118 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefix=SSE23; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=SSE414 5; Widen a v5i16 to v8i16 to do a vector sub and multiple6 7define void @update(ptr %dst, ptr %src, i32 %n) nounwind {8; SSE2-LABEL: update:9; SSE2: # %bb.0: # %entry10; SSE2-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)11; SSE2-NEXT: movq %rsi, -{{[0-9]+}}(%rsp)12; SSE2-NEXT: movl %edx, -{{[0-9]+}}(%rsp)13; SSE2-NEXT: movabsq $4295032833, %rax # imm = 0x10001000114; SSE2-NEXT: movq %rax, -{{[0-9]+}}(%rsp)15; SSE2-NEXT: movw $0, -{{[0-9]+}}(%rsp)16; SSE2-NEXT: movl $0, -{{[0-9]+}}(%rsp)17; SSE2-NEXT: movdqa {{.*#+}} xmm0 = [271,271,271,271,271,u,u,u]18; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [2,4,2,2,2,u,u,u]19; SSE2-NEXT: .p2align 420; SSE2-NEXT: .LBB0_1: # %forcond21; SSE2-NEXT: # =>This Inner Loop Header: Depth=122; SSE2-NEXT: movl -{{[0-9]+}}(%rsp), %eax23; SSE2-NEXT: cmpl -{{[0-9]+}}(%rsp), %eax24; SSE2-NEXT: jge .LBB0_325; SSE2-NEXT: # %bb.2: # %forbody26; SSE2-NEXT: # in Loop: Header=BB0_1 Depth=127; SSE2-NEXT: movslq -{{[0-9]+}}(%rsp), %rax28; SSE2-NEXT: movq -{{[0-9]+}}(%rsp), %rcx29; SSE2-NEXT: shlq $4, %rax30; SSE2-NEXT: movq -{{[0-9]+}}(%rsp), %rdx31; SSE2-NEXT: movdqa (%rdx,%rax), %xmm232; SSE2-NEXT: psubw %xmm0, %xmm233; SSE2-NEXT: pmullw %xmm1, %xmm234; SSE2-NEXT: movq %xmm2, (%rcx,%rax)35; SSE2-NEXT: pextrw $4, %xmm2, %edx36; SSE2-NEXT: movw %dx, 8(%rcx,%rax)37; SSE2-NEXT: incl -{{[0-9]+}}(%rsp)38; SSE2-NEXT: jmp .LBB0_139; SSE2-NEXT: .LBB0_3: # %afterfor40; SSE2-NEXT: retq41;42; SSE41-LABEL: update:43; SSE41: # %bb.0: # %entry44; SSE41-NEXT: movq %rdi, -{{[0-9]+}}(%rsp)45; SSE41-NEXT: movq %rsi, -{{[0-9]+}}(%rsp)46; SSE41-NEXT: movl %edx, -{{[0-9]+}}(%rsp)47; SSE41-NEXT: movabsq $4295032833, %rax # imm = 0x10001000148; SSE41-NEXT: movq %rax, -{{[0-9]+}}(%rsp)49; SSE41-NEXT: movw $0, -{{[0-9]+}}(%rsp)50; SSE41-NEXT: movl $0, -{{[0-9]+}}(%rsp)51; SSE41-NEXT: movdqa {{.*#+}} xmm0 = [271,271,271,271,271,u,u,u]52; SSE41-NEXT: .p2align 453; SSE41-NEXT: .LBB0_1: # %forcond54; SSE41-NEXT: # =>This Inner Loop Header: Depth=155; SSE41-NEXT: movl -{{[0-9]+}}(%rsp), %eax56; SSE41-NEXT: cmpl -{{[0-9]+}}(%rsp), %eax57; SSE41-NEXT: jge .LBB0_358; SSE41-NEXT: # %bb.2: # %forbody59; SSE41-NEXT: # in Loop: Header=BB0_1 Depth=160; SSE41-NEXT: movslq -{{[0-9]+}}(%rsp), %rax61; SSE41-NEXT: movq -{{[0-9]+}}(%rsp), %rcx62; SSE41-NEXT: shlq $4, %rax63; SSE41-NEXT: movq -{{[0-9]+}}(%rsp), %rdx64; SSE41-NEXT: movdqa (%rdx,%rax), %xmm165; SSE41-NEXT: psubw %xmm0, %xmm166; SSE41-NEXT: movdqa %xmm1, %xmm267; SSE41-NEXT: psllw $2, %xmm268; SSE41-NEXT: paddw %xmm1, %xmm169; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm1[0],xmm2[1],xmm1[2,3,4,5,6,7]70; SSE41-NEXT: pextrw $4, %xmm1, 8(%rcx,%rax)71; SSE41-NEXT: movq %xmm2, (%rcx,%rax)72; SSE41-NEXT: incl -{{[0-9]+}}(%rsp)73; SSE41-NEXT: jmp .LBB0_174; SSE41-NEXT: .LBB0_3: # %afterfor75; SSE41-NEXT: retq76entry:77 %dst.addr = alloca ptr78 %src.addr = alloca ptr79 %n.addr = alloca i3280 %v = alloca <5 x i16>, align 1681 %i = alloca i32, align 482 store ptr %dst, ptr %dst.addr83 store ptr %src, ptr %src.addr84 store i32 %n, ptr %n.addr85 store <5 x i16> < i16 1, i16 1, i16 1, i16 0, i16 0 >, ptr %v86 store i32 0, ptr %i87 br label %forcond88 89forcond:90 %tmp = load i32, ptr %i91 %tmp1 = load i32, ptr %n.addr92 %cmp = icmp slt i32 %tmp, %tmp193 br i1 %cmp, label %forbody, label %afterfor94 95forbody:96 %tmp2 = load i32, ptr %i97 %tmp3 = load ptr, ptr %dst.addr98 %arrayidx = getelementptr <5 x i16>, ptr %tmp3, i32 %tmp299 %tmp4 = load i32, ptr %i100 %tmp5 = load ptr, ptr %src.addr101 %arrayidx6 = getelementptr <5 x i16>, ptr %tmp5, i32 %tmp4102 %tmp7 = load <5 x i16>, ptr %arrayidx6103 %sub = sub <5 x i16> %tmp7, < i16 271, i16 271, i16 271, i16 271, i16 271 >104 %mul = mul <5 x i16> %sub, < i16 2, i16 4, i16 2, i16 2, i16 2 >105 store <5 x i16> %mul, ptr %arrayidx106 br label %forinc107 108forinc:109 %tmp8 = load i32, ptr %i110 %inc = add i32 %tmp8, 1111 store i32 %inc, ptr %i112 br label %forcond113 114afterfor:115 ret void116}117 118