1; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s 2 3define void @foo(<3 x float> %in, <4 x i8>* nocapture %out) nounwind { 4; CHECK-LABEL: foo: 5; CHECK: # BB#0: 6; CHECK-NEXT: cvttps2dq %xmm0, %xmm0 7; CHECK-NEXT: movl $255, %eax 8; CHECK-NEXT: pinsrd $3, %eax, %xmm0 9; CHECK-NEXT: pshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u] 10; CHECK-NEXT: movd %xmm0, (%rdi) 11; CHECK-NEXT: retq 12 %t0 = fptoui <3 x float> %in to <3 x i8> 13 %t1 = shufflevector <3 x i8> %t0, <3 x i8> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 undef> 14 %t2 = insertelement <4 x i8> %t1, i8 -1, i32 3 15 store <4 x i8> %t2, <4 x i8>* %out, align 4 16 ret void 17} 18 19; Verify that the DAGCombiner doesn't wrongly fold a build_vector into a 20; blend with a zero vector if the build_vector contains negative zero. 21; 22; TODO: the codegen for function 'test_negative_zero_1' is sub-optimal. 23; Ideally, we should generate a single shuffle blend operation. 24 25define <4 x float> @test_negative_zero_1(<4 x float> %A) { 26; CHECK-LABEL: test_negative_zero_1: 27; CHECK: # BB#0: # %entry 28; CHECK-NEXT: movapd %xmm0, %xmm1 29; CHECK-NEXT: shufpd {{.*#+}} xmm1 = xmm1[1,0] 30; CHECK-NEXT: xorps %xmm2, %xmm2 31; CHECK-NEXT: blendps {{.*#+}} xmm2 = xmm1[0],xmm2[1,2,3] 32; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero 33; CHECK-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1] 34; CHECK-NEXT: unpcklpd {{.*#+}} xmm0 = xmm0[0],xmm2[0] 35; CHECK-NEXT: retq 36entry: 37 %0 = extractelement <4 x float> %A, i32 0 38 %1 = insertelement <4 x float> undef, float %0, i32 0 39 %2 = insertelement <4 x float> %1, float -0.0, i32 1 40 %3 = extractelement <4 x float> %A, i32 2 41 %4 = insertelement <4 x float> %2, float %3, i32 2 42 %5 = insertelement <4 x float> %4, float 0.0, i32 3 43 ret <4 x float> %5 44} 45 46define <2 x double> @test_negative_zero_2(<2 x double> %A) { 47; CHECK-LABEL: test_negative_zero_2: 48; CHECK: # BB#0: # %entry 49; CHECK-NEXT: movhpd {{.*}}(%rip), %xmm0 50; CHECK-NEXT: retq 51entry: 52 %0 = extractelement <2 x double> %A, i32 0 53 %1 = insertelement <2 x double> undef, double %0, i32 0 54 %2 = insertelement <2 x double> %1, double -0.0, i32 1 55 ret <2 x double> %2 56} 57