On Mon, Oct 14, 2024 at 08:53:29AM +0200, Jakub Jelinek wrote: > > PR middle-end/116891 > > * match.pd ((negate (IFN_FNMS@3 @0 @1 @2)) -> (IFN_FMA @0 @1 @2)): > > Only enable for !HONOR_SIGN_DEPENDENT_ROUNDING. > > Guess it would be nice to have a testcase which FAILs without the patch and > PASSes with it, but it can be added later.
I've added such a testcase now, and additionally found the fix only fixed one of the 4 problematic similar cases. Here is a patch which fixes the others too and adds the testcases. fma-pr116891.c FAILed without your patch, FAILs with your patch too (but only due to the bar/baz/qux checks) and PASSes with the patch. Ok for trunk if it passes full bootstrap/regtest? 2024-10-15 Jakub Jelinek <ja...@redhat.com> PR middle-end/116891 * match.pd ((negate (fmas@3 @0 @1 @2)) -> (IFN_FNMS @0 @1 @2)): Only enable for !HONOR_SIGN_DEPENDENT_ROUNDING. ((negate (IFN_FMS@3 @0 @1 @2)) -> (IFN_FNMA @0 @1 @2)): Likewise. ((negate (IFN_FNMA@3 @0 @1 @2)) -> (IFN_FMS @0 @1 @2)): Likewise. * gcc.dg/pr116891.c: New test. * gcc.target/i386/fma-pr116891.c: New test. --- gcc/match.pd.jj 2024-10-15 12:50:47.699905473 +0200 +++ gcc/match.pd 2024-10-15 12:57:19.547400416 +0200 @@ -9452,7 +9452,7 @@ (define_operator_list SYNC_FETCH_AND_AND (IFN_FNMS @0 @1 @2)) (simplify (negate (fmas@3 @0 @1 @2)) - (if (single_use (@3)) + (if (!HONOR_SIGN_DEPENDENT_ROUNDING (type) && single_use (@3)) (IFN_FNMS @0 @1 @2)))) (simplify @@ -9466,7 +9466,7 @@ (define_operator_list SYNC_FETCH_AND_AND (IFN_FNMA @0 @1 @2)) (simplify (negate (IFN_FMS@3 @0 @1 @2)) - (if (single_use (@3)) + (if (!HONOR_SIGN_DEPENDENT_ROUNDING (type) && single_use (@3)) (IFN_FNMA @0 @1 @2))) (simplify @@ -9480,7 +9480,7 @@ (define_operator_list SYNC_FETCH_AND_AND (IFN_FMS @0 @1 @2)) (simplify (negate (IFN_FNMA@3 @0 @1 @2)) - (if (single_use (@3)) + (if (!HONOR_SIGN_DEPENDENT_ROUNDING (type) && single_use (@3)) (IFN_FMS @0 @1 @2))) (simplify --- gcc/testsuite/gcc.dg/pr116891.c.jj 2024-10-15 12:31:57.619723244 +0200 +++ gcc/testsuite/gcc.dg/pr116891.c 2024-10-15 12:56:37.809987933 +0200 @@ -0,0 +1,47 @@ +/* PR middle-end/116891 */ +/* { dg-do run } */ +/* { dg-require-effective-target fenv } */ +/* { dg-require-effective-target hard_float } */ +/* { dg-require-effective-target c99_runtime } */ +/* { dg-options "-O2 -frounding-math" } */ + +#include <fenv.h> + +__attribute__((noipa)) double +foo (double x, double y, double z) +{ + return -__builtin_fma (-x, y, -z); +} + +__attribute__((noipa)) double +bar (double x, double y, double z) +{ + return -__builtin_fma (-x, y, z); +} + +__attribute__((noipa)) double +baz (double x, double y, double z) +{ + return -__builtin_fma (x, y, -z); +} + +__attribute__((noipa)) double +qux (double x, double y, double z) +{ + return -__builtin_fma (x, y, z); +} + +int +main () +{ +#if defined (FE_DOWNWARD) && __DBL_MANT_DIG__ == 53 && __DBL_MAX_EXP__ == 1024 + fesetround (FE_DOWNWARD); + double a = foo (-0x1.p256, 0x1.p256, 0x1.p-256); + if (a != -__builtin_nextafter (0x1p256 * 0x1p256, 0.)) + __builtin_abort (); + if (a != bar (-0x1.p256, 0x1.p256, -0x1.p-256) + || a != baz (0x1.p256, 0x1.p256, 0x1.p-256) + || a != qux (0x1.p256, 0x1.p256, -0x1.p-256)) + __builtin_abort (); +#endif +} --- gcc/testsuite/gcc.target/i386/fma-pr116891.c.jj 2024-10-15 12:42:39.711719596 +0200 +++ gcc/testsuite/gcc.target/i386/fma-pr116891.c 2024-10-15 12:44:56.692806834 +0200 @@ -0,0 +1,19 @@ +/* PR middle-end/116891 */ +/* { dg-do run } */ +/* { dg-require-effective-target fenv } */ +/* { dg-require-effective-target hard_float } */ +/* { dg-require-effective-target c99_runtime } */ +/* { dg-require-effective-target fma } */ +/* { dg-options "-O2 -mfma -frounding-math" } */ + +#include <fenv.h> +#include "fma-check.h" + +#define main() do_main () +#include "../../gcc.dg/pr116891.c" + +static void +fma_test (void) +{ + do_main (); +} Jakub