Index: lib/Target/X86/X86InstrInfo.h =================================================================== --- lib/Target/X86/X86InstrInfo.h +++ lib/Target/X86/X86InstrInfo.h @@ -30,35 +30,44 @@ namespace X86 { // X86 specific condition code. These correspond to X86_*_COND in // X86InstrInfo.td. They must be kept in synch. - enum CondCode { - COND_A = 0, - COND_AE = 1, - COND_B = 2, - COND_BE = 3, - COND_E = 4, - COND_G = 5, - COND_GE = 6, - COND_L = 7, - COND_LE = 8, - COND_NE = 9, - COND_NO = 10, - COND_NP = 11, - COND_NS = 12, - COND_O = 13, - COND_P = 14, - COND_S = 15, - LAST_VALID_COND = COND_S, - - // Artificial condition codes. These are used by AnalyzeBranch - // to indicate a block terminated with two conditional branches to - // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE, - // which can't be represented on x86 with a single condition. These - // are never used in MachineInstrs. - COND_NE_OR_P, - COND_NP_OR_E, - - COND_INVALID - }; +enum CondCode { + COND_A = 0, + COND_AE = 1, + COND_B = 2, + COND_BE = 3, + COND_E = 4, + COND_G = 5, + COND_GE = 6, + COND_L = 7, + COND_LE = 8, + COND_NE = 9, + COND_NO = 10, + COND_NP = 11, + COND_NS = 12, + COND_O = 13, + COND_P = 14, + COND_S = 15, + LAST_VALID_COND = COND_S, + + // Artificial condition codes. These are used by AnalyzeBranch + // to indicate a block terminated with two conditional branches to + // the same location. This occurs in code using FCMP_OEQ or FCMP_UNE, + // which can't be represented on x86 with a single condition. These + // are never used in MachineInstrs. + COND_NE_OR_P, + COND_NP_OR_E, + + // Artificial condition codes. These are used to represent the negation of + // above two conditions. The only scenario we need these two conditions is + // when we try to reverse above two conditions in order to remove redundant + // unconditional jumps. Note that both true and false bodies need to be + // avaiable in order to correctly synthesize instructions for them. These are + // never used in MachineInstrs. + COND_E_AND_NP, // negate of COND_NE_OR_P + COND_P_AND_NE, // negate of COND_NP_OR_E + + COND_INVALID +}; // Turn condition code into conditional branch opcode. unsigned GetCondBranchFromCond(CondCode CC); Index: lib/Target/X86/X86InstrInfo.cpp =================================================================== --- lib/Target/X86/X86InstrInfo.cpp +++ lib/Target/X86/X86InstrInfo.cpp @@ -3793,6 +3793,10 @@ case X86::COND_NP: return X86::COND_P; case X86::COND_O: return X86::COND_NO; case X86::COND_NO: return X86::COND_O; + case X86::COND_NE_OR_P: return X86::COND_E_AND_NP; + case X86::COND_NP_OR_E: return X86::COND_P_AND_NE; + case X86::COND_E_AND_NP: return X86::COND_NE_OR_P; + case X86::COND_P_AND_NE: return X86::COND_NP_OR_E; } } @@ -3986,9 +3990,9 @@ MachineBasicBlock::iterator OldInst = I; BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(JNCC)) - .addMBB(UnCondBrIter->getOperand(0).getMBB()); + .addMBB(UnCondBrIter->getOperand(0).getMBB()); BuildMI(MBB, UnCondBrIter, MBB.findDebugLoc(I), get(X86::JMP_1)) - .addMBB(TargetBB); + .addMBB(TargetBB); OldInst->eraseFromParent(); UnCondBrIter->eraseFromParent(); @@ -4012,11 +4016,6 @@ assert(Cond.size() == 1); assert(TBB); - // Only handle the case where all conditional branches branch to the same - // destination. - if (TBB != I->getOperand(0).getMBB()) - return true; - // If the conditions are the same, we can leave them alone. X86::CondCode OldBranchCode = (X86::CondCode)Cond[0].getImm(); if (OldBranchCode == BranchCode) @@ -4025,17 +4024,40 @@ // If they differ, see if they fit one of the known patterns. Theoretically, // we could handle more patterns here, but we shouldn't expect to see them // if instruction selection has done a reasonable job. - if ((OldBranchCode == X86::COND_NP && - BranchCode == X86::COND_E) || - (OldBranchCode == X86::COND_E && - BranchCode == X86::COND_NP)) + auto NewTBB = I->getOperand(0).getMBB(); + if (TBB == NewTBB && + ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_E) || + (OldBranchCode == X86::COND_E && BranchCode == X86::COND_NP))) { BranchCode = X86::COND_NP_OR_E; - else if ((OldBranchCode == X86::COND_P && - BranchCode == X86::COND_NE) || - (OldBranchCode == X86::COND_NE && - BranchCode == X86::COND_P)) + } else if (TBB == NewTBB && + ((OldBranchCode == X86::COND_P && BranchCode == X86::COND_NE) || + (OldBranchCode == X86::COND_NE && BranchCode == X86::COND_P))) { BranchCode = X86::COND_NE_OR_P; - else + } else if ((OldBranchCode == X86::COND_NE && BranchCode == X86::COND_NP) || + (OldBranchCode == X86::COND_P && BranchCode == X86::COND_E)) { + // X86::COND_P_AND_NE usually has two different branch destinations. + // + // JNP B1 + // JNE B2 + // B1: (fall-through) + // B2: + // + // Here this condition branches to B2 only if P && NE. It has another + // equivalent form: + // + // JE B1 + // JP B2 + // B1: (fall-through) + // B2: + // + // Similarly it branches to B2 only if NE && P. That is why this condition + // is named COND_P_AND_NE. + BranchCode = X86::COND_P_AND_NE; + } else if ((OldBranchCode == X86::COND_NP && BranchCode == X86::COND_NE) || + (OldBranchCode == X86::COND_E && BranchCode == X86::COND_P)) { + // See comments above for X86::COND_P_AND_NE. + BranchCode = X86::COND_E_AND_NP; + } else return true; // Update the MachineOperand. @@ -4160,6 +4182,9 @@ return 1; } + // If FBB is null, it is implied to be a fall-through block. + bool FallThru = FBB == nullptr; + // Conditional branch. unsigned Count = 0; X86::CondCode CC = (X86::CondCode)Cond[0].getImm(); @@ -4178,13 +4203,43 @@ BuildMI(&MBB, DL, get(X86::JP_1)).addMBB(TBB); ++Count; break; + case X86::COND_P_AND_NE: + // If FBB is null, it is implied to be the next block of MBB. + if (FBB == nullptr) { + MachineFunction::iterator I = std::next(MBB.getIterator()); + FBB = &*I; + assert(I != TBB->getParent()->end() && "MBB cannot be the last block in " + "function when the false body is " + "a fall-through."); + } + // Synthesize NEG_NP_OR_E with two branches. + BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(FBB); + ++Count; + BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(TBB); + ++Count; + break; + case X86::COND_E_AND_NP: + // If FBB is null, it is implied to be the next block of MBB. + if (FBB == nullptr) { + MachineFunction::iterator I = std::next(MBB.getIterator()); + FBB = &*I; + assert(I != TBB->getParent()->end() && "MBB cannot be the last block in " + "function when the false body is " + "a fall-through."); + } + // Synthesize NEG_NE_OR_P with two branches. + BuildMI(&MBB, DL, get(X86::JNE_1)).addMBB(FBB); + ++Count; + BuildMI(&MBB, DL, get(X86::JNP_1)).addMBB(TBB); + ++Count; + break; default: { unsigned Opc = GetCondBranchFromCond(CC); BuildMI(&MBB, DL, get(Opc)).addMBB(TBB); ++Count; } } - if (FBB) { + if (!FallThru) { // Two-way Conditional branch. Insert the second branch. BuildMI(&MBB, DL, get(X86::JMP_1)).addMBB(FBB); ++Count; @@ -6724,8 +6779,6 @@ ReverseBranchCondition(SmallVectorImpl &Cond) const { assert(Cond.size() == 1 && "Invalid X86 branch condition!"); X86::CondCode CC = static_cast(Cond[0].getImm()); - if (CC == X86::COND_NE_OR_P || CC == X86::COND_NP_OR_E) - return true; Cond[0].setImm(GetOppositeBranchCondition(CC)); return false; } Index: test/CodeGen/X86/block-placement.ll =================================================================== --- test/CodeGen/X86/block-placement.ll +++ test/CodeGen/X86/block-placement.ll @@ -463,26 +463,22 @@ } define void @fpcmp_unanalyzable_branch(i1 %cond) { -; This function's CFG contains an unanalyzable branch that is likely to be -; split due to having a different high-probability predecessor. +; This function's CFG contains an once-unanalyzable branch (une on floating +; points). As now it becomes analyzable, we should get best layout in which each +; edge in 'entry' -> 'entry.if.then_crit_edge' -> 'if.then' -> 'if.end' is +; fall-through. ; CHECK: fpcmp_unanalyzable_branch ; CHECK: %entry +; CHECK: %if.then +; CHECK: %if.end ; CHECK: %exit -; CHECK-NOT: %if.then -; CHECK-NOT: %if.end -; CHECK-NOT: jne -; CHECK-NOT: jnp ; CHECK: jne ; CHECK-NEXT: jnp -; CHECK-NEXT: %if.then entry: ; Note that this branch must be strongly biased toward ; 'entry.if.then_crit_edge' to ensure that we would try to form a chain for -; 'entry' -> 'entry.if.then_crit_edge' -> 'if.then'. It is the last edge in that -; chain which would violate the unanalyzable branch in 'exit', but we won't even -; try this trick unless 'if.then' is believed to almost always be reached from -; 'entry.if.then_crit_edge'. +; 'entry' -> 'entry.if.then_crit_edge' -> 'if.then' -> 'if.end'. br i1 %cond, label %entry.if.then_crit_edge, label %lor.lhs.false, !prof !1 entry.if.then_crit_edge: @@ -665,11 +661,14 @@ ; Ensure that we can handle unanalyzable branches where the destination block ; gets selected as the optimal successor to merge. ; +; This branch is now analyzable and hence the destination block becomes the +; hotter one. +; ; CHECK: unanalyzable_branch_to_best_succ ; CHECK: %entry -; CHECK: %foo ; CHECK: %bar ; CHECK: %exit +; CHECK: %foo entry: ; Bias this branch toward bar to ensure we form that chain. Index: test/CodeGen/X86/fast-isel-cmp-branch2.ll =================================================================== --- test/CodeGen/X86/fast-isel-cmp-branch2.ll +++ test/CodeGen/X86/fast-isel-cmp-branch2.ll @@ -5,7 +5,7 @@ ; CHECK-LABEL: fcmp_oeq ; CHECK: ucomiss %xmm1, %xmm0 ; CHECK-NEXT: jne {{LBB.+_1}} -; CHECK-NEXT: jnp {{LBB.+_2}} +; CHECK-NEXT: jp {{LBB.+_1}} %1 = fcmp oeq float %x, %y br i1 %1, label %bb1, label %bb2 bb2: @@ -162,8 +162,7 @@ ; CHECK-LABEL: fcmp_une ; CHECK: ucomiss %xmm1, %xmm0 ; CHECK-NEXT: jne {{LBB.+_2}} -; CHECK-NEXT: jp {{LBB.+_2}} -; CHECK-NEXT: jmp {{LBB.+_1}} +; CHECK-NEXT: jnp {{LBB.+_1}} %1 = fcmp une float %x, %y br i1 %1, label %bb1, label %bb2 bb2: Index: test/CodeGen/X86/fast-isel-cmp-branch3.ll =================================================================== --- test/CodeGen/X86/fast-isel-cmp-branch3.ll +++ test/CodeGen/X86/fast-isel-cmp-branch3.ll @@ -17,7 +17,7 @@ ; CHECK: xorps %xmm1, %xmm1 ; CHECK-NEXT: ucomiss %xmm1, %xmm0 ; CHECK-NEXT: jne {{LBB.+_1}} -; CHECK-NEXT: jnp {{LBB.+_2}} +; CHECK-NEXT: jp {{LBB.+_1}} %1 = fcmp oeq float %x, 0.000000e+00 br i1 %1, label %bb1, label %bb2 bb2: @@ -338,8 +338,7 @@ ; CHECK: xorps %xmm1, %xmm1 ; CHECK-NEXT: ucomiss %xmm1, %xmm0 ; CHECK-NEXT: jne {{LBB.+_2}} -; CHECK-NEXT: jp {{LBB.+_2}} -; CHECK-NEXT: jmp {{LBB.+_1}} +; CHECK-NEXT: jnp {{LBB.+_1}} %1 = fcmp une float %x, 0.000000e+00 br i1 %1, label %bb1, label %bb2 bb2: Index: test/CodeGen/X86/x86-analyze-branch-jne-jp.ll =================================================================== --- /dev/null +++ test/CodeGen/X86/x86-analyze-branch-jne-jp.ll @@ -0,0 +1,21 @@ +; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux < %s | FileCheck %s -check-prefix=CHECK + +; Test if the negation of the non-equality check between floating points are +; translated to jnp followed by jne. + +; CHECK: jne +; CHECK-NEXT: jnp +define void @foo(float %f) { +entry: + %cmp = fcmp une float %f, 0.000000e+00 + br i1 %cmp, label %if.then, label %if.end + +if.then: + tail call void @a() + br label %if.end + +if.end: + ret void +} + +declare void @a()