From 48d1e4517e6cdbbcdacae6e555a0aefa6bc3975c Mon Sep 17 00:00:00 2001 From: Dale Johannesen Date: Mon, 24 Sep 2007 22:52:39 +0000 Subject: [PATCH] When mixing SSE and x87 codegen, it's possible to have situations where an SSE instruction turns into multiple blocks, with the live range of an x87 register crossing them. To do this correctly make sure we examine all blocks when inserting FP_REG_KILL. PR 1697. (This was exposed by my fix for PR 1681, but the same thing could happen mixing x87 long double with SSE.) git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@42281 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/X86/X86ISelDAGToDAG.cpp | 60 +++++++++++++++--------------- 1 file changed, 31 insertions(+), 29 deletions(-) diff --git a/lib/Target/X86/X86ISelDAGToDAG.cpp b/lib/Target/X86/X86ISelDAGToDAG.cpp index 6b3a660237b..b33ced8ba66 100644 --- a/lib/Target/X86/X86ISelDAGToDAG.cpp +++ b/lib/Target/X86/X86ISelDAGToDAG.cpp @@ -482,14 +482,17 @@ void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) { // block defines any FP values. If so, put an FP_REG_KILL instruction before // the terminator of the block. - // Note that FP stack instructions *are* used in SSE code for long double, - // so we do need this check. - bool ContainsFPCode = false; + // Note that FP stack instructions are used in all modes for long double, + // so we always need to do this check. + // Also note that it's possible for an FP stack register to be live across + // an instruction that produces multiple basic blocks (SSE CMOV) so we + // must check all the generated basic blocks. // Scan all of the machine instructions in these MBBs, checking for FP // stores. (RFP32 and RFP64 will not exist in SSE mode, but RFP80 might.) MachineFunction::iterator MBBI = FirstMBB; do { + bool ContainsFPCode = false; for (MachineBasicBlock::iterator I = MBBI->begin(), E = MBBI->end(); !ContainsFPCode && I != E; ++I) { if (I->getNumOperands() != 0 && I->getOperand(0).isRegister()) { @@ -507,35 +510,34 @@ void X86DAGToDAGISel::InstructionSelectBasicBlock(SelectionDAG &DAG) { } } } - } while (!ContainsFPCode && &*(MBBI++) != BB); - - // Check PHI nodes in successor blocks. These PHI's will be lowered to have - // a copy of the input value in this block. In SSE mode, we only care about - // 80-bit values. - if (!ContainsFPCode) { - // Final check, check LLVM BB's that are successors to the LLVM BB - // corresponding to BB for FP PHI nodes. - const BasicBlock *LLVMBB = BB->getBasicBlock(); - const PHINode *PN; - for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB); - !ContainsFPCode && SI != E; ++SI) { - for (BasicBlock::const_iterator II = SI->begin(); - (PN = dyn_cast(II)); ++II) { - if (PN->getType()==Type::X86_FP80Ty || - (!Subtarget->hasSSE2() && PN->getType()->isFloatingPoint())) { - ContainsFPCode = true; - break; + // Check PHI nodes in successor blocks. These PHI's will be lowered to have + // a copy of the input value in this block. In SSE mode, we only care about + // 80-bit values. + if (!ContainsFPCode) { + // Final check, check LLVM BB's that are successors to the LLVM BB + // corresponding to BB for FP PHI nodes. + const BasicBlock *LLVMBB = BB->getBasicBlock(); + const PHINode *PN; + for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB); + !ContainsFPCode && SI != E; ++SI) { + for (BasicBlock::const_iterator II = SI->begin(); + (PN = dyn_cast(II)); ++II) { + if (PN->getType()==Type::X86_FP80Ty || + (!Subtarget->hasSSE1() && PN->getType()->isFloatingPoint()) || + (!Subtarget->hasSSE2() && PN->getType()==Type::DoubleTy)) { + ContainsFPCode = true; + break; + } } } } - } - - // Finally, if we found any FP code, emit the FP_REG_KILL instruction. - if (ContainsFPCode) { - BuildMI(*BB, BB->getFirstTerminator(), - TM.getInstrInfo()->get(X86::FP_REG_KILL)); - ++NumFPKill; - } + // Finally, if we found any FP code, emit the FP_REG_KILL instruction. + if (ContainsFPCode) { + BuildMI(*MBBI, MBBI->getFirstTerminator(), + TM.getInstrInfo()->get(X86::FP_REG_KILL)); + ++NumFPKill; + } + } while (&*(MBBI++) != BB); } /// MatchAddress - Add the specified node to the specified addressing mode,