[X86] znver missing cmpxchg8b feature, resulting in expanded atomicrmw

[nikic]

Bugzilla Link	42935
Resolution	FIXED
Resolved on	Aug 09, 2019 02:56
Version	9.0
OS	All
Blocks	#41819
CC	@topperc,@zmodem,@RKSimon,@rotateright
Fixed by commit(s)	r368324

Extended Description

define i64 @​test(i64* %p) {
%x = atomicrmw sub i64* %p, i64 1 seq_cst
ret i64 %x
}

Under -mcpu=znver1 results in:

pushq	%rax
.cfi_def_cfa_offset 16
movl	$1, %esi
movl	$5, %edx
callq	__atomic_fetch_sub_8
popq	%rcx
.cfi_def_cfa_offset 8
retq

While it should produce a "lock xaddq" instruction.

This issue was introduced in https://reviews.llvm.org/D59576, where the CMPXCHG8B feature flag was not added to ZNFeatures.

[nikic]

assigned to @topperc

[RKSimon]

I agree this looks like a typo.

@​craig how safe do you think it would be for FeatureCMPXCHG16B to inherit FeatureCMPXCHG8B?

[topperc]

I think that's probably safe. The cmpxchg16b encoding is cmpxchg8b with a rex prefix. We didn't connect cx8 to the command line in clang. So I don't think we have to worry about something like -mno-cx8 disabling cx16.

[topperc]

Fixed in r368324 using Simon's proposal.

[zmodem]

Fixed in r368324 using Simon's proposal.

Merged to release_90 in r368423.