363 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-linux | FileCheck %s --check-prefix=CMOV3; RUN: llc < %s -mtriple=i686-unknown-linux | FileCheck %s --check-prefix=NOCMOV4 5target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"6 7; Test 2xCMOV patterns exposed after legalization.8; One way to do that is with (select (fcmp une/oeq)), which gets9; legalized to setp/setne.10 11define dso_local i32 @test_select_fcmp_oeq_i32(float %a, float %b, i32 %c, i32 %d) nounwind {12; CMOV-LABEL: test_select_fcmp_oeq_i32:13; CMOV: # %bb.0: # %entry14; CMOV-NEXT: movl %edi, %eax15; CMOV-NEXT: ucomiss %xmm1, %xmm016; CMOV-NEXT: cmovnel %esi, %eax17; CMOV-NEXT: cmovpl %esi, %eax18; CMOV-NEXT: retq19;20; NOCMOV-LABEL: test_select_fcmp_oeq_i32:21; NOCMOV: # %bb.0: # %entry22; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)23; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)24; NOCMOV-NEXT: fucompp25; NOCMOV-NEXT: fnstsw %ax26; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax27; NOCMOV-NEXT: sahf28; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %eax29; NOCMOV-NEXT: jne .LBB0_330; NOCMOV-NEXT: # %bb.1: # %entry31; NOCMOV-NEXT: jp .LBB0_332; NOCMOV-NEXT: # %bb.2: # %entry33; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %eax34; NOCMOV-NEXT: .LBB0_3: # %entry35; NOCMOV-NEXT: movl (%eax), %eax36; NOCMOV-NEXT: retl37entry:38 %cmp = fcmp oeq float %a, %b39 %r = select i1 %cmp, i32 %c, i32 %d40 ret i32 %r41}42 43define i64 @test_select_fcmp_oeq_i64(float %a, float %b, i64 %c, i64 %d) nounwind {44; CMOV-LABEL: test_select_fcmp_oeq_i64:45; CMOV: # %bb.0: # %entry46; CMOV-NEXT: movq %rdi, %rax47; CMOV-NEXT: ucomiss %xmm1, %xmm048; CMOV-NEXT: cmovneq %rsi, %rax49; CMOV-NEXT: cmovpq %rsi, %rax50; CMOV-NEXT: retq51;52; NOCMOV-LABEL: test_select_fcmp_oeq_i64:53; NOCMOV: # %bb.0: # %entry54; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)55; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)56; NOCMOV-NEXT: fucompp57; NOCMOV-NEXT: fnstsw %ax58; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax59; NOCMOV-NEXT: sahf60; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %ecx61; NOCMOV-NEXT: jne .LBB1_362; NOCMOV-NEXT: # %bb.1: # %entry63; NOCMOV-NEXT: jp .LBB1_364; NOCMOV-NEXT: # %bb.2: # %entry65; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %ecx66; NOCMOV-NEXT: .LBB1_3: # %entry67; NOCMOV-NEXT: movl (%ecx), %eax68; NOCMOV-NEXT: movl 4(%ecx), %edx69; NOCMOV-NEXT: retl70entry:71 %cmp = fcmp oeq float %a, %b72 %r = select i1 %cmp, i64 %c, i64 %d73 ret i64 %r74}75 76define i64 @test_select_fcmp_une_i64(float %a, float %b, i64 %c, i64 %d) nounwind {77; CMOV-LABEL: test_select_fcmp_une_i64:78; CMOV: # %bb.0: # %entry79; CMOV-NEXT: movq %rsi, %rax80; CMOV-NEXT: ucomiss %xmm1, %xmm081; CMOV-NEXT: cmovneq %rdi, %rax82; CMOV-NEXT: cmovpq %rdi, %rax83; CMOV-NEXT: retq84;85; NOCMOV-LABEL: test_select_fcmp_une_i64:86; NOCMOV: # %bb.0: # %entry87; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)88; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)89; NOCMOV-NEXT: fucompp90; NOCMOV-NEXT: fnstsw %ax91; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax92; NOCMOV-NEXT: sahf93; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %ecx94; NOCMOV-NEXT: jne .LBB2_395; NOCMOV-NEXT: # %bb.1: # %entry96; NOCMOV-NEXT: jp .LBB2_397; NOCMOV-NEXT: # %bb.2: # %entry98; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %ecx99; NOCMOV-NEXT: .LBB2_3: # %entry100; NOCMOV-NEXT: movl (%ecx), %eax101; NOCMOV-NEXT: movl 4(%ecx), %edx102; NOCMOV-NEXT: retl103entry:104 %cmp = fcmp une float %a, %b105 %r = select i1 %cmp, i64 %c, i64 %d106 ret i64 %r107}108 109define dso_local double @test_select_fcmp_oeq_f64(float %a, float %b, double %c, double %d) nounwind {110; CMOV-LABEL: test_select_fcmp_oeq_f64:111; CMOV: # %bb.0: # %entry112; CMOV-NEXT: ucomiss %xmm1, %xmm0113; CMOV-NEXT: jne .LBB3_3114; CMOV-NEXT: # %bb.1: # %entry115; CMOV-NEXT: jp .LBB3_3116; CMOV-NEXT: # %bb.2: # %entry117; CMOV-NEXT: movaps %xmm2, %xmm3118; CMOV-NEXT: .LBB3_3: # %entry119; CMOV-NEXT: movaps %xmm3, %xmm0120; CMOV-NEXT: retq121;122; NOCMOV-LABEL: test_select_fcmp_oeq_f64:123; NOCMOV: # %bb.0: # %entry124; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)125; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)126; NOCMOV-NEXT: fucompp127; NOCMOV-NEXT: fnstsw %ax128; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax129; NOCMOV-NEXT: sahf130; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %eax131; NOCMOV-NEXT: jne .LBB3_3132; NOCMOV-NEXT: # %bb.1: # %entry133; NOCMOV-NEXT: jp .LBB3_3134; NOCMOV-NEXT: # %bb.2: # %entry135; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %eax136; NOCMOV-NEXT: .LBB3_3: # %entry137; NOCMOV-NEXT: fldl (%eax)138; NOCMOV-NEXT: retl139entry:140 %cmp = fcmp oeq float %a, %b141 %r = select i1 %cmp, double %c, double %d142 ret double %r143}144 145define <4 x i32> @test_select_fcmp_oeq_v4i32(float %a, float %b, <4 x i32> %c, <4 x i32> %d) nounwind {146; CMOV-LABEL: test_select_fcmp_oeq_v4i32:147; CMOV: # %bb.0: # %entry148; CMOV-NEXT: ucomiss %xmm1, %xmm0149; CMOV-NEXT: jne .LBB4_3150; CMOV-NEXT: # %bb.1: # %entry151; CMOV-NEXT: jp .LBB4_3152; CMOV-NEXT: # %bb.2: # %entry153; CMOV-NEXT: movaps %xmm2, %xmm3154; CMOV-NEXT: .LBB4_3: # %entry155; CMOV-NEXT: movaps %xmm3, %xmm0156; CMOV-NEXT: retq157;158; NOCMOV-LABEL: test_select_fcmp_oeq_v4i32:159; NOCMOV: # %bb.0: # %entry160; NOCMOV-NEXT: pushl %edi161; NOCMOV-NEXT: pushl %esi162; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)163; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)164; NOCMOV-NEXT: fucompp165; NOCMOV-NEXT: fnstsw %ax166; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax167; NOCMOV-NEXT: sahf168; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %ecx169; NOCMOV-NEXT: jne .LBB4_3170; NOCMOV-NEXT: # %bb.1: # %entry171; NOCMOV-NEXT: jp .LBB4_3172; NOCMOV-NEXT: # %bb.2: # %entry173; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %ecx174; NOCMOV-NEXT: .LBB4_3: # %entry175; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %edx176; NOCMOV-NEXT: jne .LBB4_6177; NOCMOV-NEXT: # %bb.4: # %entry178; NOCMOV-NEXT: jp .LBB4_6179; NOCMOV-NEXT: # %bb.5: # %entry180; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %edx181; NOCMOV-NEXT: .LBB4_6: # %entry182; NOCMOV-NEXT: movl {{[0-9]+}}(%esp), %eax183; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %esi184; NOCMOV-NEXT: jne .LBB4_9185; NOCMOV-NEXT: # %bb.7: # %entry186; NOCMOV-NEXT: jp .LBB4_9187; NOCMOV-NEXT: # %bb.8: # %entry188; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %esi189; NOCMOV-NEXT: .LBB4_9: # %entry190; NOCMOV-NEXT: movl (%ecx), %ecx191; NOCMOV-NEXT: movl (%edx), %edx192; NOCMOV-NEXT: movl (%esi), %esi193; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %edi194; NOCMOV-NEXT: jne .LBB4_12195; NOCMOV-NEXT: # %bb.10: # %entry196; NOCMOV-NEXT: jp .LBB4_12197; NOCMOV-NEXT: # %bb.11: # %entry198; NOCMOV-NEXT: leal {{[0-9]+}}(%esp), %edi199; NOCMOV-NEXT: .LBB4_12: # %entry200; NOCMOV-NEXT: movl (%edi), %edi201; NOCMOV-NEXT: movl %edi, 12(%eax)202; NOCMOV-NEXT: movl %esi, 8(%eax)203; NOCMOV-NEXT: movl %edx, 4(%eax)204; NOCMOV-NEXT: movl %ecx, (%eax)205; NOCMOV-NEXT: popl %esi206; NOCMOV-NEXT: popl %edi207; NOCMOV-NEXT: retl $4208entry:209 %cmp = fcmp oeq float %a, %b210 %r = select i1 %cmp, <4 x i32> %c, <4 x i32> %d211 ret <4 x i32> %r212}213 214; Also make sure we catch the original code-sequence of interest:215 216define dso_local float @test_zext_fcmp_une(float %a, float %b) nounwind {217; CMOV-LABEL: test_zext_fcmp_une:218; CMOV: # %bb.0: # %entry219; CMOV-NEXT: cmpneqss %xmm1, %xmm0220; CMOV-NEXT: movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]221; CMOV-NEXT: andps %xmm1, %xmm0222; CMOV-NEXT: retq223;224; NOCMOV-LABEL: test_zext_fcmp_une:225; NOCMOV: # %bb.0: # %entry226; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)227; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)228; NOCMOV-NEXT: fucompp229; NOCMOV-NEXT: fnstsw %ax230; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax231; NOCMOV-NEXT: sahf232; NOCMOV-NEXT: fld1233; NOCMOV-NEXT: fldz234; NOCMOV-NEXT: jne .LBB5_1235; NOCMOV-NEXT: # %bb.2: # %entry236; NOCMOV-NEXT: jp .LBB5_5237; NOCMOV-NEXT: # %bb.3: # %entry238; NOCMOV-NEXT: fstp %st(1)239; NOCMOV-NEXT: jmp .LBB5_4240; NOCMOV-NEXT: .LBB5_1:241; NOCMOV-NEXT: fstp %st(0)242; NOCMOV-NEXT: .LBB5_4: # %entry243; NOCMOV-NEXT: fldz244; NOCMOV-NEXT: .LBB5_5: # %entry245; NOCMOV-NEXT: fstp %st(0)246; NOCMOV-NEXT: retl247entry:248 %cmp = fcmp une float %a, %b249 %conv1 = zext i1 %cmp to i32250 %conv2 = sitofp i32 %conv1 to float251 ret float %conv2252}253 254define dso_local float @test_zext_fcmp_oeq(float %a, float %b) nounwind {255; CMOV-LABEL: test_zext_fcmp_oeq:256; CMOV: # %bb.0: # %entry257; CMOV-NEXT: cmpeqss %xmm1, %xmm0258; CMOV-NEXT: movss {{.*#+}} xmm1 = [1.0E+0,0.0E+0,0.0E+0,0.0E+0]259; CMOV-NEXT: andps %xmm1, %xmm0260; CMOV-NEXT: retq261;262; NOCMOV-LABEL: test_zext_fcmp_oeq:263; NOCMOV: # %bb.0: # %entry264; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)265; NOCMOV-NEXT: flds {{[0-9]+}}(%esp)266; NOCMOV-NEXT: fucompp267; NOCMOV-NEXT: fnstsw %ax268; NOCMOV-NEXT: # kill: def $ah killed $ah killed $ax269; NOCMOV-NEXT: sahf270; NOCMOV-NEXT: fldz271; NOCMOV-NEXT: fld1272; NOCMOV-NEXT: jne .LBB6_1273; NOCMOV-NEXT: # %bb.2: # %entry274; NOCMOV-NEXT: jp .LBB6_5275; NOCMOV-NEXT: # %bb.3: # %entry276; NOCMOV-NEXT: fstp %st(1)277; NOCMOV-NEXT: jmp .LBB6_4278; NOCMOV-NEXT: .LBB6_1:279; NOCMOV-NEXT: fstp %st(0)280; NOCMOV-NEXT: .LBB6_4: # %entry281; NOCMOV-NEXT: fldz282; NOCMOV-NEXT: .LBB6_5: # %entry283; NOCMOV-NEXT: fstp %st(0)284; NOCMOV-NEXT: retl285entry:286 %cmp = fcmp oeq float %a, %b287 %conv1 = zext i1 %cmp to i32288 %conv2 = sitofp i32 %conv1 to float289 ret float %conv2290}291 292attributes #0 = { nounwind }293 294@g8 = dso_local global i8 0295 296; The following test failed because llvm had a bug where a structure like:297;298; %12 = CMOV_GR8 %7, %11 ... (lt)299; %13 = CMOV_GR8 %12, %11 ... (gt)300;301; was lowered to:302;303; The first two cmovs got expanded to:304; %bb.0:305; JCC_1 %bb.9, 12306; %bb.7:307; JCC_1 %bb.9, 15308; %bb.8:309; %bb.9:310; %12 = phi(%7, %bb.8, %11, %bb.0, %12, %bb.7)311; %13 = COPY %12312; Which was invalid as %12 is not the same value as %13313 314define dso_local void @no_cascade_opt(i32 %v0, i32 %v1, i32 %v2, i32 %v3) nounwind {315; CMOV-LABEL: no_cascade_opt:316; CMOV: # %bb.0: # %entry317; CMOV-NEXT: cmpl %edx, %esi318; CMOV-NEXT: movl $20, %eax319; CMOV-NEXT: cmovll %eax, %ecx320; CMOV-NEXT: cmovlel %ecx, %eax321; CMOV-NEXT: testl %edi, %edi322; CMOV-NEXT: cmovnel %ecx, %eax323; CMOV-NEXT: movb %al, g8(%rip)324; CMOV-NEXT: retq325;326; NOCMOV-LABEL: no_cascade_opt:327; NOCMOV: # %bb.0: # %entry328; NOCMOV-NEXT: movl {{[0-9]+}}(%esp), %eax329; NOCMOV-NEXT: cmpl {{[0-9]+}}(%esp), %eax330; NOCMOV-NEXT: movb $20, %al331; NOCMOV-NEXT: movb $20, %cl332; NOCMOV-NEXT: jge .LBB7_1333; NOCMOV-NEXT: # %bb.2: # %entry334; NOCMOV-NEXT: jle .LBB7_3335; NOCMOV-NEXT: .LBB7_4: # %entry336; NOCMOV-NEXT: cmpl $0, {{[0-9]+}}(%esp)337; NOCMOV-NEXT: jne .LBB7_5338; NOCMOV-NEXT: .LBB7_6: # %entry339; NOCMOV-NEXT: movb %al, g8340; NOCMOV-NEXT: retl341; NOCMOV-NEXT: .LBB7_1: # %entry342; NOCMOV-NEXT: movzbl {{[0-9]+}}(%esp), %ecx343; NOCMOV-NEXT: jg .LBB7_4344; NOCMOV-NEXT: .LBB7_3: # %entry345; NOCMOV-NEXT: movl %ecx, %eax346; NOCMOV-NEXT: cmpl $0, {{[0-9]+}}(%esp)347; NOCMOV-NEXT: je .LBB7_6348; NOCMOV-NEXT: .LBB7_5: # %entry349; NOCMOV-NEXT: movl %ecx, %eax350; NOCMOV-NEXT: movb %al, g8351; NOCMOV-NEXT: retl352entry:353 %c0 = icmp eq i32 %v0, 0354 %c1 = icmp slt i32 %v1, %v2355 %c2 = icmp sgt i32 %v1, %v2356 %trunc = trunc i32 %v3 to i8357 %sel0 = select i1 %c1, i8 20, i8 %trunc358 %sel1 = select i1 %c2, i8 20, i8 %sel0359 %sel2 = select i1 %c0, i8 %sel1, i8 %sel0360 store volatile i8 %sel2, ptr @g8361 ret void362}363