https://github.com/MitchBriles updated https://github.com/llvm/llvm-project/pull/240470
>From 4d434562d3d320ea5bd84d981936276a977bfa4f Mon Sep 17 00:00:00 2001 From: Mitch <[email protected]> Date: Fri, 23 Oct 2026 09:51:47 -0600 Subject: [PATCH 1/2] Precommit FinalizeISel FREEZE tests Assisted-By: Claude Opus 5.5 <[email protected]> --- .../CodeGen/AArch64/finalize-isel-freeze.mir | 411 ++++++++++++++++++ 1 file changed, 411 insertions(+) create mode 100644 llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir diff --git a/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir b/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir new file mode 100644 index 0000000000000..d1ae7f7230934 --- /dev/null +++ b/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir @@ -0,0 +1,411 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6 +# RUN: llc -mtriple=aarch64-- -run-pass=finalize-isel -o - %s | FileCheck %s +# RUN: llc -mtriple=aarch64-- -passes=finalize-isel -o - %s | FileCheck %s + +--- +name: freeze_implicit_def +tracksRegLiveness: true +body: | + bb.0: + ; CHECK-LABEL: name: freeze_implicit_def + ; CHECK: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[DEF]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0 + %0:gpr32 = IMPLICIT_DEF + %1:gpr32 = FREEZE %0 + $w0 = COPY %1 + RET_ReallyLR implicit $w0 +... +--- +name: freeze_copy_of_implicit_def +tracksRegLiveness: true +body: | + bb.0: + ; CHECK-LABEL: name: freeze_copy_of_implicit_def + ; CHECK: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY [[DEF]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0 + %0:gpr32 = IMPLICIT_DEF + %1:gpr32 = COPY %0 + %2:gpr32 = FREEZE %1 + $w0 = COPY %2 + RET_ReallyLR implicit $w0 +... +--- +name: freeze_instruction +tracksRegLiveness: true +body: | + bb.0: + liveins: $w0, $w1 + + ; CHECK-LABEL: name: freeze_instruction + ; CHECK: liveins: $w0, $w1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY $w1 + ; CHECK-NEXT: [[ADDWrr:%[0-9]+]]:gpr32 = ADDWrr [[COPY]], [[COPY1]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[ADDWrr]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0 + %0:gpr32 = COPY $w0 + %1:gpr32 = COPY $w1 + %2:gpr32 = ADDWrr %0, %1 + %3:gpr32 = FREEZE %2 + $w0 = COPY %3 + RET_ReallyLR implicit $w0 +... +--- +name: freeze_physreg_copy +tracksRegLiveness: true +body: | + bb.0: + liveins: $w0 + + ; CHECK-LABEL: name: freeze_physreg_copy + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0 + %0:gpr32 = COPY $w0 + %1:gpr32 = FREEZE %0 + $w0 = COPY %1 + RET_ReallyLR implicit $w0 +... +--- +name: freeze_tied_undef_input +tracksRegLiveness: true +body: | + bb.0: + liveins: $w0 + + ; CHECK-LABEL: name: freeze_tied_undef_input + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF + ; CHECK-NEXT: [[INSvi32gpr:%[0-9]+]]:fpr128 = INSvi32gpr [[DEF]], 1, [[COPY]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:fpr128 = FREEZE [[INSvi32gpr]] + ; CHECK-NEXT: $q0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $q0 + %0:gpr32 = COPY $w0 + %1:fpr128 = IMPLICIT_DEF + %2:fpr128 = INSvi32gpr %1, 1, %0 + %3:fpr128 = FREEZE %2 + $q0 = COPY %3 + RET_ReallyLR implicit $q0 +... +--- +name: freeze_remat_undef_input +tracksRegLiveness: true +body: | + bb.0: + ; CHECK-LABEL: name: freeze_remat_undef_input + ; CHECK: [[DEF:%[0-9]+]]:gpr64sp = IMPLICIT_DEF + ; CHECK-NEXT: [[ADDXri:%[0-9]+]]:gpr64sp = ADDXri [[DEF]], 1, 0 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr64 = FREEZE [[ADDXri]] + ; CHECK-NEXT: $x0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $x0 + %0:gpr64sp = IMPLICIT_DEF + %1:gpr64sp = ADDXri %0, 1, 0 + %2:gpr64 = FREEZE %1 + $x0 = COPY %2 + RET_ReallyLR implicit $x0 +... +--- +name: freeze_remat_chain_undef_input +tracksRegLiveness: true +body: | + bb.0: + ; CHECK-LABEL: name: freeze_remat_chain_undef_input + ; CHECK: [[DEF:%[0-9]+]]:gpr64sp = IMPLICIT_DEF + ; CHECK-NEXT: [[ADDXri:%[0-9]+]]:gpr64sp = ADDXri [[DEF]], 1, 0 + ; CHECK-NEXT: [[ADDXri1:%[0-9]+]]:gpr64sp = ADDXri [[ADDXri]], 1, 0 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr64 = FREEZE [[ADDXri1]] + ; CHECK-NEXT: $x0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $x0 + %0:gpr64sp = IMPLICIT_DEF + %1:gpr64sp = ADDXri %0, 1, 0 + %2:gpr64sp = ADDXri %1, 1, 0 + %3:gpr64 = FREEZE %2 + $x0 = COPY %3 + RET_ReallyLR implicit $x0 +... +--- +name: freeze_undef_use +tracksRegLiveness: true +body: | + bb.0: + liveins: $w0 + + ; CHECK-LABEL: name: freeze_undef_use + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE undef [[COPY]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0 + %0:gpr32 = COPY $w0 + %1:gpr32 = FREEZE undef %0 + $w0 = COPY %1 + RET_ReallyLR implicit $w0 +... +--- +name: freeze_of_freeze +tracksRegLiveness: true +body: | + bb.0: + liveins: $w0 + + ; CHECK-LABEL: name: freeze_of_freeze + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]] + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY [[FREEZE]] + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[COPY1]] + ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF + ; CHECK-NEXT: [[FREEZE2:%[0-9]+]]:gpr32 = FREEZE [[DEF]] + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY [[FREEZE2]] + ; CHECK-NEXT: [[FREEZE3:%[0-9]+]]:gpr32 = FREEZE [[COPY2]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE1]] + ; CHECK-NEXT: $w1 = COPY [[FREEZE3]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 + %0:gpr32 = COPY $w0 + %1:gpr32 = FREEZE %0 + %2:gpr32 = COPY %1 + %3:gpr32 = FREEZE %2 + %4:gpr32 = IMPLICIT_DEF + %5:gpr32 = FREEZE %4 + %6:gpr32 = COPY %5 + %7:gpr32 = FREEZE %6 + $w0 = COPY %3 + $w1 = COPY %7 + RET_ReallyLR implicit $w0, implicit $w1 +... +--- +name: freeze_phi +tracksRegLiveness: true +body: | + ; CHECK-LABEL: name: freeze_phi + ; CHECK: bb.0: + ; CHECK-NEXT: successors: %bb.1(0x40000000), %bb.2(0x40000000) + ; CHECK-NEXT: liveins: $w0, $w1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY $w1 + ; CHECK-NEXT: CBZW [[COPY]], %bb.2 + ; CHECK-NEXT: B %bb.1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.1: + ; CHECK-NEXT: successors: %bb.2(0x80000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF + ; CHECK-NEXT: B %bb.2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.2: + ; CHECK-NEXT: [[PHI:%[0-9]+]]:gpr32 = PHI [[COPY1]], %bb.0, [[DEF]], %bb.1 + ; CHECK-NEXT: [[PHI1:%[0-9]+]]:gpr32 = PHI [[COPY1]], %bb.0, [[COPY]], %bb.1 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[PHI]] + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[PHI1]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: $w1 = COPY [[FREEZE1]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 + bb.0: + successors: %bb.1, %bb.2 + liveins: $w0, $w1 + + %0:gpr32 = COPY $w0 + %1:gpr32 = COPY $w1 + CBZW %0, %bb.2 + B %bb.1 + + bb.1: + %2:gpr32 = IMPLICIT_DEF + B %bb.2 + + bb.2: + %3:gpr32 = PHI %1, %bb.0, %2, %bb.1 + %4:gpr32 = PHI %1, %bb.0, %0, %bb.1 + %5:gpr32 = FREEZE %3 + %6:gpr32 = FREEZE %4 + $w0 = COPY %5 + $w1 = COPY %6 + RET_ReallyLR implicit $w0, implicit $w1 +... +--- +name: freeze_loop_phi +tracksRegLiveness: true +body: | + ; CHECK-LABEL: name: freeze_loop_phi + ; CHECK: bb.0: + ; CHECK-NEXT: successors: %bb.1(0x80000000) + ; CHECK-NEXT: liveins: $w0, $w1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY $w1 + ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.1: + ; CHECK-NEXT: successors: %bb.1(0x40000000), %bb.2(0x40000000) + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[PHI:%[0-9]+]]:gpr32 = PHI [[COPY]], %bb.0, %4, %bb.1 + ; CHECK-NEXT: [[PHI1:%[0-9]+]]:gpr32 = PHI [[DEF]], %bb.0, %6, %bb.1 + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY [[PHI]] + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:gpr32 = COPY [[PHI1]] + ; CHECK-NEXT: CBZW [[COPY1]], %bb.1 + ; CHECK-NEXT: B %bb.2 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: bb.2: + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[PHI]] + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[PHI1]] + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: $w1 = COPY [[FREEZE1]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 + bb.0: + successors: %bb.1 + liveins: $w0, $w1 + + %0:gpr32 = COPY $w0 + %1:gpr32 = COPY $w1 + %2:gpr32 = IMPLICIT_DEF + + bb.1: + successors: %bb.1, %bb.2 + + %3:gpr32 = PHI %0, %bb.0, %4, %bb.1 + %5:gpr32 = PHI %2, %bb.0, %6, %bb.1 + %4:gpr32 = COPY %3 + %6:gpr32 = COPY %5 + CBZW %1, %bb.1 + B %bb.2 + + bb.2: + %7:gpr32 = FREEZE %3 + %8:gpr32 = FREEZE %5 + $w0 = COPY %7 + $w1 = COPY %8 + RET_ReallyLR implicit $w0, implicit $w1 +... +--- +name: freeze_reg_sequence +tracksRegLiveness: true +body: | + bb.0: + liveins: $q0, $q1 + + ; CHECK-LABEL: name: freeze_reg_sequence + ; CHECK: liveins: $q0, $q1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:fpr128 = COPY $q0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr128 = COPY $q1 + ; CHECK-NEXT: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF + ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:qq = REG_SEQUENCE [[COPY]], %subreg.qsub0, [[COPY1]], %subreg.qsub1 + ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:qq = REG_SEQUENCE [[COPY]], %subreg.qsub0, [[DEF]], %subreg.qsub1 + ; CHECK-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:qq = REG_SEQUENCE [[COPY]], %subreg.qsub0 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE]] + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE1]] + ; CHECK-NEXT: [[FREEZE2:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE2]] + ; CHECK-NEXT: $q0_q1 = COPY [[FREEZE]] + ; CHECK-NEXT: $q2_q3 = COPY [[FREEZE1]] + ; CHECK-NEXT: $q4_q5 = COPY [[FREEZE2]] + ; CHECK-NEXT: RET_ReallyLR implicit $q0_q1, implicit $q2_q3, implicit $q4_q5 + %0:fpr128 = COPY $q0 + %1:fpr128 = COPY $q1 + %2:fpr128 = IMPLICIT_DEF + %3:qq = REG_SEQUENCE %0, %subreg.qsub0, %1, %subreg.qsub1 + %4:qq = REG_SEQUENCE %0, %subreg.qsub0, %2, %subreg.qsub1 + %5:qq = REG_SEQUENCE %0, %subreg.qsub0 + %6:qq = FREEZE %3 + %7:qq = FREEZE %4 + %8:qq = FREEZE %5 + $q0_q1 = COPY %6 + $q2_q3 = COPY %7 + $q4_q5 = COPY %8 + RET_ReallyLR implicit $q0_q1, implicit $q2_q3, implicit $q4_q5 +... +--- +name: freeze_insert_subreg +tracksRegLiveness: true +body: | + bb.0: + liveins: $q0, $d1 + + ; CHECK-LABEL: name: freeze_insert_subreg + ; CHECK: liveins: $q0, $d1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:fpr128 = COPY $q0 + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr64 = COPY $d1 + ; CHECK-NEXT: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF + ; CHECK-NEXT: [[INSERT_SUBREG:%[0-9]+]]:fpr128 = INSERT_SUBREG [[COPY]], [[COPY1]], %subreg.dsub + ; CHECK-NEXT: [[INSERT_SUBREG1:%[0-9]+]]:fpr128 = INSERT_SUBREG [[DEF]], [[COPY1]], %subreg.dsub + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:fpr128 = FREEZE [[INSERT_SUBREG]] + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:fpr128 = FREEZE [[INSERT_SUBREG1]] + ; CHECK-NEXT: $q0 = COPY [[FREEZE]] + ; CHECK-NEXT: $q1 = COPY [[FREEZE1]] + ; CHECK-NEXT: RET_ReallyLR implicit $q0, implicit $q1 + %0:fpr128 = COPY $q0 + %1:fpr64 = COPY $d1 + %2:fpr128 = IMPLICIT_DEF + %3:fpr128 = INSERT_SUBREG %0, %1, %subreg.dsub + %4:fpr128 = INSERT_SUBREG %2, %1, %subreg.dsub + %5:fpr128 = FREEZE %3 + %6:fpr128 = FREEZE %4 + $q0 = COPY %5 + $q1 = COPY %6 + RET_ReallyLR implicit $q0, implicit $q1 +... +--- +name: freeze_subreg_to_reg +tracksRegLiveness: true +body: | + bb.0: + liveins: $w0 + + ; CHECK-LABEL: name: freeze_subreg_to_reg + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 + ; CHECK-NEXT: [[ADDWrr:%[0-9]+]]:gpr32 = ADDWrr [[COPY]], [[COPY]] + ; CHECK-NEXT: [[SUBREG_TO_REG:%[0-9]+]]:gpr64 = SUBREG_TO_REG [[ADDWrr]], %subreg.sub_32 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr64 = FREEZE [[SUBREG_TO_REG]] + ; CHECK-NEXT: $x0 = COPY [[FREEZE]] + ; CHECK-NEXT: RET_ReallyLR implicit $x0 + %0:gpr32 = COPY $w0 + %1:gpr32 = ADDWrr %0, %0 + %2:gpr64 = SUBREG_TO_REG %1, %subreg.sub_32 + %3:gpr64 = FREEZE %2 + $x0 = COPY %3 + RET_ReallyLR implicit $x0 +... +--- +name: freeze_subreg_use +tracksRegLiveness: true +body: | + bb.0: + liveins: $x0 + + ; CHECK-LABEL: name: freeze_subreg_use + ; CHECK: liveins: $x0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64 = COPY $x0 + ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF + ; CHECK-NEXT: [[INSERT_SUBREG:%[0-9]+]]:gpr64 = INSERT_SUBREG [[COPY]], [[DEF]], %subreg.sub_32 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]].sub_32 + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[INSERT_SUBREG]].sub_32 + ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: $w1 = COPY [[FREEZE1]] + ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 + %0:gpr64 = COPY $x0 + %1:gpr32 = IMPLICIT_DEF + %2:gpr64 = INSERT_SUBREG %0, %1, %subreg.sub_32 + %3:gpr32 = FREEZE %0.sub_32 + %4:gpr32 = FREEZE %2.sub_32 + $w0 = COPY %3 + $w1 = COPY %4 + RET_ReallyLR implicit $w0, implicit $w1 +... >From 7f858beaeb4fe96d955e8a61ceb05347e4d794c1 Mon Sep 17 00:00:00 2001 From: Mitch <[email protected]> Date: Mon, 9 Nov 2026 01:53:25 -0700 Subject: [PATCH 2/2] [CodeGen] Lower FREEZE of a defined value to COPY A FREEZE is only needed when its operand may be undefined. FinalizeISel now turns every other FREEZE into a COPY, right after instruction selection. Assisted-By: Claude Opus 5.5 <[email protected]> --- llvm/lib/CodeGen/FinalizeISel.cpp | 116 ++++++++++++++++++ .../lib/CodeGen/SelectionDAG/InstrEmitter.cpp | 20 +++ .../CodeGen/AArch64/finalize-isel-freeze.mir | 72 +++++------ 3 files changed, 172 insertions(+), 36 deletions(-) diff --git a/llvm/lib/CodeGen/FinalizeISel.cpp b/llvm/lib/CodeGen/FinalizeISel.cpp index da2936bdd4b11..7e8faf40f0361 100644 --- a/llvm/lib/CodeGen/FinalizeISel.cpp +++ b/llvm/lib/CodeGen/FinalizeISel.cpp @@ -15,13 +15,16 @@ //===----------------------------------------------------------------------===// #include "llvm/CodeGen/FinalizeISel.h" +#include "llvm/ADT/BitVector.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/TargetInstrInfo.h" #include "llvm/CodeGen/TargetLowering.h" #include "llvm/CodeGen/TargetSubtargetInfo.h" #include "llvm/InitializePasses.h" +#include "llvm/Support/Debug.h" using namespace llvm; #define DEBUG_TYPE "finalize-isel" @@ -41,9 +44,115 @@ namespace { }; } // end anonymous namespace +/// Return true if \p MI copies its inputs into its result, or may be +/// rematerialized. Either way, an undefined input can make different uses of +/// the result observe different values. +static bool passesUndefThrough(const MachineInstr &MI, + const TargetInstrInfo &TII) { + switch (MI.getOpcode()) { + case TargetOpcode::PHI: + case TargetOpcode::COPY: + case TargetOpcode::INSERT_SUBREG: + case TargetOpcode::SUBREG_TO_REG: + case TargetOpcode::REG_SEQUENCE: + return true; + default: + return TII.isReMaterializable(MI); + } +} + +/// Return true if the register use \p MO may read an undefined value. Either it +/// is marked undef, or it reads a virtual register without a single definition. +/// A live-in virtual register without a definition holds the incoming physical +/// register value. +static bool mayReadUndef(const MachineOperand &MO, + const MachineRegisterInfo &MRI) { + if (MO.isUndef()) + return true; + Register Reg = MO.getReg(); + return Reg.isVirtual() && !MRI.hasOneDef(Reg) && !MRI.getLiveInPhysReg(Reg); +} + +/// Return true if \p MI may define an undefined register, i.e. an IMPLICIT_DEF, +/// a partial register definition, or an instruction that copies or may +/// rematerialize an undefined operand. +static bool definesUndef(const MachineInstr &MI, const MachineRegisterInfo &MRI, + const TargetInstrInfo &TII) { + if (MI.isImplicitDef()) + return true; + for (const MachineOperand &Def : MI.all_defs()) + if (Def.getSubReg()) + return true; + if (MI.isRegSequence()) { + LaneBitmask Covered; + for (unsigned I = 2, E = MI.getNumOperands(); I < E; I += 2) + Covered |= MRI.getTargetRegisterInfo()->getSubRegIndexLaneMask( + MI.getOperand(I).getImm()); + if ((MRI.getMaxLaneMaskForVReg(MI.getOperand(0).getReg()) & ~Covered).any()) + return true; + } + return passesUndefThrough(MI, TII) && + any_of(MI.all_uses(), [&](const MachineOperand &Use) { + return mayReadUndef(Use, MRI); + }); +} + +/// Add the virtual registers defined by \p MI to \p MayBeUndef. +static void markDefsUndef(const MachineInstr &MI, BitVector &MayBeUndef, + SmallVectorImpl<Register> &Worklist) { + for (const MachineOperand &Def : MI.all_defs()) { + Register Reg = Def.getReg(); + if (Reg.isVirtual() && !MayBeUndef.test(Reg.virtRegIndex())) { + MayBeUndef.set(Reg.virtRegIndex()); + Worklist.push_back(Reg); + } + } +} + +/// A FREEZE is only needed if its input may be undefined. Turn every other +/// FREEZE into a COPY right after instruction selection. +static bool lowerDefinedFreezes(MachineFunction &MF) { + const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo(); + const MachineRegisterInfo &MRI = MF.getRegInfo(); + SmallVector<MachineInstr *, 8> Freezes; + BitVector MayBeUndef(MRI.getNumVirtRegs()); + SmallVector<Register, 16> Worklist; + for (MachineBasicBlock &MBB : MF) { + for (MachineInstr &MI : MBB) { + if (MI.isDebugInstr()) + continue; + if (MI.getOpcode() == TargetOpcode::FREEZE) + Freezes.push_back(&MI); + else if (definesUndef(MI, MRI, TII)) + markDefsUndef(MI, MayBeUndef, Worklist); + } + } + + while (!Worklist.empty()) { + Register Reg = Worklist.pop_back_val(); + for (const MachineInstr &UseMI : MRI.use_nodbg_instructions(Reg)) + if (passesUndefThrough(UseMI, TII)) + markDefsUndef(UseMI, MayBeUndef, Worklist); + } + + bool Changed = false; + for (MachineInstr *MI : Freezes) { + const MachineOperand &Src = MI->getOperand(1); + if (mayReadUndef(Src, MRI) || + (Src.getReg().isVirtual() && + MayBeUndef.test(Src.getReg().virtRegIndex()))) + continue; + LLVM_DEBUG(dbgs() << "Converting to COPY: " << *MI); + MI->setDesc(TII.get(TargetOpcode::COPY)); + Changed = true; + } + return Changed; +} + static std::pair<bool, bool> runImpl(MachineFunction &MF) { bool Changed = false; bool PreserveCFG = true; + bool SawFreeze = false; const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); const TargetLowering *TLI = MF.getSubtarget().getTargetLowering(); @@ -58,6 +167,10 @@ static std::pair<bool, bool> runImpl(MachineFunction &MF) { MBBI != MBBE; ) { MachineInstr &MI = *MBBI++; + // A FREEZE emitted by a custom inserter before MBBI is missed, and stays. + if (MI.getOpcode() == TargetOpcode::FREEZE) + SawFreeze = true; + // Set AdjustsStack to true if the instruction selector emits a stack // frame setup instruction or a stack aligning inlineasm. if (TII->isFrameInstr(MI) || MI.isStackAligningInlineAsm()) @@ -78,6 +191,9 @@ static std::pair<bool, bool> runImpl(MachineFunction &MF) { } } } + + if (SawFreeze) + Changed |= lowerDefinedFreezes(MF); return {Changed, PreserveCFG}; } diff --git a/llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp b/llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp index b100b988555e3..08871f898068e 100644 --- a/llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp +++ b/llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp @@ -1103,6 +1103,26 @@ EmitMachineNode(SDNode *Node, bool IsClone, bool IsCloned, // hook knows where in the block to insert the replacement code. MBB->insert(InsertPos, MIB); + // A FREEZE doesn't change register class. Classes without a common subclass + // are split with a COPY: + // %d:sgpr_32 = FREEZE %s:vgpr_32 + // becomes + // %f:vgpr_32 = FREEZE %s:vgpr_32 + // %d:sgpr_32 = COPY %f:vgpr_32 + if (Opc == TargetOpcode::FREEZE) { + Register DstReg = MIB->getOperand(0).getReg(); + Register SrcReg = MIB->getOperand(1).getReg(); + const TargetRegisterClass *SrcRC = + SrcReg.isVirtual() ? MRI->getRegClass(SrcReg) : nullptr; + if (SrcRC && !TRI->getCommonSubClass(MRI->getRegClass(DstReg), SrcRC)) { + Register FrozenReg = MRI->createVirtualRegister(SrcRC); + MIB->getOperand(0).setReg(FrozenReg); + BuildMI(*MBB, InsertPos, MIB->getDebugLoc(), TII->get(TargetOpcode::COPY), + DstReg) + .addReg(FrozenReg); + } + } + // The MachineInstr may also define physregs instead of virtregs. These // physreg values can reach other instructions in different ways: // diff --git a/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir b/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir index d1ae7f7230934..695a3c2602c83 100644 --- a/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir +++ b/llvm/test/CodeGen/AArch64/finalize-isel-freeze.mir @@ -47,8 +47,8 @@ body: | ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY $w1 ; CHECK-NEXT: [[ADDWrr:%[0-9]+]]:gpr32 = ADDWrr [[COPY]], [[COPY1]] - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[ADDWrr]] - ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY [[ADDWrr]] + ; CHECK-NEXT: $w0 = COPY [[COPY2]] ; CHECK-NEXT: RET_ReallyLR implicit $w0 %0:gpr32 = COPY $w0 %1:gpr32 = COPY $w1 @@ -68,8 +68,8 @@ body: | ; CHECK: liveins: $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]] - ; CHECK-NEXT: $w0 = COPY [[FREEZE]] + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY [[COPY]] + ; CHECK-NEXT: $w0 = COPY [[COPY1]] ; CHECK-NEXT: RET_ReallyLR implicit $w0 %0:gpr32 = COPY $w0 %1:gpr32 = FREEZE %0 @@ -89,8 +89,8 @@ body: | ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 ; CHECK-NEXT: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF ; CHECK-NEXT: [[INSvi32gpr:%[0-9]+]]:fpr128 = INSvi32gpr [[DEF]], 1, [[COPY]] - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:fpr128 = FREEZE [[INSvi32gpr]] - ; CHECK-NEXT: $q0 = COPY [[FREEZE]] + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:fpr128 = COPY [[INSvi32gpr]] + ; CHECK-NEXT: $q0 = COPY [[COPY1]] ; CHECK-NEXT: RET_ReallyLR implicit $q0 %0:gpr32 = COPY $w0 %1:fpr128 = IMPLICIT_DEF @@ -165,15 +165,15 @@ body: | ; CHECK: liveins: $w0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]] - ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY [[FREEZE]] - ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[COPY1]] + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY [[COPY]] + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY [[COPY1]] + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:gpr32 = COPY [[COPY2]] ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF - ; CHECK-NEXT: [[FREEZE2:%[0-9]+]]:gpr32 = FREEZE [[DEF]] - ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY [[FREEZE2]] - ; CHECK-NEXT: [[FREEZE3:%[0-9]+]]:gpr32 = FREEZE [[COPY2]] - ; CHECK-NEXT: $w0 = COPY [[FREEZE1]] - ; CHECK-NEXT: $w1 = COPY [[FREEZE3]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[DEF]] + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:gpr32 = COPY [[FREEZE]] + ; CHECK-NEXT: [[COPY5:%[0-9]+]]:gpr32 = COPY [[COPY4]] + ; CHECK-NEXT: $w0 = COPY [[COPY3]] + ; CHECK-NEXT: $w1 = COPY [[COPY5]] ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 %0:gpr32 = COPY $w0 %1:gpr32 = FREEZE %0 @@ -211,9 +211,9 @@ body: | ; CHECK-NEXT: [[PHI:%[0-9]+]]:gpr32 = PHI [[COPY1]], %bb.0, [[DEF]], %bb.1 ; CHECK-NEXT: [[PHI1:%[0-9]+]]:gpr32 = PHI [[COPY1]], %bb.0, [[COPY]], %bb.1 ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[PHI]] - ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[PHI1]] + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr32 = COPY [[PHI1]] ; CHECK-NEXT: $w0 = COPY [[FREEZE]] - ; CHECK-NEXT: $w1 = COPY [[FREEZE1]] + ; CHECK-NEXT: $w1 = COPY [[COPY2]] ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 bb.0: successors: %bb.1, %bb.2 @@ -261,10 +261,10 @@ body: | ; CHECK-NEXT: B %bb.2 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: bb.2: - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[PHI]] - ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[PHI1]] - ; CHECK-NEXT: $w0 = COPY [[FREEZE]] - ; CHECK-NEXT: $w1 = COPY [[FREEZE1]] + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:gpr32 = COPY [[PHI]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[PHI1]] + ; CHECK-NEXT: $w0 = COPY [[COPY4]] + ; CHECK-NEXT: $w1 = COPY [[FREEZE]] ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 bb.0: successors: %bb.1 @@ -307,12 +307,12 @@ body: | ; CHECK-NEXT: [[REG_SEQUENCE:%[0-9]+]]:qq = REG_SEQUENCE [[COPY]], %subreg.qsub0, [[COPY1]], %subreg.qsub1 ; CHECK-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:qq = REG_SEQUENCE [[COPY]], %subreg.qsub0, [[DEF]], %subreg.qsub1 ; CHECK-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:qq = REG_SEQUENCE [[COPY]], %subreg.qsub0 - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE]] - ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE1]] - ; CHECK-NEXT: [[FREEZE2:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE2]] - ; CHECK-NEXT: $q0_q1 = COPY [[FREEZE]] - ; CHECK-NEXT: $q2_q3 = COPY [[FREEZE1]] - ; CHECK-NEXT: $q4_q5 = COPY [[FREEZE2]] + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:qq = COPY [[REG_SEQUENCE]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE1]] + ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:qq = FREEZE [[REG_SEQUENCE2]] + ; CHECK-NEXT: $q0_q1 = COPY [[COPY2]] + ; CHECK-NEXT: $q2_q3 = COPY [[FREEZE]] + ; CHECK-NEXT: $q4_q5 = COPY [[FREEZE1]] ; CHECK-NEXT: RET_ReallyLR implicit $q0_q1, implicit $q2_q3, implicit $q4_q5 %0:fpr128 = COPY $q0 %1:fpr128 = COPY $q1 @@ -343,10 +343,10 @@ body: | ; CHECK-NEXT: [[DEF:%[0-9]+]]:fpr128 = IMPLICIT_DEF ; CHECK-NEXT: [[INSERT_SUBREG:%[0-9]+]]:fpr128 = INSERT_SUBREG [[COPY]], [[COPY1]], %subreg.dsub ; CHECK-NEXT: [[INSERT_SUBREG1:%[0-9]+]]:fpr128 = INSERT_SUBREG [[DEF]], [[COPY1]], %subreg.dsub - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:fpr128 = FREEZE [[INSERT_SUBREG]] - ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:fpr128 = FREEZE [[INSERT_SUBREG1]] - ; CHECK-NEXT: $q0 = COPY [[FREEZE]] - ; CHECK-NEXT: $q1 = COPY [[FREEZE1]] + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:fpr128 = COPY [[INSERT_SUBREG]] + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:fpr128 = FREEZE [[INSERT_SUBREG1]] + ; CHECK-NEXT: $q0 = COPY [[COPY2]] + ; CHECK-NEXT: $q1 = COPY [[FREEZE]] ; CHECK-NEXT: RET_ReallyLR implicit $q0, implicit $q1 %0:fpr128 = COPY $q0 %1:fpr64 = COPY $d1 @@ -372,8 +372,8 @@ body: | ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr32 = COPY $w0 ; CHECK-NEXT: [[ADDWrr:%[0-9]+]]:gpr32 = ADDWrr [[COPY]], [[COPY]] ; CHECK-NEXT: [[SUBREG_TO_REG:%[0-9]+]]:gpr64 = SUBREG_TO_REG [[ADDWrr]], %subreg.sub_32 - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr64 = FREEZE [[SUBREG_TO_REG]] - ; CHECK-NEXT: $x0 = COPY [[FREEZE]] + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY [[SUBREG_TO_REG]] + ; CHECK-NEXT: $x0 = COPY [[COPY1]] ; CHECK-NEXT: RET_ReallyLR implicit $x0 %0:gpr32 = COPY $w0 %1:gpr32 = ADDWrr %0, %0 @@ -395,10 +395,10 @@ body: | ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64 = COPY $x0 ; CHECK-NEXT: [[DEF:%[0-9]+]]:gpr32 = IMPLICIT_DEF ; CHECK-NEXT: [[INSERT_SUBREG:%[0-9]+]]:gpr64 = INSERT_SUBREG [[COPY]], [[DEF]], %subreg.sub_32 - ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[COPY]].sub_32 - ; CHECK-NEXT: [[FREEZE1:%[0-9]+]]:gpr32 = FREEZE [[INSERT_SUBREG]].sub_32 - ; CHECK-NEXT: $w0 = COPY [[FREEZE]] - ; CHECK-NEXT: $w1 = COPY [[FREEZE1]] + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr32 = COPY [[COPY]].sub_32 + ; CHECK-NEXT: [[FREEZE:%[0-9]+]]:gpr32 = FREEZE [[INSERT_SUBREG]].sub_32 + ; CHECK-NEXT: $w0 = COPY [[COPY1]] + ; CHECK-NEXT: $w1 = COPY [[FREEZE]] ; CHECK-NEXT: RET_ReallyLR implicit $w0, implicit $w1 %0:gpr64 = COPY $x0 %1:gpr32 = IMPLICIT_DEF _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
