* Subroutine for fold_truthop: decode a field reference.
3237 If EXP is a comparison reference, we return the innermost reference.
3239 *PBITSIZE is set to the number of bits in the reference, *PBITPOS is
3240 set to the starting bit number.
3242 If the innermost field can be completely contained in a mode-sized
3243 unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
3245 *PVOLATILEP is set to 1 if the any expression encountered is volatile;
3246 otherwise it is not changed.
3248 *PUNSIGNEDP is set to the signedness of the field.
3250 *PMASK is set to the mask used. This is either contained in a
3251 BIT_AND_EXPR or derived from the width of the field.
3253 *PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
3255 Return 0 if this is not a component reference or is one that we can't
3256 do anything with. */
3258 static tree
3263 {
3267 tree unsigned_type;
3270 /* All the optimizations using this function assume integer fields.
3271 There are problems with FP fields since the type_for_size call
3272 below can fail for, e.g., XFmode. */
3276 /* We are interested in the bare arrangement of bits, so strip everything
3277 that doesn't affect the machine mode. However, record the type of the
3278 outermost expression if it may matter below. */
3286 {
3292 }
3301 /* If the number of bits in the reference is the same as the bitsize of
3302 the outer type, then the outer type gives the signedness. Otherwise
3303 (in case of a small bitfield) the signedness is unchanged. */
3307 /* Compute the mask to access the bitfield. */
3317 /* Merge it with the mask we found in the BIT_AND_EXPR, if any. */
3325 }
3327 /* Return nonzero if MASK represents a mask of SIZE ones in the low-order
3328 bit positions. */
3330 static int
3332 {
3335 tree tmask;
3340 return
3347 }
3349 /* Subroutine for fold: determine if VAL is the INTEGER_CONST that
3350 represents the sign bit of EXP's type. If EXP represents a sign
3351 or zero extension, also test VAL against the unextended type.
3352 The return value is the (sub)expression whose sign bit is VAL,
3353 or NULL_TREE otherwise. */
3355 static tree
3357 {
3361 tree t;
3363 /* Tree EXP must have an integral type. */
3368 /* Tree VAL must be an integer constant. */
3375 {
3382 }
3383 else
3384 {
3391 }
3393 /* We mask off those bits beyond TREE_TYPE (exp) so that we can
3394 treat VAL as if it were unsigned. */
3399 /* Handle extension from a narrower type. */
3405 }
3407 /* Subroutine for fold_truthop: determine if an operand is simple enough
3408 to be evaluated unconditionally. */
3410 static int
3412 {
3413 /* Strip any conversions that don't change the machine mode. */
3425 /* Don't regard global variables as simple. They may be
3426 allocated in ways unknown to the compiler (shared memory,
3427 #pragma weak, etc). */
3430 /* Loading a static variable is unduly expensive, but global
3431 registers aren't expensive. */
3433 }
3434 \f
3435 /* The following functions are subroutines to fold_range_test and allow it to
3436 try to change a logical combination of comparisons into a range test.
3438 For example, both
3439 X == 2 || X == 3 || X == 4 || X == 5
3440 and
3441 X >= 2 && X <= 5
3442 are converted to
3443 (unsigned) (X - 2) <= 3
3445 We describe each set of comparisons as being either inside or outside
3446 a range, using a variable named like IN_P, and then describe the
3447 range with a lower and upper bound. If one of the bounds is omitted,
3448 it represents either the highest or lowest value of the type.
3450 In the comments below, we represent a range by two numbers in brackets
3451 preceded by a "+" to designate being inside that range, or a "-" to
3452 designate being outside that range, so the condition can be inverted by
3453 flipping the prefix. An omitted bound is represented by a "-". For
3454 example, "- [-, 10]" means being outside the range starting at the lowest
3455 possible value and ending at 10, in other words, being greater than 10.
3456 The range "+ [-, -]" is always true and hence the range "- [-, -]" is
3457 always false.
3459 We set up things so that the missing bounds are handled in a consistent
3460 manner so neither a missing bound nor "true" and "false" need to be
3461 handled using a special case. */
3463 /* Return the result of applying CODE to ARG0 and ARG1, but handle the case
3464 of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P
3465 and UPPER1_P are nonzero if the respective argument is an upper bound
3466 and zero for a lower. TYPE, if nonzero, is the type of the result; it
3467 must be specified for a comparison. ARG1 will be converted to ARG0's
3468 type if both are specified. */
3470 static tree
3473 {
3474 tree tem;
3478 /* If neither arg represents infinity, do the normal operation.
3479 Else, if not a comparison, return infinity. Else handle the special
3480 comparison rules. Note that most of the cases below won't occur, but
3481 are handled for consistency. */
3484 {
3489 }
3494 /* Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0
3495 for neither. In real maths, we cannot assume open ended ranges are
3496 the same. But, this is computer arithmetic, where numbers are finite.
3497 We can therefore make the transformation of any unbounded range with
3498 the value Z, Z being greater than any representable number. This permits
3499 us to treat unbounded ranges as equal. */
3503 {
3524 }
3527 }
3528 \f
3529 /* Given EXP, a logical expression, set the range it is testing into
3530 variables denoted by PIN_P, PLOW, and PHIGH. Return the expression
3531 actually being tested. *PLOW and *PHIGH will be made of the same type
3532 as the returned expression. If EXP is not a comparison, we will most
3533 likely not be returning a useful value and range. */
3535 static tree
3537 {
3544 /* Start with simply saying "EXP != 0" and then look at the code of EXP
3545 and see if we can refine the range. Some of the cases below may not
3546 happen, but it doesn't seem worth worrying about this. We "continue"
3547 the outer loop when we've changed something; otherwise we "break"
3548 the switch, which will "break" the while. */
3554 {
3559 {
3571 }
3574 {
3581 /* We can only do something if the range is testing for zero
3582 and if the second operand is an integer constant. Note that
3583 saying something is "in" the range we make is done by
3584 complementing IN_P since it will set in the initial case of
3585 being not equal to zero; "out" is leaving it alone. */
3592 {
3613 }
3615 /* If this is an unsigned comparison, we also know that EXP is
3616 greater than or equal to zero. We base the range tests we make
3617 on that fact, so we record it here so we can parse existing
3618 range tests. We test arg0_type since often the return type
3619 of, e.g. EQ_EXPR, is boolean. */
3621 {
3625 NULL_TREE))
3630 /* If the high bound is missing, but we have a nonzero low
3631 bound, reverse the range so it goes from zero to the low bound
3632 minus 1. */
3634 {
3639 }
3640 }
3646 /* (-x) IN [a,b] -> x in [-b, -a] */
3658 /* ~ X -> -X - 1 */
3667 /* If EXP is signed, any overflow in the computation is undefined,
3668 so we don't worry about it so long as our computations on
3669 the bounds don't overflow. For unsigned, overflow is defined
3670 and this is exactly the right thing. */
3679 /* Check for an unsigned range which has wrapped around the maximum
3680 value thus making n_high < n_low, and normalize it. */
3682 {
3688 /* If the range is of the form +/- [ x+1, x ], we won't
3689 be able to normalize it. But then, it represents the
3690 whole range or the empty set, so make it
3691 +/- [ -, - ]. */
3695 else
3697 }
3698 else
3722 /* If we're converting arg0 from an unsigned type, to exp,
3723 a signed type, we will be doing the comparison as unsigned.
3724 The tests above have already verified that LOW and HIGH
3725 are both positive.
3727 So we have to ensure that we will handle large unsigned
3728 values the same way that the current signed bounds treat
3729 negative values. */
3732 {
3733 tree high_positive;
3737 /* A range without an upper bound is, naturally, unbounded.
3738 Since convert would have cropped a very large value, use
3739 the max value for the destination type. */
3740 high_positive
3747 high_positive),
3749 integer_one_node)));
3751 /* If the low bound is specified, "and" the range with the
3752 range for which the original unsigned value will be
3753 positive. */
3755 {
3759 integer_zero_node),
3760 high_positive))
3764 }
3765 else
3766 {
3767 /* Otherwise, "or" the range with the range of the input
3768 that will be interpreted as negative. */
3772 integer_zero_node),
3773 high_positive))
3777 }
3778 }
3786 }
3789 }
3791 /* If EXP is a constant, we can evaluate whether this is true or false. */
3793 {
3800 }
3804 }
3805 \f
3806 /* Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result
3807 type, TYPE, return an expression to test if EXP is in (or out of, depending
3808 on IN_P) the range. Return 0 if the test couldn't be created. */
3810 static tree
3812 {
3814 tree value;
3817 {
3823 }
3838 {
3840 {
3844 }
3846 }
3848 /* Optimize (c>=1) && (c<=127) into (signed char)c > 0. */
3850 {
3852 HOST_WIDE_INT hi;
3857 {
3860 }
3861 else
3862 {
3865 }
3868 {
3870 {
3873 }
3876 }
3877 }
3881 {
3884 /* Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
3885 for the type in question, as we rely on this here. */
3887 {
3898 {
3904 }
3908 }
3909 }
3915 value);
3918 }
3919 \f
3920 /* Given two ranges, see if we can merge them into one. Return 1 if we
3921 can, 0 if we can't. Set the output range into the specified parameters. */
3923 static int
3926 {
3930 tree tem;
3940 /* Make range 0 be the range that starts first, or ends last if they
3941 start at the same value. Swap them if it isn't. */
3944 || (lowequal
3947 {
3951 }
3953 /* Now flag two cases, whether the ranges are disjoint or whether the
3954 second range is totally subsumed in the first. Note that the tests
3955 below are simplified by the ones above. */
3961 /* We now have four cases, depending on whether we are including or
3962 excluding the two ranges. */
3964 {
3965 /* If they don't overlap, the result is false. If the second range
3966 is a subset it is the result. Otherwise, the range is from the start
3967 of the second to the end of the first. */
3972 else
3974 }
3977 {
3978 /* If they don't overlap, the result is the first range. If they are
3979 equal, the result is false. If the second range is a subset of the
3980 first, and the ranges begin at the same place, we go from just after
3981 the end of the first range to the end of the second. If the second
3982 range is not a subset of the first, or if it is a subset and both
3983 ranges end at the same place, the range starts at the start of the
3984 first range and ends just before the second range.
3985 Otherwise, we can't describe this as a single range. */
3991 {
3995 }
3997 {
4001 }
4002 else
4004 }
4007 {
4008 /* If they don't overlap, the result is the second range. If the second
4009 is a subset of the first, the result is false. Otherwise,
4010 the range starts just after the first range and ends at the
4011 end of the second. */
4016 else
4017 {
4021 }
4022 }
4024 else
4025 {
4026 /* The case where we are excluding both ranges. Here the complex case
4027 is if they don't overlap. In that case, the only time we have a
4028 range is if they are adjacent. If the second is a subset of the
4029 first, the result is the first. Otherwise, the range to exclude
4030 starts at the beginning of the first range and ends at the end of the
4031 second. */
4033 {
4040 else
4041 {
4042 /* Canonicalize - [min, x] into - [-, x]. */
4045 {
4050 /* FALLTHROUGH */
4064 }
4066 /* Canonicalize - [x, max] into - [x, -]. */
4069 {
4074 /* FALLTHROUGH */
4090 }
4092 /* The ranges might be also adjacent between the maximum and
4093 minimum values of the given type. For
4094 - [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
4095 return + [x + 1, y - 1]. */
4097 {
4106 }
4107 else
4109 }
4110 }
4113 else
4115 }
4119 }
4120 \f
4122 /* Subroutine of fold, looking inside expressions of the form
4123 A op B ? A : C, where ARG0, ARG1 and ARG2 are the three operands
4124 of the COND_EXPR. This function is being used also to optimize
4125 A op B ? C : A, by reversing the comparison first.
4127 Return a folded expression whose code is not a COND_EXPR
4128 anymore, or NULL_TREE if no folding opportunity is found. */
4130 static tree
4132 {
4137 tree tem;
4142 /* If we have A op 0 ? A : -A, consider applying the following
4143 transformations:
4145 A == 0? A : -A same as -A
4146 A != 0? A : -A same as A
4147 A >= 0? A : -A same as abs (A)
4148 A > 0? A : -A same as abs (A)
4149 A <= 0? A : -A same as -abs (A)
4150 A < 0? A : -A same as -abs (A)
4152 None of these transformations work for modes with signed
4153 zeros. If A is +/-0, the first two transformations will
4154 change the sign of the result (from +0 to -0, or vice
4155 versa). The last four will fix the sign of the result,
4156 even though the original expressions could be positive or
4157 negative, depending on the sign of A.
4159 Note that all these transformations are correct if A is
4160 NaN, since the two alternatives (A and -A) are also NaNs. */
4167 {
4179 /* Fall through. */
4201 }
4203 /* A != 0 ? A : 0 is simply A, unless A is -0. Likewise
4204 A == 0 ? A : 0 is always 0 unless A is -0. Note that
4205 both transformations are correct when A is NaN: A != 0
4206 is then true, and A == 0 is false. */
4209 {
4214 }
4216 /* Try some transformations of A op B ? A : B.
4218 A == B? A : B same as B
4219 A != B? A : B same as A
4220 A >= B? A : B same as max (A, B)
4221 A > B? A : B same as max (B, A)
4222 A <= B? A : B same as min (A, B)
4223 A < B? A : B same as min (B, A)
4225 As above, these transformations don't work in the presence
4226 of signed zeros. For example, if A and B are zeros of
4227 opposite sign, the first two transformations will change
4228 the sign of the result. In the last four, the original
4229 expressions give different results for (A=+0, B=-0) and
4230 (A=-0, B=+0), but the transformed expressions do not.
4232 The first two transformations are correct if either A or B
4233 is a NaN. In the first transformation, the condition will
4234 be false, and B will indeed be chosen. In the case of the
4235 second transformation, the condition A != B will be true,
4236 and A will be chosen.
4238 The conversions to max() and min() are not correct if B is
4239 a number and A is not. The conditions in the original
4240 expressions will be false, so all four give B. The min()
4241 and max() versions would give a NaN instead. */
4243 {
4248 /* Avoid adding NOP_EXPRs in case this is an lvalue. */
4250 {
4254 }
4257 {
4266 /* In C++ a ?: expression can be an lvalue, so put the
4267 operand which will be used if they are equal first
4268 so that we can convert this back to the
4269 corresponding COND_EXPR. */
4271 {
4278 }
4285 {
4292 }
4305 }
4306 }
4308 /* If this is A op C1 ? A : C2 with C1 and C2 constant integers,
4309 we might still be able to simplify this. For example,
4310 if C1 is one less or one more than C2, this might have started
4311 out as a MIN or MAX and been transformed by this function.
4312 Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE. */
4318 {
4320 /* We can replace A with C1 in this case. */
4325 /* If C1 is C2 + 1, this is min(A, C2). */
4327 OEP_ONLY_CONST)
4331 OEP_ONLY_CONST))
4337 /* If C1 is C2 - 1, this is min(A, C2). */
4339 OEP_ONLY_CONST)
4343 OEP_ONLY_CONST))
4349 /* If C1 is C2 - 1, this is max(A, C2). */
4351 OEP_ONLY_CONST)
4355 OEP_ONLY_CONST))
4361 /* If C1 is C2 + 1, this is max(A, C2). */
4363 OEP_ONLY_CONST)
4367 OEP_ONLY_CONST))
4375 }
4378 }
4381 \f
4382 #ifndef RANGE_TEST_NON_SHORT_CIRCUIT
4383 #define RANGE_TEST_NON_SHORT_CIRCUIT (BRANCH_COST >= 2)
4384 #endif
4386 /* EXP is some logical combination of boolean tests. See if we can
4387 merge it into some range test. Return the new tree if so. */
4389 static tree
4391 {
4398 tree tem;
4400 /* If this is an OR operation, invert both sides; we will invert
4401 again at the end. */
4405 /* If both expressions are the same, if we can merge the ranges, and we
4406 can build the range test, return it or it inverted. If one of the
4407 ranges is always true or always false, consider it to be the same
4408 expression as the other. */
4418 /* On machines where the branch cost is expensive, if this is a
4419 short-circuited branch and the underlying object on both sides
4420 is the same, make a non-short-circuit operation. */
4426 {
4427 /* If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
4428 unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
4429 which cases we can't do this. */
4438 {
4450 }
4451 }
4454 }
4455 \f
4456 /* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
4457 bit value. Arrange things so the extra bits will be set to zero if and
4458 only if C is signed-extended to its full width. If MASK is nonzero,
4459 it is an INTEGER_CST that should be AND'ed with the extra bits. */
4461 static tree
4463 {
4466 tree temp;
4471 /* We work by getting just the sign bit into the low-order bit, then
4472 into the high-order bit, then sign-extend. We then XOR that value
4473 with C. */
4477 /* We must use a signed type in order to get an arithmetic right shift.
4478 However, we must also avoid introducing accidental overflows, so that
4479 a subsequent call to integer_zerop will work. Hence we must
4480 do the type conversion here. At this point, the constant is either
4481 zero or one, and the conversion to a signed type can never overflow.
4482 We could get an overflow if this conversion is done anywhere else. */
4491 /* If necessary, convert the type back to match the type of C. */
4496 }
4497 \f
4498 /* Find ways of folding logical expressions of LHS and RHS:
4499 Try to merge two comparisons to the same innermost item.
4500 Look for range tests like "ch >= '0' && ch <= '9'".
4501 Look for combinations of simple terms on machines with expensive branches
4502 and evaluate the RHS unconditionally.
4504 For example, if we have p->a == 2 && p->b == 4 and we can make an
4505 object large enough to span both A and B, we can do this with a comparison
4506 against the object ANDed with the a mask.
4508 If we have p->a == q->a && p->b == q->b, we may be able to use bit masking
4509 operations to do this with one comparison.
4511 We check for both normal comparisons and the BIT_AND_EXPRs made this by
4512 function and the one above.
4514 CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR,
4515 TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
4517 TRUTH_TYPE is the type of the logical operand and LHS and RHS are its
4518 two operands.
4520 We return the simplified tree or 0 if no optimization is possible. */
4522 static tree
4524 {
4525 /* If this is the "or" of two comparisons, we can do something if
4526 the comparisons are NE_EXPR. If this is the "and", we can do something
4527 if the comparisons are EQ_EXPR. I.e.,
4528 (a->b == 2 && a->c == 4) can become (a->new == NEW).
4530 WANTED_CODE is this operation code. For single bit fields, we can
4531 convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
4532 comparison for one-bit fields. */
4552 /* Start by getting the comparison codes. Fail if anything is volatile.
4553 If one operand is a BIT_AND_EXPR with the constant one, treat it as if
4554 it were surrounded with a NE_EXPR. */
4563 {
4567 }
4570 {
4574 }
4585 /* Simplify (x<y) && (x==y) into (x<=y) and related optimizations. */
4588 {
4589 tree result;
4592 {
4597 }
4600 {
4606 }
4607 }
4612 /* If the RHS can be evaluated unconditionally and its operands are
4613 simple, it wins to evaluate the RHS unconditionally on machines
4614 with expensive branches. In this case, this isn't a comparison
4615 that can be merged. Avoid doing this if the RHS is a floating-point
4616 comparison since those can trap. */
4622 {
4623 /* Convert (a != 0) || (b != 0) into (a | b) != 0. */
4633 /* Convert (a == 0) && (b == 0) into (a | b) == 0. */
4644 }
4646 /* See if the comparisons can be merged. Then get all the parameters for
4647 each side. */
4671 /* It must be true that the inner operation on the lhs of each
4672 comparison must be the same if we are to be able to do anything.
4673 Then see if we have constants. If not, the same must be true for
4674 the rhs's. */
4685 else
4688 /* If either comparison code is not correct for our logical operation,
4689 fail. However, we can convert a one-bit comparison against zero into
4690 the opposite comparison against that bit being set in the field. */
4694 {
4696 {
4697 /* Make the left operand unsigned, since we are only interested
4698 in the value of one bit. Otherwise we are doing the wrong
4699 thing below. */
4702 }
4703 else
4705 }
4707 /* This is analogous to the code for l_const above. */
4709 {
4711 {
4714 }
4715 else
4717 }
4719 /* After this point all optimizations will generate bit-field
4720 references, which we might not want. */
4724 /* See if we can find a mode that contains both fields being compared on
4725 the left. If we can't, fail. Otherwise, update all constants and masks
4726 to be relative to a field of that size. */
4731 volatilep);
4741 {
4744 }
4752 {
4760 {
4764 }
4765 }
4767 {
4775 {
4779 }
4780 }
4782 /* If the right sides are not constant, do the same for it. Also,
4783 disallow this optimization if a size or signedness mismatch occurs
4784 between the left and right sides. */
4786 {
4789 /* Make sure the two fields on the right
4790 correspond to the left without being swapped. */
4798 volatilep);
4808 {
4811 }
4818 /* Make a mask that corresponds to both fields being compared.
4819 Do this for both items being compared. If the operands are the
4820 same size and the bits being compared are in the same position
4821 then we can do this by masking both and comparing the masked
4822 results. */
4826 {
4838 }
4840 /* There is still another way we can do something: If both pairs of
4841 fields being compared are adjacent, we may be able to make a wider
4842 field containing them both.
4844 Note that we still must mask the lhs/rhs expressions. Furthermore,
4845 the mask must be shifted to account for the shift done by
4846 make_bit_field_ref. */
4851 {
4852 tree type;
4864 /* Convert to the smaller type before masking out unwanted bits. */
4867 {
4869 {
4873 }
4875 {
4879 }
4880 }
4889 }
4892 }
4894 /* Handle the case of comparisons with constants. If there is something in
4895 common between the masks, those bits of the constants must be the same.
4896 If not, the condition is always false. Test for this to avoid generating
4897 incorrect code below. */
4902 {
4904 {
4907 }
4908 else
4909 {
4912 }
4913 }
4915 /* Construct the expression we will return. First get the component
4916 reference we will make. Unless the mask is all ones the width of
4917 that field, perform the mask operation. Then compare with the
4918 merged constant. */
4928 }
4929 \f
4930 /* Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a
4931 constant. */
4933 static tree
4935 {
4940 tree minmax_const;
4942 tree inner;
4952 /* If something does not permit us to optimize, return the original tree. */
4960 /* Now handle all the various comparison codes. We only handle EQ_EXPR
4961 and GT_EXPR, doing the rest with recursive calls using logical
4962 simplifications. */
4964 {
4966 return
4970 return
4972 optimize_minmax_comparison
4974 optimize_minmax_comparison
4979 /* MAX (X, 0) == 0 -> X <= 0 */
4983 /* MAX (X, 0) == 5 -> X == 5 */
4987 /* MAX (X, 0) == -1 -> false */
4991 /* MIN (X, 0) == 0 -> X >= 0 */
4995 /* MIN (X, 0) == 5 -> false */
4998 else
4999 /* MIN (X, 0) == -1 -> X == -1 */
5004 /* MAX (X, 0) > 0 -> X > 0
5005 MAX (X, 0) > 5 -> X > 5 */
5009 /* MAX (X, 0) > -1 -> true */
5013 /* MIN (X, 0) > 0 -> false
5014 MIN (X, 0) > 5 -> false */
5017 else
5018 /* MIN (X, 0) > -1 -> X > -1 */
5023 }
5024 }
5025 \f
5026 /* T is an integer expression that is being multiplied, divided, or taken a
5027 modulus (CODE says which and what kind of divide or modulus) by a
5028 constant C. See if we can eliminate that operation by folding it with
5029 other operations already in T. WIDE_TYPE, if non-null, is a type that
5030 should be used for the computation if wider than our type.
5032 For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return
5033 (X * 2) + (Y * 4). We must, however, be assured that either the original
5034 expression would not overflow or that overflow is undefined for the type
5035 in the language in question.
5037 We also canonicalize (X + 7) * 4 into X * 4 + 28 in the hope that either
5038 the machine has a multiply-accumulate insn or that this is part of an
5039 addressing calculation.
5041 If we return a non-null expression, it is an equivalent form of the
5042 original computation, but need not be in the original type. */
5044 static tree
5046 {
5047 /* To avoid exponential search depth, refuse to allow recursion past
5048 three levels. Beyond that (1) it's highly unlikely that we'll find
5049 something interesting and (2) we've probably processed it before
5050 when we built the inner expression. */
5053 tree ret;
5058 depth++;
5060 depth--;
5063 }
5065 static tree
5067 {
5077 /* Don't deal with constants of zero here; they confuse the code below. */
5087 /* Note that we need not handle conditional operations here since fold
5088 already handles those cases. So just do arithmetic here. */
5090 {
5092 /* For a constant, we can always simplify if we are a multiply
5093 or (for divide and modulus) if it is a multiple of our constant. */
5101 /* If op0 is an expression ... */
5106 /* ... and is unsigned, and its type is smaller than ctype,
5107 then we cannot pass through as widening. */
5113 /* ... or this is a truncation (t is narrower than op0),
5114 then we cannot pass through this narrowing. */
5117 /* ... or signedness changes for division or modulus,
5118 then we cannot pass through this conversion. */
5124 /* Pass the constant down and see if we can make a simplification. If
5125 we can, replace this expression with the inner simplification for
5126 possible later conversion to our or some other type. */
5131 code == MULT_EXPR
5142 /* If widening the type changes the signedness, then we can't perform
5143 this optimization as that changes the result. */
5147 /* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
5150 {
5156 }
5160 /* If the second operand is constant, this is a multiplication
5161 or floor division, by a power of two, so we can treat it that
5162 way unless the multiplier or divisor overflows. Signed
5163 left-shift overflow is implementation-defined rather than
5164 undefined in C90, so do not convert signed left shift into
5165 multiplication. */
5168 /* const_binop may not detect overflow correctly,
5169 so check for it explicitly here. */
5174 size_one_node,
5184 /* See if we can eliminate the operation on both sides. If we can, we
5185 can return a new PLUS or MINUS. If we can't, the only remaining
5186 cases where we can do anything are if the second operand is a
5187 constant. */
5192 /* If not multiplication, we can only do this if both operands
5193 are divisible by c. */
5199 /* If this was a subtraction, negate OP1 and set it to be an addition.
5200 This simplifies the logic below. */
5207 /* If either OP1 or C are negative, this optimization is not safe for
5208 some of the division and remainder types while for others we need
5209 to change the code. */
5211 {
5219 }
5221 /* If it's a multiply or a division/modulus operation of a multiple
5222 of our constant, do the operation and verify it doesn't overflow. */
5225 {
5228 /* We allow the constant to overflow with wrapping semantics. */
5232 }
5233 else
5236 /* If we have an unsigned type is not a sizetype, we cannot widen
5237 the operation since it will change the result if the original
5238 computation overflowed. */
5244 /* If we were able to eliminate our operation from the first side,
5245 apply our operation to the second side and reform the PLUS. */
5249 /* The last case is if we are a multiply. In that case, we can
5250 apply the distributive law to commute the multiply and addition
5251 if the multiplication of the constants doesn't overflow. */
5257 op1));
5262 /* We have a special case here if we are doing something like
5263 (C * 8) % 4 since we know that's zero. */
5270 /* ... fall through ... */
5274 /* If we can extract our operation from the LHS, do so and return a
5275 new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
5276 do something only if the second operand is a constant. */
5288 /* If these are the same operation types, we can associate them
5289 assuming no overflow. */
5296 /* If these operations "cancel" each other, we have the main
5297 optimizations of this pass, which occur when either constant is a
5298 multiple of the other, in which case we replace this with either an
5299 operation or CODE or TCODE.
5301 If we have an unsigned type that is not a sizetype, we cannot do
5302 this since it will change the result if the original computation
5303 overflowed. */
5306 && ! flag_wrapv
5311 {
5322 }
5327 }
5330 }
5331 \f
5332 /* Return a node which has the indicated constant VALUE (either 0 or
5333 1), and is of the indicated TYPE. */
5335 tree
5337 {
5345 else
5347 }
5349 /* Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'.
5350 Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here
5351 CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)'
5352 expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the
5353 COND is the first argument to CODE; otherwise (as in the example
5354 given here), it is the second argument. TYPE is the type of the
5355 original expression. Return NULL_TREE if no simplification is
5356 possible. */
5358 static tree
5361 {
5366 /* This transformation is only worthwhile if we don't have to wrap
5367 arg in a SAVE_EXPR, and the operation can be simplified on atleast
5368 one of the branches once its pushed inside the COND_EXPR. */
5373 {
5377 /* If this operand throws an expression, then it does not make
5378 sense to try to perform a logical or arithmetic operation
5379 involving it. */
5384 }
5385 else
5386 {
5391 }
5402 }
5404 \f
5405 /* Subroutine of fold() that checks for the addition of +/- 0.0.
5407 If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type
5408 TYPE, X + ADDEND is the same as X. If NEGATE, return true if X -
5409 ADDEND is the same as X.
5411 X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero
5412 and finite. The problematic cases are when X is zero, and its mode
5413 has signed zeros. In the case of rounding towards -infinity,
5414 X - 0 is not the same as X because 0 - 0 is -0. In other rounding
5415 modes, X + 0 is not the same as X because -0 + 0 is 0. */
5417 static bool
5419 {
5423 /* Don't allow the fold with -fsignaling-nans. */
5427 /* Allow the fold if zeros aren't signed, or their sign isn't important. */
5431 /* Treat x + -0 as x - 0 and x - -0 as x + 0. */
5436 /* The mode has signed zeros, and we have to honor their sign.
5437 In this situation, there is only one case we can return true for.
5438 X - 0 is the same as X unless rounding towards -infinity is
5439 supported. */
5441 }
5443 /* Subroutine of fold() that checks comparisons of built-in math
5444 functions against real constants.
5446 FCODE is the DECL_FUNCTION_CODE of the built-in, CODE is the comparison
5447 operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE
5448 is the type of the result and ARG0 and ARG1 are the operands of the
5449 comparison. ARG1 must be a TREE_REAL_CST.
5451 The function returns the constant folded tree if a simplification
5452 can be made, and NULL_TREE otherwise. */
5454 static tree
5457 {
5458 REAL_VALUE_TYPE c;
5461 {
5467 {
5468 /* sqrt(x) < y is always false, if y is negative. */
5472 /* sqrt(x) > y is always true, if y is negative and we
5473 don't care about NaNs, i.e. negative values of x. */
5477 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
5480 }
5482 {
5483 REAL_VALUE_TYPE c2;
5489 {
5490 /* sqrt(x) > y is x == +Inf, when y is very large. */
5495 /* sqrt(x) > y is always false, when y is very large
5496 and we don't care about infinities. */
5498 }
5500 /* sqrt(x) > c is the same as x > c*c. */
5503 }
5505 {
5506 REAL_VALUE_TYPE c2;
5512 {
5513 /* sqrt(x) < y is always true, when y is a very large
5514 value and we don't care about NaNs or Infinities. */
5518 /* sqrt(x) < y is x != +Inf when y is very large and we
5519 don't care about NaNs. */
5524 /* sqrt(x) < y is x >= 0 when y is very large and we
5525 don't care about Infinities. */
5530 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
5539 dconst0))),
5542 c2)))));
5543 }
5545 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
5550 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
5553 {
5558 dconst0))),
5561 c2)))));
5562 }
5563 }
5564 }
5567 }
5569 /* Subroutine of fold() that optimizes comparisons against Infinities,
5570 either +Inf or -Inf.
5572 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5573 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5574 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5576 The function returns the constant folded tree if a simplification
5577 can be made, and NULL_TREE otherwise. */
5579 static tree
5581 {
5583 REAL_VALUE_TYPE max;
5584 tree temp;
5589 /* For negative infinity swap the sense of the comparison. */
5595 {
5597 /* x > +Inf is always false, if with ignore sNANs. */
5603 /* x <= +Inf is always true, if we don't case about NaNs. */
5607 /* x <= +Inf is the same as x == x, i.e. isfinite(x). */
5610 {
5613 }
5618 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
5624 /* x < +Inf is always equal to x <= DBL_MAX. */
5630 /* x != +Inf is always equal to !(x > DBL_MAX). */
5636 /* The transformation below creates non-gimple code and thus is
5637 not appropriate if we are in gimple form. */
5647 }
5650 }
5652 /* Subroutine of fold() that optimizes comparisons of a division by
5653 a nonzero integer constant against an integer constant, i.e.
5654 X/C1 op C2.
5656 CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR,
5657 GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1
5658 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
5660 The function returns the constant folded tree if a simplification
5661 can be made, and NULL_TREE otherwise. */
5663 static tree
5665 {
5670 HOST_WIDE_INT hpart;
5673 /* We have to do this the hard way to detect unsigned overflow.
5674 prod = int_const_binop (MULT_EXPR, arg01, arg1, 0); */
5683 {
5687 /* Likewise hi = int_const_binop (PLUS_EXPR, prod, tmp, 0). */
5696 }
5698 {
5701 {
5719 }
5720 }
5721 else
5722 {
5725 {
5743 }
5744 }
5747 {
5788 }
5791 }
5794 /* If CODE with arguments ARG0 and ARG1 represents a single bit
5795 equality/inequality test, then return a simplified form of
5796 the test using shifts and logical operations. Otherwise return
5797 NULL. TYPE is the desired result type. */
5799 tree
5801 tree result_type)
5802 {
5803 /* If this is a TRUTH_NOT_EXPR, it may have a single bit test inside
5804 operand 0. */
5806 {
5811 /* Extract the arguments of the EQ/NE. */
5815 /* This requires us to invert the code. */
5817 }
5819 /* If this is testing a single bit, we can optimize the test. */
5823 {
5830 tree arg00;
5832 /* If we have (A & C) != 0 where C is the sign bit of A, convert
5833 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
5836 /* This is only a win if casting to a signed type is cheap,
5837 i.e. when arg00's type is not a partial mode. */
5840 {
5845 }
5847 /* Otherwise we have (A & C) != 0 where C is a single bit,
5848 convert that into ((A >> C2) & 1). Where C2 = log2(C).
5849 Similarly for (A & C) == 0. */
5851 /* If INNER is a right shift of a constant and it plus BITNUM does
5852 not overflow, adjust BITNUM and INNER. */
5859 {
5862 }
5864 /* If we are going to be able to omit the AND below, we must do our
5865 operations as unsigned. If we must use the AND, we have a choice.
5866 Normally unsigned is faster, but for some machines signed is. */
5867 #ifdef LOAD_EXTEND_OP
5869 #else
5871 #endif
5886 /* Put the AND last so it can combine with more things. */
5890 /* Make sure to return the proper type. */
5894 }
5896 }
5898 /* Check whether we are allowed to reorder operands arg0 and arg1,
5899 such that the evaluation of arg1 occurs before arg0. */
5901 static bool
5903 {
5910 }
5912 /* Test whether it is preferable two swap two operands, ARG0 and
5913 ARG1, for example because ARG0 is an integer constant and ARG1
5914 isn't. If REORDER is true, only recommend swapping if we can
5915 evaluate the operands in reverse order. */
5917 bool
5919 {
5955 /* It is preferable to swap two SSA_NAME to ensure a canonical form
5956 for commutative and comparison operators. Ensuring a canonical
5957 form allows the optimizers to find additional redundancies without
5958 having to explicitly check for both orderings. */
5965 }
5967 /* Perform constant folding and related simplification of EXPR.
5968 The related simplifications include x*1 => x, x*0 => 0, etc.,
5969 and application of the associative law.
5970 NOP_EXPR conversions may be removed freely (as long as we
5971 are careful not to change the type of the overall expression).
5972 We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR,
5973 but we can constant-fold them if they have constant operands. */
5975 #ifdef ENABLE_FOLD_CHECKING
5976 # define fold(x) fold_1 (x)
5978 static
5979 #endif
5980 tree
5982 {
5986 tree tem;
5991 /* WINS will be nonzero when the switch is done
5992 if all operands are constant. */
5995 /* Return right away if a constant. */
6000 {
6001 tree subop;
6003 /* Special case for conversion ops that can have fixed point args. */
6006 /* Don't use STRIP_NOPS, because signedness of argument type matters. */
6012 else
6017 /* Note that TREE_CONSTANT isn't enough:
6018 static var addresses are constant but we can't
6019 do arithmetic on them. */
6021 }
6023 {
6027 {
6029 tree subop;
6034 /* Strip any conversions that don't change the mode. This is
6035 safe for every expression, except for a comparison expression
6036 because its signedness is derived from its operands. So, in
6037 the latter case, only strip conversions that don't change the
6038 signedness.
6040 Note that this is done as an internal manipulation within the
6041 constant folder, in order to find the simplest representation
6042 of the arguments so that their form can be studied. In any
6043 cases, the appropriate type conversions should be put back in
6044 the tree that will get out of the constant folder. */
6047 else
6052 else
6057 /* Note that TREE_CONSTANT isn't enough:
6058 static var addresses are constant but we can't
6059 do arithmetic on them. */
6066 }
6067 }
6069 /* If this is a commutative operation, and ARG0 is a constant, move it
6070 to ARG1 to reduce the number of tests below. */
6076 /* Now WINS is set as described above,
6077 ARG0 is the first operand of EXPR,
6078 and ARG1 is the second operand (if it has more than one operand).
6080 First check for cases where an arithmetic operation is applied to a
6081 compound, conditional, or comparison operation. Push the arithmetic
6082 operation inside the compound or conditional to see if any folding
6083 can then be done. Convert comparison to conditional for this purpose.
6084 The also optimizes non-constant cases that used to be done in
6085 expand_expr.
6087 Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
6088 one of the operands is a comparison and the other is a comparison, a
6089 BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
6090 code below would make the expression more complex. Change it to a
6091 TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
6092 TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. */
6104 {
6115 }
6118 {
6123 {
6133 /* If this was a conversion, and all we did was to move into
6134 inside the COND_EXPR, bring it back out. But leave it if
6135 it is a conversion from integer to integer and the
6136 result precision is no wider than a word since such a
6137 conversion is cheap and may be optimized away by combine,
6138 while it couldn't if it were outside the COND_EXPR. Then return
6139 so we don't get into an infinite recursion loop taking the
6140 conversion out and then back in. */
6152 && (INTEGRAL_TYPE_P
6163 }
6165 {
6167 {
6171 }
6175 integer_one_node)),
6177 integer_zero_node))));
6178 }
6179 }
6190 {
6194 arg1)));
6202 {
6207 }
6210 {
6215 }
6216 }
6219 {
6233 /* Handle cases of two conversions in a row. */
6236 {
6255 /* In addition to the cases of two conversions in a row
6256 handled below, if we are converting something to its own
6257 type via an object of identical or wider precision, neither
6258 conversion is needed. */
6265 /* Likewise, if the intermediate and final types are either both
6266 float or both integer, we don't need the middle conversion if
6267 it is wider than the final type and doesn't change the signedness
6268 (for integers). Avoid this if the final type is a pointer
6269 since then we sometimes need the inner conversion. Likewise if
6270 the outer has a precision not equal to the size of its mode. */
6281 /* If we have a sign-extension of a zero-extended value, we can
6282 replace that by a single zero-extension. */
6289 /* Two conversions in a row are not needed unless:
6290 - some conversion is floating-point (overstrict for now), or
6291 - the intermediate type is narrower than both initial and
6292 final, or
6293 - the intermediate type and innermost type differ in signedness,
6294 and the outermost type is wider than the intermediate, or
6295 - the initial type is a pointer type and the precisions of the
6296 intermediate and final types differ, or
6297 - the final type is a pointer type and the precisions of the
6298 initial and intermediate types differ. */
6313 }
6317 /* Detect assigning a bitfield. */
6320 {
6321 /* Don't leave an assignment inside a conversion
6322 unless assigning a bitfield. */
6326 /* First do the assignment, then return converted constant. */
6331 }
6333 /* Convert (T)(x & c) into (T)x & (T)c, if c is an integer
6334 constants (if x has signed type, the sign bit cannot be set
6335 in c). This folds extension into the BIT_AND_EXPR. */
6340 {
6350 <= HOST_BITS_PER_WIDE_INT
6352 {
6359 #ifdef LOAD_EXTEND_OP
6363 {
6367 }
6368 #endif
6369 }
6374 }
6376 /* Convert (T1)((T2)X op Y) into (T1)X op Y, for pointer types T1 and
6377 T2 being pointers to types of the same size. */
6382 {
6394 }
6408 {
6412 }
6417 {
6422 }
6435 /* Convert fabs((double)float) into (double)fabsf(float). */
6438 {
6443 targ0)));
6444 }
6477 /* A + (-B) -> A - B */
6480 /* (-A) + B -> B - A */
6485 {
6489 /* If we are adding two BIT_AND_EXPR's, both of which are and'ing
6490 with a constant, and the two constants have no bits in common,
6491 we should treat this as a BIT_IOR_EXPR since this may produce more
6492 simplifications. */
6500 {
6503 }
6505 /* Reassociate (plus (plus (mult) (foo)) (mult)) as
6506 (plus (plus (mult) (mult)) (foo)) so that we can
6507 take advantage of the factoring cases below. */
6512 {
6517 else
6538 }
6541 {
6545 /* (A * C) + (B * C) -> (A+B) * C.
6546 We are most concerned about the case where C is a constant,
6547 but other combinations show up during loop reduction. Since
6548 it is not difficult, try all four possibilities. */
6565 /* No identical multiplicands; see if we can find a common
6566 power-of-two factor in non-power-of-two multiplies. This
6567 can help in multi-dimensional array access. */
6572 {
6577 /* Move min of absolute values to int11. */
6580 {
6584 }
6587 {
6593 }
6594 }
6600 same));
6601 }
6602 }
6603 else
6604 {
6605 /* See if ARG1 is zero and X + ARG1 reduces to X. */
6609 /* Likewise if the operands are reversed. */
6613 /* Convert X + -C into X - C. */
6616 {
6622 }
6624 /* Convert x+x into x*2.0. */
6630 /* Convert x*c+x into x*(c+1). */
6636 {
6637 REAL_VALUE_TYPE c;
6643 }
6645 /* Convert x+x*c into x*(c+1). */
6651 {
6652 REAL_VALUE_TYPE c;
6658 }
6660 /* Convert x*c1+x*c2 into x*(c1+c2). */
6670 {
6679 }
6680 /* Convert a + (b*c + d*e) into (a + b*c) + d*e. */
6684 {
6689 {
6690 tree tree0;
6693 }
6694 }
6695 /* Convert (b*c + d*e) + a into b*c + (d*e +a). */
6699 {
6704 {
6705 tree tree0;
6708 }
6709 }
6710 }
6712 bit_rotate:
6713 /* (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
6714 is a rotate of A by C1 bits. */
6715 /* (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
6716 is a rotate of A by B bits. */
6717 {
6726 {
6745 {
6753 TYPE_PRECISION
6758 ? LROTATE_EXPR
6761 }
6763 {
6771 TYPE_PRECISION
6776 ? LROTATE_EXPR
6779 }
6780 }
6781 }
6783 associate:
6784 /* In most languages, can't associate operations on floats through
6785 parentheses. Rather than remember where the parentheses were, we
6786 don't associate floats at all, unless the user has specified
6787 -funsafe-math-optimizations. */
6791 {
6795 /* Split both trees into variables, constants, and literals. Then
6796 associate each group together, the constants with literals,
6797 then the result with variables. This increases the chances of
6798 literals being recombined later and of generating relocatable
6799 expressions for the sum of a constant and literal. */
6804 /* Only do something if we found more than two objects. Otherwise,
6805 nothing has changed and we risk infinite recursion. */
6810 {
6811 /* Recombine MINUS_EXPR operands by using PLUS_EXPR. */
6820 /* Preserve the MINUS_EXPR if the negative part of the literal is
6821 greater than the positive part. Otherwise, the multiplicative
6822 folding code (i.e extract_muldiv) may be fooled in case
6823 unsigned constants are subtracted, like in the following
6824 example: ((X*2 + 4) - 8U)/2. */
6826 {
6830 {
6834 }
6835 else
6836 {
6840 }
6841 }
6843 {
6848 else
6849 {
6855 }
6856 }
6861 }
6862 }
6864 binary:
6868 {
6869 /* The return value should always have
6870 the same type as the original expression. */
6875 }
6879 /* A - (-B) -> A + B */
6882 /* (-A) - B -> (-B) - A where B is easily negated and we can swap. */
6892 {
6898 /* Fold A - (A & B) into ~B & A. */
6901 {
6906 arg0));
6911 arg0));
6912 }
6914 /* Fold (A & ~B) - (A & B) into (A ^ B) - B, where B is
6915 any power of 2 minus 1. */
6920 {
6926 {
6930 }
6931 }
6932 }
6934 /* See if ARG1 is zero and X - ARG1 reduces to X. */
6938 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
6939 ARG0 is zero and X + ARG0 reduces to X, since that would mean
6940 (-ARG1 + ARG0) reduces to -ARG1. */
6944 /* Fold &x - &x. This can happen from &x.foo - &x.
6945 This is unsafe for certain floats even in non-IEEE formats.
6946 In IEEE, it is unsafe because it does wrong for NaNs.
6947 Also note that operand_equal_p is always false if an operand
6948 is volatile. */
6954 /* A - B -> A + (-B) if B is easily negatable. */
6957 /* Avoid this transformation if B is a positive REAL_CST. */
6963 /* Try folding difference of addresses. */
6964 {
6965 HOST_WIDE_INT diff;
6973 }
6978 {
6979 /* (A * C) - (B * C) -> (A-B) * C. */
6987 /* (A * C1) - (A * C2) -> A * (C1-C2). */
6995 }
7000 /* (-A) * (-B) -> A * B */
7011 {
7017 /* (a * (1 << b)) is (a << b) */
7033 }
7034 else
7035 {
7036 /* Maybe fold x * 0 to 0. The expressions aren't the same
7037 when x is NaN, since x * 0 is also NaN. Nor are they the
7038 same in modes with signed zeros, since multiplying a
7039 negative value by 0 gives -0, not +0. */
7044 /* In IEEE floating point, x*1 is not equivalent to x for snans. */
7049 /* Transform x * -1.0 into -x. */
7054 /* Convert (C1/X)*C2 into (C1*C2)/X. */
7059 {
7065 }
7068 {
7072 /* Optimizations of root(...)*root(...). */
7074 {
7079 /* Optimize sqrt(x)*sqrt(x) as x. */
7085 /* Optimize root(x)*root(y) as root(x*y). */
7090 }
7092 /* Optimize expN(x)*expN(y) as expN(x+y). */
7094 {
7101 }
7103 /* Optimizations of pow(...)*pow(...). */
7107 {
7115 /* Optimize pow(x,y)*pow(z,y) as pow(x*z,y). */
7117 {
7122 arg01));
7124 }
7126 /* Optimize pow(x,y)*pow(x,z) as pow(x,y+z). */
7128 {
7133 arg));
7135 }
7136 }
7138 /* Optimize tan(x)*cos(x) as sin(x). */
7147 {
7153 }
7155 /* Optimize x*pow(x,c) as pow(x,c+1). */
7159 {
7166 {
7168 REAL_VALUE_TYPE c;
7177 }
7178 }
7180 /* Optimize pow(x,c)*x as pow(x,c+1). */
7184 {
7191 {
7193 REAL_VALUE_TYPE c;
7202 }
7203 }
7205 /* Optimize x*x as pow(x,2.0), which is expanded as x*x. */
7208 {
7212 {
7217 }
7218 }
7219 }
7220 }
7224 bit_ior:
7232 /* ~X | X is -1. */
7235 {
7239 }
7241 /* X | ~X is -1. */
7244 {
7248 }
7254 /* Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
7256 This results in more efficient code for machines without a NAND
7257 instruction. Combine will canonicalize to the first form
7258 which will allow use of NAND instructions provided by the
7259 backend if they exist. */
7262 {
7267 }
7269 /* See if this can be simplified into a rotate first. If that
7270 is unsuccessful continue in the association code. */
7281 /* ~X ^ X is -1. */
7284 {
7288 }
7290 /* X ^ ~X is -1. */
7293 {
7297 }
7299 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
7300 with a constant, and the two constants have no bits in common,
7301 we should treat this as a BIT_IOR_EXPR since this may produce more
7302 simplifications. */
7310 {
7313 }
7315 /* See if this can be simplified into a rotate first. If that
7316 is unsuccessful continue in the association code. */
7327 /* ~X & X is always zero. */
7332 /* X & ~X is always zero. */
7340 /* Simplify ((int)c & 0377) into (int)c, if c is unsigned char. */
7343 {
7344 unsigned int prec
7351 }
7353 /* Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
7355 This results in more efficient code for machines without a NOR
7356 instruction. Combine will canonicalize to the first form
7357 which will allow use of NOR instructions provided by the
7358 backend if they exist. */
7361 {
7366 }
7371 /* Don't touch a floating-point divide by zero unless the mode
7372 of the constant can represent infinity. */
7378 /* (-A) / (-B) -> A / B */
7388 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
7393 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
7398 /* If ARG1 is a constant, we can convert this to a multiply by the
7399 reciprocal. This does not have the same rounding properties,
7400 so only do this if -funsafe-math-optimizations. We can actually
7401 always safely do it if ARG1 is a power of two, but it's hard to
7402 tell if it is or not in a portable manner. */
7404 {
7409 /* Find the reciprocal if optimizing and the result is exact. */
7411 {
7412 REAL_VALUE_TYPE r;
7415 {
7418 }
7419 }
7420 }
7421 /* Convert A/B/C to A/(B*C). */
7428 /* Convert A/(B/C) to (A/B)*C. */
7436 /* Convert C1/(X*C2) into (C1/C2)/X. */
7441 {
7447 }
7450 {
7452 /* Optimize x/expN(y) into x*expN(-y). */
7454 {
7461 }
7463 /* Optimize x/pow(y,z) into x*pow(y,-z). */
7467 {
7476 }
7477 }
7480 {
7484 /* Optimize sin(x)/cos(x) as tan(x). */
7490 {
7496 }
7498 /* Optimize cos(x)/sin(x) as 1.0/tan(x). */
7504 {
7508 {
7513 }
7514 }
7516 /* Optimize pow(x,c)/x as pow(x,c-1). */
7520 {
7526 {
7528 REAL_VALUE_TYPE c;
7537 }
7538 }
7539 }
7551 /* X / -1 is -X. */
7558 /* If arg0 is a multiple of arg1, then rewrite to the fastest div
7559 operation, EXACT_DIV_EXPR.
7561 Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
7562 At one time others generated faster code, it's not clear if they do
7563 after the last round to changes to the DIV code in expmed.c. */
7584 /* X % -1 is zero. */
7591 /* Optimize unsigned TRUNC_MOD_EXPR by a power of two into a
7592 BIT_AND_EXPR, i.e. "X % C" into "X & C2". */
7596 {
7598 tree mask;
7603 {
7607 }
7608 else
7609 {
7612 }
7617 }
7619 /* X % -C is the same as X % C. */
7624 && !flag_trapv
7625 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
7630 /* X % -Y is the same as X % Y. */
7652 /* Optimize -1 >> x for arithmetic right shifts. */
7655 /* ... fall through ... */
7658 shift:
7664 /* Since negative shift count is not well-defined,
7665 don't try to compute it in the compiler. */
7668 /* Rewrite an LROTATE_EXPR by a constant into an
7669 RROTATE_EXPR by a new constant. */
7671 {
7677 }
7679 /* If we have a rotate of a bit operation with the rotate count and
7680 the second operand of the bit operation both constant,
7681 permute the two operations. */
7693 /* Two consecutive rotates adding up to the width of the mode can
7694 be ignored. */
7725 /* The argument to invert_truthvalue must have Boolean type. */
7729 /* Note that the operand of this must be an int
7730 and its values must be 0 or 1.
7731 ("true" is a fixed value perhaps depending on the language,
7732 but we don't handle values other than 1 correctly yet.) */
7734 /* Avoid infinite recursion. */
7736 {
7741 }
7745 /* Note that the operands of this must be ints
7746 and their values must be 0 or 1.
7747 ("true" is a fixed value perhaps depending on the language.) */
7748 /* If first arg is constant zero, return it. */
7752 /* If either arg is constant true, drop it. */
7756 /* Preserve sequence points. */
7759 /* If second arg is constant zero, result is zero, but first arg
7760 must be evaluated. */
7763 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
7764 case will be handled here. */
7768 /* !X && X is always false. */
7772 /* X && !X is always false. */
7777 truth_andor:
7778 /* We only do these simplifications if we are optimizing. */
7782 /* Check for things like (A || B) && (A || C). We can convert this
7783 to A || (B && C). Note that either operator can be any of the four
7784 truth and/or operations and the transformation will still be
7785 valid. Also note that we only care about order for the
7786 ANDIF and ORIF operators. If B contains side effects, this
7787 might change the truth-value of A. */
7794 {
7814 /* This case if tricky because we must either have commutative
7815 operators or else A10 must not have side-effects. */
7821 a01));
7822 }
7824 /* See if we can build a range comparison. */
7828 /* Check for the possibility of merging component references. If our
7829 lhs is another similar operation, try to merge its rhs with our
7830 rhs. Then try to merge our lhs and rhs. */
7842 /* Note that the operands of this must be ints
7843 and their values must be 0 or true.
7844 ("true" is a fixed value perhaps depending on the language.) */
7845 /* If first arg is constant true, return it. */
7849 /* If either arg is constant zero, drop it. */
7853 /* Preserve sequence points. */
7856 /* If second arg is constant true, result is true, but we must
7857 evaluate first arg. */
7860 /* Likewise for first arg, but note this only occurs here for
7861 TRUTH_OR_EXPR. */
7865 /* !X || X is always true. */
7869 /* X || !X is always true. */
7877 /* If the second arg is constant zero, drop it. */
7880 /* If the second arg is constant true, this is a logical inversion. */
7883 /* Identical arguments cancel to zero. */
7887 /* !X ^ X is always true. */
7892 /* X ^ !X is always true. */
7905 /* If one arg is a real or integer constant, put it last. */
7909 /* If this is an equality comparison of the address of a non-weak
7910 object against zero, then we know the result. */
7918 /* If this is an equality comparison of the address of two non-weak,
7919 unaliased symbols neither of which are extern (since we do not
7920 have access to attributes for externs), then we know the result. */
7936 type);
7939 {
7947 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
7952 /* (-a) CMP (-b) -> b CMP a */
7959 {
7960 REAL_VALUE_TYPE cst;
7963 /* (-a) CMP CST -> a swap(CMP) (-CST) */
7965 return
7971 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
7972 /* a CMP (-0) -> a CMP 0 */
7977 /* x != NaN is always true, other ops are always false. */
7980 {
7983 }
7985 /* Fold comparisons against infinity. */
7987 {
7991 }
7992 }
7994 /* If this is a comparison of a real constant with a PLUS_EXPR
7995 or a MINUS_EXPR of a real constant, we can convert it into a
7996 comparison with a revised real constant as long as no overflow
7997 occurs when unsafe_math_optimizations are enabled. */
8009 /* Likewise, we can simplify a comparison of a real constant with
8010 a MINUS_EXPR whose first operand is also a real constant, i.e.
8011 (c1 - x) < c2 becomes x > c1-c2. */
8022 /* Fold comparisons against built-in math functions. */
8024 && flag_unsafe_math_optimizations
8026 {
8030 {
8034 }
8035 }
8036 }
8038 /* Convert foo++ == CONST into ++foo == CONST + INCR. */
8042 /* This optimization is invalid for ordered comparisons
8043 if CONST+INCR overflows or if foo+incr might overflow.
8044 This optimization is invalid for floating point due to rounding.
8045 For pointer types we assume overflow doesn't happen. */
8049 {
8053 {
8059 }
8060 else
8061 {
8067 }
8070 /* If VAROP is a reference to a bitfield, we must mask
8071 the constant by the width of the field. */
8076 {
8081 /* First check whether the comparison would come out
8082 always the same. If we don't do that we would
8083 change the meaning with the masking. */
8097 }
8100 }
8102 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
8103 This transformation affects the cases which are handled in later
8104 optimizations involving comparisons with non-negative constants. */
8108 {
8110 {
8121 }
8122 }
8124 /* Comparisons with the highest or lowest possible integer of
8125 the specified size will have known values.
8127 This is quite similar to fold_relational_hi_lo; however, my
8128 attempts to share the code have been nothing but trouble.
8129 I give up for now. */
8130 {
8138 {
8145 {
8148 }
8149 else
8150 {
8153 }
8158 {
8171 /* The GE_EXPR and LT_EXPR cases above are not normally
8172 reached because of previous transformations. */
8176 }
8180 {
8189 }
8193 {
8208 }
8212 {
8221 }
8227 /* signed_type does not work on pointer types. */
8229 {
8230 /* The following case also applies to X < signed_max+1
8231 and X >= signed_max+1 because previous transformations. */
8233 {
8237 return fold
8241 }
8242 }
8243 }
8244 }
8246 /* If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
8247 a MINUS_EXPR of a constant, we can convert it into a comparison with
8248 a revised constant as long as no overflow occurs. */
8260 /* Similarly for a NEGATE_EXPR. */
8269 /* If we have X - Y == 0, we can convert that to X == Y and similarly
8270 for !=. Don't do this for ordered comparisons due to overflow. */
8276 /* If we are widening one operand of an integer comparison,
8277 see if the other operand is similarly being widened. Perhaps we
8278 can do the comparison in the narrower type. */
8293 /* If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a
8294 constant, we can simplify it. */
8301 /* If we are comparing an ABS_EXPR with a constant, we can
8302 convert all the cases into explicit comparisons, but they may
8303 well not be faster than doing the ABS and one comparison.
8304 But ABS (X) <= C is a range comparison, which becomes a subtraction
8305 and a comparison, and is probably faster. */
8318 /* If this is an EQ or NE comparison with zero and ARG0 is
8319 (1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
8320 two operations, but the latter can be done in one less insn
8321 on machines that have only two-operand insns or on which a
8322 constant cannot be the first operand. */
8325 {
8330 return
8336 integer_one_node)),
8337 arg1));
8340 return
8346 integer_one_node)),
8347 arg1));
8348 }
8350 /* If this is an NE or EQ comparison of zero against the result of a
8351 signed MOD operation whose second operand is a power of 2, make
8352 the MOD operation unsigned since it is simpler and equivalent. */
8361 {
8371 }
8373 /* If this is an NE comparison of zero with an AND of one, remove the
8374 comparison since the AND will give the correct value. */
8380 /* If we have (A & C) == C where C is a power of 2, convert this into
8381 (A & C) != 0. Similarly for NE_EXPR. */
8388 integer_zero_node)));
8390 /* If we have (A & C) != 0 or (A & C) == 0 and C is a power of
8391 2, then fold the expression into shifts and logical operations. */
8396 /* If we have (A & C) == D where D & ~C != 0, convert this into 0.
8397 Similarly for NE_EXPR. */
8402 {
8411 }
8413 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
8414 Similarly for NE_EXPR. */
8419 {
8426 }
8428 /* If X is unsigned, convert X < (1 << Y) into X >> Y == 0
8429 and similarly for >= into !=. */
8445 return
8453 /* Simplify comparison of something with itself. (For IEEE
8454 floating-point, we can only do some of these simplifications.) */
8456 {
8458 {
8473 /* For NE, we can only do this simplification if integer
8474 or we don't honor IEEE floating point NaNs. */
8478 /* ... fall through ... */
8484 }
8485 }
8487 /* If we are comparing an expression that just has comparisons
8488 of two integer values, arithmetic expressions of those comparisons,
8489 and constants, we can simplify it. There are only three cases
8490 to check: the two values can either be equal, the first can be
8491 greater, or the second can be greater. Fold the expression for
8492 those three values. Since each value must be 0 or 1, we have
8493 eight possibilities, each of which corresponds to the constant 0
8494 or 1 or one of the six possible comparisons.
8496 This handles common cases like (a > b) == 0 but also handles
8497 expressions like ((x > y) - (y > x)) > 0, which supposedly
8498 occur in macroized code. */
8501 {
8506 /* Don't handle degenerate cases here; they should already
8507 have been handled anyway. */
8516 {
8520 /* We can't just pass T to eval_subst in case cval1 or cval2
8521 was the same as ARG1. */
8523 tree high_result
8527 arg1));
8528 tree equal_result
8532 arg1));
8533 tree low_result
8537 arg1));
8539 /* All three of these results should be 0 or 1. Confirm they
8540 are. Then use those values to select the proper code
8541 to use. */
8549 {
8550 /* Make a 3-bit mask with the high-order bit being the
8551 value for `>', the next for '=', and the low for '<'. */
8555 {
8557 /* Always false. */
8578 /* Always true. */
8580 }
8585 else
8587 }
8588 }
8589 }
8591 /* If this is a comparison of a field, we may be able to simplify it. */
8596 /* Handle the constant case even without -O
8597 to make sure the warnings are given. */
8599 {
8603 }
8605 /* If this is a comparison of complex values and either or both sides
8606 are a COMPLEX_EXPR or COMPLEX_CST, it is best to split up the
8607 comparisons and join them with a TRUTH_ANDIF_EXPR or TRUTH_ORIF_EXPR.
8608 This may prevent needless evaluations. */
8615 {
8628 type,
8631 }
8633 /* Optimize comparisons of strlen vs zero to a compare of the
8634 first character of the string vs zero. To wit,
8635 strlen(ptr) == 0 => *ptr == 0
8636 strlen(ptr) != 0 => *ptr != 0
8637 Other cases should reduce to one of these two (or a constant)
8638 due to the return value of strlen being unsigned. */
8642 {
8644 tree arglist;
8657 integer_zero_node)));
8658 }
8660 /* We can fold X/C1 op C2 where C1 and C2 are integer constants
8661 into a single range test. */
8668 {
8672 }
8692 {
8696 }
8698 /* If the first operand is NaN, the result is constant. */
8702 {
8704 ? integer_zero_node
8707 }
8709 /* If the second operand is NaN, the result is constant. */
8713 {
8715 ? integer_zero_node
8718 }
8720 /* Simplify unordered comparison of something with itself. */
8726 && !flag_trapping_math
8730 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
8731 {
8742 }
8747 /* Pedantic ANSI C says that a conditional expression is never an lvalue,
8748 so all simple results must be passed through pedantic_non_lvalue. */
8750 {
8752 /* Only optimize constant conditions when the selected branch
8753 has the same type as the COND_EXPR. This avoids optimizing
8754 away "c ? x : throw", where the throw has a void type. */
8759 }
8763 /* If we have A op B ? A : C, we may be able to convert this to a
8764 simpler expression, depending on the operation and the values
8765 of B and C. Signed zeros prevent all of these transformations,
8766 for reasons given above each one.
8768 Also try swapping the arguments and inverting the conditional. */
8773 {
8779 }
8786 {
8789 {
8795 }
8796 }
8798 /* If the second operand is simpler than the third, swap them
8799 since that produces better jump optimization results. */
8802 {
8803 /* See if this can be inverted. If it can't, possibly because
8804 it was a floating-point inequality comparison, don't do
8805 anything. */
8811 }
8813 /* Convert A ? 1 : 0 to simply A. */
8816 /* If we try to convert TREE_OPERAND (t, 0) to our type, the
8817 call to fold will try to move the conversion inside
8818 a COND, which will recurse. In that case, the COND_EXPR
8819 is probably the best choice, so leave it alone. */
8823 /* Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
8824 over COND_EXPR in cases such as floating point comparisons. */
8831 /* A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>). */
8839 /* (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
8840 already handled above. */
8845 {
8854 }
8856 /* A & N ? N : 0 is simply A & N if N is a power of two. This
8857 is probably obsolete because the first operand should be a
8858 truth value (that's why we have the two cases above), but let's
8859 leave it in until we can confirm this for all front-ends. */
8870 /* Convert A ? B : 0 into A && B if A and B are truth values. */
8876 /* Convert A ? B : 1 into !A || B if A and B are truth values. */
8880 {
8881 /* Only perform transformation if ARG0 is easily inverted. */
8885 }
8887 /* Convert A ? 0 : B into !A && B if A and B are truth values. */
8891 {
8892 /* Only perform transformation if ARG0 is easily inverted. */
8897 }
8899 /* Convert A ? 1 : B into A || B if A and B are truth values. */
8909 /* When pedantic, a compound expression can be neither an lvalue
8910 nor an integer constant expression. */
8913 /* Don't let (0, 0) be null pointer constant. */
8956 /* Check for a built-in function. */
8961 {
8965 }
8971 }
8973 #ifdef ENABLE_FOLD_CHECKING
8974 #undef fold
8980 /* When --enable-checking=fold, compute a digest of expr before
8981 and after actual fold call to see if fold did not accidentally
8982 change original expr. */
8984 tree
8986 {
8987 tree ret;
8990 htab_t ht;
9009 }
9011 void
9013 {
9016 htab_t ht;
9026 }
9028 static void
9030 {
9032 }
9034 static void
9036 {
9054 {
9055 /* Allow DECL_ASSEMBLER_NAME to be modified. */
9059 }
9063 {
9064 /* Allow these fields to be modified. */
9071 }
9078 {
9081 {
9095 }
9099 {
9110 }
9144 {
9147 }
9157 }
9158 }
9160 #endif
9162 /* Perform constant folding and related simplification of initializer
9163 expression EXPR. This behaves identically to "fold" but ignores
9164 potential run-time traps and exceptions that fold must preserve. */
9166 tree
9168 {
9172 tree result;
9185 }
9187 /* Determine if first argument is a multiple of second argument. Return 0 if
9188 it is not, or we cannot easily determined it to be.
9190 An example of the sort of thing we care about (at this point; this routine
9191 could surely be made more general, and expanded to do what the *_DIV_EXPR's
9192 fold cases do now) is discovering that
9194 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9196 is a multiple of
9198 SAVE_EXPR (J * 8)
9200 when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
9202 This code also handles discovering that
9204 SAVE_EXPR (I) * SAVE_EXPR (J * 8)
9206 is a multiple of 8 so we don't have to worry about dealing with a
9207 possible remainder.
9209 Note that we *look* inside a SAVE_EXPR only to determine how it was
9210 calculated; it is not safe for fold to do much of anything else with the
9211 internals of a SAVE_EXPR, since it cannot know when it will be evaluated
9212 at run time. For example, the latter example above *cannot* be implemented
9213 as SAVE_EXPR (I) * J or any variant thereof, since the value of J at
9214 evaluation time of the original SAVE_EXPR is not necessarily the same at
9215 the time the new expression is evaluated. The only optimization of this
9216 sort that would be valid is changing
9218 SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
9220 divided by 8 to
9222 SAVE_EXPR (I) * SAVE_EXPR (J)
9224 (where the same SAVE_EXPR (J) is used in the original and the
9225 transformed version). */
9227 static int
9229 {
9237 {
9249 {
9253 /* const_binop may not detect overflow correctly,
9254 so check for it explicitly here. */
9260 size_one_node,
9264 }
9268 /* Can't handle conversions from non-integral or wider integral type. */
9274 /* .. fall through ... */
9290 }
9291 }
9293 /* Return true if `t' is known to be non-negative. */
9295 int
9297 {
9299 {
9314 /* zero_extend(x) + zero_extend(y) is non-negative if x and y are
9315 both unsigned and at least 2 bits shorter than the result. */
9319 {
9324 {
9328 }
9329 }
9334 {
9335 /* x * x for floating point x is always non-negative. */
9340 }
9342 /* zero_extend(x) * zero_extend(y) is non-negative if x and y are
9343 both unsigned and their total bits is shorter than the result. */
9347 {
9354 }
9383 {
9388 {
9392 {
9396 }
9397 }
9399 {
9405 }
9406 }
9432 {
9436 /* If the initializer is non-void, then it's a normal expression
9437 that will be assigned to the slot. */
9441 /* Otherwise, the initializer sets the slot in some way. One common
9442 way is an assignment statement at the end of the initializer. */
9444 {
9452 else
9454 }
9460 }
9463 {
9470 {
9471 #define CASE_BUILTIN_F(BUILT_IN_FN) \
9472 case BUILT_IN_FN: case BUILT_IN_FN##F: case BUILT_IN_FN##L:
9473 #define CASE_BUILTIN_I(BUILT_IN_FN) \
9474 case BUILT_IN_FN: case BUILT_IN_FN##L: case BUILT_IN_FN##LL:
9492 /* Always true. */
9496 /* sqrt(-0.0) is -0.0. */
9524 /* True if the 1st argument is nonnegative. */
9528 /* True if the 1st OR 2nd arguments are nonnegative. */
9533 /* True if the 1st AND 2nd arguments are nonnegative. */
9538 /* True if the 2nd argument is nonnegative. */
9543 #undef CASE_BUILTIN_F
9544 #undef CASE_BUILTIN_I
9545 }
9546 }
9548 /* ... fall through ... */
9552 /* Truth values evaluate to 0 or 1, which is nonnegative. */
9554 }
9556 /* We don't know sign of `t', so be conservative and return false. */
9558 }
9560 /* Return true when T is an address and is known to be nonzero.
9561 For floating point we further ensure that T is not denormal.
9562 Similar logic is present in nonzero_address in rtlanal.h. */
9564 static bool
9566 {
9569 /* Doing something useful for floating point would need more work. */
9574 {
9580 /* We used to test for !integer_zerop here. This does not work correctly
9581 if TREE_CONSTANT_OVERFLOW (t). */
9587 {
9588 /* With the presence of negative values it is hard
9589 to say something. */
9593 /* One of operands must be positive and the other non-negative. */
9596 }
9601 {
9604 }
9608 {
9614 }
9618 {
9624 /* Weak declarations may link to NULL. */
9628 /* Constants are never weak. */
9633 }
9645 {
9646 /* When both operands are nonzero, then MAX must be too. */
9650 /* MAX where operand 0 is positive is positive. */
9652 }
9653 /* MAX where operand 1 is positive is positive. */
9674 }
9676 }
9678 /* See if we are applying CODE, a relational to the highest or lowest
9679 possible integer of TYPE. If so, then the result is a compile
9680 time constant. */
9682 static tree
9685 {
9696 {
9703 {
9706 }
9707 else
9708 {
9711 }
9716 {
9730 /* The GE_EXPR and LT_EXPR cases above are not normally
9731 reached because of previous transformations. */
9735 }
9739 {
9750 }
9754 {
9771 }
9775 {
9786 }
9791 /* signed_type does not work on pointer types. */
9793 {
9794 /* The following case also applies to X < signed_max+1
9795 and X >= signed_max+1 because previous transformations. */
9797 {
9803 type,
9807 retval
9813 /* If we are in gimple form, then returning EXP would create
9814 non-gimple expressions. Clearing it is safe and insures
9815 we do not allow a non-gimple expression to escape. */
9820 }
9821 }
9822 }
9825 }
9828 /* Given the components of a binary expression CODE, TYPE, OP0 and OP1,
9829 attempt to fold the expression to a constant without modifying TYPE,
9830 OP0 or OP1.
9832 If the expression could be simplified to a constant, then return
9833 the constant. If the expression would not be simplified to a
9834 constant, then return NULL_TREE.
9836 Note this is primarily designed to be called after gimplification
9837 of the tree structures and when at least one operand is a constant.
9838 As a result of those simplifying assumptions this routine is far
9839 simpler than the generic fold routine. */
9841 tree
9844 {
9846 tree subop0;
9847 tree subop1;
9848 tree tem;
9850 /* If this is a commutative operation, and ARG0 is a constant, move it
9851 to ARG1 to reduce the number of tests below. */
9854 {
9858 }
9860 /* If either operand is a complex type, extract its real component. */
9863 else
9868 else
9871 /* Note if either argument is not a real or integer constant.
9872 With a few exceptions, simplification is limited to cases
9873 where both arguments are constants. */
9881 {
9883 /* (plus (address) (const_int)) is a constant. */
9891 {
9895 }
9898 binary:
9902 /* Both arguments are constants. Simplify. */
9905 {
9906 /* The return value should always have the same type as
9907 the original expression. */
9912 }
9916 /* Fold &x - &x. This can happen from &x.foo - &x.
9917 This is unsafe for certain floats even in non-IEEE formats.
9918 In IEEE, it is unsafe because it does wrong for NaNs.
9919 Also note that operand_equal_p is always false if an
9920 operand is volatile. */
9928 /* Special case multiplication or bitwise AND where one argument
9929 is zero. */
9932 else
9941 /* Special case when we know the result will be all ones. */
9957 /* Division by zero is undefined. */
9984 /* Optimize -1 >> x for arithmetic right shifts. */
9987 /* ... fall through ... */
9993 /* Since negative shift count is not well-defined, don't
9994 try to compute it in the compiler. */
10002 /* -1 rotated either direction by any amount is still -1. */
10006 /* 0 rotated either direction by any amount is still zero. */
10023 /* If one arg is a real or integer constant, put it last. */
10028 {
10029 tree temp;
10035 }
10037 /* Change X >= C to X > (C - 1) and X < C to X <= (C - 1) if C > 0.
10038 This transformation affects the cases which are handled in later
10039 optimizations involving comparisons with non-negative constants. */
10043 {
10045 {
10058 }
10059 }
10065 /* Fall through. */
10081 /* This could probably be handled. */
10085 /* If second arg is constant zero, result is zero, but first arg
10086 must be evaluated. */
10089 /* Likewise for first arg, but note that only the TRUTH_AND_EXPR
10090 case will be handled here. */
10098 /* If second arg is constant true, result is true, but we must
10099 evaluate first arg. */
10102 /* Likewise for first arg, but note this only occurs here for
10103 TRUTH_OR_EXPR. */
10112 {
10115 }
10120 }
10121 }
10123 /* Given the components of a unary expression CODE, TYPE and OP0,
10124 attempt to fold the expression to a constant without modifying
10125 TYPE or OP0.
10127 If the expression could be simplified to a constant, then return
10128 the constant. If the expression would not be simplified to a
10129 constant, then return NULL_TREE.
10131 Note this is primarily designed to be called after gimplification
10132 of the tree structures and when op0 is a constant. As a result
10133 of those simplifying assumptions this routine is far simpler than
10134 the generic fold routine. */
10136 tree
10138 tree op0)
10139 {
10140 /* Make sure we have a suitable constant argument. */
10142 {
10143 tree subop;
10147 else
10152 }
10155 {
10167 else
10173 else
10179 else
10185 else
10191 else
10203 }
10204 }
10206 /* If EXP represents referencing an element in a constant string
10207 (either via pointer arithmetic or array indexing), return the
10208 tree representing the value accessed, otherwise return NULL. */
10210 tree
10212 {
10214 {
10216 tree index;
10217 tree string;
10221 else
10222 {
10226 /* Optimize the special-case of a zero lower bound.
10228 We convert the low_bound to sizetype to avoid some problems
10229 with constant folding. (E.g. suppose the lower bound is 1,
10230 and its mode is QI. Without the conversion,l (ARRAY
10231 +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1))
10232 +INDEX), which becomes (ARRAY+255+INDEX). Opps!) */
10237 }
10251 }
10253 }
10255 /* Return the tree for neg (ARG0) when ARG0 is known to be either
10256 an integer constant or real constant.
10258 TYPE is the type of the result. */
10260 static tree
10262 {
10266 {
10268 {
10270 HOST_WIDE_INT high;
10280 }
10288 }
10291 }
10293 /* Return the tree for abs (ARG0) when ARG0 is known to be either
10294 an integer constant or real constant.
10296 TYPE is the type of the result. */
10298 tree
10300 {
10304 {
10306 /* If the value is unsigned, then the absolute value is
10307 the same as the ordinary value. */
10310 /* Similarly, if the value is non-negative. */
10313 /* If the value is negative, then the absolute value is
10314 its negation. */
10315 else
10316 {
10318 HOST_WIDE_INT high;
10325 }
10331 else
10337 }
10340 }
10342 /* Return the tree for not (ARG0) when ARG0 is known to be an integer
10343 constant. TYPE is the type of the result. */
10345 static tree
10347 {
10359 }
10361 /* Given CODE, a relational operator, the target type, TYPE and two
10362 constant operands OP0 and OP1, return the result of the
10363 relational operation. If the result is not a compile time
10364 constant, then return NULL_TREE. */
10366 static tree
10368 {
10371 /* From here on, the only cases we handle are when the result is
10372 known to be a constant. */
10375 {
10379 /* Handle the cases where either operand is a NaN. */
10381 {
10383 {
10411 }
10414 }
10417 }
10419 /* From here on we only handle LT, LE, GT, GE, EQ and NE.
10421 To compute GT, swap the arguments and do LT.
10422 To compute GE, do LT and invert the result.
10423 To compute LE, swap the arguments, do LT and invert the result.
10424 To compute NE, do EQ and invert the result.
10426 Therefore, the code below must handle only EQ and LT. */
10429 {
10434 }
10436 /* Note that it is safe to invert for real values here because we
10437 have already handled the one case that it matters. */
10441 {
10444 }
10446 /* Compute a result for LT or EQ if args permit;
10447 Otherwise return T. */
10449 {
10454 else
10456 }
10457 else
10463 }
10465 /* Build an expression for the address of T. Folds away INDIRECT_REF to
10466 avoid confusing the gimplify process. */
10468 tree
10470 {
10471 /* The size of the object is not relevant when talking about its address. */
10475 /* Note: doesn't apply to ALIGN_INDIRECT_REF */
10478 {
10482 }
10483 else
10484 {
10495 }
10498 }
10500 tree
10502 {
10504 }
10506 /* Builds an expression for an indirection through T, simplifying some
10507 cases. */
10509 tree
10511 {
10514 tree subtype;
10518 {
10521 /* *&p => p */
10524 /* *(foo *)&fooarray => fooarray[0] */
10528 }
10530 /* *(foo *)fooarrptr => (*fooarrptr)[0] */
10534 {
10537 }
10540 }
10542 /* Strip non-trapping, non-side-effecting tree nodes from an expression
10543 whose result is ignored. The type of the returned tree need not be
10544 the same as the original expression. */
10546 tree
10548 {
10554 {
10565 else
10571 {
10587 }
10592 }
10593 }
10595 /* Return the value of VALUE, rounded up to a multiple of DIVISOR.
10596 This can only be applied to objects of a sizetype. */
10598 tree
10600 {
10607 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10608 have to do anything. Only do this when we are not given a const,
10609 because in that case, this check is more expensive than just
10610 doing it. */
10612 {
10617 }
10619 /* If divisor is a power of two, simplify this to bit manipulation. */
10621 {
10622 tree t;
10628 }
10629 else
10630 {
10635 }
10638 }
10640 /* Likewise, but round down. */
10642 tree
10644 {
10651 /* See if VALUE is already a multiple of DIVISOR. If so, we don't
10652 have to do anything. Only do this when we are not given a const,
10653 because in that case, this check is more expensive than just
10654 doing it. */
10656 {
10661 }
10663 /* If divisor is a power of two, simplify this to bit manipulation. */
10665 {
10666 tree t;
10670 }
10671 else
10672 {
10677 }
10680 }
10682 /* Returns true if addresses of E1 and E2 differ by a constant, false
10683 otherwise. If they do, &E1 - &E2 is stored in *DIFF. */
10685 bool
10687 {
10706 {
10716 }
10718 {
10719 /* If only one of the offsets is non-constant, the difference cannot
10720 be a constant. */
10722 }
10723 else
10728 }