https://llvm.godbolt.org/z/dsxerr9Mq
target triple = "x86_64-unknown-linux-gnu"
define i32 @_mm_testc_ps(ptr %a, ptr %b) #0 {
start:
%0 = load <2 x i64>, ptr %a, align 16
%1 = load <2 x i64>, ptr %b, align 16
%_8.i = xor <2 x i64> %0, splat (i64 -1)
%_7.i = and <2 x i64> %1, %_8.i
%2 = bitcast <2 x i64> %_7.i to <4 x i32>
%_51.lobit = ashr <4 x i32> %2, splat (i32 31)
%_7 = tail call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> %_51.lobit)
%_6 = icmp eq i32 %_7, 0
%_0 = zext i1 %_6 to i32
ret i32 %_0
}
attributes #0 = { "target-features"="+avx"}VectorCombine transforms this to:
define i32 @_mm_testc_ps(ptr %a, ptr %b) #0 {
start:
%0 = load <2 x i64>, ptr %a, align 16
%1 = load <2 x i64>, ptr %b, align 16
%_8.i = xor <2 x i64> %0, splat (i64 -1)
%_7.i = and <2 x i64> %1, %_8.i
%2 = bitcast <2 x i64> %_7.i to <4 x i32>
%3 = call i32 @llvm.vector.reduce.umax.v4i32(<4 x i32> %2)
%_6 = icmp sgt i32 %3, -1
%_0 = zext i1 %_6 to i32
ret i32 %_0
}Previous lowering:
_mm_testc_ps: # @_mm_testc_ps
vmovaps xmm0, xmmword ptr [rdi]
vandnps xmm0, xmm0, xmmword ptr [rsi]
xor eax, eax
vtestps xmm0, xmm0
sete al
retNew lowering:
_mm_testc_ps: # @_mm_testc_ps
vmovdqa xmm0, xmmword ptr [rdi]
vpandn xmm0, xmm0, xmmword ptr [rsi]
vpshufd xmm1, xmm0, 238 # xmm1 = xmm0[2,3,2,3]
vpmaxud xmm0, xmm0, xmm1
vpshufd xmm1, xmm0, 85 # xmm1 = xmm0[1,1,1,1]
vpor xmm0, xmm0, xmm1
vmovd eax, xmm0
not eax
shr eax, 31
retThe expected vtestps is no longer produced.
Relevant VectorCombine changes are presumably #175194 #181998. I think those transforms are fine by themselves though, so this probably needs a backend fix.
https://llvm.godbolt.org/z/dsxerr9Mq
VectorCombine transforms this to:
Previous lowering:
New lowering:
The expected vtestps is no longer produced.
Relevant VectorCombine changes are presumably #175194 #181998. I think those transforms are fine by themselves though, so this probably needs a backend fix.