From: Dale Johannesen Date: Thu, 14 Jun 2007 22:03:45 +0000 (+0000) Subject: Do not treat FP_REG_KILL as terminator in branch analysis (X86). X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=318093b6f8d21ac8eab34573b0526984895fe941;p=oota-llvm.git Do not treat FP_REG_KILL as terminator in branch analysis (X86). git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@37578 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/include/llvm/Target/TargetInstrInfo.h b/include/llvm/Target/TargetInstrInfo.h index e221466606b..2d73dbf83c8 100644 --- a/include/llvm/Target/TargetInstrInfo.h +++ b/include/llvm/Target/TargetInstrInfo.h @@ -417,7 +417,7 @@ public: /// isUnpredicatedTerminator - Returns true if the instruction is a /// terminator instruction that has not been predicated. - bool isUnpredicatedTerminator(const MachineInstr *MI) const; + virtual bool isUnpredicatedTerminator(const MachineInstr *MI) const; /// PredicateInstruction - Convert the instruction into a predicated /// instruction. It returns true if the operation was successful. diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp index f06827522c3..c8621ad1b73 100644 --- a/lib/Target/X86/X86InstrInfo.cpp +++ b/lib/Target/X86/X86InstrInfo.cpp @@ -398,13 +398,20 @@ X86::CondCode X86::GetOppositeBranchCondition(X86::CondCode CC) { } } +// For purposes of branch analysis do not count FP_REG_KILL as a terminator. +bool X86InstrInfo::isUnpredicatedTerminator(const MachineInstr *MI) const { + const TargetInstrDescriptor *TID = MI->getInstrDescriptor(); + if (MI->getOpcode() == X86::FP_REG_KILL) + return false; + if (TID->Flags & M_TERMINATOR_FLAG) + return !isPredicated(MI); + return false; +} bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB, std::vector &Cond) const { - // TODO: If FP_REG_KILL is around, ignore it. - // If the block has no terminators, it just falls into the block after it. MachineBasicBlock::iterator I = MBB.end(); if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) @@ -439,8 +446,7 @@ bool X86InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, MachineInstr *SecondLastInst = I; // If there are three terminators, we don't know what sort of block this is. - if (SecondLastInst && I != MBB.begin() && - isTerminatorInstr((--I)->getOpcode())) + if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I)) return true; // If the block ends with X86::JMP and a conditional branch, handle it. diff --git a/lib/Target/X86/X86InstrInfo.h b/lib/Target/X86/X86InstrInfo.h index 3d3952a6ee3..75cce47401d 100644 --- a/lib/Target/X86/X86InstrInfo.h +++ b/lib/Target/X86/X86InstrInfo.h @@ -261,6 +261,7 @@ public: virtual MachineInstr *commuteInstruction(MachineInstr *MI) const; // Branch analysis. + virtual bool isUnpredicatedTerminator(const MachineInstr* MI) const; virtual bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB, std::vector &Cond) const; diff --git a/test/CodeGen/X86/2007-06-14-branchfold.ll b/test/CodeGen/X86/2007-06-14-branchfold.ll new file mode 100644 index 00000000000..a7194a08ff8 --- /dev/null +++ b/test/CodeGen/X86/2007-06-14-branchfold.ll @@ -0,0 +1,137 @@ +; RUN: llvm-as < %s | llc -mcpu=i686 | grep jmp | wc -l | grep 1 +; check that branch folding understands FP_REG_KILL is not a branch +; the remaining jmp can be removed if we take advantage of knowing +; abort does not return + +; ModuleID = 'g.bc' +target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64" +target triple = "i686-pc-linux-gnu" + %struct.FRAME.c34003a = type { float, float } +@report_E = global i8 0 ; [#uses=0] + +define void @main() { +entry: + %FRAME.31 = alloca %struct.FRAME.c34003a, align 8 ; <%struct.FRAME.c34003a*> [#uses=4] + %tmp20 = call i32 @report__ident_int( i32 -50 ) ; [#uses=1] + %tmp2021 = sitofp i32 %tmp20 to float ; [#uses=5] + %tmp23 = fcmp ult float %tmp2021, 0xC7EFFFFFE0000000 ; [#uses=1] + %tmp26 = fcmp ugt float %tmp2021, 0x47EFFFFFE0000000 ; [#uses=1] + %bothcond = or i1 %tmp23, %tmp26 ; [#uses=1] + br i1 %bothcond, label %bb, label %bb30 + +bb: ; preds = %entry + unwind + +bb30: ; preds = %entry + %tmp35 = call i32 @report__ident_int( i32 50 ) ; [#uses=1] + %tmp3536 = sitofp i32 %tmp35 to float ; [#uses=4] + %tmp38 = fcmp ult float %tmp3536, 0xC7EFFFFFE0000000 ; [#uses=1] + %tmp44 = fcmp ugt float %tmp3536, 0x47EFFFFFE0000000 ; [#uses=1] + %bothcond226 = or i1 %tmp38, %tmp44 ; [#uses=1] + br i1 %bothcond226, label %bb47, label %bb49 + +bb47: ; preds = %bb30 + unwind + +bb49: ; preds = %bb30 + %tmp60 = fcmp ult float %tmp3536, %tmp2021 ; [#uses=1] + %tmp60.not = xor i1 %tmp60, true ; [#uses=1] + %tmp65 = fcmp olt float %tmp2021, 0xC7EFFFFFE0000000 ; [#uses=1] + %bothcond227 = and i1 %tmp65, %tmp60.not ; [#uses=1] + br i1 %bothcond227, label %cond_true68, label %cond_next70 + +cond_true68: ; preds = %bb49 + unwind + +cond_next70: ; preds = %bb49 + %tmp71 = call i32 @report__ident_int( i32 -30 ) ; [#uses=1] + %tmp7172 = sitofp i32 %tmp71 to float ; [#uses=3] + %tmp74 = fcmp ult float %tmp7172, 0xC7EFFFFFE0000000 ; [#uses=1] + %tmp80 = fcmp ugt float %tmp7172, 0x47EFFFFFE0000000 ; [#uses=1] + %bothcond228 = or i1 %tmp74, %tmp80 ; [#uses=1] + br i1 %bothcond228, label %bb83, label %bb85 + +bb83: ; preds = %cond_next70 + unwind + +bb85: ; preds = %cond_next70 + %tmp90 = getelementptr %struct.FRAME.c34003a* %FRAME.31, i32 0, i32 1 ; [#uses=3] + store float %tmp7172, float* %tmp90 + %tmp92 = call i32 @report__ident_int( i32 30 ) ; [#uses=1] + %tmp9293 = sitofp i32 %tmp92 to float ; [#uses=7] + %tmp95 = fcmp ult float %tmp9293, 0xC7EFFFFFE0000000 ; [#uses=1] + %tmp101 = fcmp ugt float %tmp9293, 0x47EFFFFFE0000000 ; [#uses=1] + %bothcond229 = or i1 %tmp95, %tmp101 ; [#uses=1] + br i1 %bothcond229, label %bb104, label %bb106 + +bb104: ; preds = %bb85 + unwind + +bb106: ; preds = %bb85 + %tmp111 = getelementptr %struct.FRAME.c34003a* %FRAME.31, i32 0, i32 0 ; [#uses=2] + store float %tmp9293, float* %tmp111 + %tmp123 = load float* %tmp90 ; [#uses=4] + %tmp125 = fcmp ult float %tmp9293, %tmp123 ; [#uses=1] + br i1 %tmp125, label %cond_next147, label %cond_true128 + +cond_true128: ; preds = %bb106 + %tmp133 = fcmp olt float %tmp123, %tmp2021 ; [#uses=1] + %tmp142 = fcmp ogt float %tmp9293, %tmp3536 ; [#uses=1] + %bothcond230 = or i1 %tmp133, %tmp142 ; [#uses=1] + br i1 %bothcond230, label %bb145, label %cond_next147 + +bb145: ; preds = %cond_true128 + unwind + +cond_next147: ; preds = %cond_true128, %bb106 + %tmp157 = fcmp ugt float %tmp123, -3.000000e+01 ; [#uses=1] + %tmp165 = fcmp ult float %tmp9293, -3.000000e+01 ; [#uses=1] + %bothcond231 = or i1 %tmp157, %tmp165 ; [#uses=1] + br i1 %bothcond231, label %bb168, label %bb169 + +bb168: ; preds = %cond_next147 + unwind + +bb169: ; preds = %cond_next147 + %tmp176 = fcmp ugt float %tmp123, 3.000000e+01 ; [#uses=1] + %tmp184 = fcmp ult float %tmp9293, 3.000000e+01 ; [#uses=1] + %bothcond232 = or i1 %tmp176, %tmp184 ; [#uses=1] + br i1 %bothcond232, label %bb187, label %bb188 + +bb187: ; preds = %bb169 + unwind + +bb188: ; preds = %bb169 + %tmp192 = call fastcc float @c34003a__ident.154( %struct.FRAME.c34003a* %FRAME.31, float 3.000000e+01 ) ; [#uses=2] + %tmp194 = load float* %tmp90 ; [#uses=1] + %tmp196 = fcmp ugt float %tmp194, 0.000000e+00 ; [#uses=1] + br i1 %tmp196, label %bb207, label %cond_next200 + +cond_next200: ; preds = %bb188 + %tmp202 = load float* %tmp111 ; [#uses=1] + %tmp204 = fcmp ult float %tmp202, 0.000000e+00 ; [#uses=1] + br i1 %tmp204, label %bb207, label %bb208 + +bb207: ; preds = %cond_next200, %bb188 + unwind + +bb208: ; preds = %cond_next200 + %tmp212 = call fastcc float @c34003a__ident.154( %struct.FRAME.c34003a* %FRAME.31, float 0.000000e+00 ) ; [#uses=1] + %tmp214 = fcmp oge float %tmp212, %tmp192 ; [#uses=1] + %tmp217 = fcmp oge float %tmp192, 1.000000e+02 ; [#uses=1] + %tmp221 = or i1 %tmp214, %tmp217 ; [#uses=1] + br i1 %tmp221, label %cond_true224, label %UnifiedReturnBlock + +cond_true224: ; preds = %bb208 + call void @abort( ) noreturn + ret void + +UnifiedReturnBlock: ; preds = %bb208 + ret void +} + +declare fastcc float @c34003a__ident.154(%struct.FRAME.c34003a* %CHAIN.32, float %x) + +declare i32 @report__ident_int(i32 %x) + +declare void @abort() noreturn