On Tue, 5 May 2020, Jakub Jelinek wrote:
> Hi!
>
> On x86 (the only target with umulv4_optab) one can use mull; seto to check
> for overflow instead of performing wider multiplication and performing
> comparison on the high bits.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Hmm. The preceeding pattern doesn't check for umulv4 availability, so
why do it here? I suppose an alternative simplification would be
to use a highpart multiply? Do you intentionally not consider
((type)A * CST) >> prec?
Thanks,
Richard.
> 2020-05-05 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/94914
> * match.pd ((((type)A * B) >> prec) != 0 to .MUL_OVERFLOW(A, B) != 0):
> New simplification.
>
> * gcc.target/i386/pr94914.c: New test.
>
> --- gcc/match.pd.jj 2020-05-04 11:02:14.288865592 +0200
> +++ gcc/match.pd 2020-05-04 12:15:33.220799388 +0200
> @@ -4776,6 +4776,27 @@ (define_operator_list COND_TERNARY
> (with { tree t = TREE_TYPE (@0), cpx = build_complex_type (t); }
> (out (imagpart (IFN_MUL_OVERFLOW:cpx @0 @1)) { build_zero_cst (t); })))))
>
> +/* Similarly, for unsigned operands, (((type) A * B) >> prec) != 0 where type
> + is at least twice as wide as type of A and B, simplify to
> + __builtin_mul_overflow (A, B, <unused>). */
> +(for cmp (eq ne)
> + (simplify
> + (cmp (rshift (mult:s (convert@3 @0) (convert @1)) INTEGER_CST@2)
> + integer_zerop)
> + (if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
> + && INTEGRAL_TYPE_P (TREE_TYPE (@3))
> + && TYPE_UNSIGNED (TREE_TYPE (@0))
> + && (TYPE_PRECISION (TREE_TYPE (@3))
> + >= 2 * TYPE_PRECISION (TREE_TYPE (@0)))
> + && tree_fits_uhwi_p (@2)
> + && tree_to_uhwi (@2) == TYPE_PRECISION (TREE_TYPE (@0))
> + && types_match (@0, @1)
> + && type_has_mode_precision_p (TREE_TYPE (@0))
> + && (optab_handler (umulv4_optab, TYPE_MODE (TREE_TYPE (@0)))
> + != CODE_FOR_nothing))
> + (with { tree t = TREE_TYPE (@0), cpx = build_complex_type (t); }
> + (cmp (imagpart (IFN_MUL_OVERFLOW:cpx @0 @1)) { build_zero_cst (t); })))))
> +
> /* Simplification of math builtins. These rules must all be optimizations
> as well as IL simplifications. If there is a possibility that the new
> form could be a pessimization, the rule should go in the canonicalization
> --- gcc/testsuite/gcc.target/i386/pr94914.c.jj 2020-05-04
> 12:58:04.435775670 +0200
> +++ gcc/testsuite/gcc.target/i386/pr94914.c 2020-05-04 12:57:39.307152707
> +0200
> @@ -0,0 +1,17 @@
> +/* PR tree-optimization/94914 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +/* { dg-final { scan-assembler "\tseto\t" } } */
> +/* { dg-final { scan-assembler "\tsetno\t" } } */
> +
> +int
> +foo (unsigned int x, unsigned int y)
> +{
> + return (((unsigned long long)x * y) >> 32) != 0;
> +}
> +
> +int
> +bar (unsigned int x, unsigned int y)
> +{
> + return (((unsigned long long)x * y) >> 32) == 0;
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)
