assigned to @topperc

There's something very strange going on in the IR out of the middle end

%37 = shl <2 x i64> %24, <i64 32, i64 32>, !dbg !​327
%38 = ashr exact <2 x i64> %37, <i64 32, i64 32>, !dbg !​327
%39 = shl <2 x i64> %36, <i64 32, i64 32>, !dbg !​327
%40 = ashr exact <2 x i64> %39, <i64 32, i64 32>, !dbg !​327
%41 = mul nsw <2 x i64> %40, %38, !dbg !​327

I assume InstCombine has gotten clever, but I'm not sure why yet.

Disregard that last comment. That's our IR for pmuldq that we should pattern match. Will keep digging.

I think because warping is loop invariant, it got pulled out of the loop along with the _mm_set1 and the 2xi64 shl+ashr we insert to sign extend the mul input for pmuldq pattern match.

Because we don't have a vsraq instruction we emit this

    pshufd  $0, %xmm0, %xmm0                # xmm0 = xmm0[0,0,0,0]                                                                                                                                    
    movdqa  %xmm0, %xmm1                                                                                                                                                                              
    psllq   $32, %xmm1                                                                                                                                                                                
    psrad   $31, %xmm1                                                                                                                                                                                
    pblendw $51, %xmm0, %xmm1               # xmm1 = xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]  

And we aren't able to create an AssertSExt from that to tell the block inside the loop that the input is already sign extended to allow pmuldq to be match. And the sign extend outside the loop would have been unnecessary if we formed pmuldq.

Maybe we need an X86CodeGenPrepare to bring them back into the same basic block? I assume the same thing can happen to pmuludq.

I think because warping is loop invariant, it got pulled out of the loop
along with the _mm_set1 and the 2xi64 shl+ashr we insert to sign extend the
mul input for pmuldq pattern match.

Because we don't have a vsraq instruction we emit this

    pshufd  $0, %xmm0, %xmm0                # xmm0 = xmm0[0,0,0,0]      

    movdqa  %xmm0, %xmm1                                                

    psllq   $32, %xmm1                                                  

    psrad   $31, %xmm1                                                  

    pblendw $51, %xmm0, %xmm1               # xmm1 =

xmm0[0,1],xmm1[2,3],xmm0[4,5],xmm1[6,7]

And we aren't able to create an AssertSExt from that to tell the block
inside the loop that the input is already sign extended to allow pmuldq to
be match. And the sign extend outside the loop would have been unnecessary
if we formed pmuldq.

Thanks for your analysis, and I also want to know is there any solution from code level to avoid this problem?

To fix this in the compiler, I think we just need to add this to X86TargetLowering::shouldSinkOperands

I'm not sure how to work around it in code.

Candidate patch https://reviews.llvm.org/D107689

Anything left to do here?

Not sure if this qualifies for backport to the release branch.

Anything left to do here?

Not sure if this qualifies for backport to the release branch.

I don't think so - its a rather old perf regression.

Tentatively resolving this.

Anything left to do here?

Not sure if this qualifies for backport to the release branch.

I don't think so - its a rather old perf regression.

Tentatively resolving this.

Thank you for resolving this issue, so what's the release version of llvm which can avoid this issue?

Anything left to do here?

Not sure if this qualifies for backport to the release branch.

I don't think so - its a rather old perf regression.

Tentatively resolving this.

Thank you for resolving this issue, so what's the release version of llvm
which can avoid this issue?

It will be in 14.0 (about 6 months from now).