[x86] After 16f3d698f2af, -march=amdfam10 incorrectly implies SSSE3 instructions

[DimitryAndric]

Bugzilla Link	47464
Resolution	FIXED
Resolved on	Sep 08, 2020 12:00
Version	11.0
OS	All
Blocks	#46070
CC	@avg-I,@topperc,@zmodem,@LebedevRI,@RKSimon,@rotateright
Fixed by commit(s)	e6bb4c8

Extended Description

After https://reviews.llvm.org/D83273 /
16f3d698f2af ("[X86] Move
the feature dependency handling in X86TargetInfo::setFeatureEnabledImpl
to a table based lookup in X86TargetParser.cpp"), compiling with -march=amdfam10 incorrectly starts using SSSE3 instructions, in particular pshufb.

Test case:

#include <xmmintrin.h>

__m128i foo(unsigned char c) {
return _mm_set1_epi8(c);
}

Asm output before:

foo: # @​foo
.cfi_startproc

%bb.0: # %entry

    movd    %edi, %xmm0
    punpcklbw       %xmm0, %xmm0            # xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
    pshuflw $224, %xmm0, %xmm0              # xmm0 = xmm0[0,0,2,3,4,5,6,7]
    pshufd  $0, %xmm0, %xmm0                # xmm0 = xmm0[0,0,0,0]
    retq

After:

foo: # @​foo
.cfi_startproc

%bb.0: # %entry

    movd    %edi, %xmm0
    pxor    %xmm1, %xmm1
    pshufb  %xmm1, %xmm0
    retq

Apparently 16f3d69#diff-045a11495d5fad4e01af59c60f6a387aR500 makes SSE4_A now imply SSSE3, which is not correct (at least not always, I think).

[zmodem]

Craig, can you take a look?

[avg-I]

Perhaps it was just a typo -- SSE3 vs SSSE3?

[topperc]

Fixed by e6bb4c8

[zmodem]

Fixed by e6bb4c8

Thanks! Pushed to 11.x as 6f1dbbc.