aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/CodeGen
diff options
context:
space:
mode:
authorChristudasan Devadasan <Christudasan.Devadasan@amd.com>2026-01-08 07:53:18 +0000
committerChristudasan Devadasan <Christudasan.Devadasan@amd.com>2026-01-13 03:50:24 +0000
commit070b3e99ac751c9e6cb38201e0eec68d72a55542 (patch)
treeb798e5d2717b69aa89cccd322f5518d37f9e47c0 /llvm/lib/CodeGen
parent12d1aa0c8430c9d8015bfb285aae7d5e260db8ad (diff)
downloadllvm-users/cdevadas/subreg-reload.zip
llvm-users/cdevadas/subreg-reload.tar.gz
llvm-users/cdevadas/subreg-reload.tar.bz2
[InlineSpiller][AMDGPU] Implement subreg reload during RA spillusers/cdevadas/subreg-reload
Currently, when a virtual register is partially used, the entire tuple is restored from the spilled location, even if only a subset of its sub-registers is needed. This patch introduces support for partial reloads by analyzing actual register usage and restoring only the required sub-registers. This improvement enhances register allocation efficiency, particularly for cases involving tuple virtual registers. For AMDGPU, this change brings considerable improvements in workloads that involve matrix operations, large vectors, and complex control flows.
Diffstat (limited to 'llvm/lib/CodeGen')
-rw-r--r--llvm/lib/CodeGen/InlineSpiller.cpp57
1 files changed, 51 insertions, 6 deletions
diff --git a/llvm/lib/CodeGen/InlineSpiller.cpp b/llvm/lib/CodeGen/InlineSpiller.cpp
index 6837030..c567b88f 100644
--- a/llvm/lib/CodeGen/InlineSpiller.cpp
+++ b/llvm/lib/CodeGen/InlineSpiller.cpp
@@ -217,7 +217,8 @@ private:
bool coalesceStackAccess(MachineInstr *MI, Register Reg);
bool foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>>,
MachineInstr *LoadMI = nullptr);
- void insertReload(Register VReg, SlotIndex, MachineBasicBlock::iterator MI);
+ void insertReload(Register VReg, unsigned SubReg, SlotIndex,
+ MachineBasicBlock::iterator MI);
void insertSpill(Register VReg, bool isKill, MachineBasicBlock::iterator MI);
void spillAroundUses(Register Reg);
@@ -1112,14 +1113,14 @@ foldMemoryOperand(ArrayRef<std::pair<MachineInstr *, unsigned>> Ops,
return true;
}
-void InlineSpiller::insertReload(Register NewVReg,
+void InlineSpiller::insertReload(Register NewVReg, unsigned SubReg,
SlotIndex Idx,
MachineBasicBlock::iterator MI) {
MachineBasicBlock &MBB = *MI->getParent();
MachineInstrSpan MIS(MI, &MBB);
TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot,
- MRI.getRegClass(NewVReg), Register());
+ MRI.getRegClass(NewVReg), Register(), SubReg);
LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI);
@@ -1248,10 +1249,51 @@ void InlineSpiller::spillAroundUses(Register Reg) {
// Create a new virtual register for spill/fill.
// FIXME: Infer regclass from instruction alone.
- Register NewVReg = Edit->createFrom(Reg);
+
+ unsigned SubReg = 0;
+ LaneBitmask CoveringLanes = LaneBitmask::getNone();
+ // If the subreg liveness is enabled, identify the subreg use(s) to try
+ // subreg reload. Skip if the instruction also defines the register.
+ // For copy bundles, get the covering lane masks.
+ if (MRI.subRegLivenessEnabled() && !RI.Writes) {
+ for (auto [MI, OpIdx] : Ops) {
+ const MachineOperand &MO = MI->getOperand(OpIdx);
+ assert(MO.isReg() && MO.getReg() == Reg);
+ if (MO.isUse()) {
+ SubReg = MO.getSubReg();
+ if (SubReg)
+ CoveringLanes |= TRI.getSubRegIndexLaneMask(SubReg);
+ }
+ }
+ }
+
+ if (MI.isBundled() && CoveringLanes.any()) {
+ CoveringLanes = LaneBitmask(bit_ceil(CoveringLanes.getAsInteger()) - 1);
+ // Obtain the covering subregister index, including any missing indices
+ // within the identified small range. Although this may be suboptimal due
+ // to gaps in the subregisters that are not part of the copy bundle, it is
+ // benificial when components outside this range of the original tuple can
+ // be completely skipped from the reload.
+ SubReg = TRI.getSubRegIdxFromLaneMask(CoveringLanes);
+ }
+
+ // If the target doesn't support subreg reload, fallback to restoring the
+ // full tuple.
+ if (SubReg && !TRI.shouldEnableSubRegReload(SubReg))
+ SubReg = 0;
+
+ const TargetRegisterClass *OrigRC = MRI.getRegClass(Reg);
+ const TargetRegisterClass *NewRC =
+ SubReg ? TRI.getSubRegisterClass(OrigRC, SubReg) : nullptr;
+
+ // Check if the target needs to constrain the RC further.
+ if (NewRC)
+ NewRC = TRI.getConstrainedRegClass(NewRC);
+
+ Register NewVReg = Edit->createFrom(Reg, NewRC);
if (RI.Reads)
- insertReload(NewVReg, Idx, &MI);
+ insertReload(NewVReg, SubReg, Idx, &MI);
// Rewrite instruction operands.
bool hasLiveDef = false;
@@ -1259,7 +1301,10 @@ void InlineSpiller::spillAroundUses(Register Reg) {
MachineOperand &MO = OpPair.first->getOperand(OpPair.second);
MO.setReg(NewVReg);
if (MO.isUse()) {
- if (!OpPair.first->isRegTiedToDefOperand(OpPair.second))
+ if (SubReg && !MI.isBundled())
+ MO.setSubReg(0);
+ if (!OpPair.first->isRegTiedToDefOperand(OpPair.second) ||
+ (SubReg && !MI.isBundled()))
MO.setIsKill();
} else {
if (!MO.isDead())