1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=i686-unknown-linux-gnu -mattr=+sse2,-sse4.1 | FileCheck %s --check-prefix=X32 3; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+sse2,-sse4.1 | FileCheck %s --check-prefix=X64 4 5define void @test1(<4 x float>* %F, float* %f) nounwind { 6; X32-LABEL: test1: 7; X32: # %bb.0: # %entry 8; X32-NEXT: movl {{[0-9]+}}(%esp), %eax 9; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx 10; X32-NEXT: movaps (%ecx), %xmm0 11; X32-NEXT: addps %xmm0, %xmm0 12; X32-NEXT: movss %xmm0, (%eax) 13; X32-NEXT: retl 14; 15; X64-LABEL: test1: 16; X64: # %bb.0: # %entry 17; X64-NEXT: movaps (%rdi), %xmm0 18; X64-NEXT: addps %xmm0, %xmm0 19; X64-NEXT: movss %xmm0, (%rsi) 20; X64-NEXT: retq 21entry: 22 %tmp = load <4 x float>, <4 x float>* %F 23 %tmp7 = fadd <4 x float> %tmp, %tmp 24 %tmp2 = extractelement <4 x float> %tmp7, i32 0 25 store float %tmp2, float* %f 26 ret void 27} 28 29define float @test2(<4 x float>* %F, float* %f) nounwind { 30; X32-LABEL: test2: 31; X32: # %bb.0: # %entry 32; X32-NEXT: pushl %eax 33; X32-NEXT: movl {{[0-9]+}}(%esp), %eax 34; X32-NEXT: movaps (%eax), %xmm0 35; X32-NEXT: addps %xmm0, %xmm0 36; X32-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1] 37; X32-NEXT: movss %xmm0, (%esp) 38; X32-NEXT: flds (%esp) 39; X32-NEXT: popl %eax 40; X32-NEXT: retl 41; 42; X64-LABEL: test2: 43; X64: # %bb.0: # %entry 44; X64-NEXT: movaps (%rdi), %xmm0 45; X64-NEXT: addps %xmm0, %xmm0 46; X64-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1] 47; X64-NEXT: retq 48entry: 49 %tmp = load <4 x float>, <4 x float>* %F 50 %tmp7 = fadd <4 x float> %tmp, %tmp 51 %tmp2 = extractelement <4 x float> %tmp7, i32 2 52 ret float %tmp2 53} 54 55define void @test3(float* %R, <4 x float>* %P1) nounwind { 56; X32-LABEL: test3: 57; X32: # %bb.0: # %entry 58; X32-NEXT: movl {{[0-9]+}}(%esp), %eax 59; X32-NEXT: movl {{[0-9]+}}(%esp), %ecx 60; X32-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 61; X32-NEXT: movss %xmm0, (%eax) 62; X32-NEXT: retl 63; 64; X64-LABEL: test3: 65; X64: # %bb.0: # %entry 66; X64-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero 67; X64-NEXT: movss %xmm0, (%rdi) 68; X64-NEXT: retq 69entry: 70 %X = load <4 x float>, <4 x float>* %P1 71 %tmp = extractelement <4 x float> %X, i32 3 72 store float %tmp, float* %R 73 ret void 74} 75 76define double @test4(double %A) nounwind { 77; X32-LABEL: test4: 78; X32: # %bb.0: # %entry 79; X32-NEXT: subl $12, %esp 80; X32-NEXT: calll foo 81; X32-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1] 82; X32-NEXT: addsd {{[0-9]+}}(%esp), %xmm0 83; X32-NEXT: movsd %xmm0, (%esp) 84; X32-NEXT: fldl (%esp) 85; X32-NEXT: addl $12, %esp 86; X32-NEXT: retl 87; 88; X64-LABEL: test4: 89; X64: # %bb.0: # %entry 90; X64-NEXT: pushq %rax 91; X64-NEXT: movsd %xmm0, (%rsp) # 8-byte Spill 92; X64-NEXT: callq foo 93; X64-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1] 94; X64-NEXT: addsd (%rsp), %xmm0 # 8-byte Folded Reload 95; X64-NEXT: popq %rax 96; X64-NEXT: retq 97entry: 98 %tmp1 = call <2 x double> @foo( ) 99 %tmp2 = extractelement <2 x double> %tmp1, i32 1 100 %tmp3 = fadd double %tmp2, %A 101 ret double %tmp3 102} 103 104declare <2 x double> @foo() 105