[llvm] a6dabed - [AMDGPU] Fix nondeterminism in SIFixSGPRCopies (#70644)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 31 04:47:46 PDT 2023
Author: Jay Foad
Date: 2023-10-31T11:47:42Z
New Revision: a6dabed3483c60469ff53d51622b22efc4b7b7d2
URL: https://github.com/llvm/llvm-project/commit/a6dabed3483c60469ff53d51622b22efc4b7b7d2
DIFF: https://github.com/llvm/llvm-project/commit/a6dabed3483c60469ff53d51622b22efc4b7b7d2.diff
LOG: [AMDGPU] Fix nondeterminism in SIFixSGPRCopies (#70644)
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.
Added:
llvm/test/CodeGen/AMDGPU/fix-sgpr-copies-nondeterminism.ll
Modified:
llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
llvm/test/CodeGen/AMDGPU/branch-folding-implicit-def-subreg.ll
llvm/test/CodeGen/AMDGPU/carryout-selection.ll
llvm/test/CodeGen/AMDGPU/sdwa-peephole.ll
llvm/test/CodeGen/AMDGPU/vgpr-liverange-ir.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp b/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
index b7ac90e33f65e00..86980ee851bb1fc 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);
@@ -1040,6 +1040,8 @@ void SIFixSGPRCopies::lowerVGPR2SGPRCopies(MachineFunction &MF) {
}
LLVM_DEBUG(dbgs() << "V2S copy " << *C.Copy
<< " is being turned to VALU\n");
+ // TODO: MapVector::erase is inefficient. Do bulk removal with remove_if
+ // instead.
V2SCopies.erase(C.ID);
Copies.insert(C.Copy);
}
diff --git a/llvm/test/CodeGen/AMDGPU/branch-folding-implicit-def-subreg.ll b/llvm/test/CodeGen/AMDGPU/branch-folding-implicit-def-subreg.ll
index b5a823355543c9c..96dbb03cded5588 100644
--- a/llvm/test/CodeGen/AMDGPU/branch-folding-implicit-def-subreg.ll
+++ b/llvm/test/CodeGen/AMDGPU/branch-folding-implicit-def-subreg.ll
@@ -948,12 +948,12 @@ define amdgpu_kernel void @f1(ptr addrspace(1) %arg, ptr addrspace(1) %arg1, i64
; GFX90A-NEXT: liveins: $sgpr14, $sgpr15, $sgpr16, $vgpr15, $vgpr17, $vgpr30, $vgpr31, $vgpr52, $vgpr53, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9:0x000000000000000F, $sgpr10_sgpr11, $sgpr12_sgpr13, $sgpr18_sgpr19, $sgpr24_sgpr25, $sgpr28_sgpr29, $sgpr34_sgpr35, $sgpr38_sgpr39, $sgpr40_sgpr41, $sgpr42_sgpr43, $sgpr44_sgpr45, $sgpr46_sgpr47, $sgpr48_sgpr49, $sgpr50_sgpr51, $sgpr52_sgpr53, $sgpr54_sgpr55, $sgpr58_sgpr59, $sgpr20_sgpr21_sgpr22_sgpr23:0x000000000000003C, $sgpr24_sgpr25_sgpr26_sgpr27:0x00000000000000F0, $vgpr0_vgpr1:0x000000000000000F, $vgpr2_vgpr3:0x000000000000000F, $vgpr4_vgpr5:0x000000000000000F, $vgpr6_vgpr7:0x000000000000000F, $vgpr8_vgpr9:0x000000000000000F, $vgpr10_vgpr11:0x000000000000000F, $vgpr12_vgpr13:0x000000000000000F, $vgpr14_vgpr15:0x0000000000000003, $vgpr16_vgpr17:0x0000000000000003, $vgpr18_vgpr19:0x000000000000000F, $vgpr20_vgpr21:0x000000000000000F, $vgpr22_vgpr23:0x000000000000000F, $vgpr24_vgpr25:0x000000000000000F, $vgpr40_vgpr41:0x000000000000000F, $vgpr42_vgpr43:0x000000000000000F, $vgpr44_vgpr45:0x000000000000000F, $vgpr46_vgpr47:0x000000000000000F, $vgpr56_vgpr57:0x000000000000000F, $vgpr58_vgpr59:0x000000000000000F, $vgpr60_vgpr61:0x000000000000000F, $vgpr62_vgpr63:0x000000000000000F, $sgpr0_sgpr1_sgpr2_sgpr3
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: renamable $vgpr26 = V_OR_B32_e32 1, $vgpr24, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr38 = V_OR_B32_e32 $vgpr26, $vgpr22, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr34 = V_OR_B32_e32 $vgpr38, $vgpr20, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr48 = V_OR_B32_e32 $vgpr26, $vgpr22, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr34 = V_OR_B32_e32 $vgpr48, $vgpr20, implicit $exec
; GFX90A-NEXT: renamable $vgpr28 = V_CNDMASK_B32_e64 0, $vgpr34, 0, 0, $sgpr12_sgpr13, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr36 = V_OR_B32_e32 $vgpr28, $vgpr18, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr48 = V_OR_B32_e32 $vgpr36, $vgpr10, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr32 = V_OR_B32_e32 $vgpr48, $vgpr12, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr38 = V_OR_B32_e32 $vgpr28, $vgpr18, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr36 = V_OR_B32_e32 $vgpr38, $vgpr10, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr32 = V_OR_B32_e32 $vgpr36, $vgpr12, implicit $exec
; GFX90A-NEXT: renamable $vgpr50 = V_CNDMASK_B32_e64 0, 0, 0, $vgpr32, killed $sgpr12_sgpr13, implicit $exec
; GFX90A-NEXT: renamable $sgpr12_sgpr13 = S_MOV_B64 -1
; GFX90A-NEXT: renamable $vcc = S_AND_B64 $exec, killed renamable $sgpr28_sgpr29, implicit-def dead $scc
@@ -975,20 +975,20 @@ define amdgpu_kernel void @f1(ptr addrspace(1) %arg, ptr addrspace(1) %arg1, i64
; GFX90A-NEXT: renamable $vgpr2, renamable $vcc = V_ADD_CO_U32_e64 killed $sgpr26, $vgpr2, 0, implicit $exec
; GFX90A-NEXT: renamable $vgpr3, dead renamable $vcc = V_ADDC_U32_e64 killed $vgpr10, killed $vgpr3, killed $vcc, 0, implicit $exec
; GFX90A-NEXT: renamable $vgpr27 = V_MOV_B32_e32 0, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr39 = COPY renamable $vgpr27, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr49 = COPY renamable $vgpr27, implicit $exec
; GFX90A-NEXT: renamable $vgpr35 = COPY renamable $vgpr27, implicit $exec
+ ; GFX90A-NEXT: renamable $vgpr39 = COPY renamable $vgpr27, implicit $exec
; GFX90A-NEXT: renamable $vgpr37 = COPY renamable $vgpr27, implicit $exec
- ; GFX90A-NEXT: renamable $vgpr49 = COPY renamable $vgpr27, implicit $exec
; GFX90A-NEXT: renamable $vgpr29 = COPY renamable $vgpr27, implicit $exec
; GFX90A-NEXT: renamable $vgpr51 = COPY renamable $vgpr27, implicit $exec
; GFX90A-NEXT: renamable $vgpr33 = COPY renamable $vgpr27, implicit $exec
; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr27, renamable $vgpr26_vgpr27, 0, 0, implicit $exec :: (store (s64) into `ptr addrspace(3) null`, addrspace 3)
; GFX90A-NEXT: renamable $vgpr10 = COPY renamable $sgpr21, implicit $exec
- ; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr10, killed renamable $vgpr38_vgpr39, 0, 0, implicit $exec :: (store (s64) into %ir.7, addrspace 3)
+ ; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr10, killed renamable $vgpr48_vgpr49, 0, 0, implicit $exec :: (store (s64) into %ir.7, addrspace 3)
; GFX90A-NEXT: renamable $vgpr12 = COPY killed renamable $sgpr22, implicit $exec
; GFX90A-NEXT: DS_WRITE_B64_gfx9 killed renamable $vgpr12, killed renamable $vgpr34_vgpr35, 0, 0, implicit $exec :: (store (s64) into %ir.8, addrspace 3)
- ; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr27, killed renamable $vgpr36_vgpr37, 0, 0, implicit $exec :: (store (s64) into `ptr addrspace(3) null`, addrspace 3)
- ; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr10, killed renamable $vgpr48_vgpr49, 0, 0, implicit $exec :: (store (s64) into %ir.7, addrspace 3)
+ ; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr27, killed renamable $vgpr38_vgpr39, 0, 0, implicit $exec :: (store (s64) into `ptr addrspace(3) null`, addrspace 3)
+ ; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr10, killed renamable $vgpr36_vgpr37, 0, 0, implicit $exec :: (store (s64) into %ir.7, addrspace 3)
; GFX90A-NEXT: DS_WRITE_B64_gfx9 renamable $vgpr27, killed renamable $vgpr28_vgpr29, 0, 0, implicit $exec :: (store (s64) into `ptr addrspace(3) null`, addrspace 3)
; GFX90A-NEXT: DS_WRITE_B64_gfx9 killed renamable $vgpr10, killed renamable $vgpr50_vgpr51, 0, 0, implicit $exec :: (store (s64) into %ir.7, addrspace 3)
; GFX90A-NEXT: DS_WRITE_B64_gfx9 killed renamable $vgpr27, killed renamable $vgpr32_vgpr33, 0, 0, implicit $exec :: (store (s64) into `ptr addrspace(3) null`, addrspace 3)
diff --git a/llvm/test/CodeGen/AMDGPU/carryout-selection.ll b/llvm/test/CodeGen/AMDGPU/carryout-selection.ll
index 50693a92bc92cb9..c6320658e7800b6 100644
--- a/llvm/test/CodeGen/AMDGPU/carryout-selection.ll
+++ b/llvm/test/CodeGen/AMDGPU/carryout-selection.ll
@@ -1965,7 +1965,7 @@ define amdgpu_kernel void @sudiv64(ptr addrspace(1) %out, i64 %x, i64 %y) {
; VI-NEXT: v_mul_lo_u32 v2, s8, v4
; VI-NEXT: v_mad_u64_u32 v[0:1], s[0:1], s8, v5, 0
; VI-NEXT: v_mul_lo_u32 v3, s9, v5
-; VI-NEXT: v_add_u32_e32 v1, vcc, v2, v1
+; VI-NEXT: v_add_u32_e32 v1, vcc, v1, v2
; VI-NEXT: v_add_u32_e32 v3, vcc, v1, v3
; VI-NEXT: v_mul_hi_u32 v6, v5, v0
; VI-NEXT: v_mad_u64_u32 v[1:2], s[0:1], v5, v3, 0
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..a78a8a2ed8fe664
--- /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 v5, v0
+; CHECK-NEXT: s_branch .LBB0_3
+; CHECK-NEXT: .LBB0_2:
+; 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 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[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
+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)
diff --git a/llvm/test/CodeGen/AMDGPU/sdwa-peephole.ll b/llvm/test/CodeGen/AMDGPU/sdwa-peephole.ll
index 086fab3f5217589..680c27b69770903 100644
--- a/llvm/test/CodeGen/AMDGPU/sdwa-peephole.ll
+++ b/llvm/test/CodeGen/AMDGPU/sdwa-peephole.ll
@@ -16,7 +16,7 @@ define amdgpu_kernel void @add_shr_i32(ptr addrspace(1) %out, ptr addrspace(1) %
; NOSDWA-NEXT: v_mov_b32_e32 v1, s1
; NOSDWA-NEXT: s_waitcnt vmcnt(0)
; NOSDWA-NEXT: v_lshrrev_b32_e32 v3, 16, v2
-; NOSDWA-NEXT: v_add_u32_e32 v2, vcc, v2, v3
+; NOSDWA-NEXT: v_add_u32_e32 v2, vcc, v3, v2
; NOSDWA-NEXT: flat_store_dword v[0:1], v2
; NOSDWA-NEXT: s_endpgm
;
@@ -30,7 +30,7 @@ define amdgpu_kernel void @add_shr_i32(ptr addrspace(1) %out, ptr addrspace(1) %
; GFX89-NEXT: v_mov_b32_e32 v0, s0
; GFX89-NEXT: v_mov_b32_e32 v1, s1
; GFX89-NEXT: s_waitcnt vmcnt(0)
-; GFX89-NEXT: v_add_u32_sdwa v2, vcc, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX89-NEXT: v_add_u32_sdwa v2, vcc, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; GFX89-NEXT: flat_store_dword v[0:1], v2
; GFX89-NEXT: s_endpgm
;
diff --git a/llvm/test/CodeGen/AMDGPU/vgpr-liverange-ir.ll b/llvm/test/CodeGen/AMDGPU/vgpr-liverange-ir.ll
index aa65a62e242daaa..60f309c0c1882fb 100644
--- a/llvm/test/CodeGen/AMDGPU/vgpr-liverange-ir.ll
+++ b/llvm/test/CodeGen/AMDGPU/vgpr-liverange-ir.ll
@@ -509,7 +509,7 @@ define amdgpu_kernel void @livevariables_update_missed_block(ptr addrspace(1) %s
; SI-NEXT: bb.6.sw.bb18:
; SI-NEXT: successors: %bb.5(0x80000000)
; SI-NEXT: {{ $}}
- ; SI-NEXT: [[PHI1:%[0-9]+]]:vgpr_32 = PHI undef %33:vgpr_32, %bb.3, [[GLOBAL_LOAD_UBYTE1]], %bb.4
+ ; SI-NEXT: [[PHI1:%[0-9]+]]:vgpr_32 = PHI undef %35:vgpr_32, %bb.3, [[GLOBAL_LOAD_UBYTE1]], %bb.4
; SI-NEXT: [[V_MOV_B2:%[0-9]+]]:vreg_64 = V_MOV_B64_PSEUDO 0, implicit $exec
; SI-NEXT: GLOBAL_STORE_BYTE killed [[V_MOV_B2]], killed [[PHI1]], 0, 0, implicit $exec :: (store (s8) into `ptr addrspace(1) null`, addrspace 1)
; SI-NEXT: S_BRANCH %bb.5
More information about the llvm-commits
mailing list