| Message ID | CAFULd4a2Y4QPfhAPOpF7dudkZR7U6qDmtBCuRSNUw_pSE+seaA@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
> Attached patch fixes PR 59021, where new vzeroupper instructions are > generated for -mavx after Vlad's useless insn removal patch. > > The problem was, that we depent on (useless) moves to AVX256 function > argument registers in front of the function call to switch the mode to > DIRTY mode. This is not true anymore, so call_insn has to switch to > DIRTY mode by itself. Thanks for fixing this! > The patch will be backported to 4.7 and 4.8 branch in a couple of days. The 4.7 branch uses a different mechanism so I'm not sure that's needed.
On Wed, Nov 6, 2013 at 11:06 PM, Eric Botcazou <ebotcazou@adacore.com> wrote: >> Attached patch fixes PR 59021, where new vzeroupper instructions are >> generated for -mavx after Vlad's useless insn removal patch. >> >> The problem was, that we depent on (useless) moves to AVX256 function >> argument registers in front of the function call to switch the mode to >> DIRTY mode. This is not true anymore, so call_insn has to switch to >> DIRTY mode by itself. > > Thanks for fixing this! > >> The patch will be backported to 4.7 and 4.8 branch in a couple of days. > > The 4.7 branch uses a different mechanism so I'm not sure that's needed. Oh, I remembered this fact when the patch didn't apply to 4.7 branch... the patch is not needed on 4.7 branch. Thanks, Uros.
Index: config/i386/i386.c =================================================================== --- config/i386/i386.c (revision 204459) +++ config/i386/i386.c (working copy) @@ -15708,7 +15708,7 @@ ix86_avx_u128_mode_needed (rtx insn) rtx arg = XEXP (XEXP (link, 0), 0); if (ix86_check_avx256_register (&arg, NULL)) - return AVX_U128_ANY; + return AVX_U128_DIRTY; } } @@ -15828,8 +15828,8 @@ ix86_avx_u128_mode_after (int mode, rtx insn) { bool avx_reg256_found = false; note_stores (pat, ix86_check_avx256_stores, &avx_reg256_found); - if (!avx_reg256_found) - return AVX_U128_CLEAN; + + return avx_reg256_found ? AVX_U128_DIRTY : AVX_U128_CLEAN; } /* Otherwise, return current mode. Remember that if insn Index: testsuite/gcc.target/i386/pr59021.c =================================================================== --- testsuite/gcc.target/i386/pr59021.c (revision 0) +++ testsuite/gcc.target/i386/pr59021.c (working copy) @@ -0,0 +1,23 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -mavx -mvzeroupper" } */ + +extern void abort (void); + +struct S { + int i1; + int i2; + int i3; +}; + +typedef double v4df __attribute__ ((vector_size (32))); + +extern int foo (v4df, int i1, int i2, int i3, int i4, int i5, struct S s); + +void bar (v4df v, struct S s) +{ + int r = foo (v, 1, 2, 3, 4, 5, s); + if (r) + abort (); +} + +/* { dg-final { scan-assembler-not "vzeroupper" } } */
Hello! Attached patch fixes PR 59021, where new vzeroupper instructions are generated for -mavx after Vlad's useless insn removal patch. The problem was, that we depent on (useless) moves to AVX256 function argument registers in front of the function call to switch the mode to DIRTY mode. This is not true anymore, so call_insn has to switch to DIRTY mode by itself. While looking at this area - always set the mode after the call_insn in MODE_AFTER function. 2013-11-06 Uros Bizjak <ubizjak@gmail.com> PR target/59021 * config/i386/i386.c (ix86_avx_u128_mode_needed): Require AVX_U128_DIRTY mode for call_insn RTXes that use AVX256 registers. (ix86_avx_u128_mode_needed): Return AVX_U128_DIRTY mode for call_insn RTXes that return in AVX256 register. testsuite/ChangeLog: 2013-11-06 Uros Bizjak <ubizjak@gmail.com> PR target/59021 * gcc.target/i386/pr59021.c: New test. Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}, configured with "--with-arch=core-avx-i --with-cpu=core-avx-i --with-fpmath=avx". The patch will be backported to 4.7 and 4.8 branch in a couple of days. Uros.