91 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 %s --check-prefix=WIDE3 4; FIXME: We shouldn't require both a movd and an insert in the wide version.5 6define void @update(ptr %dst_i, ptr %src_i, i32 %n) nounwind {7; WIDE-LABEL: update:8; WIDE: # %bb.0: # %entry9; WIDE-NEXT: subl $12, %esp10; WIDE-NEXT: movl $0, (%esp)11; WIDE-NEXT: pcmpeqd %xmm0, %xmm012; WIDE-NEXT: movdqa {{.*#+}} xmm1 = [63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63]13; WIDE-NEXT: movdqa {{.*#+}} xmm2 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32]14; WIDE-NEXT: .p2align 415; WIDE-NEXT: .LBB0_1: # %forcond16; WIDE-NEXT: # =>This Inner Loop Header: Depth=117; WIDE-NEXT: movl (%esp), %eax18; WIDE-NEXT: cmpl {{[0-9]+}}(%esp), %eax19; WIDE-NEXT: jge .LBB0_320; WIDE-NEXT: # %bb.2: # %forbody21; WIDE-NEXT: # in Loop: Header=BB0_1 Depth=122; WIDE-NEXT: movl (%esp), %eax23; WIDE-NEXT: leal (,%eax,8), %edx24; WIDE-NEXT: movl {{[0-9]+}}(%esp), %ecx25; WIDE-NEXT: addl %edx, %ecx26; WIDE-NEXT: movl %ecx, {{[0-9]+}}(%esp)27; WIDE-NEXT: addl {{[0-9]+}}(%esp), %edx28; WIDE-NEXT: movl %edx, {{[0-9]+}}(%esp)29; WIDE-NEXT: movq {{.*#+}} xmm3 = mem[0],zero30; WIDE-NEXT: psubb %xmm0, %xmm331; WIDE-NEXT: psrlw $2, %xmm332; WIDE-NEXT: pand %xmm1, %xmm333; WIDE-NEXT: pxor %xmm2, %xmm334; WIDE-NEXT: psubb %xmm2, %xmm335; WIDE-NEXT: movq %xmm3, (%ecx,%eax,8)36; WIDE-NEXT: incl (%esp)37; WIDE-NEXT: jmp .LBB0_138; WIDE-NEXT: .LBB0_3: # %afterfor39; WIDE-NEXT: addl $12, %esp40; WIDE-NEXT: retl41entry:42 %dst_i.addr = alloca ptr43 %src_i.addr = alloca ptr44 %n.addr = alloca i3245 %i = alloca i32, align 446 %dst = alloca ptr, align 447 %src = alloca ptr, align 448 store ptr %dst_i, ptr %dst_i.addr49 store ptr %src_i, ptr %src_i.addr50 store i32 %n, ptr %n.addr51 store i32 0, ptr %i52 br label %forcond53 54forcond:55 %tmp = load i32, ptr %i56 %tmp1 = load i32, ptr %n.addr57 %cmp = icmp slt i32 %tmp, %tmp158 br i1 %cmp, label %forbody, label %afterfor59 60forbody:61 %tmp2 = load i32, ptr %i62 %tmp3 = load ptr, ptr %dst_i.addr63 %arrayidx = getelementptr i64, ptr %tmp3, i32 %tmp264 store ptr %arrayidx, ptr %dst65 %tmp4 = load i32, ptr %i66 %tmp5 = load ptr, ptr %src_i.addr67 %arrayidx6 = getelementptr i64, ptr %tmp5, i32 %tmp468 store ptr %arrayidx6, ptr %src69 %tmp8 = load i32, ptr %i70 %tmp9 = load ptr, ptr %dst71 %arrayidx10 = getelementptr <8 x i8>, ptr %tmp9, i32 %tmp872 %tmp11 = load i32, ptr %i73 %tmp12 = load ptr, ptr %src74 %arrayidx13 = getelementptr <8 x i8>, ptr %tmp12, i32 %tmp1175 %tmp14 = load <8 x i8>, ptr %arrayidx1376 %add = add <8 x i8> %tmp14, < i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1 >77 %shr = ashr <8 x i8> %add, < i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2, i8 2 >78 store <8 x i8> %shr, ptr %arrayidx1079 br label %forinc80 81forinc:82 %tmp15 = load i32, ptr %i83 %inc = add i32 %tmp15, 184 store i32 %inc, ptr %i85 br label %forcond86 87afterfor:88 ret void89}90 91