760 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-linux | FileCheck %s --check-prefix=X643; RUN: llc < %s -mtriple=i686 -mattr=cmov | FileCheck %s --check-prefix=X864 5declare i4 @llvm.smul.fix.sat.i4 (i4, i4, i32)6declare i32 @llvm.smul.fix.sat.i32 (i32, i32, i32)7declare i64 @llvm.smul.fix.sat.i64 (i64, i64, i32)8declare <4 x i32> @llvm.smul.fix.sat.v4i32(<4 x i32>, <4 x i32>, i32)9 10define i32 @func(i32 %x, i32 %y) nounwind {11; X64-LABEL: func:12; X64: # %bb.0:13; X64-NEXT: movslq %esi, %rax14; X64-NEXT: movslq %edi, %rcx15; X64-NEXT: imulq %rax, %rcx16; X64-NEXT: movq %rcx, %rax17; X64-NEXT: shrq $32, %rax18; X64-NEXT: shrdl $2, %eax, %ecx19; X64-NEXT: cmpl $2, %eax20; X64-NEXT: movl $2147483647, %edx # imm = 0x7FFFFFFF21; X64-NEXT: cmovll %ecx, %edx22; X64-NEXT: cmpl $-2, %eax23; X64-NEXT: movl $-2147483648, %eax # imm = 0x8000000024; X64-NEXT: cmovgel %edx, %eax25; X64-NEXT: retq26;27; X86-LABEL: func:28; X86: # %bb.0:29; X86-NEXT: movl {{[0-9]+}}(%esp), %eax30; X86-NEXT: imull {{[0-9]+}}(%esp)31; X86-NEXT: shrdl $2, %edx, %eax32; X86-NEXT: cmpl $2, %edx33; X86-NEXT: movl $2147483647, %ecx # imm = 0x7FFFFFFF34; X86-NEXT: cmovgel %ecx, %eax35; X86-NEXT: cmpl $-2, %edx36; X86-NEXT: movl $-2147483648, %ecx # imm = 0x8000000037; X86-NEXT: cmovll %ecx, %eax38; X86-NEXT: retl39 %tmp = call i32 @llvm.smul.fix.sat.i32(i32 %x, i32 %y, i32 2)40 ret i32 %tmp41}42 43define i64 @func2(i64 %x, i64 %y) nounwind {44; X64-LABEL: func2:45; X64: # %bb.0:46; X64-NEXT: movq %rdi, %rax47; X64-NEXT: imulq %rsi48; X64-NEXT: shrdq $2, %rdx, %rax49; X64-NEXT: cmpq $2, %rdx50; X64-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF51; X64-NEXT: cmovgeq %rcx, %rax52; X64-NEXT: cmpq $-2, %rdx53; X64-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x800000000000000054; X64-NEXT: cmovlq %rcx, %rax55; X64-NEXT: retq56;57; X86-LABEL: func2:58; X86: # %bb.0:59; X86-NEXT: pushl %ebp60; X86-NEXT: pushl %ebx61; X86-NEXT: pushl %edi62; X86-NEXT: pushl %esi63; X86-NEXT: subl $8, %esp64; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx65; X86-NEXT: movl {{[0-9]+}}(%esp), %ebx66; X86-NEXT: movl {{[0-9]+}}(%esp), %esi67; X86-NEXT: movl %ecx, %eax68; X86-NEXT: mull %esi69; X86-NEXT: movl %edx, %edi70; X86-NEXT: movl %eax, %ebp71; X86-NEXT: movl %ecx, %eax72; X86-NEXT: mull %ebx73; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill74; X86-NEXT: addl %edx, %ebp75; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx76; X86-NEXT: adcl $0, %edi77; X86-NEXT: movl %ecx, %eax78; X86-NEXT: imull %esi79; X86-NEXT: movl %edx, %esi80; X86-NEXT: movl %eax, (%esp) # 4-byte Spill81; X86-NEXT: movl %ecx, %eax82; X86-NEXT: mull %ebx83; X86-NEXT: movl %eax, %ebx84; X86-NEXT: addl %ebp, %ebx85; X86-NEXT: adcl %edi, %edx86; X86-NEXT: adcl $0, %esi87; X86-NEXT: addl (%esp), %edx # 4-byte Folded Reload88; X86-NEXT: adcl $0, %esi89; X86-NEXT: movl %edx, %edi90; X86-NEXT: subl {{[0-9]+}}(%esp), %edi91; X86-NEXT: movl %esi, %ebp92; X86-NEXT: sbbl $0, %ebp93; X86-NEXT: testl %ecx, %ecx94; X86-NEXT: cmovnsl %esi, %ebp95; X86-NEXT: cmovnsl %edx, %edi96; X86-NEXT: movl %edi, %ecx97; X86-NEXT: subl {{[0-9]+}}(%esp), %ecx98; X86-NEXT: movl %ebp, %edx99; X86-NEXT: sbbl $0, %edx100; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)101; X86-NEXT: cmovnsl %ebp, %edx102; X86-NEXT: cmovnsl %edi, %ecx103; X86-NEXT: testl %edx, %edx104; X86-NEXT: setg %ah105; X86-NEXT: sete (%esp) # 1-byte Folded Spill106; X86-NEXT: cmpl $2, %ecx107; X86-NEXT: setae %al108; X86-NEXT: andb (%esp), %al # 1-byte Folded Reload109; X86-NEXT: orb %ah, %al110; X86-NEXT: movl {{[-0-9]+}}(%e{{[sb]}}p), %ebp # 4-byte Reload111; X86-NEXT: shrdl $2, %ebx, %ebp112; X86-NEXT: shrdl $2, %ecx, %ebx113; X86-NEXT: testb %al, %al114; X86-NEXT: movl $2147483647, %esi # imm = 0x7FFFFFFF115; X86-NEXT: cmovel %ebx, %esi116; X86-NEXT: movl $-1, %edi117; X86-NEXT: cmovel %ebp, %edi118; X86-NEXT: cmpl $-1, %edx119; X86-NEXT: setl %dl120; X86-NEXT: sete %al121; X86-NEXT: cmpl $-2, %ecx122; X86-NEXT: setb %cl123; X86-NEXT: andb %al, %cl124; X86-NEXT: xorl %eax, %eax125; X86-NEXT: orb %dl, %cl126; X86-NEXT: cmovel %edi, %eax127; X86-NEXT: movl $-2147483648, %edx # imm = 0x80000000128; X86-NEXT: cmovel %esi, %edx129; X86-NEXT: addl $8, %esp130; X86-NEXT: popl %esi131; X86-NEXT: popl %edi132; X86-NEXT: popl %ebx133; X86-NEXT: popl %ebp134; X86-NEXT: retl135 %tmp = call i64 @llvm.smul.fix.sat.i64(i64 %x, i64 %y, i32 2)136 ret i64 %tmp137}138 139define i4 @func3(i4 %x, i4 %y) nounwind {140; X64-LABEL: func3:141; X64: # %bb.0:142; X64-NEXT: shlb $4, %sil143; X64-NEXT: sarb $4, %sil144; X64-NEXT: shlb $4, %dil145; X64-NEXT: movsbl %dil, %eax146; X64-NEXT: movsbl %sil, %ecx147; X64-NEXT: imull %eax, %ecx148; X64-NEXT: movl %ecx, %eax149; X64-NEXT: shrb $2, %al150; X64-NEXT: shrl $8, %ecx151; X64-NEXT: movl %ecx, %edx152; X64-NEXT: shlb $6, %dl153; X64-NEXT: orb %al, %dl154; X64-NEXT: movzbl %dl, %eax155; X64-NEXT: cmpb $2, %cl156; X64-NEXT: movl $127, %edx157; X64-NEXT: cmovll %eax, %edx158; X64-NEXT: cmpb $-2, %cl159; X64-NEXT: movl $128, %eax160; X64-NEXT: cmovgel %edx, %eax161; X64-NEXT: sarb $4, %al162; X64-NEXT: # kill: def $al killed $al killed $eax163; X64-NEXT: retq164;165; X86-LABEL: func3:166; X86: # %bb.0:167; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax168; X86-NEXT: shlb $4, %al169; X86-NEXT: sarb $4, %al170; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx171; X86-NEXT: shlb $4, %cl172; X86-NEXT: movsbl %cl, %ecx173; X86-NEXT: movsbl %al, %eax174; X86-NEXT: imull %ecx, %eax175; X86-NEXT: movb %ah, %cl176; X86-NEXT: shlb $6, %cl177; X86-NEXT: shrb $2, %al178; X86-NEXT: orb %cl, %al179; X86-NEXT: movzbl %al, %ecx180; X86-NEXT: cmpb $2, %ah181; X86-NEXT: movl $127, %edx182; X86-NEXT: cmovll %ecx, %edx183; X86-NEXT: cmpb $-2, %ah184; X86-NEXT: movl $128, %eax185; X86-NEXT: cmovgel %edx, %eax186; X86-NEXT: sarb $4, %al187; X86-NEXT: # kill: def $al killed $al killed $eax188; X86-NEXT: retl189 %tmp = call i4 @llvm.smul.fix.sat.i4(i4 %x, i4 %y, i32 2)190 ret i4 %tmp191}192 193define <4 x i32> @vec(<4 x i32> %x, <4 x i32> %y) nounwind {194; X64-LABEL: vec:195; X64: # %bb.0:196; X64-NEXT: pshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]197; X64-NEXT: movd %xmm2, %eax198; X64-NEXT: cltq199; X64-NEXT: pshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]200; X64-NEXT: movd %xmm2, %ecx201; X64-NEXT: movslq %ecx, %rdx202; X64-NEXT: imulq %rax, %rdx203; X64-NEXT: movq %rdx, %rcx204; X64-NEXT: shrq $32, %rcx205; X64-NEXT: shrdl $2, %ecx, %edx206; X64-NEXT: cmpl $2, %ecx207; X64-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF208; X64-NEXT: cmovgel %eax, %edx209; X64-NEXT: cmpl $-2, %ecx210; X64-NEXT: movl $-2147483648, %ecx # imm = 0x80000000211; X64-NEXT: cmovll %ecx, %edx212; X64-NEXT: movd %edx, %xmm2213; X64-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]214; X64-NEXT: movd %xmm3, %edx215; X64-NEXT: movslq %edx, %rdx216; X64-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]217; X64-NEXT: movd %xmm3, %esi218; X64-NEXT: movslq %esi, %rsi219; X64-NEXT: imulq %rdx, %rsi220; X64-NEXT: movq %rsi, %rdx221; X64-NEXT: shrq $32, %rdx222; X64-NEXT: shrdl $2, %edx, %esi223; X64-NEXT: cmpl $2, %edx224; X64-NEXT: cmovgel %eax, %esi225; X64-NEXT: cmpl $-2, %edx226; X64-NEXT: cmovll %ecx, %esi227; X64-NEXT: movd %esi, %xmm3228; X64-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]229; X64-NEXT: movd %xmm1, %edx230; X64-NEXT: movslq %edx, %rdx231; X64-NEXT: movd %xmm0, %esi232; X64-NEXT: movslq %esi, %rsi233; X64-NEXT: imulq %rdx, %rsi234; X64-NEXT: movq %rsi, %rdx235; X64-NEXT: shrq $32, %rdx236; X64-NEXT: shrdl $2, %edx, %esi237; X64-NEXT: cmpl $2, %edx238; X64-NEXT: cmovgel %eax, %esi239; X64-NEXT: cmpl $-2, %edx240; X64-NEXT: cmovll %ecx, %esi241; X64-NEXT: movd %esi, %xmm2242; X64-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]243; X64-NEXT: movd %xmm1, %edx244; X64-NEXT: movslq %edx, %rdx245; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]246; X64-NEXT: movd %xmm0, %esi247; X64-NEXT: movslq %esi, %rsi248; X64-NEXT: imulq %rdx, %rsi249; X64-NEXT: movq %rsi, %rdx250; X64-NEXT: shrq $32, %rdx251; X64-NEXT: shrdl $2, %edx, %esi252; X64-NEXT: cmpl $2, %edx253; X64-NEXT: cmovgel %eax, %esi254; X64-NEXT: cmpl $-2, %edx255; X64-NEXT: cmovll %ecx, %esi256; X64-NEXT: movd %esi, %xmm0257; X64-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]258; X64-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]259; X64-NEXT: movdqa %xmm2, %xmm0260; X64-NEXT: retq261;262; X86-LABEL: vec:263; X86: # %bb.0:264; X86-NEXT: pushl %ebp265; X86-NEXT: pushl %ebx266; X86-NEXT: pushl %edi267; X86-NEXT: pushl %esi268; X86-NEXT: movl {{[0-9]+}}(%esp), %ebx269; X86-NEXT: movl {{[0-9]+}}(%esp), %edi270; X86-NEXT: movl {{[0-9]+}}(%esp), %eax271; X86-NEXT: imull {{[0-9]+}}(%esp)272; X86-NEXT: movl %eax, %ecx273; X86-NEXT: shrdl $2, %edx, %ecx274; X86-NEXT: cmpl $2, %edx275; X86-NEXT: movl $2147483647, %ebp # imm = 0x7FFFFFFF276; X86-NEXT: cmovgel %ebp, %ecx277; X86-NEXT: cmpl $-2, %edx278; X86-NEXT: movl $-2147483648, %esi # imm = 0x80000000279; X86-NEXT: cmovll %esi, %ecx280; X86-NEXT: movl %edi, %eax281; X86-NEXT: imull {{[0-9]+}}(%esp)282; X86-NEXT: movl %eax, %edi283; X86-NEXT: shrdl $2, %edx, %edi284; X86-NEXT: cmpl $2, %edx285; X86-NEXT: cmovgel %ebp, %edi286; X86-NEXT: cmpl $-2, %edx287; X86-NEXT: cmovll %esi, %edi288; X86-NEXT: movl %ebx, %eax289; X86-NEXT: imull {{[0-9]+}}(%esp)290; X86-NEXT: movl %eax, %ebx291; X86-NEXT: shrdl $2, %edx, %ebx292; X86-NEXT: cmpl $2, %edx293; X86-NEXT: cmovgel %ebp, %ebx294; X86-NEXT: cmpl $-2, %edx295; X86-NEXT: cmovll %esi, %ebx296; X86-NEXT: movl {{[0-9]+}}(%esp), %eax297; X86-NEXT: imull {{[0-9]+}}(%esp)298; X86-NEXT: shrdl $2, %edx, %eax299; X86-NEXT: cmpl $2, %edx300; X86-NEXT: cmovgel %ebp, %eax301; X86-NEXT: cmpl $-2, %edx302; X86-NEXT: cmovll %esi, %eax303; X86-NEXT: movl {{[0-9]+}}(%esp), %edx304; X86-NEXT: movl %eax, 12(%edx)305; X86-NEXT: movl %ebx, 8(%edx)306; X86-NEXT: movl %edi, 4(%edx)307; X86-NEXT: movl %ecx, (%edx)308; X86-NEXT: movl %edx, %eax309; X86-NEXT: popl %esi310; X86-NEXT: popl %edi311; X86-NEXT: popl %ebx312; X86-NEXT: popl %ebp313; X86-NEXT: retl $4314 %tmp = call <4 x i32> @llvm.smul.fix.sat.v4i32(<4 x i32> %x, <4 x i32> %y, i32 2)315 ret <4 x i32> %tmp316}317 318; These result in regular integer multiplication319define i32 @func4(i32 %x, i32 %y) nounwind {320; X64-LABEL: func4:321; X64: # %bb.0:322; X64-NEXT: xorl %eax, %eax323; X64-NEXT: movl %edi, %ecx324; X64-NEXT: xorl %esi, %ecx325; X64-NEXT: sets %al326; X64-NEXT: addl $2147483647, %eax # imm = 0x7FFFFFFF327; X64-NEXT: imull %esi, %edi328; X64-NEXT: cmovnol %edi, %eax329; X64-NEXT: retq330;331; X86-LABEL: func4:332; X86: # %bb.0:333; X86-NEXT: pushl %esi334; X86-NEXT: movl {{[0-9]+}}(%esp), %edx335; X86-NEXT: movl {{[0-9]+}}(%esp), %eax336; X86-NEXT: xorl %ecx, %ecx337; X86-NEXT: movl %eax, %esi338; X86-NEXT: xorl %edx, %esi339; X86-NEXT: sets %cl340; X86-NEXT: addl $2147483647, %ecx # imm = 0x7FFFFFFF341; X86-NEXT: imull %edx, %eax342; X86-NEXT: cmovol %ecx, %eax343; X86-NEXT: popl %esi344; X86-NEXT: retl345 %tmp = call i32 @llvm.smul.fix.sat.i32(i32 %x, i32 %y, i32 0)346 ret i32 %tmp347}348 349define i64 @func5(i64 %x, i64 %y) {350; X64-LABEL: func5:351; X64: # %bb.0:352; X64-NEXT: xorl %ecx, %ecx353; X64-NEXT: movq %rdi, %rax354; X64-NEXT: xorq %rsi, %rax355; X64-NEXT: sets %cl356; X64-NEXT: movabsq $9223372036854775807, %rax # imm = 0x7FFFFFFFFFFFFFFF357; X64-NEXT: addq %rcx, %rax358; X64-NEXT: imulq %rsi, %rdi359; X64-NEXT: cmovnoq %rdi, %rax360; X64-NEXT: retq361;362; X86-LABEL: func5:363; X86: # %bb.0:364; X86-NEXT: pushl %ebp365; X86-NEXT: .cfi_def_cfa_offset 8366; X86-NEXT: pushl %ebx367; X86-NEXT: .cfi_def_cfa_offset 12368; X86-NEXT: pushl %edi369; X86-NEXT: .cfi_def_cfa_offset 16370; X86-NEXT: pushl %esi371; X86-NEXT: .cfi_def_cfa_offset 20372; X86-NEXT: subl $8, %esp373; X86-NEXT: .cfi_def_cfa_offset 28374; X86-NEXT: .cfi_offset %esi, -20375; X86-NEXT: .cfi_offset %edi, -16376; X86-NEXT: .cfi_offset %ebx, -12377; X86-NEXT: .cfi_offset %ebp, -8378; X86-NEXT: movl {{[0-9]+}}(%esp), %edi379; X86-NEXT: movl {{[0-9]+}}(%esp), %eax380; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx381; X86-NEXT: movl %ecx, %esi382; X86-NEXT: sarl $31, %esi383; X86-NEXT: imull %edi, %esi384; X86-NEXT: mull %edi385; X86-NEXT: movl %edx, %ebx386; X86-NEXT: movl %eax, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill387; X86-NEXT: movl %ecx, %eax388; X86-NEXT: mull %edi389; X86-NEXT: movl %edx, %ebp390; X86-NEXT: addl %eax, %ebx391; X86-NEXT: movl %ebx, (%esp) # 4-byte Spill392; X86-NEXT: movl {{[0-9]+}}(%esp), %edx393; X86-NEXT: adcl %esi, %ebp394; X86-NEXT: movl %ebp, %edi395; X86-NEXT: sarl $31, %edi396; X86-NEXT: movl %edx, %esi397; X86-NEXT: sarl $31, %esi398; X86-NEXT: movl {{[0-9]+}}(%esp), %eax399; X86-NEXT: imull %eax, %esi400; X86-NEXT: mull %edx401; X86-NEXT: movl %edx, %ebx402; X86-NEXT: addl %eax, (%esp) # 4-byte Folded Spill403; X86-NEXT: adcl %esi, %ebx404; X86-NEXT: movl %ebx, %esi405; X86-NEXT: sarl $31, %esi406; X86-NEXT: addl %ebp, %ebx407; X86-NEXT: adcl %edi, %esi408; X86-NEXT: movl %ecx, %eax409; X86-NEXT: movl {{[0-9]+}}(%esp), %ebp410; X86-NEXT: imull %ebp411; X86-NEXT: addl %ebx, %eax412; X86-NEXT: adcl %esi, %edx413; X86-NEXT: movl (%esp), %edi # 4-byte Reload414; X86-NEXT: movl %edi, %esi415; X86-NEXT: sarl $31, %esi416; X86-NEXT: xorl %esi, %edx417; X86-NEXT: xorl %eax, %esi418; X86-NEXT: xorl %ebp, %ecx419; X86-NEXT: sarl $31, %ecx420; X86-NEXT: movl %ecx, %ebx421; X86-NEXT: xorl $2147483647, %ebx # imm = 0x7FFFFFFF422; X86-NEXT: orl %edx, %esi423; X86-NEXT: notl %ecx424; X86-NEXT: cmovel {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Folded Reload425; X86-NEXT: cmovel %edi, %ebx426; X86-NEXT: movl %ecx, %eax427; X86-NEXT: movl %ebx, %edx428; X86-NEXT: addl $8, %esp429; X86-NEXT: .cfi_def_cfa_offset 20430; X86-NEXT: popl %esi431; X86-NEXT: .cfi_def_cfa_offset 16432; X86-NEXT: popl %edi433; X86-NEXT: .cfi_def_cfa_offset 12434; X86-NEXT: popl %ebx435; X86-NEXT: .cfi_def_cfa_offset 8436; X86-NEXT: popl %ebp437; X86-NEXT: .cfi_def_cfa_offset 4438; X86-NEXT: retl439 %tmp = call i64 @llvm.smul.fix.sat.i64(i64 %x, i64 %y, i32 0)440 ret i64 %tmp441}442 443define i4 @func6(i4 %x, i4 %y) nounwind {444; X64-LABEL: func6:445; X64: # %bb.0:446; X64-NEXT: movl %edi, %eax447; X64-NEXT: shlb $4, %sil448; X64-NEXT: sarb $4, %sil449; X64-NEXT: shlb $4, %al450; X64-NEXT: xorl %ecx, %ecx451; X64-NEXT: movl %eax, %edx452; X64-NEXT: xorb %sil, %dl453; X64-NEXT: sets %cl454; X64-NEXT: addl $127, %ecx455; X64-NEXT: # kill: def $al killed $al killed $eax456; X64-NEXT: imulb %sil457; X64-NEXT: movzbl %al, %eax458; X64-NEXT: cmovol %ecx, %eax459; X64-NEXT: sarb $4, %al460; X64-NEXT: # kill: def $al killed $al killed $eax461; X64-NEXT: retq462;463; X86-LABEL: func6:464; X86: # %bb.0:465; X86-NEXT: movzbl {{[0-9]+}}(%esp), %ecx466; X86-NEXT: shlb $4, %cl467; X86-NEXT: sarb $4, %cl468; X86-NEXT: movzbl {{[0-9]+}}(%esp), %eax469; X86-NEXT: shlb $4, %al470; X86-NEXT: xorl %edx, %edx471; X86-NEXT: movb %al, %ah472; X86-NEXT: xorb %cl, %ah473; X86-NEXT: sets %dl474; X86-NEXT: addl $127, %edx475; X86-NEXT: imulb %cl476; X86-NEXT: movzbl %al, %eax477; X86-NEXT: cmovol %edx, %eax478; X86-NEXT: sarb $4, %al479; X86-NEXT: # kill: def $al killed $al killed $eax480; X86-NEXT: retl481 %tmp = call i4 @llvm.smul.fix.sat.i4(i4 %x, i4 %y, i32 0)482 ret i4 %tmp483}484 485define <4 x i32> @vec2(<4 x i32> %x, <4 x i32> %y) nounwind {486; X64-LABEL: vec2:487; X64: # %bb.0:488; X64-NEXT: pshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]489; X64-NEXT: movd %xmm2, %eax490; X64-NEXT: pshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]491; X64-NEXT: movd %xmm2, %ecx492; X64-NEXT: xorl %edx, %edx493; X64-NEXT: movl %ecx, %esi494; X64-NEXT: xorl %eax, %esi495; X64-NEXT: sets %dl496; X64-NEXT: addl $2147483647, %edx # imm = 0x7FFFFFFF497; X64-NEXT: imull %eax, %ecx498; X64-NEXT: cmovol %edx, %ecx499; X64-NEXT: movd %ecx, %xmm2500; X64-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]501; X64-NEXT: movd %xmm3, %eax502; X64-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]503; X64-NEXT: movd %xmm3, %ecx504; X64-NEXT: xorl %edx, %edx505; X64-NEXT: movl %ecx, %esi506; X64-NEXT: xorl %eax, %esi507; X64-NEXT: sets %dl508; X64-NEXT: addl $2147483647, %edx # imm = 0x7FFFFFFF509; X64-NEXT: imull %eax, %ecx510; X64-NEXT: cmovol %edx, %ecx511; X64-NEXT: movd %ecx, %xmm3512; X64-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]513; X64-NEXT: movd %xmm1, %eax514; X64-NEXT: movd %xmm0, %ecx515; X64-NEXT: xorl %edx, %edx516; X64-NEXT: movl %ecx, %esi517; X64-NEXT: xorl %eax, %esi518; X64-NEXT: sets %dl519; X64-NEXT: addl $2147483647, %edx # imm = 0x7FFFFFFF520; X64-NEXT: imull %eax, %ecx521; X64-NEXT: cmovol %edx, %ecx522; X64-NEXT: movd %ecx, %xmm2523; X64-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]524; X64-NEXT: movd %xmm1, %eax525; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]526; X64-NEXT: movd %xmm0, %ecx527; X64-NEXT: xorl %edx, %edx528; X64-NEXT: movl %ecx, %esi529; X64-NEXT: xorl %eax, %esi530; X64-NEXT: sets %dl531; X64-NEXT: addl $2147483647, %edx # imm = 0x7FFFFFFF532; X64-NEXT: imull %eax, %ecx533; X64-NEXT: cmovol %edx, %ecx534; X64-NEXT: movd %ecx, %xmm0535; X64-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]536; X64-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]537; X64-NEXT: movdqa %xmm2, %xmm0538; X64-NEXT: retq539;540; X86-LABEL: vec2:541; X86: # %bb.0:542; X86-NEXT: pushl %ebp543; X86-NEXT: pushl %ebx544; X86-NEXT: pushl %edi545; X86-NEXT: pushl %esi546; X86-NEXT: movl {{[0-9]+}}(%esp), %ebp547; X86-NEXT: movl {{[0-9]+}}(%esp), %esi548; X86-NEXT: movl {{[0-9]+}}(%esp), %ebx549; X86-NEXT: movl {{[0-9]+}}(%esp), %edi550; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx551; X86-NEXT: xorl %eax, %eax552; X86-NEXT: movl %ecx, %edx553; X86-NEXT: xorl %edi, %edx554; X86-NEXT: movl {{[0-9]+}}(%esp), %edx555; X86-NEXT: sets %al556; X86-NEXT: addl $2147483647, %eax # imm = 0x7FFFFFFF557; X86-NEXT: imull %edi, %ecx558; X86-NEXT: cmovol %eax, %ecx559; X86-NEXT: xorl %eax, %eax560; X86-NEXT: movl %edx, %edi561; X86-NEXT: xorl %ebx, %edi562; X86-NEXT: movl {{[0-9]+}}(%esp), %edi563; X86-NEXT: sets %al564; X86-NEXT: addl $2147483647, %eax # imm = 0x7FFFFFFF565; X86-NEXT: imull %ebx, %edx566; X86-NEXT: cmovol %eax, %edx567; X86-NEXT: xorl %eax, %eax568; X86-NEXT: movl %edi, %ebx569; X86-NEXT: xorl %esi, %ebx570; X86-NEXT: movl {{[0-9]+}}(%esp), %ebx571; X86-NEXT: sets %al572; X86-NEXT: addl $2147483647, %eax # imm = 0x7FFFFFFF573; X86-NEXT: imull %esi, %edi574; X86-NEXT: cmovol %eax, %edi575; X86-NEXT: xorl %eax, %eax576; X86-NEXT: movl %ebx, %esi577; X86-NEXT: xorl %ebp, %esi578; X86-NEXT: sets %al579; X86-NEXT: addl $2147483647, %eax # imm = 0x7FFFFFFF580; X86-NEXT: imull %ebp, %ebx581; X86-NEXT: cmovol %eax, %ebx582; X86-NEXT: movl {{[0-9]+}}(%esp), %eax583; X86-NEXT: movl %ebx, 12(%eax)584; X86-NEXT: movl %edi, 8(%eax)585; X86-NEXT: movl %edx, 4(%eax)586; X86-NEXT: movl %ecx, (%eax)587; X86-NEXT: popl %esi588; X86-NEXT: popl %edi589; X86-NEXT: popl %ebx590; X86-NEXT: popl %ebp591; X86-NEXT: retl $4592 %tmp = call <4 x i32> @llvm.smul.fix.sat.v4i32(<4 x i32> %x, <4 x i32> %y, i32 0)593 ret <4 x i32> %tmp594}595 596define i64 @func7(i64 %x, i64 %y) nounwind {597; X64-LABEL: func7:598; X64: # %bb.0:599; X64-NEXT: movq %rdi, %rax600; X64-NEXT: imulq %rsi601; X64-NEXT: shrdq $32, %rdx, %rax602; X64-NEXT: cmpq $2147483647, %rdx # imm = 0x7FFFFFFF603; X64-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF604; X64-NEXT: cmovgq %rcx, %rax605; X64-NEXT: cmpq $-2147483648, %rdx # imm = 0x80000000606; X64-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000607; X64-NEXT: cmovlq %rcx, %rax608; X64-NEXT: retq609;610; X86-LABEL: func7:611; X86: # %bb.0:612; X86-NEXT: pushl %ebp613; X86-NEXT: pushl %ebx614; X86-NEXT: pushl %edi615; X86-NEXT: pushl %esi616; X86-NEXT: pushl %eax617; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx618; X86-NEXT: movl {{[0-9]+}}(%esp), %esi619; X86-NEXT: movl %ecx, %eax620; X86-NEXT: mull {{[0-9]+}}(%esp)621; X86-NEXT: movl %edx, %edi622; X86-NEXT: movl %eax, %ebp623; X86-NEXT: movl %ecx, %eax624; X86-NEXT: mull {{[0-9]+}}(%esp)625; X86-NEXT: addl %edx, %ebp626; X86-NEXT: adcl $0, %edi627; X86-NEXT: movl %esi, %eax628; X86-NEXT: imull {{[0-9]+}}(%esp)629; X86-NEXT: movl %edx, %ebx630; X86-NEXT: movl %eax, %ecx631; X86-NEXT: movl %esi, %eax632; X86-NEXT: mull {{[0-9]+}}(%esp)633; X86-NEXT: addl %ebp, %eax634; X86-NEXT: adcl %edi, %edx635; X86-NEXT: adcl $0, %ebx636; X86-NEXT: addl %ecx, %edx637; X86-NEXT: adcl $0, %ebx638; X86-NEXT: movl %edx, %ecx639; X86-NEXT: subl {{[0-9]+}}(%esp), %ecx640; X86-NEXT: movl %ebx, %esi641; X86-NEXT: sbbl $0, %esi642; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)643; X86-NEXT: cmovnsl %ebx, %esi644; X86-NEXT: cmovsl %ecx, %edx645; X86-NEXT: movl %edx, %edi646; X86-NEXT: subl {{[0-9]+}}(%esp), %edi647; X86-NEXT: movl %esi, %ecx648; X86-NEXT: sbbl $0, %ecx649; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)650; X86-NEXT: cmovnsl %esi, %ecx651; X86-NEXT: cmovsl %edi, %edx652; X86-NEXT: testl %edx, %edx653; X86-NEXT: setns {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill654; X86-NEXT: sets %bh655; X86-NEXT: testl %ecx, %ecx656; X86-NEXT: setg {{[-0-9]+}}(%e{{[sb]}}p) # 1-byte Folded Spill657; X86-NEXT: sete %bl658; X86-NEXT: andb %bh, %bl659; X86-NEXT: orb {{[-0-9]+}}(%e{{[sb]}}p), %bl # 1-byte Folded Reload660; X86-NEXT: movl $2147483647, %esi # imm = 0x7FFFFFFF661; X86-NEXT: cmovnel %esi, %edx662; X86-NEXT: movl $-1, %esi663; X86-NEXT: cmovnel %esi, %eax664; X86-NEXT: cmpl $-1, %ecx665; X86-NEXT: setl %cl666; X86-NEXT: sete %ch667; X86-NEXT: andb {{[-0-9]+}}(%e{{[sb]}}p), %ch # 1-byte Folded Reload668; X86-NEXT: xorl %esi, %esi669; X86-NEXT: orb %cl, %ch670; X86-NEXT: cmovnel %esi, %eax671; X86-NEXT: movl $-2147483648, %ecx # imm = 0x80000000672; X86-NEXT: cmovnel %ecx, %edx673; X86-NEXT: addl $4, %esp674; X86-NEXT: popl %esi675; X86-NEXT: popl %edi676; X86-NEXT: popl %ebx677; X86-NEXT: popl %ebp678; X86-NEXT: retl679 %tmp = call i64 @llvm.smul.fix.sat.i64(i64 %x, i64 %y, i32 32)680 ret i64 %tmp681}682 683define i64 @func8(i64 %x, i64 %y) nounwind {684; X64-LABEL: func8:685; X64: # %bb.0:686; X64-NEXT: movq %rdi, %rax687; X64-NEXT: imulq %rsi688; X64-NEXT: shrdq $63, %rdx, %rax689; X64-NEXT: movabsq $4611686018427387903, %rcx # imm = 0x3FFFFFFFFFFFFFFF690; X64-NEXT: cmpq %rcx, %rdx691; X64-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF692; X64-NEXT: cmovgq %rcx, %rax693; X64-NEXT: movabsq $-4611686018427387904, %rcx # imm = 0xC000000000000000694; X64-NEXT: cmpq %rcx, %rdx695; X64-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000696; X64-NEXT: cmovlq %rcx, %rax697; X64-NEXT: retq698;699; X86-LABEL: func8:700; X86: # %bb.0:701; X86-NEXT: pushl %ebp702; X86-NEXT: pushl %ebx703; X86-NEXT: pushl %edi704; X86-NEXT: pushl %esi705; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx706; X86-NEXT: movl {{[0-9]+}}(%esp), %esi707; X86-NEXT: movl %ecx, %eax708; X86-NEXT: mull {{[0-9]+}}(%esp)709; X86-NEXT: movl %edx, %edi710; X86-NEXT: movl %eax, %ebp711; X86-NEXT: movl %ecx, %eax712; X86-NEXT: mull {{[0-9]+}}(%esp)713; X86-NEXT: addl %edx, %ebp714; X86-NEXT: adcl $0, %edi715; X86-NEXT: movl %esi, %eax716; X86-NEXT: imull {{[0-9]+}}(%esp)717; X86-NEXT: movl %edx, %ebx718; X86-NEXT: movl %eax, %ecx719; X86-NEXT: movl %esi, %eax720; X86-NEXT: mull {{[0-9]+}}(%esp)721; X86-NEXT: addl %ebp, %eax722; X86-NEXT: adcl %edx, %edi723; X86-NEXT: adcl $0, %ebx724; X86-NEXT: addl %ecx, %edi725; X86-NEXT: adcl $0, %ebx726; X86-NEXT: movl %edi, %edx727; X86-NEXT: subl {{[0-9]+}}(%esp), %edx728; X86-NEXT: movl %ebx, %ebp729; X86-NEXT: sbbl $0, %ebp730; X86-NEXT: testl %esi, %esi731; X86-NEXT: cmovnsl %ebx, %ebp732; X86-NEXT: cmovnsl %edi, %edx733; X86-NEXT: movl %edx, %esi734; X86-NEXT: subl {{[0-9]+}}(%esp), %esi735; X86-NEXT: movl %ebp, %ecx736; X86-NEXT: sbbl $0, %ecx737; X86-NEXT: cmpl $0, {{[0-9]+}}(%esp)738; X86-NEXT: cmovnsl %ebp, %ecx739; X86-NEXT: cmovnsl %edx, %esi740; X86-NEXT: shrdl $31, %esi, %eax741; X86-NEXT: shrdl $31, %ecx, %esi742; X86-NEXT: cmpl $1073741824, %ecx # imm = 0x40000000743; X86-NEXT: movl $2147483647, %edi # imm = 0x7FFFFFFF744; X86-NEXT: cmovll %esi, %edi745; X86-NEXT: movl $-1, %edx746; X86-NEXT: cmovgel %edx, %eax747; X86-NEXT: xorl %edx, %edx748; X86-NEXT: cmpl $-1073741824, %ecx # imm = 0xC0000000749; X86-NEXT: cmovll %edx, %eax750; X86-NEXT: movl $-2147483648, %edx # imm = 0x80000000751; X86-NEXT: cmovgel %edi, %edx752; X86-NEXT: popl %esi753; X86-NEXT: popl %edi754; X86-NEXT: popl %ebx755; X86-NEXT: popl %ebp756; X86-NEXT: retl757 %tmp = call i64 @llvm.smul.fix.sat.i64(i64 %x, i64 %y, i32 63)758 ret i64 %tmp759}760