From 495de3b783c239b4b3244a1e739b2a5b82561b1b Mon Sep 17 00:00:00 2001 From: Bob Wilson Date: Fri, 17 Dec 2010 01:21:12 +0000 Subject: [PATCH] Fix crash compiling a QQQQ REG_SEQUENCE for a Neon vld3_lane operation. Radar 8776599 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122018 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/CodeGen/SelectionDAG/InstrEmitter.cpp | 4 +--- lib/CodeGen/TwoAddressInstructionPass.cpp | 7 +++++-- test/CodeGen/ARM/vldlane.ll | 19 +++++++++++++++++++ 3 files changed, 25 insertions(+), 5 deletions(-) diff --git a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp index 3ba8b3690ad..86f88f1f179 100644 --- a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp +++ b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp @@ -543,9 +543,7 @@ void InstrEmitter::EmitRegSequence(SDNode *Node, const TargetRegisterClass *TRC = MRI->getRegClass(SubReg); const TargetRegisterClass *SRC = TRI->getMatchingSuperRegClass(RC, TRC, SubIdx); - if (!SRC) - llvm_unreachable("Invalid subregister index in REG_SEQUENCE"); - if (SRC != RC) { + if (SRC && SRC != RC) { MRI->setRegClass(NewVReg, SRC); RC = SRC; } diff --git a/lib/CodeGen/TwoAddressInstructionPass.cpp b/lib/CodeGen/TwoAddressInstructionPass.cpp index 60086ce224e..d74d2da6b77 100644 --- a/lib/CodeGen/TwoAddressInstructionPass.cpp +++ b/lib/CodeGen/TwoAddressInstructionPass.cpp @@ -1417,6 +1417,7 @@ bool TwoAddressInstructionPass::EliminateRegSequences() { SmallSet Seen; for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) { unsigned SrcReg = MI->getOperand(i).getReg(); + unsigned SubIdx = MI->getOperand(i+1).getImm(); if (MI->getOperand(i).getSubReg() || TargetRegisterInfo::isPhysicalRegister(SrcReg)) { DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << *MI); @@ -1436,7 +1437,9 @@ bool TwoAddressInstructionPass::EliminateRegSequences() { bool isKill = MI->getOperand(i).isKill(); if (!Seen.insert(SrcReg) || MI->getParent() != DefMI->getParent() || - !isKill || HasOtherRegSequenceUses(SrcReg, MI, MRI)) { + !isKill || HasOtherRegSequenceUses(SrcReg, MI, MRI) || + !TRI->getMatchingSuperRegClass(MRI->getRegClass(DstReg), + MRI->getRegClass(SrcReg), SubIdx)) { // REG_SEQUENCE cannot have duplicated operands, add a copy. // Also add an copy if the source is live-in the block. We don't want // to end up with a partial-redef of a livein, e.g. @@ -1465,7 +1468,7 @@ bool TwoAddressInstructionPass::EliminateRegSequences() { MachineBasicBlock::iterator InsertLoc = MI; MachineInstr *CopyMI = BuildMI(*MI->getParent(), InsertLoc, MI->getDebugLoc(), TII->get(TargetOpcode::COPY)) - .addReg(DstReg, RegState::Define, MI->getOperand(i+1).getImm()) + .addReg(DstReg, RegState::Define, SubIdx) .addReg(SrcReg, getKillRegState(isKill)); MI->getOperand(i).setReg(0); if (LV && isKill) diff --git a/test/CodeGen/ARM/vldlane.ll b/test/CodeGen/ARM/vldlane.ll index fe2a0dc707f..c5514a63fd2 100644 --- a/test/CodeGen/ARM/vldlane.ll +++ b/test/CodeGen/ARM/vldlane.ll @@ -432,3 +432,22 @@ declare %struct.__neon_float32x2x4_t @llvm.arm.neon.vld4lane.v2f32(i8*, <2 x flo declare %struct.__neon_int16x8x4_t @llvm.arm.neon.vld4lane.v8i16(i8*, <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16>, i32, i32) nounwind readonly declare %struct.__neon_int32x4x4_t @llvm.arm.neon.vld4lane.v4i32(i8*, <4 x i32>, <4 x i32>, <4 x i32>, <4 x i32>, i32, i32) nounwind readonly declare %struct.__neon_float32x4x4_t @llvm.arm.neon.vld4lane.v4f32(i8*, <4 x float>, <4 x float>, <4 x float>, <4 x float>, i32, i32) nounwind readonly + +; Radar 8776599: If one of the operands to a QQQQ REG_SEQUENCE is a register +; in the QPR_VFP2 regclass, it needs to be copied to a QPR regclass because +; we don't currently have a QQQQ_VFP2 super-regclass. (The "0" for the low +; part of %ins67 is supposed to be loaded by a VLDRS instruction in this test.) +define void @test_qqqq_regsequence_subreg([6 x i64] %b) nounwind { +;CHECK: test_qqqq_regsequence_subreg +;CHECK: vld3.16 + %tmp63 = extractvalue [6 x i64] %b, 5 + %tmp64 = zext i64 %tmp63 to i128 + %tmp65 = shl i128 %tmp64, 64 + %ins67 = or i128 %tmp65, 0 + %tmp78 = bitcast i128 %ins67 to <8 x i16> + %vld3_lane = tail call %struct.__neon_int16x8x3_t @llvm.arm.neon.vld3lane.v8i16(i8* undef, <8 x i16> undef, <8 x i16> undef, <8 x i16> %tmp78, i32 1, i32 2) + call void @llvm.trap() + unreachable +} + +declare void @llvm.trap() nounwind