91 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=i686-unknown-unknown -mattr=+sse4.2 < %s | FileCheck %s3; RUN: llc -mtriple=i686-unknown-unknown -mcpu=atom < %s | FileCheck -check-prefix=ATOM %s4 5; Scheduler causes produce a different instruction order6 7; bitcast a v4i16 to v2i328 9define void @convert(ptr %dst, ptr %src) nounwind {10; CHECK-LABEL: convert:11; CHECK: # %bb.0: # %entry12; CHECK-NEXT: pushl %eax13; CHECK-NEXT: movl $0, (%esp)14; CHECK-NEXT: pcmpeqd %xmm0, %xmm015; CHECK-NEXT: cmpl $3, (%esp)16; CHECK-NEXT: jg .LBB0_317; CHECK-NEXT: .p2align 418; CHECK-NEXT: .LBB0_2: # %forbody19; CHECK-NEXT: # =>This Inner Loop Header: Depth=120; CHECK-NEXT: movl (%esp), %eax21; CHECK-NEXT: movl {{[0-9]+}}(%esp), %ecx22; CHECK-NEXT: movl {{[0-9]+}}(%esp), %edx23; CHECK-NEXT: movq {{.*#+}} xmm1 = mem[0],zero24; CHECK-NEXT: psubw %xmm0, %xmm125; CHECK-NEXT: movq %xmm1, (%ecx,%eax,8)26; CHECK-NEXT: incl (%esp)27; CHECK-NEXT: cmpl $3, (%esp)28; CHECK-NEXT: jle .LBB0_229; CHECK-NEXT: .LBB0_3: # %afterfor30; CHECK-NEXT: popl %eax31; CHECK-NEXT: retl32;33; ATOM-LABEL: convert:34; ATOM: # %bb.0: # %entry35; ATOM-NEXT: pushl %eax36; ATOM-NEXT: pcmpeqd %xmm0, %xmm037; ATOM-NEXT: movl $0, (%esp)38; ATOM-NEXT: cmpl $3, (%esp)39; ATOM-NEXT: jg .LBB0_340; ATOM-NEXT: .p2align 441; ATOM-NEXT: .LBB0_2: # %forbody42; ATOM-NEXT: # =>This Inner Loop Header: Depth=143; ATOM-NEXT: movl (%esp), %eax44; ATOM-NEXT: movl {{[0-9]+}}(%esp), %ecx45; ATOM-NEXT: movq {{.*#+}} xmm1 = mem[0],zero46; ATOM-NEXT: movl {{[0-9]+}}(%esp), %ecx47; ATOM-NEXT: psubw %xmm0, %xmm148; ATOM-NEXT: movq %xmm1, (%ecx,%eax,8)49; ATOM-NEXT: incl (%esp)50; ATOM-NEXT: cmpl $3, (%esp)51; ATOM-NEXT: jle .LBB0_252; ATOM-NEXT: .LBB0_3: # %afterfor53; ATOM-NEXT: popl %eax54; ATOM-NEXT: retl55entry:56 %dst.addr = alloca ptr57 %src.addr = alloca ptr58 %i = alloca i32, align 459 store ptr %dst, ptr %dst.addr60 store ptr %src, ptr %src.addr61 store i32 0, ptr %i62 br label %forcond63 64forcond:65 %tmp = load i32, ptr %i66 %cmp = icmp slt i32 %tmp, 467 br i1 %cmp, label %forbody, label %afterfor68 69forbody:70 %tmp1 = load i32, ptr %i71 %tmp2 = load ptr, ptr %dst.addr72 %arrayidx = getelementptr <2 x i32>, ptr %tmp2, i32 %tmp173 %tmp3 = load i32, ptr %i74 %tmp4 = load ptr, ptr %src.addr75 %arrayidx5 = getelementptr <4 x i16>, ptr %tmp4, i32 %tmp376 %tmp6 = load <4 x i16>, ptr %arrayidx577 %add = add <4 x i16> %tmp6, < i16 1, i16 1, i16 1, i16 1 >78 %conv = bitcast <4 x i16> %add to <2 x i32>79 store <2 x i32> %conv, ptr %arrayidx80 br label %forinc81 82forinc:83 %tmp7 = load i32, ptr %i84 %inc = add i32 %tmp7, 185 store i32 %inc, ptr %i86 br label %forcond87 88afterfor:89 ret void90}91