[llvm] [AMDGPU] Fix nondeterminism in SIFixSGPRCopies (PR #70644)
Jay Foad via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 30 08:28:12 PDT 2023
https://github.com/jayfoad updated https://github.com/llvm/llvm-project/pull/70644
>From bfc7b2041f5a05105808b0b1ee0427d9c9eb9f4b Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Mon, 30 Oct 2023 15:23:48 +0000
Subject: [PATCH 1/2] Precommit test
---
.../AMDGPU/fix-sgpr-copies-nondeterminism.ll | 52 +++++++++++++++++++
1 file changed, 52 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
diff --git a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
new file mode 100644
index 000000000000000..8b7e691dbddeae5
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
@@ -0,0 +1,52 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck %s
+
+define amdgpu_gs void @f(i32 inreg %arg, i32 %arg1, i32 %arg2) {
+; CHECK-LABEL: f:
+; CHECK: ; %bb.0: ; %bb
+; CHECK-NEXT: s_cmp_eq_u32 s0, 0
+; CHECK-NEXT: s_mov_b32 s0, 0
+; CHECK-NEXT: s_cbranch_scc1 .LBB0_2
+; CHECK-NEXT: ; %bb.1: ; %bb3
+; CHECK-NEXT: v_mov_b32_e32 v4, v1
+; CHECK-NEXT: s_branch .LBB0_3
+; CHECK-NEXT: .LBB0_2:
+; CHECK-NEXT: v_mov_b32_e32 v0, 1
+; CHECK-NEXT: v_mov_b32_e32 v4, 0
+; CHECK-NEXT: .LBB0_3: ; %bb4
+; CHECK-NEXT: v_mov_b32_e32 v1, 0
+; CHECK-NEXT: s_mov_b32 s1, s0
+; CHECK-NEXT: s_mov_b32 s2, s0
+; CHECK-NEXT: s_mov_b32 s3, s0
+; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; CHECK-NEXT: v_mov_b32_e32 v2, v1
+; CHECK-NEXT: v_mov_b32_e32 v3, v1
+; CHECK-NEXT: v_mov_b32_e32 v5, v1
+; CHECK-NEXT: v_mov_b32_e32 v6, v1
+; CHECK-NEXT: v_mov_b32_e32 v7, v1
+; CHECK-NEXT: s_clause 0x1
+; CHECK-NEXT: buffer_store_b128 v[0:3], v1, s[0:3], 0 idxen
+; CHECK-NEXT: buffer_store_b128 v[4:7], v1, s[0:3], 0 idxen
+; CHECK-NEXT: s_nop 0
+; CHECK-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; CHECK-NEXT: s_endpgm
+bb:
+ %i = icmp eq i32 %arg, 0
+ br i1 %i, label %bb4, label %bb3
+
+bb3:
+ br label %bb4
+
+bb4:
+ %i5 = phi i32 [ %arg1, %bb3 ], [ 1, %bb ]
+ %i6 = phi i32 [ %arg2, %bb3 ], [ 0, %bb ]
+ %i7 = insertelement <4 x i32> zeroinitializer, i32 %i5, i64 0
+ %i8 = bitcast <4 x i32> %i7 to <4 x float>
+ call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %i8, <4 x i32> zeroinitializer, i32 0, i32 0, i32 0, i32 0)
+ %i9 = insertelement <4 x i32> zeroinitializer, i32 %i6, i64 0
+ %i10 = bitcast <4 x i32> %i9 to <4 x float>
+ call void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float> %i10, <4 x i32> zeroinitializer, i32 0, i32 0, i32 0, i32 0)
+ ret void
+}
+
+declare void @llvm.amdgcn.struct.buffer.store.v4f32(<4 x float>, <4 x i32>, i32, i32, i32, i32 immarg)
>From aa050e8d720150b97d7af18d97d1d7f5d010bedc Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Mon, 30 Oct 2023 10:40:22 +0000
Subject: [PATCH 2/2] [AMDGPU] Fix nondeterminism in SIFixSGPRCopies
There are a couple of loops that iterate over V2SCopies. The iteration
order needs to be deterministic, otherwise we can call moveToVALU in
different orders, which causes temporary vregs to be allocated in
different orders, which can affect register allocation heuristics.
---
llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp | 8 +++----
.../AMDGPU/fix-sgpr-copies-nondeterminism.ll | 22 +++++++++----------
2 files changed, 15 insertions(+), 15 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
index b32ed9fef5dd34e..3e6ed2d793ae563 100644
--- a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
@@ -125,7 +125,7 @@ class SIFixSGPRCopies : public MachineFunctionPass {
SmallVector<MachineInstr*, 4> PHINodes;
SmallVector<MachineInstr*, 4> S2VCopies;
unsigned NextVGPRToSGPRCopyID;
- DenseMap<unsigned, V2SCopyInfo> V2SCopies;
+ MapVector<unsigned, V2SCopyInfo> V2SCopies;
DenseMap<MachineInstr *, SetVector<unsigned>> SiblingPenalty;
public:
@@ -988,7 +988,7 @@ bool SIFixSGPRCopies::needToBeConvertedToVALU(V2SCopyInfo *Info) {
for (auto J : Info->Siblings) {
auto InfoIt = V2SCopies.find(J);
if (InfoIt != V2SCopies.end()) {
- MachineInstr *SiblingCopy = InfoIt->getSecond().Copy;
+ MachineInstr *SiblingCopy = InfoIt->second.Copy;
if (SiblingCopy->isImplicitDef())
// the COPY has already been MoveToVALUed
continue;
@@ -1023,12 +1023,12 @@ void SIFixSGPRCopies::lowerVGPR2SGPRCopies(MachineFunction &MF) {
unsigned CurID = LoweringWorklist.pop_back_val();
auto CurInfoIt = V2SCopies.find(CurID);
if (CurInfoIt != V2SCopies.end()) {
- V2SCopyInfo C = CurInfoIt->getSecond();
+ V2SCopyInfo C = CurInfoIt->second;
LLVM_DEBUG(dbgs() << "Processing ...\n"; C.dump());
for (auto S : C.Siblings) {
auto SibInfoIt = V2SCopies.find(S);
if (SibInfoIt != V2SCopies.end()) {
- V2SCopyInfo &SI = SibInfoIt->getSecond();
+ V2SCopyInfo &SI = SibInfoIt->second;
LLVM_DEBUG(dbgs() << "Sibling:\n"; SI.dump());
if (!SI.NeedToBeConvertedToVALU) {
SI.SChain.set_subtract(C.SChain);
diff --git a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
index 8b7e691dbddeae5..a78a8a2ed8fe664 100644
--- a/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
+++ b/llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
@@ -8,25 +8,25 @@ define amdgpu_gs void @f(i32 inreg %arg, i32 %arg1, i32 %arg2) {
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_cbranch_scc1 .LBB0_2
; CHECK-NEXT: ; %bb.1: ; %bb3
-; CHECK-NEXT: v_mov_b32_e32 v4, v1
+; CHECK-NEXT: v_mov_b32_e32 v5, v0
; CHECK-NEXT: s_branch .LBB0_3
; CHECK-NEXT: .LBB0_2:
-; CHECK-NEXT: v_mov_b32_e32 v0, 1
-; CHECK-NEXT: v_mov_b32_e32 v4, 0
-; CHECK-NEXT: .LBB0_3: ; %bb4
+; CHECK-NEXT: v_mov_b32_e32 v5, 1
; CHECK-NEXT: v_mov_b32_e32 v1, 0
+; CHECK-NEXT: .LBB0_3: ; %bb4
+; CHECK-NEXT: v_mov_b32_e32 v6, 0
; CHECK-NEXT: s_mov_b32 s1, s0
; CHECK-NEXT: s_mov_b32 s2, s0
; CHECK-NEXT: s_mov_b32 s3, s0
; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; CHECK-NEXT: v_mov_b32_e32 v2, v1
-; CHECK-NEXT: v_mov_b32_e32 v3, v1
-; CHECK-NEXT: v_mov_b32_e32 v5, v1
-; CHECK-NEXT: v_mov_b32_e32 v6, v1
-; CHECK-NEXT: v_mov_b32_e32 v7, v1
+; CHECK-NEXT: v_mov_b32_e32 v7, v6
+; CHECK-NEXT: v_mov_b32_e32 v8, v6
+; CHECK-NEXT: v_mov_b32_e32 v2, v6
+; CHECK-NEXT: v_mov_b32_e32 v3, v6
+; CHECK-NEXT: v_mov_b32_e32 v4, v6
; CHECK-NEXT: s_clause 0x1
-; CHECK-NEXT: buffer_store_b128 v[0:3], v1, s[0:3], 0 idxen
-; CHECK-NEXT: buffer_store_b128 v[4:7], v1, s[0:3], 0 idxen
+; CHECK-NEXT: buffer_store_b128 v[5:8], v6, s[0:3], 0 idxen
+; CHECK-NEXT: buffer_store_b128 v[1:4], v6, s[0:3], 0 idxen
; CHECK-NEXT: s_nop 0
; CHECK-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; CHECK-NEXT: s_endpgm
More information about the llvm-commits
mailing list