| Message ID | 20141001121715.GR1986@tucnak.redhat.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Oct 1, 2014 at 2:17 PM, Jakub Jelinek <jakub@redhat.com> wrote: > On Wed, Oct 01, 2014 at 01:45:54PM +0200, Uros Bizjak wrote: >> OK. > > Thanks. Second step is a tiny optimization, for the > simplified 122 (now 24) vshuf-v4di.c testcase: > typedef unsigned long long V __attribute__ ((vector_size (32))); > V a, b, c, d; > > int > main () > { > int i; > for (i = 0; i < 4; ++i) > { > a[i] = i + 2; > b[i] = 4 + i + 2; > } > asm volatile ("" : : : "memory"); > c = __builtin_shuffle (a, b, (V) { 2, 5, 6, 3 }); > d = __builtin_shuffle ((V) { 2, 3, 4, 5 }, (V) { 6, 7, 8, 9 }, (V) { 2, 5, 6, 3 }); > if (__builtin_memcmp (&c, &d, sizeof (c))) > __builtin_abort (); > return 0; > } > > this patch allows better code to be generated: > - vmovdqa b(%rip), %ymm0 > + vpermq $238, a(%rip), %ymm1 > movl $32, %edx > - movl $d, %esi > - vmovdqa a(%rip), %ymm1 > + vmovdqa b(%rip), %ymm0 > + movl $d, %esi > movl $c, %edi > - vperm2i128 $17, %ymm0, %ymm1, %ymm1 > vpblendd $195, %ymm1, %ymm0, %ymm0 > vmovdqa %ymm0, c(%rip) > > That is because vperm2i128 $17 unnecessarily uses > two operands when all the data it grabs are from a single one. > So, by canonicalizing the permutation we can emit > vpermq $238 instead. Perhaps more places might benefit from > extra canonicalize_perm calls (two spots already use that beyond > the single one on the expansion/testing entry point). > > Tested again with > GCC_TEST_RUN_EXPENSIVE=1 make check-gcc \ > RUNTESTFLAGS='--target_board=unix/-mavx2 dg-torture.exp=vshuf*.c' > on x86_64-linux. Ok for trunk? > > 2014-10-01 Jakub Jelinek <jakub@redhat.com> > > * config/i386/i386.c (expand_vec_perm_vperm2f128): Canonicalize > dfirst permutation. OK. Thanks, Uros.
--- gcc/config/i386/i386.c.jj 2014-10-01 13:00:30.000000000 +0200 +++ gcc/config/i386/i386.c 2014-10-01 13:59:40.061956852 +0200 @@ -43905,15 +43905,16 @@ expand_vec_perm_vperm2f128 (struct expan dfirst.perm[i] = (i & (nelt2 - 1)) + ((perm >> (2 * (i >= nelt2))) & 3) * nelt2; + canonicalize_perm (&dfirst); ok = expand_vec_perm_1 (&dfirst); gcc_assert (ok); /* And dsecond is some single insn shuffle, taking d->op0 and result of vperm2f128 (if perm < 16) or d->op1 and result of vperm2f128 (otherwise). */ - dsecond.op1 = dfirst.target; if (perm >= 16) - dsecond.op0 = dfirst.op1; + dsecond.op0 = dsecond.op1; + dsecond.op1 = dfirst.target; ok = expand_vec_perm_1 (&dsecond); gcc_assert (ok);
On Wed, Oct 01, 2014 at 01:45:54PM +0200, Uros Bizjak wrote: > OK. Thanks. Second step is a tiny optimization, for the simplified 122 (now 24) vshuf-v4di.c testcase: typedef unsigned long long V __attribute__ ((vector_size (32))); V a, b, c, d; int main () { int i; for (i = 0; i < 4; ++i) { a[i] = i + 2; b[i] = 4 + i + 2; } asm volatile ("" : : : "memory"); c = __builtin_shuffle (a, b, (V) { 2, 5, 6, 3 }); d = __builtin_shuffle ((V) { 2, 3, 4, 5 }, (V) { 6, 7, 8, 9 }, (V) { 2, 5, 6, 3 }); if (__builtin_memcmp (&c, &d, sizeof (c))) __builtin_abort (); return 0; } this patch allows better code to be generated: - vmovdqa b(%rip), %ymm0 + vpermq $238, a(%rip), %ymm1 movl $32, %edx - movl $d, %esi - vmovdqa a(%rip), %ymm1 + vmovdqa b(%rip), %ymm0 + movl $d, %esi movl $c, %edi - vperm2i128 $17, %ymm0, %ymm1, %ymm1 vpblendd $195, %ymm1, %ymm0, %ymm0 vmovdqa %ymm0, c(%rip) That is because vperm2i128 $17 unnecessarily uses two operands when all the data it grabs are from a single one. So, by canonicalizing the permutation we can emit vpermq $238 instead. Perhaps more places might benefit from extra canonicalize_perm calls (two spots already use that beyond the single one on the expansion/testing entry point). Tested again with GCC_TEST_RUN_EXPENSIVE=1 make check-gcc \ RUNTESTFLAGS='--target_board=unix/-mavx2 dg-torture.exp=vshuf*.c' on x86_64-linux. Ok for trunk? 2014-10-01 Jakub Jelinek <jakub@redhat.com> * config/i386/i386.c (expand_vec_perm_vperm2f128): Canonicalize dfirst permutation. Jakub