221 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s3 4define <8 x i32> @cmp00(<8 x float> %a, <8 x float> %b) nounwind {5; CHECK-LABEL: cmp00:6; CHECK: # %bb.0:7; CHECK-NEXT: vcmpltps %ymm1, %ymm0, %ymm08; CHECK-NEXT: retq9 %bincmp = fcmp olt <8 x float> %a, %b10 %s = sext <8 x i1> %bincmp to <8 x i32>11 ret <8 x i32> %s12}13 14define <4 x i64> @cmp01(<4 x double> %a, <4 x double> %b) nounwind {15; CHECK-LABEL: cmp01:16; CHECK: # %bb.0:17; CHECK-NEXT: vcmpltpd %ymm1, %ymm0, %ymm018; CHECK-NEXT: retq19 %bincmp = fcmp olt <4 x double> %a, %b20 %s = sext <4 x i1> %bincmp to <4 x i64>21 ret <4 x i64> %s22}23 24declare void @scale() nounwind25 26define void @render(double %a0) nounwind {27; CHECK-LABEL: render:28; CHECK: # %bb.0: # %entry29; CHECK-NEXT: pushq %rbx30; CHECK-NEXT: subq $16, %rsp31; CHECK-NEXT: vmovsd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill32; CHECK-NEXT: xorl %eax, %eax33; CHECK-NEXT: testb %al, %al34; CHECK-NEXT: jne .LBB2_535; CHECK-NEXT: # %bb.1: # %for.cond5.preheader36; CHECK-NEXT: movb $1, %bl37; CHECK-NEXT: .p2align 438; CHECK-NEXT: .LBB2_2: # %for.cond539; CHECK-NEXT: # =>This Inner Loop Header: Depth=140; CHECK-NEXT: testb %bl, %bl41; CHECK-NEXT: jne .LBB2_242; CHECK-NEXT: # %bb.3: # %for.body33.preheader43; CHECK-NEXT: # in Loop: Header=BB2_2 Depth=144; CHECK-NEXT: vmovsd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 8-byte Reload45; CHECK-NEXT: # xmm0 = mem[0],zero46; CHECK-NEXT: vxorpd %xmm1, %xmm1, %xmm147; CHECK-NEXT: vucomisd %xmm1, %xmm048; CHECK-NEXT: jne .LBB2_449; CHECK-NEXT: jnp .LBB2_250; CHECK-NEXT: .LBB2_4: # %if.then51; CHECK-NEXT: # in Loop: Header=BB2_2 Depth=152; CHECK-NEXT: callq scale@PLT53; CHECK-NEXT: jmp .LBB2_254; CHECK-NEXT: .LBB2_5: # %for.end5255; CHECK-NEXT: addq $16, %rsp56; CHECK-NEXT: popq %rbx57; CHECK-NEXT: retq58entry:59 br i1 undef, label %for.cond5, label %for.end5260 61for.cond5:62 %or.cond = and i1 undef, false63 br i1 %or.cond, label %for.body33, label %for.cond564 65for.cond30:66 br i1 false, label %for.body33, label %for.cond567 68for.body33:69 %tobool = fcmp une double %a0, 0.000000e+0070 br i1 %tobool, label %if.then, label %for.cond3071 72if.then:73 call void @scale()74 br label %for.cond3075 76for.end52:77 ret void78}79 80define <8 x i32> @int256_cmp(<8 x i32> %i, <8 x i32> %j) nounwind {81; CHECK-LABEL: int256_cmp:82; CHECK: # %bb.0:83; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm284; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm385; CHECK-NEXT: vpcmpgtd %xmm2, %xmm3, %xmm286; CHECK-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm087; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm088; CHECK-NEXT: retq89 %bincmp = icmp slt <8 x i32> %i, %j90 %x = sext <8 x i1> %bincmp to <8 x i32>91 ret <8 x i32> %x92}93 94define <4 x i64> @v4i64_cmp(<4 x i64> %i, <4 x i64> %j) nounwind {95; CHECK-LABEL: v4i64_cmp:96; CHECK: # %bb.0:97; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm298; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm399; CHECK-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm2100; CHECK-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0101; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0102; CHECK-NEXT: retq103 %bincmp = icmp slt <4 x i64> %i, %j104 %x = sext <4 x i1> %bincmp to <4 x i64>105 ret <4 x i64> %x106}107 108define <16 x i16> @v16i16_cmp(<16 x i16> %i, <16 x i16> %j) nounwind {109; CHECK-LABEL: v16i16_cmp:110; CHECK: # %bb.0:111; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm2112; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm3113; CHECK-NEXT: vpcmpgtw %xmm2, %xmm3, %xmm2114; CHECK-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0115; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0116; CHECK-NEXT: retq117 %bincmp = icmp slt <16 x i16> %i, %j118 %x = sext <16 x i1> %bincmp to <16 x i16>119 ret <16 x i16> %x120}121 122define <32 x i8> @v32i8_cmp(<32 x i8> %i, <32 x i8> %j) nounwind {123; CHECK-LABEL: v32i8_cmp:124; CHECK: # %bb.0:125; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm2126; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm3127; CHECK-NEXT: vpcmpgtb %xmm2, %xmm3, %xmm2128; CHECK-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0129; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0130; CHECK-NEXT: retq131 %bincmp = icmp slt <32 x i8> %i, %j132 %x = sext <32 x i1> %bincmp to <32 x i8>133 ret <32 x i8> %x134}135 136define <8 x i32> @int256_cmpeq(<8 x i32> %i, <8 x i32> %j) nounwind {137; CHECK-LABEL: int256_cmpeq:138; CHECK: # %bb.0:139; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm2140; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm3141; CHECK-NEXT: vpcmpeqd %xmm2, %xmm3, %xmm2142; CHECK-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0143; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0144; CHECK-NEXT: retq145 %bincmp = icmp eq <8 x i32> %i, %j146 %x = sext <8 x i1> %bincmp to <8 x i32>147 ret <8 x i32> %x148}149 150define <4 x i64> @v4i64_cmpeq(<4 x i64> %i, <4 x i64> %j) nounwind {151; CHECK-LABEL: v4i64_cmpeq:152; CHECK: # %bb.0:153; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm2154; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm3155; CHECK-NEXT: vpcmpeqq %xmm2, %xmm3, %xmm2156; CHECK-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0157; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0158; CHECK-NEXT: retq159 %bincmp = icmp eq <4 x i64> %i, %j160 %x = sext <4 x i1> %bincmp to <4 x i64>161 ret <4 x i64> %x162}163 164define <16 x i16> @v16i16_cmpeq(<16 x i16> %i, <16 x i16> %j) nounwind {165; CHECK-LABEL: v16i16_cmpeq:166; CHECK: # %bb.0:167; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm2168; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm3169; CHECK-NEXT: vpcmpeqw %xmm2, %xmm3, %xmm2170; CHECK-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0171; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0172; CHECK-NEXT: retq173 %bincmp = icmp eq <16 x i16> %i, %j174 %x = sext <16 x i1> %bincmp to <16 x i16>175 ret <16 x i16> %x176}177 178define <32 x i8> @v32i8_cmpeq(<32 x i8> %i, <32 x i8> %j) nounwind {179; CHECK-LABEL: v32i8_cmpeq:180; CHECK: # %bb.0:181; CHECK-NEXT: vextractf128 $1, %ymm1, %xmm2182; CHECK-NEXT: vextractf128 $1, %ymm0, %xmm3183; CHECK-NEXT: vpcmpeqb %xmm2, %xmm3, %xmm2184; CHECK-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0185; CHECK-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0186; CHECK-NEXT: retq187 %bincmp = icmp eq <32 x i8> %i, %j188 %x = sext <32 x i1> %bincmp to <32 x i8>189 ret <32 x i8> %x190}191 192;; Scalar comparison193 194define i32 @scalarcmpA(double %a0) uwtable ssp {195; CHECK-LABEL: scalarcmpA:196; CHECK: # %bb.0:197; CHECK-NEXT: vxorpd %xmm1, %xmm1, %xmm1198; CHECK-NEXT: vcmpeqsd %xmm1, %xmm0, %xmm0199; CHECK-NEXT: vmovq %xmm0, %rax200; CHECK-NEXT: andl $1, %eax201; CHECK-NEXT: # kill: def $eax killed $eax killed $rax202; CHECK-NEXT: retq203 %cmp29 = fcmp oeq double %a0, 0.000000e+00204 %res = zext i1 %cmp29 to i32205 ret i32 %res206}207 208define i32 @scalarcmpB(float %a0) uwtable ssp {209; CHECK-LABEL: scalarcmpB:210; CHECK: # %bb.0:211; CHECK-NEXT: vxorps %xmm1, %xmm1, %xmm1212; CHECK-NEXT: vcmpeqss %xmm1, %xmm0, %xmm0213; CHECK-NEXT: vmovd %xmm0, %eax214; CHECK-NEXT: andl $1, %eax215; CHECK-NEXT: retq216 %cmp29 = fcmp oeq float %a0, 0.000000e+00217 %res = zext i1 %cmp29 to i32218 ret i32 %res219}220 221