84 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse4.2 | FileCheck %s3 4; widen a v3f32 to vfi32 to do a vector multiple and an add5 6define void @update(ptr %dst, ptr %src, i32 %n) nounwind {7; CHECK-LABEL: update:8; CHECK: # %bb.0: # %entry9; CHECK-NEXT: pushl %ebp10; CHECK-NEXT: movl %esp, %ebp11; CHECK-NEXT: andl $-16, %esp12; CHECK-NEXT: subl $48, %esp13; CHECK-NEXT: movl $1077936128, {{[0-9]+}}(%esp) # imm = 0x4040000014; CHECK-NEXT: movl $1073741824, {{[0-9]+}}(%esp) # imm = 0x4000000015; CHECK-NEXT: movl $1065353216, {{[0-9]+}}(%esp) # imm = 0x3F80000016; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)17; CHECK-NEXT: movaps {{.*#+}} xmm0 = [1.97604004E+3,1.97604004E+3,1.97604004E+3,u]18; CHECK-NEXT: .p2align 419; CHECK-NEXT: .LBB0_1: # %forcond20; CHECK-NEXT: # =>This Inner Loop Header: Depth=121; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax22; CHECK-NEXT: cmpl 16(%ebp), %eax23; CHECK-NEXT: jge .LBB0_324; CHECK-NEXT: # %bb.2: # %forbody25; CHECK-NEXT: # in Loop: Header=BB0_1 Depth=126; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax27; CHECK-NEXT: movl 8(%ebp), %ecx28; CHECK-NEXT: shll $4, %eax29; CHECK-NEXT: movl 12(%ebp), %edx30; CHECK-NEXT: movaps (%edx,%eax), %xmm131; CHECK-NEXT: mulps {{[0-9]+}}(%esp), %xmm132; CHECK-NEXT: addps %xmm0, %xmm133; CHECK-NEXT: extractps $2, %xmm1, 8(%ecx,%eax)34; CHECK-NEXT: extractps $1, %xmm1, 4(%ecx,%eax)35; CHECK-NEXT: movss %xmm1, (%ecx,%eax)36; CHECK-NEXT: incl {{[0-9]+}}(%esp)37; CHECK-NEXT: jmp .LBB0_138; CHECK-NEXT: .LBB0_3: # %afterfor39; CHECK-NEXT: movl %ebp, %esp40; CHECK-NEXT: popl %ebp41; CHECK-NEXT: retl42entry:43 %dst.addr = alloca ptr44 %src.addr = alloca ptr45 %n.addr = alloca i3246 %v = alloca <3 x float>, align 1647 %i = alloca i32, align 448 store ptr %dst, ptr %dst.addr49 store ptr %src, ptr %src.addr50 store i32 %n, ptr %n.addr51 store <3 x float> < float 1.000000e+00, float 2.000000e+00, float 3.000000e+00 >, ptr %v52 store i32 0, ptr %i53 br label %forcond54 55forcond:56 %tmp = load i32, ptr %i57 %tmp1 = load i32, ptr %n.addr58 %cmp = icmp slt i32 %tmp, %tmp159 br i1 %cmp, label %forbody, label %afterfor60 61forbody:62 %tmp2 = load i32, ptr %i63 %tmp3 = load ptr, ptr %dst.addr64 %arrayidx = getelementptr <3 x float>, ptr %tmp3, i32 %tmp265 %tmp4 = load i32, ptr %i66 %tmp5 = load ptr, ptr %src.addr67 %arrayidx6 = getelementptr <3 x float>, ptr %tmp5, i32 %tmp468 %tmp7 = load <3 x float>, ptr %arrayidx669 %tmp8 = load <3 x float>, ptr %v70 %mul = fmul <3 x float> %tmp7, %tmp871 %add = fadd <3 x float> %mul, < float 0x409EE02900000000, float 0x409EE02900000000, float 0x409EE02900000000 >72 store <3 x float> %add, ptr %arrayidx73 br label %forinc74 75forinc:76 %tmp9 = load i32, ptr %i77 %inc = add i32 %tmp9, 178 store i32 %inc, ptr %i79 br label %forcond80 81afterfor:82 ret void83}84