256 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i386-unknown-unknown -mattr=+sse -O3 | FileCheck %s --check-prefixes=CHECK,X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-sse2,+sse -O3 | FileCheck %s --check-prefixes=CHECK,X644 5; Tests for SSE1 and below, without SSE2+.6 7; PR79938;define <4 x i32> @test3(<4 x i16> %a) nounwind {9; %c = sext <4 x i16> %a to <4 x i32> ; <<4 x i32>> [#uses=1]10; ret <4 x i32> %c11;}12 13; This should not emit shuffles to populate the top 2 elements of the 4-element14; vector that this ends up returning.15; rdar://836841416define <2 x float> @test4(<2 x float> %A, <2 x float> %B) nounwind {17; CHECK-LABEL: test4:18; CHECK: # %bb.0: # %entry19; CHECK-NEXT: movaps %xmm0, %xmm220; CHECK-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1],xmm0[1,1]21; CHECK-NEXT: addss %xmm1, %xmm022; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1,1,1]23; CHECK-NEXT: subss %xmm1, %xmm224; CHECK-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]25; CHECK-NEXT: ret{{[l|q]}}26entry:27 %tmp7 = extractelement <2 x float> %A, i32 028 %tmp5 = extractelement <2 x float> %A, i32 129 %tmp3 = extractelement <2 x float> %B, i32 030 %tmp1 = extractelement <2 x float> %B, i32 131 %add.r = fadd float %tmp7, %tmp332 %add.i = fsub float %tmp5, %tmp133 %tmp11 = insertelement <2 x float> undef, float %add.r, i32 034 %tmp9 = insertelement <2 x float> %tmp11, float %add.i, i32 135 ret <2 x float> %tmp936}37 38; We used to get stuck in type legalization for this example when lowering the39; vselect. With SSE1 v4f32 is a legal type but v4i1 (or any vector integer type)40; is not. We used to ping pong between splitting the vselect for the v4i41; condition operand and widening the resulting vselect for the v4f32 result.42; PR1803643 44define <4 x float> @vselect(ptr%p, <4 x i32> %q) {45; X86-LABEL: vselect:46; X86: # %bb.0: # %entry47; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)48; X86-NEXT: xorps %xmm0, %xmm049; X86-NEXT: je .LBB1_150; X86-NEXT: # %bb.2: # %entry51; X86-NEXT: xorps %xmm1, %xmm152; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)53; X86-NEXT: jne .LBB1_554; X86-NEXT: .LBB1_4:55; X86-NEXT: movss {{.*#+}} xmm2 = [3.0E+0,0.0E+0,0.0E+0,0.0E+0]56; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)57; X86-NEXT: jne .LBB1_858; X86-NEXT: .LBB1_7:59; X86-NEXT: movss {{.*#+}} xmm3 = [4.0E+0,0.0E+0,0.0E+0,0.0E+0]60; X86-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]61; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)62; X86-NEXT: je .LBB1_1063; X86-NEXT: jmp .LBB1_1164; X86-NEXT: .LBB1_1:65; X86-NEXT: movss {{.*#+}} xmm1 = [2.0E+0,0.0E+0,0.0E+0,0.0E+0]66; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)67; X86-NEXT: je .LBB1_468; X86-NEXT: .LBB1_5: # %entry69; X86-NEXT: xorps %xmm2, %xmm270; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)71; X86-NEXT: je .LBB1_772; X86-NEXT: .LBB1_8: # %entry73; X86-NEXT: xorps %xmm3, %xmm374; X86-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]75; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)76; X86-NEXT: jne .LBB1_1177; X86-NEXT: .LBB1_10:78; X86-NEXT: movss {{.*#+}} xmm0 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]79; X86-NEXT: .LBB1_11: # %entry80; X86-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]81; X86-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]82; X86-NEXT: retl83;84; X64-LABEL: vselect:85; X64: # %bb.0: # %entry86; X64-NEXT: testl %edx, %edx87; X64-NEXT: xorps %xmm0, %xmm088; X64-NEXT: je .LBB1_189; X64-NEXT: # %bb.2: # %entry90; X64-NEXT: xorps %xmm1, %xmm191; X64-NEXT: testl %ecx, %ecx92; X64-NEXT: jne .LBB1_593; X64-NEXT: .LBB1_4:94; X64-NEXT: movss {{.*#+}} xmm2 = [3.0E+0,0.0E+0,0.0E+0,0.0E+0]95; X64-NEXT: testl %r8d, %r8d96; X64-NEXT: jne .LBB1_897; X64-NEXT: .LBB1_7:98; X64-NEXT: movss {{.*#+}} xmm3 = [4.0E+0,0.0E+0,0.0E+0,0.0E+0]99; X64-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]100; X64-NEXT: testl %esi, %esi101; X64-NEXT: je .LBB1_10102; X64-NEXT: jmp .LBB1_11103; X64-NEXT: .LBB1_1:104; X64-NEXT: movss {{.*#+}} xmm1 = [2.0E+0,0.0E+0,0.0E+0,0.0E+0]105; X64-NEXT: testl %ecx, %ecx106; X64-NEXT: je .LBB1_4107; X64-NEXT: .LBB1_5: # %entry108; X64-NEXT: xorps %xmm2, %xmm2109; X64-NEXT: testl %r8d, %r8d110; X64-NEXT: je .LBB1_7111; X64-NEXT: .LBB1_8: # %entry112; X64-NEXT: xorps %xmm3, %xmm3113; X64-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]114; X64-NEXT: testl %esi, %esi115; X64-NEXT: jne .LBB1_11116; X64-NEXT: .LBB1_10:117; X64-NEXT: movss {{.*#+}} xmm0 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]118; X64-NEXT: .LBB1_11: # %entry119; X64-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]120; X64-NEXT: movlhps {{.*#+}} xmm0 = xmm0[0],xmm2[0]121; X64-NEXT: retq122entry:123 %a1 = icmp eq <4 x i32> %q, zeroinitializer124 %a14 = select <4 x i1> %a1, <4 x float> <float 1.000000e+00, float 2.000000e+00, float 3.000000e+00, float 4.000000e+0> , <4 x float> zeroinitializer125 ret <4 x float> %a14126}127 128; v4i32 isn't legal for SSE1, but this should be cmpps.129 130define <4 x float> @PR28044(<4 x float> %a0, <4 x float> %a1) nounwind {131; CHECK-LABEL: PR28044:132; CHECK: # %bb.0:133; CHECK-NEXT: cmpeqps %xmm1, %xmm0134; CHECK-NEXT: ret{{[l|q]}}135 %cmp = fcmp oeq <4 x float> %a0, %a1136 %sext = sext <4 x i1> %cmp to <4 x i32>137 %res = bitcast <4 x i32> %sext to <4 x float>138 ret <4 x float> %res139}140 141; Don't crash trying to do the impossible: an integer vector comparison doesn't exist, so we must scalarize.142; https://llvm.org/bugs/show_bug.cgi?id=30512143 144define <4 x i32> @PR30512(<4 x i32> %x, <4 x i32> %y) nounwind {145; X86-LABEL: PR30512:146; X86: # %bb.0:147; X86-NEXT: pushl %ebx148; X86-NEXT: pushl %edi149; X86-NEXT: pushl %esi150; X86-NEXT: subl $16, %esp151; X86-NEXT: movl {{[0-9]+}}(%esp), %eax152; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx153; X86-NEXT: movl {{[0-9]+}}(%esp), %edx154; X86-NEXT: movl {{[0-9]+}}(%esp), %esi155; X86-NEXT: movl {{[0-9]+}}(%esp), %edi156; X86-NEXT: xorl %ebx, %ebx157; X86-NEXT: cmpl {{[0-9]+}}(%esp), %edi158; X86-NEXT: sete %bl159; X86-NEXT: negl %ebx160; X86-NEXT: movl %ebx, {{[0-9]+}}(%esp)161; X86-NEXT: xorl %ebx, %ebx162; X86-NEXT: cmpl {{[0-9]+}}(%esp), %esi163; X86-NEXT: sete %bl164; X86-NEXT: negl %ebx165; X86-NEXT: movl %ebx, {{[0-9]+}}(%esp)166; X86-NEXT: xorl %ebx, %ebx167; X86-NEXT: cmpl {{[0-9]+}}(%esp), %edx168; X86-NEXT: sete %bl169; X86-NEXT: negl %ebx170; X86-NEXT: movl %ebx, {{[0-9]+}}(%esp)171; X86-NEXT: xorl %edx, %edx172; X86-NEXT: cmpl {{[0-9]+}}(%esp), %ecx173; X86-NEXT: sete %dl174; X86-NEXT: negl %edx175; X86-NEXT: movl %edx, (%esp)176; X86-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero177; X86-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero178; X86-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]179; X86-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero180; X86-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero181; X86-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]182; X86-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]183; X86-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm2184; X86-NEXT: movaps %xmm2, (%eax)185; X86-NEXT: addl $16, %esp186; X86-NEXT: popl %esi187; X86-NEXT: popl %edi188; X86-NEXT: popl %ebx189; X86-NEXT: retl $4190;191; X64-LABEL: PR30512:192; X64: # %bb.0:193; X64-NEXT: movq %rdi, %rax194; X64-NEXT: xorl %edi, %edi195; X64-NEXT: cmpl {{[0-9]+}}(%rsp), %r8d196; X64-NEXT: sete %dil197; X64-NEXT: negl %edi198; X64-NEXT: movl %edi, -{{[0-9]+}}(%rsp)199; X64-NEXT: xorl %edi, %edi200; X64-NEXT: cmpl {{[0-9]+}}(%rsp), %ecx201; X64-NEXT: sete %dil202; X64-NEXT: negl %edi203; X64-NEXT: movl %edi, -{{[0-9]+}}(%rsp)204; X64-NEXT: xorl %ecx, %ecx205; X64-NEXT: cmpl {{[0-9]+}}(%rsp), %edx206; X64-NEXT: sete %cl207; X64-NEXT: negl %ecx208; X64-NEXT: movl %ecx, -{{[0-9]+}}(%rsp)209; X64-NEXT: xorl %ecx, %ecx210; X64-NEXT: cmpl %r9d, %esi211; X64-NEXT: sete %cl212; X64-NEXT: negl %ecx213; X64-NEXT: movl %ecx, -{{[0-9]+}}(%rsp)214; X64-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero215; X64-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero216; X64-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]217; X64-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero218; X64-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero219; X64-NEXT: unpcklps {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]220; X64-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm1[0]221; X64-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2222; X64-NEXT: movaps %xmm2, (%rax)223; X64-NEXT: retq224 %cmp = icmp eq <4 x i32> %x, %y225 %zext = zext <4 x i1> %cmp to <4 x i32>226 ret <4 x i32> %zext227}228 229; Fragile test warning - we need to induce the generation of a vselect230; post-legalization to cause the crash seen in:231; https://llvm.org/bugs/show_bug.cgi?id=31672232; Is there a way to do that without an unsafe/fast sqrt intrinsic call?233;234; We now no longer try to lower sqrt using rsqrt with SSE1 only as the235; v4i32 vselect mentioned above should never have been created. We ended up236; scalarizing it anyway.237 238define <2 x float> @PR31672() #0 {239; X86-LABEL: PR31672:240; X86: # %bb.0:241; X86-NEXT: sqrtps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0242; X86-NEXT: retl243;244; X64-LABEL: PR31672:245; X64: # %bb.0:246; X64-NEXT: sqrtps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0247; X64-NEXT: retq248 %t0 = call fast <2 x float> @llvm.sqrt.v2f32(<2 x float> <float 42.0, float 3.0>)249 ret <2 x float> %t0250}251 252declare <2 x float> @llvm.sqrt.v2f32(<2 x float>) #1253 254attributes #0 = { nounwind }255 256