| Message ID | CAFULd4Z2Se+KmiUWf9=DsiDeQ0ST1BXSNkw70BD-TWrx7A0BXQ@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [AVX512] : Fix cvtusi2<ssescalarmodesuffix>64<round_name> insn mnemonic | expand |
On 05/23/2018 04:08 AM, Uros Bizjak wrote: > Hello! > > With current insn mnemonic and ATT assembler dialect, there is no way > for the assembler to distinguish between DImode and SImode instruction > when memory input operand is used. The dump for 32bit memory reads as: > > 0: 62 f1 7e 08 7b 05 00 vcvtusi2ss 0x0(%rip),%xmm0,%xmm0 > 7: 00 00 00 > 10: 62 f1 7f 08 7b 05 00 vcvtusi2sd 0x0(%rip),%xmm0,%xmm0 > 17: 00 00 00 > > And for 64bit access: > > 20: 62 f1 fe 08 7b 05 00 vcvtusi2ss 0x0(%rip),%xmm0,%xmm0 > 27: 00 00 00 > 30: 62 f1 ff 08 7b 05 00 vcvtusi2sd 0x0(%rip),%xmm0,%xmm0 > 37: 00 00 00 > > (Note the difference in the 3rd byte. On a related note, binutils > should also be fixed to dump vcvtsi2sdq in the 64bit case to avoid > ambiguity) > > We should use "q" suffix with the ATT dialect in the Dimode insn mnemonic. > > 2018-05-23 Uros Bizjak <ubizjak@gmail.com> > > * config/i386/sse.md (cvtusi2<ssescalarmodesuffix>64<round_name>): > Add {q} suffix to insn mnemonic. > > testsuite/Changelog: > > 2018-05-23 Uros Bizjak <ubizjak@gmail.com> > > * gcc.target/i386/avx512f-vcvtusi2sd64-1.c: Update scan string. > * gcc.target/i386/avx512f-vcvtusi2ss64-1.c: Ditto. > > Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}. > > OK for mainline and backports? OK. jeff
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md index 8a80fa35067..30411b15493 100644 --- a/gcc/config/i386/sse.md +++ b/gcc/config/i386/sse.md @@ -4463,7 +4463,7 @@ (match_operand:VF_128 1 "register_operand" "v") (const_int 1)))] "TARGET_AVX512F && TARGET_64BIT" - "vcvtusi2<ssescalarmodesuffix>\t{%2, <round_op3>%1, %0|%0, %1<round_op3>, %2}" + "vcvtusi2<ssescalarmodesuffix>{q}\t{%2, <round_op3>%1, %0|%0, %1<round_op3>, %2}" [(set_attr "type" "sseicvt") (set_attr "prefix" "evex") (set_attr "mode" "<ssescalarmode>")]) diff --git a/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2sd64-1.c b/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2sd64-1.c index 8675450f0c4..66476c3013f 100644 --- a/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2sd64-1.c +++ b/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2sd64-1.c @@ -1,7 +1,7 @@ /* { dg-do compile { target { ! ia32 } } } */ /* { dg-options "-mavx512f -O2" } */ -/* { dg-final { scan-assembler-times "vcvtusi2sd\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ -/* { dg-final { scan-assembler-times "vcvtusi2sd\[ \\t\]+\[^%\n\]*%r\[^\{\n\]*\{ru-sae\}\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ +/* { dg-final { scan-assembler-times "vcvtusi2sdq\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ +/* { dg-final { scan-assembler-times "vcvtusi2sdq\[ \\t\]+\[^%\n\]*%r\[^\{\n\]*\{ru-sae\}\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ #include <immintrin.h> diff --git a/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2ss64-1.c b/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2ss64-1.c index 38ecf39ad65..f4dae536873 100644 --- a/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2ss64-1.c +++ b/gcc/testsuite/gcc.target/i386/avx512f-vcvtusi2ss64-1.c @@ -1,7 +1,7 @@ /* { dg-do compile { target { ! ia32 } } } */ /* { dg-options "-mavx512f -O2" } */ -/* { dg-final { scan-assembler-times "vcvtusi2ss\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ -/* { dg-final { scan-assembler-times "vcvtusi2ss\[ \\t\]+\[^%\n\]*%r\[^\{\n\]*\{rz-sae\}\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ +/* { dg-final { scan-assembler-times "vcvtusi2ssq\[ \\t\]+\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ +/* { dg-final { scan-assembler-times "vcvtusi2ssq\[ \\t\]+\[^%\n\]*%r\[^\{\n\]*\{rz-sae\}\[^\{\n\]*%xmm\[0-9\]+(?:\n|\[ \\t\]+#)" 1 } } */ #include <immintrin.h>
Hello! With current insn mnemonic and ATT assembler dialect, there is no way for the assembler to distinguish between DImode and SImode instruction when memory input operand is used. The dump for 32bit memory reads as: 0: 62 f1 7e 08 7b 05 00 vcvtusi2ss 0x0(%rip),%xmm0,%xmm0 7: 00 00 00 10: 62 f1 7f 08 7b 05 00 vcvtusi2sd 0x0(%rip),%xmm0,%xmm0 17: 00 00 00 And for 64bit access: 20: 62 f1 fe 08 7b 05 00 vcvtusi2ss 0x0(%rip),%xmm0,%xmm0 27: 00 00 00 30: 62 f1 ff 08 7b 05 00 vcvtusi2sd 0x0(%rip),%xmm0,%xmm0 37: 00 00 00 (Note the difference in the 3rd byte. On a related note, binutils should also be fixed to dump vcvtsi2sdq in the 64bit case to avoid ambiguity) We should use "q" suffix with the ATT dialect in the Dimode insn mnemonic. 2018-05-23 Uros Bizjak <ubizjak@gmail.com> * config/i386/sse.md (cvtusi2<ssescalarmodesuffix>64<round_name>): Add {q} suffix to insn mnemonic. testsuite/Changelog: 2018-05-23 Uros Bizjak <ubizjak@gmail.com> * gcc.target/i386/avx512f-vcvtusi2sd64-1.c: Update scan string. * gcc.target/i386/avx512f-vcvtusi2ss64-1.c: Ditto. Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}. OK for mainline and backports? Uros.