1078 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefixes=X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=-slow-incdec | FileCheck %s --check-prefixes=X64,X64-FASTINC4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+slow-incdec | FileCheck %s --check-prefixes=X64,X64-SLOWINC5 6; Select of constants: control flow / conditional moves can always be replaced by logic+math (but may not be worth it?).7; Test the zeroext/signext variants of each pattern to see if that makes a difference.8 9; select Cond, 0, 1 --> zext (!Cond)10 11define i32 @select_0_or_1(i1 %cond) {12; X86-LABEL: select_0_or_1:13; X86: # %bb.0:14; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax15; X86-NEXT: notb %al16; X86-NEXT: movzbl %al, %eax17; X86-NEXT: andl $1, %eax18; X86-NEXT: retl19;20; X64-LABEL: select_0_or_1:21; X64: # %bb.0:22; X64-NEXT: notb %dil23; X64-NEXT: movzbl %dil, %eax24; X64-NEXT: andl $1, %eax25; X64-NEXT: retq26 %sel = select i1 %cond, i32 0, i32 127 ret i32 %sel28}29 30define i32 @select_0_or_1_zeroext(i1 zeroext %cond) {31; X86-LABEL: select_0_or_1_zeroext:32; X86: # %bb.0:33; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax34; X86-NEXT: xorb $1, %al35; X86-NEXT: movzbl %al, %eax36; X86-NEXT: retl37;38; X64-LABEL: select_0_or_1_zeroext:39; X64: # %bb.0:40; X64-NEXT: xorb $1, %dil41; X64-NEXT: movzbl %dil, %eax42; X64-NEXT: retq43 %sel = select i1 %cond, i32 0, i32 144 ret i32 %sel45}46 47define i32 @select_0_or_1_signext(i1 signext %cond) {48; X86-LABEL: select_0_or_1_signext:49; X86: # %bb.0:50; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax51; X86-NEXT: notb %al52; X86-NEXT: movzbl %al, %eax53; X86-NEXT: andl $1, %eax54; X86-NEXT: retl55;56; X64-LABEL: select_0_or_1_signext:57; X64: # %bb.0:58; X64-NEXT: notb %dil59; X64-NEXT: movzbl %dil, %eax60; X64-NEXT: andl $1, %eax61; X64-NEXT: retq62 %sel = select i1 %cond, i32 0, i32 163 ret i32 %sel64}65 66; select Cond, 1, 0 --> zext (Cond)67 68define i32 @select_1_or_0(i1 %cond) {69; X86-LABEL: select_1_or_0:70; X86: # %bb.0:71; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax72; X86-NEXT: andl $1, %eax73; X86-NEXT: retl74;75; X64-LABEL: select_1_or_0:76; X64: # %bb.0:77; X64-NEXT: movl %edi, %eax78; X64-NEXT: andl $1, %eax79; X64-NEXT: retq80 %sel = select i1 %cond, i32 1, i32 081 ret i32 %sel82}83 84define i32 @select_1_or_0_zeroext(i1 zeroext %cond) {85; X86-LABEL: select_1_or_0_zeroext:86; X86: # %bb.0:87; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax88; X86-NEXT: retl89;90; X64-LABEL: select_1_or_0_zeroext:91; X64: # %bb.0:92; X64-NEXT: movl %edi, %eax93; X64-NEXT: retq94 %sel = select i1 %cond, i32 1, i32 095 ret i32 %sel96}97 98define i32 @select_1_or_0_signext(i1 signext %cond) {99; X86-LABEL: select_1_or_0_signext:100; X86: # %bb.0:101; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax102; X86-NEXT: andl $1, %eax103; X86-NEXT: retl104;105; X64-LABEL: select_1_or_0_signext:106; X64: # %bb.0:107; X64-NEXT: movl %edi, %eax108; X64-NEXT: andl $1, %eax109; X64-NEXT: retq110 %sel = select i1 %cond, i32 1, i32 0111 ret i32 %sel112}113 114; select Cond, 0, -1 --> sext (!Cond)115 116define i32 @select_0_or_neg1(i1 %cond) {117; X86-LABEL: select_0_or_neg1:118; X86: # %bb.0:119; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax120; X86-NEXT: andl $1, %eax121; X86-NEXT: decl %eax122; X86-NEXT: retl123;124; X64-LABEL: select_0_or_neg1:125; X64: # %bb.0:126; X64-NEXT: # kill: def $edi killed $edi def $rdi127; X64-NEXT: andl $1, %edi128; X64-NEXT: leal -1(%rdi), %eax129; X64-NEXT: retq130 %sel = select i1 %cond, i32 0, i32 -1131 ret i32 %sel132}133 134define i32 @select_0_or_neg1_zeroext(i1 zeroext %cond) {135; X86-LABEL: select_0_or_neg1_zeroext:136; X86: # %bb.0:137; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax138; X86-NEXT: decl %eax139; X86-NEXT: retl140;141; X64-LABEL: select_0_or_neg1_zeroext:142; X64: # %bb.0:143; X64-NEXT: # kill: def $edi killed $edi def $rdi144; X64-NEXT: leal -1(%rdi), %eax145; X64-NEXT: retq146 %sel = select i1 %cond, i32 0, i32 -1147 ret i32 %sel148}149 150define i32 @select_0_or_neg1_signext(i1 signext %cond) {151; X86-LABEL: select_0_or_neg1_signext:152; X86: # %bb.0:153; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax154; X86-NEXT: andl $1, %eax155; X86-NEXT: decl %eax156; X86-NEXT: retl157;158; X64-LABEL: select_0_or_neg1_signext:159; X64: # %bb.0:160; X64-NEXT: movl %edi, %eax161; X64-NEXT: notl %eax162; X64-NEXT: retq163 %sel = select i1 %cond, i32 0, i32 -1164 ret i32 %sel165}166 167; select Cond, -1, 0 --> sext (Cond)168 169define i32 @select_neg1_or_0(i1 %cond) {170; X86-LABEL: select_neg1_or_0:171; X86: # %bb.0:172; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax173; X86-NEXT: andl $1, %eax174; X86-NEXT: negl %eax175; X86-NEXT: retl176;177; X64-LABEL: select_neg1_or_0:178; X64: # %bb.0:179; X64-NEXT: movl %edi, %eax180; X64-NEXT: andl $1, %eax181; X64-NEXT: negl %eax182; X64-NEXT: retq183 %sel = select i1 %cond, i32 -1, i32 0184 ret i32 %sel185}186 187define i32 @select_neg1_or_0_zeroext(i1 zeroext %cond) {188; X86-LABEL: select_neg1_or_0_zeroext:189; X86: # %bb.0:190; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax191; X86-NEXT: negl %eax192; X86-NEXT: retl193;194; X64-LABEL: select_neg1_or_0_zeroext:195; X64: # %bb.0:196; X64-NEXT: movl %edi, %eax197; X64-NEXT: negl %eax198; X64-NEXT: retq199 %sel = select i1 %cond, i32 -1, i32 0200 ret i32 %sel201}202 203define i32 @select_neg1_or_0_signext(i1 signext %cond) {204; X86-LABEL: select_neg1_or_0_signext:205; X86: # %bb.0:206; X86-NEXT: movsbl {{[0-9]+}}(%esp), %eax207; X86-NEXT: retl208;209; X64-LABEL: select_neg1_or_0_signext:210; X64: # %bb.0:211; X64-NEXT: movl %edi, %eax212; X64-NEXT: retq213 %sel = select i1 %cond, i32 -1, i32 0214 ret i32 %sel215}216 217; select Cond, C+1, C --> add (zext Cond), C218 219define i32 @select_Cplus1_C(i1 %cond) {220; X86-LABEL: select_Cplus1_C:221; X86: # %bb.0:222; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax223; X86-NEXT: andl $1, %eax224; X86-NEXT: addl $41, %eax225; X86-NEXT: retl226;227; X64-LABEL: select_Cplus1_C:228; X64: # %bb.0:229; X64-NEXT: # kill: def $edi killed $edi def $rdi230; X64-NEXT: andl $1, %edi231; X64-NEXT: leal 41(%rdi), %eax232; X64-NEXT: retq233 %sel = select i1 %cond, i32 42, i32 41234 ret i32 %sel235}236 237define i32 @select_Cplus1_C_zeroext(i1 zeroext %cond) {238; X86-LABEL: select_Cplus1_C_zeroext:239; X86: # %bb.0:240; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax241; X86-NEXT: addl $41, %eax242; X86-NEXT: retl243;244; X64-LABEL: select_Cplus1_C_zeroext:245; X64: # %bb.0:246; X64-NEXT: # kill: def $edi killed $edi def $rdi247; X64-NEXT: leal 41(%rdi), %eax248; X64-NEXT: retq249 %sel = select i1 %cond, i32 42, i32 41250 ret i32 %sel251}252 253define i32 @select_Cplus1_C_signext(i1 signext %cond) {254; X86-LABEL: select_Cplus1_C_signext:255; X86: # %bb.0:256; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax257; X86-NEXT: andl $1, %eax258; X86-NEXT: addl $41, %eax259; X86-NEXT: retl260;261; X64-LABEL: select_Cplus1_C_signext:262; X64: # %bb.0:263; X64-NEXT: movl $41, %eax264; X64-NEXT: subl %edi, %eax265; X64-NEXT: retq266 %sel = select i1 %cond, i32 42, i32 41267 ret i32 %sel268}269 270; select Cond, C, C+1 --> add (sext Cond), C271 272define i32 @select_C_Cplus1(i1 %cond) {273; X86-LABEL: select_C_Cplus1:274; X86: # %bb.0:275; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx276; X86-NEXT: andl $1, %ecx277; X86-NEXT: movl $42, %eax278; X86-NEXT: subl %ecx, %eax279; X86-NEXT: retl280;281; X64-LABEL: select_C_Cplus1:282; X64: # %bb.0:283; X64-NEXT: andl $1, %edi284; X64-NEXT: movl $42, %eax285; X64-NEXT: subl %edi, %eax286; X64-NEXT: retq287 %sel = select i1 %cond, i32 41, i32 42288 ret i32 %sel289}290 291define i32 @select_C_Cplus1_zeroext(i1 zeroext %cond) {292; X86-LABEL: select_C_Cplus1_zeroext:293; X86: # %bb.0:294; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx295; X86-NEXT: movl $42, %eax296; X86-NEXT: subl %ecx, %eax297; X86-NEXT: retl298;299; X64-LABEL: select_C_Cplus1_zeroext:300; X64: # %bb.0:301; X64-NEXT: movl $42, %eax302; X64-NEXT: subl %edi, %eax303; X64-NEXT: retq304 %sel = select i1 %cond, i32 41, i32 42305 ret i32 %sel306}307 308define i32 @select_C_Cplus1_signext(i1 signext %cond) {309; X86-LABEL: select_C_Cplus1_signext:310; X86: # %bb.0:311; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx312; X86-NEXT: andl $1, %ecx313; X86-NEXT: movl $42, %eax314; X86-NEXT: subl %ecx, %eax315; X86-NEXT: retl316;317; X64-LABEL: select_C_Cplus1_signext:318; X64: # %bb.0:319; X64-NEXT: # kill: def $edi killed $edi def $rdi320; X64-NEXT: leal 42(%rdi), %eax321; X64-NEXT: retq322 %sel = select i1 %cond, i32 41, i32 42323 ret i32 %sel324}325 326; If the constants differ by a small multiplier, use LEA.327; select Cond, C1, C2 --> add (mul (zext Cond), C1-C2), C2 --> LEA C2(Cond * (C1-C2))328 329define i32 @select_lea_2(i1 zeroext %cond) {330; X86-LABEL: select_lea_2:331; X86: # %bb.0:332; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax333; X86-NEXT: negl %eax334; X86-NEXT: orl $1, %eax335; X86-NEXT: retl336;337; X64-LABEL: select_lea_2:338; X64: # %bb.0:339; X64-NEXT: movl %edi, %eax340; X64-NEXT: negl %eax341; X64-NEXT: orl $1, %eax342; X64-NEXT: retq343 %sel = select i1 %cond, i32 -1, i32 1344 ret i32 %sel345}346 347define i64 @select_lea_3(i1 zeroext %cond) {348; X86-LABEL: select_lea_3:349; X86: # %bb.0:350; X86-NEXT: cmpb $0, {{[0-9]+}}(%esp)351; X86-NEXT: movl $-2, %eax352; X86-NEXT: je .LBB19_1353; X86-NEXT: # %bb.2:354; X86-NEXT: movl $-1, %edx355; X86-NEXT: je .LBB19_3356; X86-NEXT: .LBB19_4:357; X86-NEXT: retl358; X86-NEXT: .LBB19_1:359; X86-NEXT: movl $1, %eax360; X86-NEXT: movl $-1, %edx361; X86-NEXT: jne .LBB19_4362; X86-NEXT: .LBB19_3:363; X86-NEXT: xorl %edx, %edx364; X86-NEXT: retl365;366; X64-LABEL: select_lea_3:367; X64: # %bb.0:368; X64-NEXT: xorb $1, %dil369; X64-NEXT: movzbl %dil, %eax370; X64-NEXT: leaq -2(%rax,%rax,2), %rax371; X64-NEXT: retq372 %sel = select i1 %cond, i64 -2, i64 1373 ret i64 %sel374}375 376define i32 @select_lea_5(i1 zeroext %cond) {377; X86-LABEL: select_lea_5:378; X86: # %bb.0:379; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax380; X86-NEXT: xorb $1, %al381; X86-NEXT: movzbl %al, %eax382; X86-NEXT: leal -2(%eax,%eax,4), %eax383; X86-NEXT: retl384;385; X64-LABEL: select_lea_5:386; X64: # %bb.0:387; X64-NEXT: xorb $1, %dil388; X64-NEXT: movzbl %dil, %eax389; X64-NEXT: leal -2(%rax,%rax,4), %eax390; X64-NEXT: retq391 %sel = select i1 %cond, i32 -2, i32 3392 ret i32 %sel393}394 395define i64 @select_lea_9(i1 zeroext %cond) {396; X86-LABEL: select_lea_9:397; X86: # %bb.0:398; X86-NEXT: cmpb $0, {{[0-9]+}}(%esp)399; X86-NEXT: movl $-7, %eax400; X86-NEXT: je .LBB21_1401; X86-NEXT: # %bb.2:402; X86-NEXT: movl $-1, %edx403; X86-NEXT: je .LBB21_3404; X86-NEXT: .LBB21_4:405; X86-NEXT: retl406; X86-NEXT: .LBB21_1:407; X86-NEXT: movl $2, %eax408; X86-NEXT: movl $-1, %edx409; X86-NEXT: jne .LBB21_4410; X86-NEXT: .LBB21_3:411; X86-NEXT: xorl %edx, %edx412; X86-NEXT: retl413;414; X64-LABEL: select_lea_9:415; X64: # %bb.0:416; X64-NEXT: xorb $1, %dil417; X64-NEXT: movzbl %dil, %eax418; X64-NEXT: leaq -7(%rax,%rax,8), %rax419; X64-NEXT: retq420 %sel = select i1 %cond, i64 -7, i64 2421 ret i64 %sel422}423 424; Should this be 'sbb x,x' or 'sbb 0,x' with simpler LEA or add?425 426define i64 @sel_1_2(i64 %x, i64 %y) {427; X86-LABEL: sel_1_2:428; X86: # %bb.0:429; X86-NEXT: movl {{[0-9]+}}(%esp), %eax430; X86-NEXT: movl {{[0-9]+}}(%esp), %edx431; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx432; X86-NEXT: cmpl $42, {{[0-9]+}}(%esp)433; X86-NEXT: sbbl $0, %ecx434; X86-NEXT: sbbl $0, %eax435; X86-NEXT: sbbl $0, %edx436; X86-NEXT: addl $2, %eax437; X86-NEXT: adcl $0, %edx438; X86-NEXT: retl439;440; X64-LABEL: sel_1_2:441; X64: # %bb.0:442; X64-NEXT: cmpq $42, %rdi443; X64-NEXT: sbbq $0, %rsi444; X64-NEXT: leaq 2(%rsi), %rax445; X64-NEXT: retq446 %cmp = icmp ult i64 %x, 42447 %sel = select i1 %cmp, i64 1, i64 2448 %sub = add i64 %sel, %y449 ret i64 %sub450}451 452; No LEA with 8-bit, but this shouldn't need branches or cmov.453 454define i8 @sel_1_neg1(i32 %x) {455; X86-LABEL: sel_1_neg1:456; X86: # %bb.0:457; X86-NEXT: cmpl $43, {{[0-9]+}}(%esp)458; X86-NEXT: setl %al459; X86-NEXT: negb %al460; X86-NEXT: orb $3, %al461; X86-NEXT: retl462;463; X64-LABEL: sel_1_neg1:464; X64: # %bb.0:465; X64-NEXT: cmpl $43, %edi466; X64-NEXT: setl %al467; X64-NEXT: negb %al468; X64-NEXT: orb $3, %al469; X64-NEXT: retq470 %cmp = icmp sgt i32 %x, 42471 %sel = select i1 %cmp, i8 3, i8 -1472 ret i8 %sel473}474 475; We get an LEA for 16-bit because we ignore the high-bits.476 477define i16 @sel_neg1_1(i32 %x) {478; X86-LABEL: sel_neg1_1:479; X86: # %bb.0:480; X86-NEXT: xorl %eax, %eax481; X86-NEXT: cmpl $43, {{[0-9]+}}(%esp)482; X86-NEXT: setl %al483; X86-NEXT: leal -1(,%eax,4), %eax484; X86-NEXT: # kill: def $ax killed $ax killed $eax485; X86-NEXT: retl486;487; X64-LABEL: sel_neg1_1:488; X64: # %bb.0:489; X64-NEXT: xorl %eax, %eax490; X64-NEXT: cmpl $43, %edi491; X64-NEXT: setl %al492; X64-NEXT: leal -1(,%rax,4), %eax493; X64-NEXT: # kill: def $ax killed $ax killed $eax494; X64-NEXT: retq495 %cmp = icmp sgt i32 %x, 42496 %sel = select i1 %cmp, i16 -1, i16 3497 ret i16 %sel498}499 500; If the comparison is available, the predicate can be inverted.501 502define i32 @sel_1_neg1_32(i32 %x) {503; X86-LABEL: sel_1_neg1_32:504; X86: # %bb.0:505; X86-NEXT: xorl %eax, %eax506; X86-NEXT: cmpl $43, {{[0-9]+}}(%esp)507; X86-NEXT: setge %al508; X86-NEXT: leal -1(%eax,%eax,8), %eax509; X86-NEXT: retl510;511; X64-LABEL: sel_1_neg1_32:512; X64: # %bb.0:513; X64-NEXT: xorl %eax, %eax514; X64-NEXT: cmpl $43, %edi515; X64-NEXT: setge %al516; X64-NEXT: leal -1(%rax,%rax,8), %eax517; X64-NEXT: retq518 %cmp = icmp sgt i32 %x, 42519 %sel = select i1 %cmp, i32 8, i32 -1520 ret i32 %sel521}522 523define i32 @sel_neg1_1_32(i32 %x) {524; X86-LABEL: sel_neg1_1_32:525; X86: # %bb.0:526; X86-NEXT: xorl %eax, %eax527; X86-NEXT: cmpl $43, {{[0-9]+}}(%esp)528; X86-NEXT: setl %al529; X86-NEXT: leal -7(%eax,%eax,8), %eax530; X86-NEXT: retl531;532; X64-LABEL: sel_neg1_1_32:533; X64: # %bb.0:534; X64-NEXT: xorl %eax, %eax535; X64-NEXT: cmpl $43, %edi536; X64-NEXT: setl %al537; X64-NEXT: leal -7(%rax,%rax,8), %eax538; X64-NEXT: retq539 %cmp = icmp sgt i32 %x, 42540 %sel = select i1 %cmp, i32 -7, i32 2541 ret i32 %sel542}543 544 545; If the constants differ by a large power-of-2, that can be a shift of the difference plus the smaller constant.546; select Cond, C1, C2 --> add (mul (zext Cond), C1-C2), C2547 548define i8 @select_pow2_diff(i1 zeroext %cond) {549; X86-LABEL: select_pow2_diff:550; X86: # %bb.0:551; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax552; X86-NEXT: shlb $4, %al553; X86-NEXT: orb $3, %al554; X86-NEXT: retl555;556; X64-LABEL: select_pow2_diff:557; X64: # %bb.0:558; X64-NEXT: # kill: def $edi killed $edi def $rdi559; X64-NEXT: shlb $4, %dil560; X64-NEXT: leal 3(%rdi), %eax561; X64-NEXT: # kill: def $al killed $al killed $eax562; X64-NEXT: retq563 %sel = select i1 %cond, i8 19, i8 3564 ret i8 %sel565}566 567define i16 @select_pow2_diff_invert(i1 zeroext %cond) {568; X86-LABEL: select_pow2_diff_invert:569; X86: # %bb.0:570; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax571; X86-NEXT: xorb $1, %al572; X86-NEXT: movzbl %al, %eax573; X86-NEXT: shll $6, %eax574; X86-NEXT: orl $7, %eax575; X86-NEXT: # kill: def $ax killed $ax killed $eax576; X86-NEXT: retl577;578; X64-LABEL: select_pow2_diff_invert:579; X64: # %bb.0:580; X64-NEXT: xorb $1, %dil581; X64-NEXT: movzbl %dil, %eax582; X64-NEXT: shll $6, %eax583; X64-NEXT: orl $7, %eax584; X64-NEXT: # kill: def $ax killed $ax killed $eax585; X64-NEXT: retq586 %sel = select i1 %cond, i16 7, i16 71587 ret i16 %sel588}589 590define i32 @select_pow2_diff_neg(i1 zeroext %cond) {591; X86-LABEL: select_pow2_diff_neg:592; X86: # %bb.0:593; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax594; X86-NEXT: shll $4, %eax595; X86-NEXT: orl $-25, %eax596; X86-NEXT: retl597;598; X64-LABEL: select_pow2_diff_neg:599; X64: # %bb.0:600; X64-NEXT: # kill: def $edi killed $edi def $rdi601; X64-NEXT: shll $4, %edi602; X64-NEXT: leal -25(%rdi), %eax603; X64-NEXT: retq604 %sel = select i1 %cond, i32 -9, i32 -25605 ret i32 %sel606}607 608define i64 @select_pow2_diff_neg_invert(i1 zeroext %cond) {609; X86-LABEL: select_pow2_diff_neg_invert:610; X86: # %bb.0:611; X86-NEXT: cmpb $0, {{[0-9]+}}(%esp)612; X86-NEXT: movl $-99, %eax613; X86-NEXT: je .LBB30_1614; X86-NEXT: # %bb.2:615; X86-NEXT: movl $-1, %edx616; X86-NEXT: je .LBB30_3617; X86-NEXT: .LBB30_4:618; X86-NEXT: retl619; X86-NEXT: .LBB30_1:620; X86-NEXT: movl $29, %eax621; X86-NEXT: movl $-1, %edx622; X86-NEXT: jne .LBB30_4623; X86-NEXT: .LBB30_3:624; X86-NEXT: xorl %edx, %edx625; X86-NEXT: retl626;627; X64-LABEL: select_pow2_diff_neg_invert:628; X64: # %bb.0:629; X64-NEXT: xorb $1, %dil630; X64-NEXT: movzbl %dil, %eax631; X64-NEXT: shll $7, %eax632; X64-NEXT: addq $-99, %rax633; X64-NEXT: retq634 %sel = select i1 %cond, i64 -99, i64 29635 ret i64 %sel636}637 638; This doesn't need a branch, but don't do the wrong thing if subtraction of the constants overflows.639 640define i8 @sel_67_neg125(i32 %x) {641; X86-LABEL: sel_67_neg125:642; X86: # %bb.0:643; X86-NEXT: cmpl $43, {{[0-9]+}}(%esp)644; X86-NEXT: movb $67, %al645; X86-NEXT: jge .LBB31_2646; X86-NEXT: # %bb.1:647; X86-NEXT: movb $-125, %al648; X86-NEXT: .LBB31_2:649; X86-NEXT: retl650;651; X64-LABEL: sel_67_neg125:652; X64: # %bb.0:653; X64-NEXT: cmpl $43, %edi654; X64-NEXT: movl $67, %ecx655; X64-NEXT: movl $131, %eax656; X64-NEXT: cmovgel %ecx, %eax657; X64-NEXT: # kill: def $al killed $al killed $eax658; X64-NEXT: retq659 %cmp = icmp sgt i32 %x, 42660 %sel = select i1 %cmp, i8 67, i8 -125661 ret i8 %sel662}663 664 665; In general, select of 2 constants could be:666; select Cond, C1, C2 --> add (mul (zext Cond), C1-C2), C2 --> add (and (sext Cond), C1-C2), C2667 668define i32 @select_C1_C2(i1 %cond) {669; X86-LABEL: select_C1_C2:670; X86: # %bb.0:671; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax672; X86-NEXT: andl $1, %eax673; X86-NEXT: negl %eax674; X86-NEXT: andl $399, %eax # imm = 0x18F675; X86-NEXT: xorl $42, %eax676; X86-NEXT: retl677;678; X64-LABEL: select_C1_C2:679; X64: # %bb.0:680; X64-NEXT: testb $1, %dil681; X64-NEXT: movl $421, %ecx # imm = 0x1A5682; X64-NEXT: movl $42, %eax683; X64-NEXT: cmovnel %ecx, %eax684; X64-NEXT: retq685 %sel = select i1 %cond, i32 421, i32 42686 ret i32 %sel687}688 689define i32 @select_C1_C2_zeroext(i1 zeroext %cond) {690; X86-LABEL: select_C1_C2_zeroext:691; X86: # %bb.0:692; X86-NEXT: cmpb $0, {{[0-9]+}}(%esp)693; X86-NEXT: movl $421, %eax # imm = 0x1A5694; X86-NEXT: jne .LBB33_2695; X86-NEXT: # %bb.1:696; X86-NEXT: movl $42, %eax697; X86-NEXT: .LBB33_2:698; X86-NEXT: retl699;700; X64-LABEL: select_C1_C2_zeroext:701; X64: # %bb.0:702; X64-NEXT: testl %edi, %edi703; X64-NEXT: movl $421, %ecx # imm = 0x1A5704; X64-NEXT: movl $42, %eax705; X64-NEXT: cmovnel %ecx, %eax706; X64-NEXT: retq707 %sel = select i1 %cond, i32 421, i32 42708 ret i32 %sel709}710 711define i32 @select_C1_C2_signext(i1 signext %cond) {712; X86-LABEL: select_C1_C2_signext:713; X86: # %bb.0:714; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax715; X86-NEXT: andl $1, %eax716; X86-NEXT: negl %eax717; X86-NEXT: andl $399, %eax # imm = 0x18F718; X86-NEXT: xorl $42, %eax719; X86-NEXT: retl720;721; X64-LABEL: select_C1_C2_signext:722; X64: # %bb.0:723; X64-NEXT: testb $1, %dil724; X64-NEXT: movl $421, %ecx # imm = 0x1A5725; X64-NEXT: movl $42, %eax726; X64-NEXT: cmovnel %ecx, %eax727; X64-NEXT: retq728 %sel = select i1 %cond, i32 421, i32 42729 ret i32 %sel730}731 732define i32 @select_n_or_minus1(i1 signext %cond) {733; X86-LABEL: select_n_or_minus1:734; X86: # %bb.0:735; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax736; X86-NEXT: andl $1, %eax737; X86-NEXT: decl %eax738; X86-NEXT: orl $12414, %eax # imm = 0x307E739; X86-NEXT: retl740;741; X64-LABEL: select_n_or_minus1:742; X64: # %bb.0:743; X64-NEXT: movl %edi, %eax744; X64-NEXT: notl %eax745; X64-NEXT: orl $12414, %eax # imm = 0x307E746; X64-NEXT: retq747 %sel = select i1 %cond, i32 12414, i32 -1748 ret i32 %sel749}750 751; select (x == 2), 2, (x + 1) --> select (x == 2), x, (x + 1)752 753define i64 @select_2_or_inc(i64 %x) {754; X86-LABEL: select_2_or_inc:755; X86: # %bb.0:756; X86-NEXT: pushl %edi757; X86-NEXT: .cfi_def_cfa_offset 8758; X86-NEXT: pushl %esi759; X86-NEXT: .cfi_def_cfa_offset 12760; X86-NEXT: .cfi_offset %esi, -12761; X86-NEXT: .cfi_offset %edi, -8762; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx763; X86-NEXT: movl {{[0-9]+}}(%esp), %esi764; X86-NEXT: movl %ecx, %edi765; X86-NEXT: xorl $2, %edi766; X86-NEXT: addl $1, %ecx767; X86-NEXT: movl %esi, %eax768; X86-NEXT: adcl $0, %eax769; X86-NEXT: xorl %edx, %edx770; X86-NEXT: orl %esi, %edi771; X86-NEXT: je .LBB36_2772; X86-NEXT: # %bb.1:773; X86-NEXT: movl %eax, %edx774; X86-NEXT: .LBB36_2:775; X86-NEXT: movl $2, %eax776; X86-NEXT: je .LBB36_4777; X86-NEXT: # %bb.3:778; X86-NEXT: movl %ecx, %eax779; X86-NEXT: .LBB36_4:780; X86-NEXT: popl %esi781; X86-NEXT: .cfi_def_cfa_offset 8782; X86-NEXT: popl %edi783; X86-NEXT: .cfi_def_cfa_offset 4784; X86-NEXT: retl785;786; X64-LABEL: select_2_or_inc:787; X64: # %bb.0:788; X64-NEXT: leaq 1(%rdi), %rax789; X64-NEXT: cmpq $2, %rdi790; X64-NEXT: cmoveq %rdi, %rax791; X64-NEXT: retq792 %cmp = icmp eq i64 %x, 2793 %add = add i64 %x, 1794 %retval.0 = select i1 %cmp, i64 2, i64 %add795 ret i64 %retval.0796}797 798define <4 x i32> @sel_constants_add_constant_vec(i1 %cond) {799; X86-LABEL: sel_constants_add_constant_vec:800; X86: # %bb.0:801; X86-NEXT: pushl %edi802; X86-NEXT: .cfi_def_cfa_offset 8803; X86-NEXT: pushl %esi804; X86-NEXT: .cfi_def_cfa_offset 12805; X86-NEXT: .cfi_offset %esi, -12806; X86-NEXT: .cfi_offset %edi, -8807; X86-NEXT: movl {{[0-9]+}}(%esp), %eax808; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx809; X86-NEXT: andl $1, %ecx810; X86-NEXT: negl %ecx811; X86-NEXT: movl %ecx, %edx812; X86-NEXT: andl $-15, %edx813; X86-NEXT: orl $12, %edx814; X86-NEXT: movl %ecx, %esi815; X86-NEXT: andl $3, %esi816; X86-NEXT: xorl $13, %esi817; X86-NEXT: movl %ecx, %edi818; X86-NEXT: andl $10, %edi819; X86-NEXT: xorl $14, %edi820; X86-NEXT: andl $11, %ecx821; X86-NEXT: xorl $15, %ecx822; X86-NEXT: movl %ecx, 12(%eax)823; X86-NEXT: movl %edi, 8(%eax)824; X86-NEXT: movl %esi, 4(%eax)825; X86-NEXT: movl %edx, (%eax)826; X86-NEXT: popl %esi827; X86-NEXT: .cfi_def_cfa_offset 8828; X86-NEXT: popl %edi829; X86-NEXT: .cfi_def_cfa_offset 4830; X86-NEXT: retl $4831;832; X64-LABEL: sel_constants_add_constant_vec:833; X64: # %bb.0:834; X64-NEXT: testb $1, %dil835; X64-NEXT: jne .LBB37_1836; X64-NEXT: # %bb.2:837; X64-NEXT: movaps {{.*#+}} xmm0 = [12,13,14,15]838; X64-NEXT: retq839; X64-NEXT: .LBB37_1:840; X64-NEXT: movaps {{.*#+}} xmm0 = [4294967293,14,4,4]841; X64-NEXT: retq842 %sel = select i1 %cond, <4 x i32> <i32 -4, i32 12, i32 1, i32 0>, <4 x i32> <i32 11, i32 11, i32 11, i32 11>843 %bo = add <4 x i32> %sel, <i32 1, i32 2, i32 3, i32 4>844 ret <4 x i32> %bo845}846 847define <2 x double> @sel_constants_fmul_constant_vec(i1 %cond) {848; X86-LABEL: sel_constants_fmul_constant_vec:849; X86: # %bb.0:850; X86-NEXT: testb $1, {{[0-9]+}}(%esp)851; X86-NEXT: fldl {{\.?LCPI[0-9]+_[0-9]+}}852; X86-NEXT: fldl {{\.?LCPI[0-9]+_[0-9]+}}853; X86-NEXT: jne .LBB38_2854; X86-NEXT: # %bb.1:855; X86-NEXT: fstp %st(1)856; X86-NEXT: fldz857; X86-NEXT: .LBB38_2:858; X86-NEXT: fstp %st(0)859; X86-NEXT: fldl {{\.?LCPI[0-9]+_[0-9]+}}860; X86-NEXT: fldl {{\.?LCPI[0-9]+_[0-9]+}}861; X86-NEXT: jne .LBB38_4862; X86-NEXT: # %bb.3:863; X86-NEXT: fstp %st(1)864; X86-NEXT: fldz865; X86-NEXT: .LBB38_4:866; X86-NEXT: fstp %st(0)867; X86-NEXT: retl868;869; X64-LABEL: sel_constants_fmul_constant_vec:870; X64: # %bb.0:871; X64-NEXT: testb $1, %dil872; X64-NEXT: jne .LBB38_1873; X64-NEXT: # %bb.2:874; X64-NEXT: movaps {{.*#+}} xmm0 = [1.1883E+2,3.4539999999999999E+1]875; X64-NEXT: retq876; X64-NEXT: .LBB38_1:877; X64-NEXT: movaps {{.*#+}} xmm0 = [-2.0399999999999999E+1,3.768E+1]878; X64-NEXT: retq879 %sel = select i1 %cond, <2 x double> <double -4.0, double 12.0>, <2 x double> <double 23.3, double 11.0>880 %bo = fmul <2 x double> %sel, <double 5.1, double 3.14>881 ret <2 x double> %bo882}883 884; 4294967297 = 0x100000001.885; This becomes an opaque constant via ConstantHoisting, so we don't fold it into the select.886 887define i64 @opaque_constant(i1 %cond, i64 %x) {888; X86-LABEL: opaque_constant:889; X86: # %bb.0:890; X86-NEXT: pushl %ebx891; X86-NEXT: .cfi_def_cfa_offset 8892; X86-NEXT: pushl %esi893; X86-NEXT: .cfi_def_cfa_offset 12894; X86-NEXT: .cfi_offset %esi, -12895; X86-NEXT: .cfi_offset %ebx, -8896; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx897; X86-NEXT: movl {{[0-9]+}}(%esp), %esi898; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax899; X86-NEXT: movl %eax, %edx900; X86-NEXT: andl $1, %edx901; X86-NEXT: negl %edx902; X86-NEXT: andl $1, %edx903; X86-NEXT: decl %eax904; X86-NEXT: andl $1, %eax905; X86-NEXT: xorl $1, %esi906; X86-NEXT: xorl $1, %ecx907; X86-NEXT: xorl %ebx, %ebx908; X86-NEXT: orl %esi, %ecx909; X86-NEXT: sete %bl910; X86-NEXT: subl %ebx, %eax911; X86-NEXT: sbbl $0, %edx912; X86-NEXT: popl %esi913; X86-NEXT: .cfi_def_cfa_offset 8914; X86-NEXT: popl %ebx915; X86-NEXT: .cfi_def_cfa_offset 4916; X86-NEXT: retl917;918; X64-LABEL: opaque_constant:919; X64: # %bb.0:920; X64-NEXT: testb $1, %dil921; X64-NEXT: movq $-4, %rcx922; X64-NEXT: movl $23, %eax923; X64-NEXT: cmovneq %rcx, %rax924; X64-NEXT: movabsq $4294967297, %rcx # imm = 0x100000001925; X64-NEXT: andq %rcx, %rax926; X64-NEXT: xorl %edx, %edx927; X64-NEXT: cmpq %rcx, %rsi928; X64-NEXT: sete %dl929; X64-NEXT: subq %rdx, %rax930; X64-NEXT: retq931 %sel = select i1 %cond, i64 -4, i64 23932 %bo = and i64 %sel, 4294967297933 %cmp = icmp eq i64 %x, 4294967297934 %sext = sext i1 %cmp to i64935 %add = add i64 %bo, %sext936 ret i64 %add937}938 939define float @select_undef_fp(float %x) {940; X86-LABEL: select_undef_fp:941; X86: # %bb.0:942; X86-NEXT: flds {{\.?LCPI[0-9]+_[0-9]+}}943; X86-NEXT: retl944;945; X64-LABEL: select_undef_fp:946; X64: # %bb.0:947; X64-NEXT: movss {{.*#+}} xmm0 = [4.0E+0,0.0E+0,0.0E+0,0.0E+0]948; X64-NEXT: retq949 %f = select i1 undef, float 4.0, float %x950 ret float %f951}952 953define i32 @select_eq0_3_2(i32 %X) {954; X86-LABEL: select_eq0_3_2:955; X86: # %bb.0:956; X86-NEXT: xorl %eax, %eax957; X86-NEXT: cmpl $1, {{[0-9]+}}(%esp)958; X86-NEXT: adcl $2, %eax959; X86-NEXT: retl960;961; X64-LABEL: select_eq0_3_2:962; X64: # %bb.0:963; X64-NEXT: xorl %eax, %eax964; X64-NEXT: cmpl $1, %edi965; X64-NEXT: adcl $2, %eax966; X64-NEXT: retq967 %cmp = icmp eq i32 %X, 0968 %sel = select i1 %cmp, i32 3, i32 2969 ret i32 %sel970}971 972define i32 @select_ugt3_2_3(i32 %X) {973; X86-LABEL: select_ugt3_2_3:974; X86: # %bb.0:975; X86-NEXT: xorl %eax, %eax976; X86-NEXT: cmpl $4, {{[0-9]+}}(%esp)977; X86-NEXT: adcl $2, %eax978; X86-NEXT: retl979;980; X64-LABEL: select_ugt3_2_3:981; X64: # %bb.0:982; X64-NEXT: xorl %eax, %eax983; X64-NEXT: cmpl $4, %edi984; X64-NEXT: adcl $2, %eax985; X64-NEXT: retq986 %cmp = icmp ugt i32 %X, 3987 %sel = select i1 %cmp, i32 2, i32 3988 ret i32 %sel989}990 991define i32 @select_ult9_7_6(i32 %X) {992; X86-LABEL: select_ult9_7_6:993; X86: # %bb.0:994; X86-NEXT: xorl %eax, %eax995; X86-NEXT: cmpl $9, {{[0-9]+}}(%esp)996; X86-NEXT: adcl $6, %eax997; X86-NEXT: retl998;999; X64-LABEL: select_ult9_7_6:1000; X64: # %bb.0:1001; X64-NEXT: xorl %eax, %eax1002; X64-NEXT: cmpl $9, %edi1003; X64-NEXT: adcl $6, %eax1004; X64-NEXT: retq1005 %cmp = icmp ult i32 %X, 91006 %sel = select i1 %cmp, i32 7, i32 61007 ret i32 %sel1008}1009 1010define i32 @select_ult2_2_3(i32 %X) {1011; X86-LABEL: select_ult2_2_3:1012; X86: # %bb.0:1013; X86-NEXT: cmpl $2, {{[0-9]+}}(%esp)1014; X86-NEXT: movl $3, %eax1015; X86-NEXT: sbbl $0, %eax1016; X86-NEXT: retl1017;1018; X64-LABEL: select_ult2_2_3:1019; X64: # %bb.0:1020; X64-NEXT: cmpl $2, %edi1021; X64-NEXT: movl $3, %eax1022; X64-NEXT: sbbl $0, %eax1023; X64-NEXT: retq1024 %cmp = icmp ult i32 %X, 21025 %cond = select i1 %cmp, i32 2, i32 31026 ret i32 %cond1027}1028 1029define i32 @select_ugt3_3_2(i32 %X) {1030; X86-LABEL: select_ugt3_3_2:1031; X86: # %bb.0:1032; X86-NEXT: cmpl $4, {{[0-9]+}}(%esp)1033; X86-NEXT: movl $2, %eax1034; X86-NEXT: sbbl $-1, %eax1035; X86-NEXT: retl1036;1037; X64-LABEL: select_ugt3_3_2:1038; X64: # %bb.0:1039; X64-NEXT: cmpl $4, %edi1040; X64-NEXT: movl $2, %eax1041; X64-NEXT: sbbl $-1, %eax1042; X64-NEXT: retq1043 %cmp.inv = icmp ugt i32 %X, 31044 %cond = select i1 %cmp.inv, i32 3, i32 21045 ret i32 %cond1046}1047 1048define i32 @select_eq_1_2(i32 %a, i32 %b) {1049; X86-LABEL: select_eq_1_2:1050; X86: # %bb.0:1051; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx1052; X86-NEXT: xorl %eax, %eax1053; X86-NEXT: cmpl {{[0-9]+}}(%esp), %ecx1054; X86-NEXT: setne %al1055; X86-NEXT: incl %eax1056; X86-NEXT: retl1057;1058; X64-FASTINC-LABEL: select_eq_1_2:1059; X64-FASTINC: # %bb.0:1060; X64-FASTINC-NEXT: xorl %eax, %eax1061; X64-FASTINC-NEXT: cmpl %esi, %edi1062; X64-FASTINC-NEXT: setne %al1063; X64-FASTINC-NEXT: incl %eax1064; X64-FASTINC-NEXT: retq1065;1066; X64-SLOWINC-LABEL: select_eq_1_2:1067; X64-SLOWINC: # %bb.0:1068; X64-SLOWINC-NEXT: xorl %eax, %eax1069; X64-SLOWINC-NEXT: cmpl %esi, %edi1070; X64-SLOWINC-NEXT: setne %al1071; X64-SLOWINC-NEXT: addl $1, %eax1072; X64-SLOWINC-NEXT: retq1073 1074 %cmp = icmp eq i32 %a, %b1075 %cond = select i1 %cmp, i32 1, i32 21076 ret i32 %cond1077}1078