[llvm] RegisterCoalescer: Fix assert on remat to copy-to-physreg with subregs (PR #121734)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Sun Jan 5 23:51:33 PST 2025
https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/121734
Do not try to rematerialize a super-register def used by a subregister
extract copy into a copy to a physical register if the other pieces of the
full physreg are live at the rematerialization point. It would insert the
super-register def at the rematerialization point, and assert since the
other half of the register was already live.
This is analagous to the undef subregister def handling above,
which handled the virtual register case.
Fixes #120970
>From fc0b6a774ac3e5144ee1f20e268b2a779fba37f2 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Fri, 3 Jan 2025 12:34:41 +0700
Subject: [PATCH] RegisterCoalescer: Fix assert on remat to copy-to-physreg
with subregs
Do not try to rematerialize a super-register def used by a subregister
extract copy into a copy to a physical register if the other pieces of the
full physreg are live at the rematerialization point. It would insert the
super-register def at the rematerialization point, and assert since the
other half of the register was already live.
This is analagous to the undef subregister def handling above,
which handled the virtual register case.
Fixes #120970
---
llvm/lib/CodeGen/RegisterCoalescer.cpp | 31 +++++--
...xtract-already-live-at-def-issue120970.mir | 85 +++++++++++++++++++
2 files changed, 111 insertions(+), 5 deletions(-)
create mode 100644 llvm/test/CodeGen/AMDGPU/remat-physreg-copy-subreg-extract-already-live-at-def-issue120970.mir
diff --git a/llvm/lib/CodeGen/RegisterCoalescer.cpp b/llvm/lib/CodeGen/RegisterCoalescer.cpp
index 20ad6445344d83..c9e6ed03ed42f3 100644
--- a/llvm/lib/CodeGen/RegisterCoalescer.cpp
+++ b/llvm/lib/CodeGen/RegisterCoalescer.cpp
@@ -1325,11 +1325,6 @@ bool RegisterCoalescer::reMaterializeTrivialDef(const CoalescerPair &CP,
const MCInstrDesc &MCID = DefMI->getDesc();
if (MCID.getNumDefs() != 1)
return false;
- // Only support subregister destinations when the def is read-undef.
- MachineOperand &DstOperand = CopyMI->getOperand(0);
- Register CopyDstReg = DstOperand.getReg();
- if (DstOperand.getSubReg() && !DstOperand.isUndef())
- return false;
// If both SrcIdx and DstIdx are set, correct rematerialization would widen
// the register substantially (beyond both source and dest size). This is bad
@@ -1339,6 +1334,32 @@ bool RegisterCoalescer::reMaterializeTrivialDef(const CoalescerPair &CP,
if (SrcIdx && DstIdx)
return false;
+ // Only support subregister destinations when the def is read-undef.
+ MachineOperand &DstOperand = CopyMI->getOperand(0);
+ Register CopyDstReg = DstOperand.getReg();
+ if (DstOperand.getSubReg() && !DstOperand.isUndef())
+ return false;
+
+ // Only support subregister destinations when the def is read-undef, in the
+ // physical register case. We're widening the def and need to avoid clobbering
+ // other live values in the unused register pieces.
+ //
+ // TODO: Targets may support rewriting the rematerialized instruction to only
+ // touch relevant lanes, in which case we don't need any liveness check.
+ if (CopyDstReg.isPhysical() && CP.isPartial()) {
+ for (MCRegUnit Unit : TRI->regunits(DstReg)) {
+ // Ignore the register units we are writing anyway.
+ if (is_contained(TRI->regunits(CopyDstReg), Unit))
+ continue;
+
+ // Check if the other lanes we are defining are live at the
+ // rematerialization point.
+ LiveRange &LR = LIS->getRegUnit(Unit);
+ if (LR.liveAt(CopyIdx))
+ return false;
+ }
+ }
+
const unsigned DefSubIdx = DefMI->getOperand(0).getSubReg();
const TargetRegisterClass *DefRC = TII->getRegClass(MCID, 0, TRI, *MF);
if (!DefMI->isImplicitDef()) {
diff --git a/llvm/test/CodeGen/AMDGPU/remat-physreg-copy-subreg-extract-already-live-at-def-issue120970.mir b/llvm/test/CodeGen/AMDGPU/remat-physreg-copy-subreg-extract-already-live-at-def-issue120970.mir
new file mode 100644
index 00000000000000..3879f6dccf9d69
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/remat-physreg-copy-subreg-extract-already-live-at-def-issue120970.mir
@@ -0,0 +1,85 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -run-pass=register-coalescer -o - %s | FileCheck %s
+
+# This used to assert due to trying to rematerialize V_MOV_B64_PSEUDO
+# at copy to $vgpr1. This would assert since this would clobber the
+# live value in $vgpr0.
+
+---
+name: rematerialize_physreg_sub_def_already_live_at_def_assert
+tracksRegLiveness: true
+body: |
+ bb.0:
+ liveins: $vgpr0
+
+ ; CHECK-LABEL: name: rematerialize_physreg_sub_def_already_live_at_def_assert
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[V_MOV_B:%[0-9]+]]:vreg_64 = V_MOV_B64_PSEUDO 1, implicit $exec
+ ; CHECK-NEXT: $vgpr0 = V_MOV_B32_e32 0, implicit $exec
+ ; CHECK-NEXT: $vgpr1 = COPY [[V_MOV_B]].sub1
+ ; CHECK-NEXT: SI_RETURN implicit $vgpr0, implicit killed $vgpr1
+ %0:vreg_64 = V_MOV_B64_PSEUDO 1, implicit $exec
+ %1:vgpr_32 = COPY %0.sub1
+ $vgpr0 = V_MOV_B32_e32 0, implicit $exec
+ $vgpr1 = COPY %1
+ SI_RETURN implicit $vgpr0, implicit killed $vgpr1
+...
+
+# Same as previous, except with an IMPLICIT_DEF
+---
+name: rematerialize_physreg_sub_def_already_live_at_def_assert_implicit_def
+tracksRegLiveness: true
+body: |
+ bb.0:
+ liveins: $vgpr0
+
+ ; CHECK-LABEL: name: rematerialize_physreg_sub_def_already_live_at_def_assert_implicit_def
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[DEF:%[0-9]+]]:vreg_64 = IMPLICIT_DEF
+ ; CHECK-NEXT: $vgpr0 = V_MOV_B32_e32 0, implicit $exec
+ ; CHECK-NEXT: $vgpr1 = COPY [[DEF]].sub1
+ ; CHECK-NEXT: SI_RETURN implicit $vgpr0, implicit killed $vgpr1
+ %0:vreg_64 = IMPLICIT_DEF
+ %1:vgpr_32 = COPY %0.sub1
+ $vgpr0 = V_MOV_B32_e32 0, implicit $exec
+ $vgpr1 = COPY %1
+ SI_RETURN implicit $vgpr0, implicit killed $vgpr1
+...
+
+---
+name: rematerialize_physreg_sub_def_no_live_sub_def_0
+tracksRegLiveness: true
+body: |
+ bb.0:
+ liveins: $vgpr0
+
+ ; CHECK-LABEL: name: rematerialize_physreg_sub_def_no_live_sub_def_0
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: dead $vgpr0_vgpr1 = V_MOV_B64_PSEUDO 1, implicit $exec, implicit-def $vgpr1
+ ; CHECK-NEXT: SI_RETURN implicit killed $vgpr1
+ %0:vreg_64 = V_MOV_B64_PSEUDO 1, implicit $exec
+ %1:vgpr_32 = COPY %0.sub1
+ $vgpr1 = COPY %1
+ SI_RETURN implicit killed $vgpr1
+...
+
+---
+name: rematerialize_physreg_sub_def_no_live_sub_def_1
+tracksRegLiveness: true
+body: |
+ bb.0:
+ liveins: $vgpr0
+
+ ; CHECK-LABEL: name: rematerialize_physreg_sub_def_no_live_sub_def_1
+ ; CHECK: liveins: $vgpr0
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: dead $vgpr1_vgpr2 = V_MOV_B64_PSEUDO 1, implicit $exec, implicit-def $vgpr1
+ ; CHECK-NEXT: SI_RETURN implicit killed $vgpr1
+ %0:vreg_64 = V_MOV_B64_PSEUDO 1, implicit $exec
+ %1:vgpr_32 = COPY %0.sub0
+ $vgpr1 = COPY %1
+ SI_RETURN implicit killed $vgpr1
+...
More information about the llvm-commits
mailing list