Hi!
As I've tried to explain in the comments, the extract_muldiv_1
MIN/MAX_EXPR optimization is wrong for code == MULT_EXPR.
If the multiplication is done in unsigned type or in signed
type with -fwrapv, it is fairly obvious that max (a, b) * c
in many cases isn't equivalent to max (a * c, b * c) (or min if c is
negative) due to overflows, but even for signed with undefined overflow,
the optimization could turn something without UB in it (where
say a * c invokes UB, but max (or min) picks the other operand where
b * c doesn't).
As for division/modulo, I think it is in most cases safe, except if
the problematic INT_MIN / -1 case could be triggered, but we can
just punt for MAX_EXPR because for MIN_EXPR if one operand is INT_MIN,
we'd pick that operand already. It is just for completeness, match.pd
already has an optimization which turns x / -1 into -x, so the division
by zero is mostly theoretical. That is also why in the testcase the
i case isn't actually miscompiled without the patch, while the c and f
cases are.
Bootstrapped/regtested on x86_64-linux and i686-linux, additionally
bootstrapped/regtested on x86_64-linux with statistics gathering when
the patch changes behavior and it is solely on the new testcase and
nothing else during the bootstrap/regtest. Ok for trunk?
2024-03-26 Jakub Jelinek <ja...@redhat.com>
PR middle-end/111151
* fold-const.cc (extract_muldiv_1) <case MAX_EXPR>: Punt for
MULT_EXPR altogether, or for MAX_EXPR if c is -1.
* gcc.c-torture/execute/pr111151.c: New test.
--- gcc/fold-const.cc.jj 2024-03-11 09:42:04.544588951 +0100
+++ gcc/fold-const.cc 2024-03-25 11:48:12.133625285 +0100
@@ -7104,6 +7104,27 @@ extract_muldiv_1 (tree t, tree c, enum t
if (TYPE_UNSIGNED (ctype) != TYPE_UNSIGNED (type))
break;
+ /* Punt for multiplication altogether.
+ MAX (1U + INT_MAX, 1U) * 2U is not equivalent to
+ MAX ((1U + INT_MAX) * 2U, 1U * 2U), the former is
+ 0U, the latter is 2U.
+ MAX (INT_MIN / 2, 0) * -2 is not equivalent to
+ MIN (INT_MIN / 2 * -2, 0 * -2), the former is
+ well defined 0, the latter invokes UB.
+ MAX (INT_MIN / 2, 5) * 5 is not equivalent to
+ MAX (INT_MIN / 2 * 5, 5 * 5), the former is
+ well defined 25, the latter invokes UB. */
+ if (code == MULT_EXPR)
+ break;
+ /* For division/modulo, punt on c being -1 for MAX, as
+ MAX (INT_MIN, 0) / -1 is not equivalent to
+ MIN (INT_MIN / -1, 0 / -1), the former is well defined
+ 0, the latter invokes UB (or for -fwrapv is INT_MIN).
+ MIN (INT_MIN, 0) / -1 already invokes UB, so the
+ transformation won't make it worse. */
+ else if (tcode == MAX_EXPR && integer_minus_onep (c))
+ break;
+
/* MIN (a, b) / 5 -> MIN (a / 5, b / 5) */
sub_strict_overflow_p = false;
if ((t1 = extract_muldiv (op0, c, code, wide_type,
--- gcc/testsuite/gcc.c-torture/execute/pr111151.c.jj 2024-03-25
11:50:27.199744988 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr111151.c 2024-03-26
10:41:51.003384032 +0100
@@ -0,0 +1,21 @@
+/* PR middle-end/111151 */
+
+int
+main ()
+{
+ unsigned a = (1U + __INT_MAX__) / 2U;
+ unsigned b = 1U;
+ unsigned c = (a * 2U > b * 2U ? a * 2U : b * 2U) * 2U;
+ if (c != 0U)
+ __builtin_abort ();
+ int d = (-__INT_MAX__ - 1) / 2;
+ int e = 10;
+ int f = (d * 2 > e * 5 ? d * 2 : e * 5) * 6;
+ if (f != 120)
+ __builtin_abort ();
+ int g = (-__INT_MAX__ - 1) / 2;
+ int h = 0;
+ int i = (g * 2 > h * 5 ? g * 2 : h * 5) / -1;
+ if (i != 0)
+ __builtin_abort ();
+}
Jakub
