83 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 -post-RA-scheduler=true | FileCheck %s3 4; Widen a v3i16 to v8i16 to do a vector add5 6@.str = internal constant [4 x i8] c"%d \00"7@.str1 = internal constant [2 x i8] c"\0A\00"8 9define void @update(ptr %dst, ptr %src, i32 %n) nounwind {10; CHECK-LABEL: update:11; CHECK: # %bb.0: # %entry12; CHECK-NEXT: pushl %ebp13; CHECK-NEXT: movl %esp, %ebp14; CHECK-NEXT: andl $-8, %esp15; CHECK-NEXT: subl $16, %esp16; CHECK-NEXT: pcmpeqd %xmm0, %xmm017; CHECK-NEXT: movw $1, {{[0-9]+}}(%esp)18; CHECK-NEXT: movl $65537, {{[0-9]+}}(%esp) # imm = 0x1000119; CHECK-NEXT: movl $0, {{[0-9]+}}(%esp)20; CHECK-NEXT: .p2align 421; CHECK-NEXT: .LBB0_1: # %forcond22; CHECK-NEXT: # =>This Inner Loop Header: Depth=123; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax24; CHECK-NEXT: cmpl 16(%ebp), %eax25; CHECK-NEXT: jge .LBB0_326; CHECK-NEXT: # %bb.2: # %forbody27; CHECK-NEXT: # in Loop: Header=BB0_1 Depth=128; CHECK-NEXT: movl {{[0-9]+}}(%esp), %eax29; CHECK-NEXT: movl 12(%ebp), %edx30; CHECK-NEXT: movl 8(%ebp), %ecx31; CHECK-NEXT: movd {{.*#+}} xmm1 = mem[0],zero,zero,zero32; CHECK-NEXT: pinsrw $2, 4(%edx,%eax,8), %xmm133; CHECK-NEXT: psubw %xmm0, %xmm134; CHECK-NEXT: pextrw $2, %xmm1, 4(%ecx,%eax,8)35; CHECK-NEXT: movd %xmm1, (%ecx,%eax,8)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 i16>, align 847 %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 i16> < i16 1, i16 1, i16 1 >, 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 i16>, ptr %tmp3, i32 %tmp265 %tmp4 = load i32, ptr %i66 %tmp5 = load ptr, ptr %src.addr67 %arrayidx6 = getelementptr <3 x i16>, ptr %tmp5, i32 %tmp468 %tmp7 = load <3 x i16>, ptr %arrayidx669 %add = add <3 x i16> %tmp7, < i16 1, i16 1, i16 1 >70 store <3 x i16> %add, ptr %arrayidx71 br label %forinc72 73forinc:74 %tmp8 = load i32, ptr %i75 %inc = add i32 %tmp8, 176 store i32 %inc, ptr %i77 br label %forcond78 79afterfor:80 ret void81}82 83