[X86] Using _mm512_maskz_set1_epi64 on a 32-bit target causes cannot select intrinsic error

[topperc]

Bugzilla Link	34631
Resolution	FIXED
Resolved on	Sep 29, 2017 16:52
Version	trunk
OS	All
Blocks	#33840
CC	@RKSimon,@tstellar
Fixed by commit(s)	r313392 r314569

Extended Description

This intrinsic is currently implemented in clang by selecting between two different builtins, __builtin_ia32_pbroadcastq512_gpr_mask(64-bit) and __builtin_ia32_pbroadcastq512_mem_mask(32-bit).

But it seems __builtin_ia32_pbroadcastq512_mem_mask was never implemented in the X86 backend. So using it just causes an isel error. And worse, on a debug build it hits an llvm_unreachable trying to expand a 64-bit operand on an intrinsic.

Test case

#include <x86intrin.h>

__m512i test_mm512_maskz_set1_epi64 (__mmask8 __M, long long __A)
{
//CHECK-LABEL: @​test_mm512_maskz_set1_epi64
//CHECK: @​llvm.x86.avx512.mask.pbroadcast.q.gpr.512
return _mm512_maskz_set1_epi64 (__M, __A);
}

[topperc]

Workaround to remove the intrinsic from 32-bit mode until the backend can handle it submitted in r313392. D37668 should make the backend able to handle this.

[topperc]

This has probably been broken for a while, but the workaround to disable the intrinsic entirely in 32-bit mode is simple. Nominating for 5.0.1

[RKSimon]

Not great codegen, but in the meantime would we be better off at least providing SOMETHING on 32-bit? GCC/ICC both support the intrinsic on -m32.

__m512i _mm512_maskz_set1_epi64 (__mmask8 __M, long long __A)
{
return (__m512i)__builtin_ia32_selectq_512(
__M,
(__v8di) (__m512i{__A,__A,__A,__A,__A,__A,__A,__A},
_mm512_setzero_si512());
}

[topperc]

Isn't that the equivalent of what D37668 will do? If that doesn't generate good code should we still be proceeding with that patch?

All of the other masked set1_epi64 intrinsics are removed from 32-bit mode currently, and gcc supports all of them in -m32.

[RKSimon]

Isn't that the equivalent of what D37668 will do? If that doesn't generate
good code should we still be proceeding with that patch?

All of the other masked set1_epi64 intrinsics are removed from 32-bit mode
currently, and gcc supports all of them in -m32.

So do we want D37668 to be merged into 5.0.1 (or at least the epi64 intrinsic changes)?

[tstellar]

Any more thoughts on which commit to merge for 5.0.1?

[topperc]

I believe D37668 landed, but didn't enable 32-bit mode support due to codegen deficiencies.

So for now we should merge r313392 to at least hide the broken intrinsics from 32-bit mode.

[tstellar]

Merged: r314569