From 6ef06f9159950b14db7f40f1fd37aa53ca2e955b Mon Sep 17 00:00:00 2001 From: Juergen Ributzka Date: Mon, 23 Jun 2014 21:55:36 +0000 Subject: [PATCH] [FastISel][X86] Optimize selects when the condition comes from a compare. Optimize the select instructions sequence to use the EFLAGS directly from a compare when possible. llvm-svn: 211543 --- llvm/lib/Target/X86/X86FastISel.cpp | 180 +++++++++++++---- llvm/lib/Target/X86/X86InstrInfo.cpp | 4 +- llvm/lib/Target/X86/X86InstrInfo.h | 5 + llvm/test/CodeGen/X86/fast-isel-select-cmov.ll | 62 ++++++ llvm/test/CodeGen/X86/fast-isel-select-cmov2.ll | 255 ++++++++++++++++++++++++ llvm/test/CodeGen/X86/fast-isel-select.ll | 4 +- 6 files changed, 471 insertions(+), 39 deletions(-) create mode 100644 llvm/test/CodeGen/X86/fast-isel-select-cmov.ll create mode 100644 llvm/test/CodeGen/X86/fast-isel-select-cmov2.ll diff --git a/llvm/lib/Target/X86/X86FastISel.cpp b/llvm/lib/Target/X86/X86FastISel.cpp index 6bccd12..3c9acba 100644 --- a/llvm/lib/Target/X86/X86FastISel.cpp +++ b/llvm/lib/Target/X86/X86FastISel.cpp @@ -111,6 +111,8 @@ private: bool X86SelectDivRem(const Instruction *I); + bool X86FastEmitCMoveSelect(const Instruction *I); + bool X86SelectSelect(const Instruction *I); bool X86SelectTrunc(const Instruction *I); @@ -1611,50 +1613,158 @@ bool X86FastISel::X86SelectDivRem(const Instruction *I) { return true; } -bool X86FastISel::X86SelectSelect(const Instruction *I) { - MVT VT; - if (!isTypeLegal(I->getType(), VT)) +/// \brief Emit a conditional move instruction (if the are supported) to lower +/// the select. +bool X86FastISel::X86FastEmitCMoveSelect(const Instruction *I) { + MVT RetVT; + if (!isTypeLegal(I->getType(), RetVT)) return false; - // We only use cmov here, if we don't have a cmov instruction bail. - if (!Subtarget->hasCMov()) return false; - - unsigned Opc = 0; - const TargetRegisterClass *RC = nullptr; - if (VT == MVT::i16) { - Opc = X86::CMOVE16rr; - RC = &X86::GR16RegClass; - } else if (VT == MVT::i32) { - Opc = X86::CMOVE32rr; - RC = &X86::GR32RegClass; - } else if (VT == MVT::i64) { - Opc = X86::CMOVE64rr; - RC = &X86::GR64RegClass; - } else { + // Check if the subtarget supports these instructions. + if (!Subtarget->hasCMov()) return false; + + // FIXME: Add support for i8. + unsigned Opc; + switch (RetVT.SimpleTy) { + default: return false; + case MVT::i16: Opc = X86::CMOVNE16rr; break; + case MVT::i32: Opc = X86::CMOVNE32rr; break; + case MVT::i64: Opc = X86::CMOVNE64rr; break; } - unsigned Op0Reg = getRegForValue(I->getOperand(0)); - if (Op0Reg == 0) return false; - unsigned Op1Reg = getRegForValue(I->getOperand(1)); - if (Op1Reg == 0) return false; - unsigned Op2Reg = getRegForValue(I->getOperand(2)); - if (Op2Reg == 0) return false; - - // Selects operate on i1, however, Op0Reg is 8 bits width and may contain - // garbage. Indeed, only the less significant bit is supposed to be accurate. - // If we read more than the lsb, we may see non-zero values whereas lsb - // is zero. Therefore, we have to truncate Op0Reg to i1 for the select. - // This is achieved by performing TEST against 1. - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) - .addReg(Op0Reg).addImm(1); - unsigned ResultReg = createResultReg(RC); - BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg) - .addReg(Op1Reg).addReg(Op2Reg); + const Value *Cond = I->getOperand(0); + const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT); + bool NeedTest = true; + + // Optimize conditons coming from a compare. + if (const auto *CI = dyn_cast(Cond)) { + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + + // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. + static unsigned SETFOpcTable[2][3] = { + { X86::SETNPr, X86::SETEr , X86::TEST8rr }, + { X86::SETPr, X86::SETNEr, X86::OR8rr } + }; + unsigned *SETFOpc = nullptr; + switch (Predicate) { + default: break; + case CmpInst::FCMP_OEQ: + SETFOpc = &SETFOpcTable[0][0]; + Predicate = CmpInst::ICMP_NE; + break; + case CmpInst::FCMP_UNE: + SETFOpc = &SETFOpcTable[1][0]; + Predicate = CmpInst::ICMP_NE; + break; + } + + X86::CondCode CC; + bool NeedSwap; + std::tie(CC, NeedSwap) = getX86ConditonCode(Predicate); + assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code."); + Opc = X86::getCMovFromCond(CC, RC->getSize()); + + const Value *CmpLHS = CI->getOperand(0); + const Value *CmpRHS = CI->getOperand(1); + if (NeedSwap) + std::swap(CmpLHS, CmpRHS); + + EVT CmpVT = TLI.getValueType(CmpLHS->getType()); + // Emit a compare of the LHS and RHS, setting the flags. + if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT)) + return false; + + if (SETFOpc) { + unsigned FlagReg1 = createResultReg(&X86::GR8RegClass); + unsigned FlagReg2 = createResultReg(&X86::GR8RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[0]), + FlagReg1); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[1]), + FlagReg2); + auto const &II = TII.get(SETFOpc[2]); + if (II.getNumDefs()) { + unsigned TmpReg = createResultReg(&X86::GR8RegClass); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, TmpReg) + .addReg(FlagReg2).addReg(FlagReg1); + } else { + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II) + .addReg(FlagReg2).addReg(FlagReg1); + } + } + NeedTest = false; + } + + if (NeedTest) { + // Selects operate on i1, however, CondReg is 8 bits width and may contain + // garbage. Indeed, only the less significant bit is supposed to be + // accurate. If we read more than the lsb, we may see non-zero values + // whereas lsb is zero. Therefore, we have to truncate Op0Reg to i1 for + // the select. This is achieved by performing TEST against 1. + unsigned CondReg = getRegForValue(Cond); + if (CondReg == 0) + return false; + bool CondIsKill = hasTrivialKill(Cond); + + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri)) + .addReg(CondReg, getKillRegState(CondIsKill)).addImm(1); + } + + const Value *LHS = I->getOperand(1); + const Value *RHS = I->getOperand(2); + + unsigned RHSReg = getRegForValue(RHS); + bool RHSIsKill = hasTrivialKill(RHS); + + unsigned LHSReg = getRegForValue(LHS); + bool LHSIsKill = hasTrivialKill(LHS); + + if (!LHSReg || !RHSReg) + return false; + + unsigned ResultReg = FastEmitInst_rr(Opc, RC, RHSReg, RHSIsKill, + LHSReg, LHSIsKill); UpdateValueMap(I, ResultReg); return true; } +bool X86FastISel::X86SelectSelect(const Instruction *I) { + MVT RetVT; + if (!isTypeLegal(I->getType(), RetVT)) + return false; + + // Check if we can fold the select. + if (const auto *CI = dyn_cast(I->getOperand(0))) { + CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); + const Value *Opnd = nullptr; + switch (Predicate) { + default: break; + case CmpInst::FCMP_FALSE: Opnd = I->getOperand(2); break; + case CmpInst::FCMP_TRUE: Opnd = I->getOperand(1); break; + } + // No need for a select anymore - this is an unconditional move. + if (Opnd) { + unsigned OpReg = getRegForValue(Opnd); + if (OpReg == 0) + return false; + bool OpIsKill = hasTrivialKill(Opnd); + const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT); + unsigned ResultReg = createResultReg(RC); + BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, + TII.get(TargetOpcode::COPY), ResultReg) + .addReg(OpReg, getKillRegState(OpIsKill)); + UpdateValueMap(I, ResultReg); + return true; + } + } + + // First try to use real conditional move instructions. + if (X86FastEmitCMoveSelect(I)) + return true; + + return false; +} + bool X86FastISel::X86SelectFPExt(const Instruction *I) { // fpext from float to double. if (X86ScalarSSEf64 && diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp index bfc8e27..0797fc6 100644 --- a/llvm/lib/Target/X86/X86InstrInfo.cpp +++ b/llvm/lib/Target/X86/X86InstrInfo.cpp @@ -2696,8 +2696,8 @@ unsigned X86::getSETFromCond(CondCode CC, bool HasMemoryOperand) { /// getCMovFromCond - Return a cmov opcode for the given condition, /// register size in bytes, and operand type. -static unsigned getCMovFromCond(X86::CondCode CC, unsigned RegBytes, - bool HasMemoryOperand) { +unsigned X86::getCMovFromCond(CondCode CC, unsigned RegBytes, + bool HasMemoryOperand) { static const uint16_t Opc[32][3] = { { X86::CMOVA16rr, X86::CMOVA32rr, X86::CMOVA64rr }, { X86::CMOVAE16rr, X86::CMOVAE32rr, X86::CMOVAE64rr }, diff --git a/llvm/lib/Target/X86/X86InstrInfo.h b/llvm/lib/Target/X86/X86InstrInfo.h index d76c52c..c177e3a 100644 --- a/llvm/lib/Target/X86/X86InstrInfo.h +++ b/llvm/lib/Target/X86/X86InstrInfo.h @@ -66,6 +66,11 @@ namespace X86 { /// a memory operand. unsigned getSETFromCond(CondCode CC, bool HasMemoryOperand = false); + /// \brief Return a cmov opcode for the given condition, register size in + /// bytes, and operand type. + unsigned getCMovFromCond(CondCode CC, unsigned RegBytes, + bool HasMemoryOperand = false); + // Turn CMov opcode into condition code. CondCode getCondFromCMovOpc(unsigned Opc); diff --git a/llvm/test/CodeGen/X86/fast-isel-select-cmov.ll b/llvm/test/CodeGen/X86/fast-isel-select-cmov.ll new file mode 100644 index 0000000..8008e28 --- /dev/null +++ b/llvm/test/CodeGen/X86/fast-isel-select-cmov.ll @@ -0,0 +1,62 @@ +; RUN: llc < %s -fast-isel -fast-isel-abort -mtriple=x86_64-apple-darwin10 | FileCheck %s + +; Test conditional move for the supported types (i16, i32, and i32) and +; conditon input (argument or cmp). Currently i8 is not supported. + +define zeroext i16 @select_cmov_i16(i1 zeroext %cond, i16 zeroext %a, i16 zeroext %b) { +; CHECK-LABEL: select_cmov_i16 +; CHECK: testb $1, %dil +; CHECK-NEXT: cmovew %dx, %si +; CHECK-NEXT: movzwl %si, %eax + %1 = select i1 %cond, i16 %a, i16 %b + ret i16 %1 +} + +define zeroext i16 @select_cmp_cmov_i16(i16 zeroext %a, i16 zeroext %b) { +; CHECK-LABEL: select_cmp_cmov_i16 +; CHECK: cmpw %si, %di +; CHECK-NEXT: cmovbw %di, %si +; CHECK-NEXT: movzwl %si, %eax + %1 = icmp ult i16 %a, %b + %2 = select i1 %1, i16 %a, i16 %b + ret i16 %2 +} + +define i32 @select_cmov_i32(i1 zeroext %cond, i32 %a, i32 %b) { +; CHECK-LABEL: select_cmov_i32 +; CHECK: testb $1, %dil +; CHECK-NEXT: cmovel %edx, %esi +; CHECK-NEXT: movl %esi, %eax + %1 = select i1 %cond, i32 %a, i32 %b + ret i32 %1 +} + +define i32 @select_cmp_cmov_i32(i32 %a, i32 %b) { +; CHECK-LABEL: select_cmp_cmov_i32 +; CHECK: cmpl %esi, %edi +; CHECK-NEXT: cmovbl %edi, %esi +; CHECK-NEXT: movl %esi, %eax + %1 = icmp ult i32 %a, %b + %2 = select i1 %1, i32 %a, i32 %b + ret i32 %2 +} + +define i64 @select_cmov_i64(i1 zeroext %cond, i64 %a, i64 %b) { +; CHECK-LABEL: select_cmov_i64 +; CHECK: testb $1, %dil +; CHECK-NEXT: cmoveq %rdx, %rsi +; CHECK-NEXT: movq %rsi, %rax + %1 = select i1 %cond, i64 %a, i64 %b + ret i64 %1 +} + +define i64 @select_cmp_cmov_i64(i64 %a, i64 %b) { +; CHECK-LABEL: select_cmp_cmov_i64 +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovbq %rdi, %rsi +; CHECK-NEXT: movq %rsi, %rax + %1 = icmp ult i64 %a, %b + %2 = select i1 %1, i64 %a, i64 %b + ret i64 %2 +} + diff --git a/llvm/test/CodeGen/X86/fast-isel-select-cmov2.ll b/llvm/test/CodeGen/X86/fast-isel-select-cmov2.ll new file mode 100644 index 0000000..658098f --- /dev/null +++ b/llvm/test/CodeGen/X86/fast-isel-select-cmov2.ll @@ -0,0 +1,255 @@ +; RUN: llc < %s -mtriple=x86_64-apple-darwin10 | FileCheck %s +; RUN: llc < %s -mtriple=x86_64-apple-darwin10 -fast-isel -fast-isel-abort | FileCheck %s + +; Test all the cmp predicates that can feed an integer conditional move. + +define i64 @select_fcmp_false_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_false_cmov +; CHECK: movq %rsi, %rax +; CHECK-NEXT: retq + %1 = fcmp false double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_oeq_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_oeq_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: setnp %al +; CHECK-NEXT: sete %cl +; CHECK-NEXT: testb %al, %cl +; CHECK-NEXT: cmoveq %rsi, %rdi + %1 = fcmp oeq double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ogt_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ogt_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovbeq %rsi, %rdi + %1 = fcmp ogt double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_oge_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_oge_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovbq %rsi, %rdi + %1 = fcmp oge double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_olt_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_olt_cmov +; CHECK: ucomisd %xmm0, %xmm1 +; CHECK-NEXT: cmovbeq %rsi, %rdi + %1 = fcmp olt double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ole_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ole_cmov +; CHECK: ucomisd %xmm0, %xmm1 +; CHECK-NEXT: cmovbq %rsi, %rdi + %1 = fcmp ole double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_one_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_one_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmoveq %rsi, %rdi + %1 = fcmp one double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ord_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ord_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovpq %rsi, %rdi + %1 = fcmp ord double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_uno_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_uno_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovnpq %rsi, %rdi + %1 = fcmp uno double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ueq_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ueq_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovneq %rsi, %rdi + %1 = fcmp ueq double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ugt_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ugt_cmov +; CHECK: ucomisd %xmm0, %xmm1 +; CHECK-NEXT: cmovaeq %rsi, %rdi + %1 = fcmp ugt double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_uge_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_uge_cmov +; CHECK: ucomisd %xmm0, %xmm1 +; CHECK-NEXT: cmovaq %rsi, %rdi + %1 = fcmp uge double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ult_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ult_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovaeq %rsi, %rdi + %1 = fcmp ult double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_ule_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_ule_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: cmovaq %rsi, %rdi + %1 = fcmp ule double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_une_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_une_cmov +; CHECK: ucomisd %xmm1, %xmm0 +; CHECK-NEXT: setp %al +; CHECK-NEXT: setne %cl +; CHECK-NEXT: orb %al, %cl +; CHECK-NEXT: cmoveq %rsi, %rdi + %1 = fcmp une double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_fcmp_true_cmov(double %a, double %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_fcmp_true_cmov +; CHECK: movq %rdi, %rax + %1 = fcmp true double %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_eq_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_eq_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovneq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp eq i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_ne_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_ne_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmoveq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp ne i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_ugt_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_ugt_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovbeq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp ugt i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + + +define i64 @select_icmp_uge_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_uge_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovbq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp uge i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_ult_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_ult_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovaeq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp ult i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_ule_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_ule_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovaq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp ule i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_sgt_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_sgt_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovleq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp sgt i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_sge_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_sge_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovlq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp sge i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_slt_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_slt_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovgeq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp slt i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + +define i64 @select_icmp_sle_cmov(i64 %a, i64 %b, i64 %c, i64 %d) { +; CHECK-LABEL: select_icmp_sle_cmov +; CHECK: cmpq %rsi, %rdi +; CHECK-NEXT: cmovgq %rcx, %rdx +; CHECK-NEXT: movq %rdx, %rax + %1 = icmp sle i64 %a, %b + %2 = select i1 %1, i64 %c, i64 %d + ret i64 %2 +} + diff --git a/llvm/test/CodeGen/X86/fast-isel-select.ll b/llvm/test/CodeGen/X86/fast-isel-select.ll index 53158bc..7b3c99f 100644 --- a/llvm/test/CodeGen/X86/fast-isel-select.ll +++ b/llvm/test/CodeGen/X86/fast-isel-select.ll @@ -4,10 +4,10 @@ ; lsb is zero. ; -; CHECK-LABEL: fastisel_select: +; CHECK-LABEL: fastisel_select: ; CHECK: subb {{%[a-z0-9]+}}, [[RES:%[a-z0-9]+]] ; CHECK: testb $1, [[RES]] -; CHECK: cmovel +; CHECK: cmovnel %edi, %esi define i32 @fastisel_select(i1 %exchSub2211_, i1 %trunc_8766) { %shuffleInternal15257_8932 = sub i1 %exchSub2211_, %trunc_8766 %counter_diff1345 = select i1 %shuffleInternal15257_8932, i32 1204476887, i32 0 -- 2.7.4