[llvm] [AMDGPU] Add DAG mutation to improve scheduling before barriers (PR #142716)
Carl Ritson via llvm-commits
llvm-commits at lists.llvm.org
Thu Oct 16 00:48:19 PDT 2025
https://github.com/perlfu updated https://github.com/llvm/llvm-project/pull/142716
>From b593a9838d3ab8b45544da8c30969a08048c3777 Mon Sep 17 00:00:00 2001
From: Carl Ritson <carl.ritson at amd.com>
Date: Tue, 3 Jun 2025 13:22:41 +0900
Subject: [PATCH 1/2] [AMDGPU] Add DAG mutation to improve scheduling before
barriers
Add scheduler DAG mutation to add data dependencies between
atomic fences and preceding memory reads. This allows some
modelling of the impact an atomic fence can have on outstanding
memory accesses.
This is beneficial when a fence would cause wait count insertion,
as more instructions will be scheduled before the fence hiding
memory latency. It also reduces the risk of a fence causing a
premature wait on all active memory operations.
---
.../Target/AMDGPU/AMDGPUBarrierLatency.cpp | 73 +++++++++++++++++
llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.h | 21 +++++
.../lib/Target/AMDGPU/AMDGPUTargetMachine.cpp | 4 +
llvm/lib/Target/AMDGPU/CMakeLists.txt | 1 +
.../CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll | 35 ++++----
.../AMDGPU/schedule-barrier-latency.mir | 82 +++++++++++++++++++
llvm/test/CodeGen/AMDGPU/waitcnt-vscnt.ll | 20 ++---
7 files changed, 207 insertions(+), 29 deletions(-)
create mode 100644 llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
create mode 100644 llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.h
create mode 100644 llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp b/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
new file mode 100644
index 0000000000000..088660c0a3e26
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
@@ -0,0 +1,73 @@
+//===--- AMDGPUBarrierLatency.cpp - AMDGPU Barrier Latency ----------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file contains a DAG scheduling mutation to add latency to
+/// barrier edges between ATOMIC_FENCE instructions and preceeding
+/// memory accesses potentially affected by the fence.
+/// This is beneficial when a fence would cause wait count insertion,
+/// as more instructions will be scheduled before the fence hiding
+/// memory latency.
+/// It also reduces the risk of a fence causing a premature wait
+/// on all active memory operations.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUBarrierLatency.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "SIInstrInfo.h"
+#include "llvm/CodeGen/ScheduleDAGInstrs.h"
+
+using namespace llvm;
+
+namespace {
+
+class BarrierLatency : public ScheduleDAGMutation {
+public:
+ BarrierLatency() = default;
+ void apply(ScheduleDAGInstrs *DAG) override;
+};
+
+void BarrierLatency::apply(ScheduleDAGInstrs *DAG) {
+ const unsigned SyntheticLatency = 2000;
+ for (SUnit &SU : DAG->SUnits) {
+ const MachineInstr *MI = SU.getInstr();
+ if (MI->getOpcode() != AMDGPU::ATOMIC_FENCE)
+ continue;
+
+ // Update latency on barrier edges of ATOMIC_FENCE.
+ // We don't consider the scope of the fence or type of instruction
+ // involved in the barrier edge.
+ for (SDep &PredDep : SU.Preds) {
+ if (!PredDep.isBarrier())
+ continue;
+ SUnit *PredSU = PredDep.getSUnit();
+ MachineInstr *MI = PredSU->getInstr();
+ // Only consider memory loads
+ if (!MI->mayLoad() || MI->mayStore())
+ continue;
+ SDep ForwardD = PredDep;
+ ForwardD.setSUnit(&SU);
+ for (SDep &SuccDep : PredSU->Succs) {
+ if (SuccDep == ForwardD) {
+ SuccDep.setLatency(SuccDep.getLatency() + SyntheticLatency);
+ break;
+ }
+ }
+ PredDep.setLatency(PredDep.getLatency() + SyntheticLatency);
+ PredSU->setDepthDirty();
+ SU.setDepthDirty();
+ }
+ }
+}
+
+} // end namespace
+
+std::unique_ptr<ScheduleDAGMutation>
+llvm::createAMDGPUBarrierLatencyDAGMutation() {
+ return std::make_unique<BarrierLatency>();
+}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.h b/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.h
new file mode 100644
index 0000000000000..c23f0b99fe822
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.h
@@ -0,0 +1,21 @@
+//===- AMDGPUBarrierLatency.h - AMDGPU Export Clustering --------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUBARRIERLATENCY_H
+#define LLVM_LIB_TARGET_AMDGPU_AMDGPUBARRIERLATENCY_H
+
+#include "llvm/CodeGen/ScheduleDAGMutation.h"
+#include <memory>
+
+namespace llvm {
+
+std::unique_ptr<ScheduleDAGMutation> createAMDGPUBarrierLatencyDAGMutation();
+
+} // namespace llvm
+
+#endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUBARRIERLATENCY_H
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 4958a200de4e0..996b55f42fd0b 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -17,6 +17,7 @@
#include "AMDGPUTargetMachine.h"
#include "AMDGPU.h"
#include "AMDGPUAliasAnalysis.h"
+#include "AMDGPUBarrierLatency.h"
#include "AMDGPUCtorDtorLowering.h"
#include "AMDGPUExportClustering.h"
#include "AMDGPUExportKernelRuntimeHandles.h"
@@ -639,6 +640,7 @@ createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
DAG->addMutation(createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::Initial));
DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
+ DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation());
return DAG;
}
@@ -659,6 +661,7 @@ createGCNMaxMemoryClauseMachineScheduler(MachineSchedContext *C) {
if (ST.shouldClusterStores())
DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
+ DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation());
return DAG;
}
@@ -1197,6 +1200,7 @@ GCNTargetMachine::createPostMachineScheduler(MachineSchedContext *C) const {
EnableVOPD)
DAG->addMutation(createVOPDPairingMutation());
DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
+ DAG->addMutation(createAMDGPUBarrierLatencyDAGMutation());
return DAG;
}
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/CMakeLists.txt b/llvm/lib/Target/AMDGPU/CMakeLists.txt
index 13f727b68c0d9..a1e0e5293c706 100644
--- a/llvm/lib/Target/AMDGPU/CMakeLists.txt
+++ b/llvm/lib/Target/AMDGPU/CMakeLists.txt
@@ -52,6 +52,7 @@ add_llvm_target(AMDGPUCodeGen
AMDGPUAsmPrinter.cpp
AMDGPUAtomicOptimizer.cpp
AMDGPUAttributor.cpp
+ AMDGPUBarrierLatency.cpp
AMDGPUCallLowering.cpp
AMDGPUCodeGenPrepare.cpp
AMDGPUCombinerHelper.cpp
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll
index e8b8d05ab26ad..e8eccb0e408c1 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll
@@ -147,14 +147,13 @@ define weak_odr amdgpu_kernel void @dpp_test1(ptr %arg) local_unnamed_addr {
; GFX8-OPT-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX8-OPT-NEXT: v_mov_b32_e32 v2, 0
; GFX8-OPT-NEXT: s_waitcnt lgkmcnt(0)
-; GFX8-OPT-NEXT: s_barrier
-; GFX8-OPT-NEXT: v_add_u32_e32 v1, vcc, v1, v1
-; GFX8-OPT-NEXT: s_nop 1
-; GFX8-OPT-NEXT: v_mov_b32_dpp v2, v1 quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
-; GFX8-OPT-NEXT: v_add_u32_e32 v2, vcc, v2, v1
-; GFX8-OPT-NEXT: v_mov_b32_e32 v1, s1
+; GFX8-OPT-NEXT: v_add_u32_e32 v4, vcc, v1, v1
+; GFX8-OPT-NEXT: v_mov_b32_e32 v3, s1
; GFX8-OPT-NEXT: v_add_u32_e32 v0, vcc, s0, v0
-; GFX8-OPT-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-OPT-NEXT: v_mov_b32_dpp v2, v4 quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
+; GFX8-OPT-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc
+; GFX8-OPT-NEXT: v_add_u32_e32 v2, vcc, v2, v4
+; GFX8-OPT-NEXT: s_barrier
; GFX8-OPT-NEXT: flat_store_dword v[0:1], v2
; GFX8-OPT-NEXT: s_endpgm
;
@@ -194,14 +193,14 @@ define weak_odr amdgpu_kernel void @dpp_test1(ptr %arg) local_unnamed_addr {
; GFX10-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX10-NEXT: v_mov_b32_e32 v2, 0
; GFX10-NEXT: ds_read_b32 v1, v0
-; GFX10-NEXT: s_barrier
-; GFX10-NEXT: buffer_gl0_inv
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: v_add_co_u32 v0, s0, s0, v0
-; GFX10-NEXT: v_add_nc_u32_e32 v1, v1, v1
-; GFX10-NEXT: v_mov_b32_dpp v2, v1 quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
-; GFX10-NEXT: v_add_nc_u32_e32 v2, v2, v1
+; GFX10-NEXT: v_add_nc_u32_e32 v3, v1, v1
; GFX10-NEXT: v_add_co_ci_u32_e64 v1, s0, s1, 0, s0
+; GFX10-NEXT: v_mov_b32_dpp v2, v3 quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
+; GFX10-NEXT: v_add_nc_u32_e32 v2, v2, v3
+; GFX10-NEXT: s_barrier
+; GFX10-NEXT: buffer_gl0_inv
; GFX10-NEXT: flat_store_dword v[0:1], v2
; GFX10-NEXT: s_endpgm
;
@@ -213,15 +212,15 @@ define weak_odr amdgpu_kernel void @dpp_test1(ptr %arg) local_unnamed_addr {
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT: v_and_b32_e32 v0, 0xffc, v0
; GFX11-NEXT: ds_load_b32 v1, v0
-; GFX11-NEXT: s_barrier
-; GFX11-NEXT: buffer_gl0_inv
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: v_add_co_u32 v0, s0, s0, v0
-; GFX11-NEXT: v_add_nc_u32_e32 v1, v1, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_mov_b32_dpp v2, v1 quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
-; GFX11-NEXT: v_add_nc_u32_e32 v2, v2, v1
+; GFX11-NEXT: v_add_nc_u32_e32 v3, v1, v1
; GFX11-NEXT: v_add_co_ci_u32_e64 v1, null, s1, 0, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT: v_mov_b32_dpp v2, v3 quad_perm:[1,0,3,2] row_mask:0xf bank_mask:0xf
+; GFX11-NEXT: v_add_nc_u32_e32 v2, v2, v3
+; GFX11-NEXT: s_barrier
+; GFX11-NEXT: buffer_gl0_inv
; GFX11-NEXT: flat_store_b32 v[0:1], v2
; GFX11-NEXT: s_endpgm
bb:
diff --git a/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir b/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir
new file mode 100644
index 0000000000000..fc736d055a3a5
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir
@@ -0,0 +1,82 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -passes=postmisched -o - %s | FileCheck %s
+
+# Ensure WMMA operations stay before the final atomic fence and barrier group.
+---
+name: test
+tracksRegLiveness: true
+body: |
+ bb.0:
+ liveins: $sgpr0, $sgpr12, $vgpr36, $vgpr37, $vgpr38, $vgpr39, $vgpr40, $vgpr1_vgpr2_vgpr3_vgpr4, $vgpr5_vgpr6_vgpr7_vgpr8, $vgpr9_vgpr10_vgpr11_vgpr12, $vgpr13_vgpr14_vgpr15_vgpr16, $vgpr17_vgpr18_vgpr19_vgpr20, $vgpr21_vgpr22_vgpr23_vgpr24, $vgpr25_vgpr26_vgpr27_vgpr28, $vgpr29_vgpr30_vgpr31_vgpr32
+
+ ; CHECK-LABEL: name: test
+ ; CHECK: liveins: $sgpr0, $sgpr12, $vgpr36, $vgpr37, $vgpr38, $vgpr39, $vgpr40, $vgpr1_vgpr2_vgpr3_vgpr4, $vgpr5_vgpr6_vgpr7_vgpr8, $vgpr9_vgpr10_vgpr11_vgpr12, $vgpr13_vgpr14_vgpr15_vgpr16, $vgpr17_vgpr18_vgpr19_vgpr20, $vgpr21_vgpr22_vgpr23_vgpr24, $vgpr25_vgpr26_vgpr27_vgpr28, $vgpr29_vgpr30_vgpr31_vgpr32
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: ATOMIC_FENCE 5, 2
+ ; CHECK-NEXT: S_BARRIER
+ ; CHECK-NEXT: ATOMIC_FENCE 4, 2
+ ; CHECK-NEXT: BUNDLE implicit-def $vgpr41_vgpr42_vgpr43_vgpr44, implicit-def $vgpr41, implicit-def $vgpr41_lo16, implicit-def $vgpr41_hi16, implicit-def $vgpr42, implicit-def $vgpr42_lo16, implicit-def $vgpr42_hi16, implicit-def $vgpr43, implicit-def $vgpr43_lo16, implicit-def $vgpr43_hi16, implicit-def $vgpr44, implicit-def $vgpr44_lo16, implicit-def $vgpr44_hi16, implicit-def $vgpr41_vgpr42, implicit-def $vgpr41_vgpr42_vgpr43, implicit-def $vgpr42_vgpr43, implicit-def $vgpr42_vgpr43_vgpr44, implicit-def $vgpr43_vgpr44, implicit-def $vgpr45_vgpr46_vgpr47_vgpr48, implicit-def $vgpr45, implicit-def $vgpr45_lo16, implicit-def $vgpr45_hi16, implicit-def $vgpr46, implicit-def $vgpr46_lo16, implicit-def $vgpr46_hi16, implicit-def $vgpr47, implicit-def $vgpr47_lo16, implicit-def $vgpr47_hi16, implicit-def $vgpr48, implicit-def $vgpr48_lo16, implicit-def $vgpr48_hi16, implicit-def $vgpr45_vgpr46, implicit-def $vgpr45_vgpr46_vgpr47, implicit-def $vgpr46_vgpr47, implicit-def $vgpr46_vgpr47_vgpr48, implicit-def $vgpr47_vgpr48, implicit-def $vgpr49_vgpr50_vgpr51_vgpr52, implicit-def $vgpr49, implicit-def $vgpr49_lo16, implicit-def $vgpr49_hi16, implicit-def $vgpr50, implicit-def $vgpr50_lo16, implicit-def $vgpr50_hi16, implicit-def $vgpr51, implicit-def $vgpr51_lo16, implicit-def $vgpr51_hi16, implicit-def $vgpr52, implicit-def $vgpr52_lo16, implicit-def $vgpr52_hi16, implicit-def $vgpr49_vgpr50, implicit-def $vgpr49_vgpr50_vgpr51, implicit-def $vgpr50_vgpr51, implicit-def $vgpr50_vgpr51_vgpr52, implicit-def $vgpr51_vgpr52, implicit-def $vgpr53_vgpr54_vgpr55_vgpr56, implicit-def $vgpr53, implicit-def $vgpr53_lo16, implicit-def $vgpr53_hi16, implicit-def $vgpr54, implicit-def $vgpr54_lo16, implicit-def $vgpr54_hi16, implicit-def $vgpr55, implicit-def $vgpr55_lo16, implicit-def $vgpr55_hi16, implicit-def $vgpr56, implicit-def $vgpr56_lo16, implicit-def $vgpr56_hi16, implicit-def $vgpr53_vgpr54, implicit-def $vgpr53_vgpr54_vgpr55, implicit-def $vgpr54_vgpr55, implicit-def $vgpr54_vgpr55_vgpr56, implicit-def $vgpr55_vgpr56, implicit-def $vgpr57_vgpr58_vgpr59_vgpr60, implicit-def $vgpr57, implicit-def $vgpr57_lo16, implicit-def $vgpr57_hi16, implicit-def $vgpr58, implicit-def $vgpr58_lo16, implicit-def $vgpr58_hi16, implicit-def $vgpr59, implicit-def $vgpr59_lo16, implicit-def $vgpr59_hi16, implicit-def $vgpr60, implicit-def $vgpr60_lo16, implicit-def $vgpr60_hi16, implicit-def $vgpr57_vgpr58, implicit-def $vgpr57_vgpr58_vgpr59, implicit-def $vgpr58_vgpr59, implicit-def $vgpr58_vgpr59_vgpr60, implicit-def $vgpr59_vgpr60, implicit-def $vgpr61_vgpr62_vgpr63_vgpr64, implicit-def $vgpr61, implicit-def $vgpr61_lo16, implicit-def $vgpr61_hi16, implicit-def $vgpr62, implicit-def $vgpr62_lo16, implicit-def $vgpr62_hi16, implicit-def $vgpr63, implicit-def $vgpr63_lo16, implicit-def $vgpr63_hi16, implicit-def $vgpr64, implicit-def $vgpr64_lo16, implicit-def $vgpr64_hi16, implicit-def $vgpr61_vgpr62, implicit-def $vgpr61_vgpr62_vgpr63, implicit-def $vgpr62_vgpr63, implicit-def $vgpr62_vgpr63_vgpr64, implicit-def $vgpr63_vgpr64, implicit-def $vgpr65_vgpr66_vgpr67_vgpr68, implicit-def $vgpr65, implicit-def $vgpr65_lo16, implicit-def $vgpr65_hi16, implicit-def $vgpr66, implicit-def $vgpr66_lo16, implicit-def $vgpr66_hi16, implicit-def $vgpr67, implicit-def $vgpr67_lo16, implicit-def $vgpr67_hi16, implicit-def $vgpr68, implicit-def $vgpr68_lo16, implicit-def $vgpr68_hi16, implicit-def $vgpr65_vgpr66, implicit-def $vgpr65_vgpr66_vgpr67, implicit-def $vgpr66_vgpr67, implicit-def $vgpr66_vgpr67_vgpr68, implicit-def $vgpr67_vgpr68, implicit-def $vgpr69_vgpr70_vgpr71_vgpr72, implicit-def $vgpr69, implicit-def $vgpr69_lo16, implicit-def $vgpr69_hi16, implicit-def $vgpr70, implicit-def $vgpr70_lo16, implicit-def $vgpr70_hi16, implicit-def $vgpr71, implicit-def $vgpr71_lo16, implicit-def $vgpr71_hi16, implicit-def $vgpr72, implicit-def $vgpr72_lo16, implicit-def $vgpr72_hi16, implicit-def $vgpr69_vgpr70, implicit-def $vgpr69_vgpr70_vgpr71, implicit-def $vgpr70_vgpr71, implicit-def $vgpr70_vgpr71_vgpr72, implicit-def $vgpr71_vgpr72, implicit-def $vgpr73_vgpr74_vgpr75_vgpr76, implicit-def $vgpr73, implicit-def $vgpr73_lo16, implicit-def $vgpr73_hi16, implicit-def $vgpr74, implicit-def $vgpr74_lo16, implicit-def $vgpr74_hi16, implicit-def $vgpr75, implicit-def $vgpr75_lo16, implicit-def $vgpr75_hi16, implicit-def $vgpr76, implicit-def $vgpr76_lo16, implicit-def $vgpr76_hi16, implicit-def $vgpr73_vgpr74, implicit-def $vgpr73_vgpr74_vgpr75, implicit-def $vgpr74_vgpr75, implicit-def $vgpr74_vgpr75_vgpr76, implicit-def $vgpr75_vgpr76, implicit-def $vgpr77_vgpr78_vgpr79_vgpr80, implicit-def $vgpr77, implicit-def $vgpr77_lo16, implicit-def $vgpr77_hi16, implicit-def $vgpr78, implicit-def $vgpr78_lo16, implicit-def $vgpr78_hi16, implicit-def $vgpr79, implicit-def $vgpr79_lo16, implicit-def $vgpr79_hi16, implicit-def $vgpr80, implicit-def $vgpr80_lo16, implicit-def $vgpr80_hi16, implicit-def $vgpr77_vgpr78, implicit-def $vgpr77_vgpr78_vgpr79, implicit-def $vgpr78_vgpr79, implicit-def $vgpr78_vgpr79_vgpr80, implicit-def $vgpr79_vgpr80, implicit-def $vgpr81_vgpr82_vgpr83_vgpr84, implicit-def $vgpr81, implicit-def $vgpr81_lo16, implicit-def $vgpr81_hi16, implicit-def $vgpr82, implicit-def $vgpr82_lo16, implicit-def $vgpr82_hi16, implicit-def $vgpr83, implicit-def $vgpr83_lo16, implicit-def $vgpr83_hi16, implicit-def $vgpr84, implicit-def $vgpr84_lo16, implicit-def $vgpr84_hi16, implicit-def $vgpr81_vgpr82, implicit-def $vgpr81_vgpr82_vgpr83, implicit-def $vgpr82_vgpr83, implicit-def $vgpr82_vgpr83_vgpr84, implicit-def $vgpr83_vgpr84, implicit-def $vgpr85_vgpr86_vgpr87_vgpr88, implicit-def $vgpr85, implicit-def $vgpr85_lo16, implicit-def $vgpr85_hi16, implicit-def $vgpr86, implicit-def $vgpr86_lo16, implicit-def $vgpr86_hi16, implicit-def $vgpr87, implicit-def $vgpr87_lo16, implicit-def $vgpr87_hi16, implicit-def $vgpr88, implicit-def $vgpr88_lo16, implicit-def $vgpr88_hi16, implicit-def $vgpr85_vgpr86, implicit-def $vgpr85_vgpr86_vgpr87, implicit-def $vgpr86_vgpr87, implicit-def $vgpr86_vgpr87_vgpr88, implicit-def $vgpr87_vgpr88, implicit killed $vgpr36, implicit $exec, implicit killed $vgpr37, implicit killed $vgpr38 {
+ ; CHECK-NEXT: $vgpr41_vgpr42_vgpr43_vgpr44 = DS_READ_B128_gfx9 $vgpr36, 0, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr45_vgpr46_vgpr47_vgpr48 = DS_READ2_B64_gfx9 $vgpr36, 2, 3, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr49_vgpr50_vgpr51_vgpr52 = DS_READ_B128_gfx9 $vgpr37, 0, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr53_vgpr54_vgpr55_vgpr56 = DS_READ2_B64_gfx9 $vgpr37, 2, 3, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr57_vgpr58_vgpr59_vgpr60 = DS_READ_B128_gfx9 $vgpr37, 768, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr61_vgpr62_vgpr63_vgpr64 = DS_READ2_B64_gfx9 killed $vgpr37, 98, 99, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr65_vgpr66_vgpr67_vgpr68 = DS_READ_B128_gfx9 $vgpr36, 768, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr69_vgpr70_vgpr71_vgpr72 = DS_READ2_B64_gfx9 $vgpr36, 98, 99, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr73_vgpr74_vgpr75_vgpr76 = DS_READ_B128_gfx9 $vgpr36, 1536, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr77_vgpr78_vgpr79_vgpr80 = DS_READ2_B64_gfx9 $vgpr36, 194, 195, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr81_vgpr82_vgpr83_vgpr84 = DS_READ_B128_gfx9 killed $vgpr36, 2304, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: $vgpr85_vgpr86_vgpr87_vgpr88 = DS_READ2_B64_gfx9 killed $vgpr38, 0, 1, 0, implicit $exec :: (load (s128), addrspace 3)
+ ; CHECK-NEXT: }
+ ; CHECK-NEXT: $sgpr1 = S_ADD_I32 $sgpr0, 16, implicit-def dead $scc
+ ; CHECK-NEXT: $vgpr39 = V_ADD_U32_e32 32, killed $vgpr39, implicit $exec
+ ; CHECK-NEXT: $vgpr40 = V_ADD_U32_e32 32, killed $vgpr40, implicit $exec
+ ; CHECK-NEXT: S_CMP_LT_U32 killed $sgpr0, killed $sgpr12, implicit-def $scc
+ ; CHECK-NEXT: $sgpr0 = S_MOV_B32 killed $sgpr1
+ ; CHECK-NEXT: early-clobber $vgpr29_vgpr30_vgpr31_vgpr32 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48, 8, $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr29_vgpr30_vgpr31_vgpr32, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr25_vgpr26_vgpr27_vgpr28 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48, 8, $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr25_vgpr26_vgpr27_vgpr28, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr21_vgpr22_vgpr23_vgpr24 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72, 8, $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr21_vgpr22_vgpr23_vgpr24, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr17_vgpr18_vgpr19_vgpr20 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72, 8, $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr17_vgpr18_vgpr19_vgpr20, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr13_vgpr14_vgpr15_vgpr16 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80, 8, $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr13_vgpr14_vgpr15_vgpr16, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr9_vgpr10_vgpr11_vgpr12 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80, 8, $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr9_vgpr10_vgpr11_vgpr12, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr5_vgpr6_vgpr7_vgpr8 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88, 8, killed $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr5_vgpr6_vgpr7_vgpr8, 0, 0, implicit $exec
+ ; CHECK-NEXT: early-clobber $vgpr1_vgpr2_vgpr3_vgpr4 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88, 8, killed $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr1_vgpr2_vgpr3_vgpr4, 0, 0, implicit $exec
+ ; CHECK-NEXT: ATOMIC_FENCE 5, 2
+ ; CHECK-NEXT: S_BARRIER
+ ; CHECK-NEXT: ATOMIC_FENCE 4, 2
+ ATOMIC_FENCE 5, 2
+ S_BARRIER
+ ATOMIC_FENCE 4, 2
+ BUNDLE implicit-def $vgpr41_vgpr42_vgpr43_vgpr44, implicit-def $vgpr41, implicit-def $vgpr41_lo16, implicit-def $vgpr41_hi16, implicit-def $vgpr42, implicit-def $vgpr42_lo16, implicit-def $vgpr42_hi16, implicit-def $vgpr43, implicit-def $vgpr43_lo16, implicit-def $vgpr43_hi16, implicit-def $vgpr44, implicit-def $vgpr44_lo16, implicit-def $vgpr44_hi16, implicit-def $vgpr41_vgpr42, implicit-def $vgpr41_vgpr42_vgpr43, implicit-def $vgpr42_vgpr43, implicit-def $vgpr42_vgpr43_vgpr44, implicit-def $vgpr43_vgpr44, implicit-def $vgpr45_vgpr46_vgpr47_vgpr48, implicit-def $vgpr45, implicit-def $vgpr45_lo16, implicit-def $vgpr45_hi16, implicit-def $vgpr46, implicit-def $vgpr46_lo16, implicit-def $vgpr46_hi16, implicit-def $vgpr47, implicit-def $vgpr47_lo16, implicit-def $vgpr47_hi16, implicit-def $vgpr48, implicit-def $vgpr48_lo16, implicit-def $vgpr48_hi16, implicit-def $vgpr45_vgpr46, implicit-def $vgpr45_vgpr46_vgpr47, implicit-def $vgpr46_vgpr47, implicit-def $vgpr46_vgpr47_vgpr48, implicit-def $vgpr47_vgpr48, implicit-def $vgpr49_vgpr50_vgpr51_vgpr52, implicit-def $vgpr49, implicit-def $vgpr49_lo16, implicit-def $vgpr49_hi16, implicit-def $vgpr50, implicit-def $vgpr50_lo16, implicit-def $vgpr50_hi16, implicit-def $vgpr51, implicit-def $vgpr51_lo16, implicit-def $vgpr51_hi16, implicit-def $vgpr52, implicit-def $vgpr52_lo16, implicit-def $vgpr52_hi16, implicit-def $vgpr49_vgpr50, implicit-def $vgpr49_vgpr50_vgpr51, implicit-def $vgpr50_vgpr51, implicit-def $vgpr50_vgpr51_vgpr52, implicit-def $vgpr51_vgpr52, implicit-def $vgpr53_vgpr54_vgpr55_vgpr56, implicit-def $vgpr53, implicit-def $vgpr53_lo16, implicit-def $vgpr53_hi16, implicit-def $vgpr54, implicit-def $vgpr54_lo16, implicit-def $vgpr54_hi16, implicit-def $vgpr55, implicit-def $vgpr55_lo16, implicit-def $vgpr55_hi16, implicit-def $vgpr56, implicit-def $vgpr56_lo16, implicit-def $vgpr56_hi16, implicit-def $vgpr53_vgpr54, implicit-def $vgpr53_vgpr54_vgpr55, implicit-def $vgpr54_vgpr55, implicit-def $vgpr54_vgpr55_vgpr56, implicit-def $vgpr55_vgpr56, implicit-def $vgpr57_vgpr58_vgpr59_vgpr60, implicit-def $vgpr57, implicit-def $vgpr57_lo16, implicit-def $vgpr57_hi16, implicit-def $vgpr58, implicit-def $vgpr58_lo16, implicit-def $vgpr58_hi16, implicit-def $vgpr59, implicit-def $vgpr59_lo16, implicit-def $vgpr59_hi16, implicit-def $vgpr60, implicit-def $vgpr60_lo16, implicit-def $vgpr60_hi16, implicit-def $vgpr57_vgpr58, implicit-def $vgpr57_vgpr58_vgpr59, implicit-def $vgpr58_vgpr59, implicit-def $vgpr58_vgpr59_vgpr60, implicit-def $vgpr59_vgpr60, implicit-def $vgpr61_vgpr62_vgpr63_vgpr64, implicit-def $vgpr61, implicit-def $vgpr61_lo16, implicit-def $vgpr61_hi16, implicit-def $vgpr62, implicit-def $vgpr62_lo16, implicit-def $vgpr62_hi16, implicit-def $vgpr63, implicit-def $vgpr63_lo16, implicit-def $vgpr63_hi16, implicit-def $vgpr64, implicit-def $vgpr64_lo16, implicit-def $vgpr64_hi16, implicit-def $vgpr61_vgpr62, implicit-def $vgpr61_vgpr62_vgpr63, implicit-def $vgpr62_vgpr63, implicit-def $vgpr62_vgpr63_vgpr64, implicit-def $vgpr63_vgpr64, implicit-def $vgpr65_vgpr66_vgpr67_vgpr68, implicit-def $vgpr65, implicit-def $vgpr65_lo16, implicit-def $vgpr65_hi16, implicit-def $vgpr66, implicit-def $vgpr66_lo16, implicit-def $vgpr66_hi16, implicit-def $vgpr67, implicit-def $vgpr67_lo16, implicit-def $vgpr67_hi16, implicit-def $vgpr68, implicit-def $vgpr68_lo16, implicit-def $vgpr68_hi16, implicit-def $vgpr65_vgpr66, implicit-def $vgpr65_vgpr66_vgpr67, implicit-def $vgpr66_vgpr67, implicit-def $vgpr66_vgpr67_vgpr68, implicit-def $vgpr67_vgpr68, implicit-def $vgpr69_vgpr70_vgpr71_vgpr72, implicit-def $vgpr69, implicit-def $vgpr69_lo16, implicit-def $vgpr69_hi16, implicit-def $vgpr70, implicit-def $vgpr70_lo16, implicit-def $vgpr70_hi16, implicit-def $vgpr71, implicit-def $vgpr71_lo16, implicit-def $vgpr71_hi16, implicit-def $vgpr72, implicit-def $vgpr72_lo16, implicit-def $vgpr72_hi16, implicit-def $vgpr69_vgpr70, implicit-def $vgpr69_vgpr70_vgpr71, implicit-def $vgpr70_vgpr71, implicit-def $vgpr70_vgpr71_vgpr72, implicit-def $vgpr71_vgpr72, implicit-def $vgpr73_vgpr74_vgpr75_vgpr76, implicit-def $vgpr73, implicit-def $vgpr73_lo16, implicit-def $vgpr73_hi16, implicit-def $vgpr74, implicit-def $vgpr74_lo16, implicit-def $vgpr74_hi16, implicit-def $vgpr75, implicit-def $vgpr75_lo16, implicit-def $vgpr75_hi16, implicit-def $vgpr76, implicit-def $vgpr76_lo16, implicit-def $vgpr76_hi16, implicit-def $vgpr73_vgpr74, implicit-def $vgpr73_vgpr74_vgpr75, implicit-def $vgpr74_vgpr75, implicit-def $vgpr74_vgpr75_vgpr76, implicit-def $vgpr75_vgpr76, implicit-def $vgpr77_vgpr78_vgpr79_vgpr80, implicit-def $vgpr77, implicit-def $vgpr77_lo16, implicit-def $vgpr77_hi16, implicit-def $vgpr78, implicit-def $vgpr78_lo16, implicit-def $vgpr78_hi16, implicit-def $vgpr79, implicit-def $vgpr79_lo16, implicit-def $vgpr79_hi16, implicit-def $vgpr80, implicit-def $vgpr80_lo16, implicit-def $vgpr80_hi16, implicit-def $vgpr77_vgpr78, implicit-def $vgpr77_vgpr78_vgpr79, implicit-def $vgpr78_vgpr79, implicit-def $vgpr78_vgpr79_vgpr80, implicit-def $vgpr79_vgpr80, implicit-def $vgpr81_vgpr82_vgpr83_vgpr84, implicit-def $vgpr81, implicit-def $vgpr81_lo16, implicit-def $vgpr81_hi16, implicit-def $vgpr82, implicit-def $vgpr82_lo16, implicit-def $vgpr82_hi16, implicit-def $vgpr83, implicit-def $vgpr83_lo16, implicit-def $vgpr83_hi16, implicit-def $vgpr84, implicit-def $vgpr84_lo16, implicit-def $vgpr84_hi16, implicit-def $vgpr81_vgpr82, implicit-def $vgpr81_vgpr82_vgpr83, implicit-def $vgpr82_vgpr83, implicit-def $vgpr82_vgpr83_vgpr84, implicit-def $vgpr83_vgpr84, implicit-def $vgpr85_vgpr86_vgpr87_vgpr88, implicit-def $vgpr85, implicit-def $vgpr85_lo16, implicit-def $vgpr85_hi16, implicit-def $vgpr86, implicit-def $vgpr86_lo16, implicit-def $vgpr86_hi16, implicit-def $vgpr87, implicit-def $vgpr87_lo16, implicit-def $vgpr87_hi16, implicit-def $vgpr88, implicit-def $vgpr88_lo16, implicit-def $vgpr88_hi16, implicit-def $vgpr85_vgpr86, implicit-def $vgpr85_vgpr86_vgpr87, implicit-def $vgpr86_vgpr87, implicit-def $vgpr86_vgpr87_vgpr88, implicit-def $vgpr87_vgpr88, implicit $vgpr36, implicit $exec, implicit $vgpr37, implicit $vgpr38 {
+ $vgpr41_vgpr42_vgpr43_vgpr44 = DS_READ_B128_gfx9 $vgpr36, 0, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr45_vgpr46_vgpr47_vgpr48 = DS_READ2_B64_gfx9 $vgpr36, 2, 3, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr49_vgpr50_vgpr51_vgpr52 = DS_READ_B128_gfx9 $vgpr37, 0, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr53_vgpr54_vgpr55_vgpr56 = DS_READ2_B64_gfx9 $vgpr37, 2, 3, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr57_vgpr58_vgpr59_vgpr60 = DS_READ_B128_gfx9 $vgpr37, 768, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr61_vgpr62_vgpr63_vgpr64 = DS_READ2_B64_gfx9 $vgpr37, 98, 99, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr65_vgpr66_vgpr67_vgpr68 = DS_READ_B128_gfx9 $vgpr36, 768, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr69_vgpr70_vgpr71_vgpr72 = DS_READ2_B64_gfx9 $vgpr36, 98, 99, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr73_vgpr74_vgpr75_vgpr76 = DS_READ_B128_gfx9 $vgpr36, 1536, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr77_vgpr78_vgpr79_vgpr80 = DS_READ2_B64_gfx9 $vgpr36, 194, 195, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr81_vgpr82_vgpr83_vgpr84 = DS_READ_B128_gfx9 $vgpr36, 2304, 0, implicit $exec :: (load (s128), addrspace 3)
+ $vgpr85_vgpr86_vgpr87_vgpr88 = DS_READ2_B64_gfx9 $vgpr38, 0, 1, 0, implicit $exec :: (load (s128), addrspace 3)
+ }
+ $sgpr1 = S_ADD_I32 $sgpr0, 16, implicit-def dead $scc
+ $vgpr39 = V_ADD_U32_e32 32, killed $vgpr39, implicit $exec
+ $vgpr40 = V_ADD_U32_e32 32, killed $vgpr40, implicit $exec
+ S_CMP_LT_U32 killed $sgpr0, $sgpr12, implicit-def $scc
+ early-clobber $vgpr29_vgpr30_vgpr31_vgpr32 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48, 8, $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr29_vgpr30_vgpr31_vgpr32, 0, 0, implicit $exec
+ early-clobber $vgpr25_vgpr26_vgpr27_vgpr28 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48, 8, $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr25_vgpr26_vgpr27_vgpr28, 0, 0, implicit $exec
+ early-clobber $vgpr21_vgpr22_vgpr23_vgpr24 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72, 8, $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr21_vgpr22_vgpr23_vgpr24, 0, 0, implicit $exec
+ early-clobber $vgpr17_vgpr18_vgpr19_vgpr20 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr65_vgpr66_vgpr67_vgpr68_vgpr69_vgpr70_vgpr71_vgpr72, 8, $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr17_vgpr18_vgpr19_vgpr20, 0, 0, implicit $exec
+ early-clobber $vgpr13_vgpr14_vgpr15_vgpr16 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80, 8, $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr13_vgpr14_vgpr15_vgpr16, 0, 0, implicit $exec
+ early-clobber $vgpr9_vgpr10_vgpr11_vgpr12 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr73_vgpr74_vgpr75_vgpr76_vgpr77_vgpr78_vgpr79_vgpr80, 8, $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr9_vgpr10_vgpr11_vgpr12, 0, 0, implicit $exec
+ early-clobber $vgpr5_vgpr6_vgpr7_vgpr8 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, $vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88, 8, killed $vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56, 8, killed $vgpr5_vgpr6_vgpr7_vgpr8, 0, 0, implicit $exec
+ early-clobber $vgpr1_vgpr2_vgpr3_vgpr4 = V_WMMA_F32_16X16X16_F16_twoaddr_w64 8, killed $vgpr81_vgpr82_vgpr83_vgpr84_vgpr85_vgpr86_vgpr87_vgpr88, 8, killed $vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63_vgpr64, 8, killed $vgpr1_vgpr2_vgpr3_vgpr4, 0, 0, implicit $exec
+ $sgpr0 = S_MOV_B32 killed $sgpr1
+ ATOMIC_FENCE 5, 2
+ S_BARRIER
+ ATOMIC_FENCE 4, 2
+
+...
diff --git a/llvm/test/CodeGen/AMDGPU/waitcnt-vscnt.ll b/llvm/test/CodeGen/AMDGPU/waitcnt-vscnt.ll
index f3cb5a783a08c..30f52773bbce5 100644
--- a/llvm/test/CodeGen/AMDGPU/waitcnt-vscnt.ll
+++ b/llvm/test/CodeGen/AMDGPU/waitcnt-vscnt.ll
@@ -26,17 +26,17 @@ define amdgpu_kernel void @barrier_vmcnt_global(ptr addrspace(1) %arg) {
; GFX9-LABEL: barrier_vmcnt_global:
; GFX9: s_load_dwordx2 s[0:1], s[4:5], 0x24
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 2, v0
+; GFX9-NEXT: v_add_u32_e32 v2, 1, v0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dword v2, v1, s[0:1]
-; GFX9-NEXT: v_add_u32_e32 v1, 1, v0
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: v_lshrrev_b64 v[0:1], 30, v[0:1]
-; GFX9-NEXT: v_mov_b32_e32 v3, s1
+; GFX9-NEXT: global_load_dword v3, v1, s[0:1]
+; GFX9-NEXT: v_mov_b32_e32 v1, 0
+; GFX9-NEXT: v_lshrrev_b64 v[0:1], 30, v[1:2]
+; GFX9-NEXT: v_mov_b32_e32 v2, s1
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, s0, v0
-; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, v3, v1, vcc
+; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, v2, v1, vcc
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_barrier
-; GFX9-NEXT: global_store_dword v[0:1], v2, off
+; GFX9-NEXT: global_store_dword v[0:1], v3, off
; GFX9-NEXT: s_endpgm
bb:
%tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
@@ -369,10 +369,9 @@ define amdgpu_kernel void @barrier_vmcnt_vscnt_flat_workgroup(ptr %arg) {
; GFX8-NEXT: flat_load_dword v3, v[2:3]
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 1, v0
; GFX8-NEXT: v_lshrrev_b64 v[0:1], 30, v[1:2]
-; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v0, vcc, s0, v0
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, v4, v1, vcc
-; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_barrier
; GFX8-NEXT: flat_store_dword v[0:1], v3
; GFX8-NEXT: s_endpgm
@@ -393,10 +392,9 @@ define amdgpu_kernel void @barrier_vmcnt_vscnt_flat_workgroup(ptr %arg) {
; GFX9-NEXT: flat_load_dword v3, v[2:3]
; GFX9-NEXT: v_add_u32_e32 v2, 1, v0
; GFX9-NEXT: v_lshrrev_b64 v[0:1], 30, v[1:2]
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_add_co_u32_e32 v0, vcc, s0, v0
; GFX9-NEXT: v_addc_co_u32_e32 v1, vcc, v4, v1, vcc
-; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_barrier
; GFX9-NEXT: flat_store_dword v[0:1], v3
; GFX9-NEXT: s_endpgm
>From 9affc3aaa38ed2bfa25531fecaea51f8ff87a0fa Mon Sep 17 00:00:00 2001
From: Carl Ritson <carl.ritson at amd.com>
Date: Thu, 16 Oct 2025 16:47:51 +0900
Subject: [PATCH 2/2] - Reword \file description
---
llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp | 12 ++++++------
.../test/CodeGen/AMDGPU/schedule-barrier-latency.mir | 1 +
2 files changed, 7 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp b/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
index 088660c0a3e26..9d6a2964a1889 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
@@ -7,13 +7,13 @@
//===----------------------------------------------------------------------===//
//
/// \file This file contains a DAG scheduling mutation to add latency to
-/// barrier edges between ATOMIC_FENCE instructions and preceeding
+/// barrier edges between ATOMIC_FENCE instructions and preceding
/// memory accesses potentially affected by the fence.
-/// This is beneficial when a fence would cause wait count insertion,
-/// as more instructions will be scheduled before the fence hiding
-/// memory latency.
-/// It also reduces the risk of a fence causing a premature wait
-/// on all active memory operations.
+/// This encourages the scheduling of more instructions before
+/// ATOMIC_FENCE instructions. ATOMIC_FENCE instructions may
+/// introduce wait counting or indicate an impending S_BARRIER
+/// wait. Having more instructions in-flight across these
+/// constructs improves latency hiding.
//
//===----------------------------------------------------------------------===//
diff --git a/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir b/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir
index fc736d055a3a5..93f7bcc478737 100644
--- a/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir
+++ b/llvm/test/CodeGen/AMDGPU/schedule-barrier-latency.mir
@@ -2,6 +2,7 @@
# RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -passes=postmisched -o - %s | FileCheck %s
# Ensure WMMA operations stay before the final atomic fence and barrier group.
+# This allows the latency of the WMMA operations to be hidden by barrier wait.
---
name: test
tracksRegLiveness: true
More information about the llvm-commits
mailing list