[llvm-branch-commits] [llvm] [AMDGPU] Add MemoryPipeline scheduling to Coexec sched (PR #192325)

Jeffrey Byrnes via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Fri Apr 24 09:56:13 PDT 2026


https://github.com/jrbyrnes updated https://github.com/llvm/llvm-project/pull/192325

>From 6edc82a25c6ca54cc4b20c0cec42cce4d614e6f4 Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Fri, 20 Mar 2026 14:58:27 -0700
Subject: [PATCH 1/7] [AMDGPU] Add MemoryPipeline scheduling to Coexec sched

Change-Id: I52c476834155823d1ba998cdbbcb3ad6a7e6f2f5
---
 .../AMDGPU/AMDGPUCoExecSchedStrategy.cpp      | 100 ++++--
 .../Target/AMDGPU/AMDGPUCoExecSchedStrategy.h |  18 +
 .../AMDGPU/coexec-sched-effective-stall.mir   | 323 ++++++++++++++++++
 3 files changed, 418 insertions(+), 23 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
index 71d016dfea6ae..0292deee4ae9c 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
@@ -52,7 +52,7 @@ InstructionFlavor llvm::AMDGPU::classifyFlavor(const MachineInstr &MI,
   // Check for specific opcodes first.
   if (Opc == AMDGPU::ATOMIC_FENCE || Opc == AMDGPU::S_WAIT_ASYNCCNT ||
       Opc == AMDGPU::S_WAIT_TENSORCNT || Opc == AMDGPU::S_BARRIER_WAIT ||
-      Opc == AMDGPU::S_BARRIER_SIGNAL_IMM)
+      Opc == AMDGPU::S_BARRIER_SIGNAL_IMM || SII.isWaitcnt(Opc))
     return InstructionFlavor::Fence;
 
   if (SII.isLDSDMA(MI))
@@ -409,19 +409,8 @@ unsigned CandidateHeuristics::getStructuralStallCycles(SchedBoundary &Zone,
 }
 
 bool CandidateHeuristics::tryEffectiveStall(
-    GenericSchedulerBase::SchedCandidate &Cand,
-    GenericSchedulerBase::SchedCandidate &TryCand, SchedBoundary &Zone) {
-  // Treat structural and latency stalls as a single scheduling cost for the
-  // current cycle.
-  struct StallCosts {
-    unsigned Ready = 0;
-    unsigned Structural = 0;
-    unsigned Latency = 0;
-    unsigned Effective = 0;
-    unsigned Carried = 0;
-    unsigned Buffer = 0;
-  };
-
+    GenericSchedulerBase::SchedCandidate &TryCand,
+    GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary &Zone) {
   auto getBufferFullStalls = [this, &Zone](SUnit *SU) -> unsigned {
     InstructionFlavor Flavor = classifyFlavor(
         *SU->getInstr(), *static_cast<const SIInstrInfo *>(DAG->TII));
@@ -442,6 +431,43 @@ bool CandidateHeuristics::tryEffectiveStall(
   };
 
   unsigned CurrCycle = Zone.getCurrCycle();
+
+  auto getFenceStalls = [this, &CurrCycle](SUnit *SU) -> unsigned {
+    InstructionFlavor Flavor = classifyFlavor(
+        *SU->getInstr(), *static_cast<const SIInstrInfo *>(DAG->TII));
+
+    if (Flavor != InstructionFlavor::Fence)
+      return 0;
+
+    HardwareUnitInfo *FenceHWUI = getHWUIFromFlavor(Flavor);
+    HardwareUnitInfo *DSHWUI = getHWUIFromFlavor(InstructionFlavor::DS);
+
+    SUnit *LastDS = DSHWUI->getLastScheduledSU();
+    if (!LastDS)
+      return 0;
+
+    SUnit *LastFence = FenceHWUI->getLastScheduledSU();
+    unsigned LastFenceCycle = LastFence ? LastFence->TopReadyCycle : 0;
+    unsigned LastDSCycle = LastDS->TopReadyCycle;
+
+    if (LastDSCycle < LastFenceCycle)
+      return 0;
+
+    unsigned LastDSFinish = LastDSCycle + getHWUICyclesForSU(LastDS);
+    return LastDSFinish <= CurrCycle ? 0 : LastDSFinish - CurrCycle;
+  };
+
+  // Treat stalls as a single scheduling cost for the current cycle.
+  struct StallCosts {
+    unsigned Ready = 0;
+    unsigned Structural = 0;
+    unsigned Latency = 0;
+    unsigned Carried = 0;
+    unsigned Buffer = 0;
+    unsigned Fence = 0;
+    unsigned Effective = 0;
+  };
+
   auto GetStallCosts = [&](SUnit *SU) {
     unsigned ReadyCycle = Zone.isTop() ? SU->TopReadyCycle : SU->BotReadyCycle;
     StallCosts Costs;
@@ -451,9 +477,9 @@ bool CandidateHeuristics::tryEffectiveStall(
     unsigned CarriedLatency = CarriedLatencies.lookup_or(SU->getInstr(), 0);
     Costs.Carried = CarriedLatency > CurrCycle ? CarriedLatency - CurrCycle : 0;
     Costs.Buffer = getBufferFullStalls(SU);
-
+    Costs.Fence = getFenceStalls(SU);
     Costs.Effective = std::max({Costs.Ready, Costs.Structural, Costs.Latency,
-                                Costs.Carried, Costs.Buffer});
+                                Costs.Carried, Costs.Buffer, Costs.Fence});
     return Costs;
   };
 
@@ -475,6 +501,33 @@ bool CandidateHeuristics::tryEffectiveStall(
                  AMDGPUCoExecSchedStrategy::Stall);
 }
 
+bool CandidateHeuristics::tryMemoryPipeline(
+    GenericSchedulerBase::SchedCandidate &TryCand,
+    GenericSchedulerBase::SchedCandidate &Cand) {
+
+  InstructionFlavor TryFlavor = classifyFlavor(*TryCand.SU->getInstr(), *SII);
+
+  InstructionFlavor CandFlavor = classifyFlavor(*Cand.SU->getInstr(), *SII);
+
+  bool TryIsMemoryPipeline = TryFlavor == InstructionFlavor::DMA ||
+                             TryFlavor == InstructionFlavor::Fence;
+  bool CandIsMemoryPipeline = CandFlavor == InstructionFlavor::DMA ||
+                              CandFlavor == InstructionFlavor::Fence;
+
+  if (TryIsMemoryPipeline == CandIsMemoryPipeline)
+    return false;
+
+  if (CandIsMemoryPipeline) {
+    if (Cand.Reason > GenericSchedulerBase::RegCritical)
+      Cand.Reason = GenericSchedulerBase::RegCritical;
+
+    return true;
+  }
+
+  TryCand.Reason = GenericSchedulerBase::RegCritical;
+  return true;
+}
+
 bool CandidateHeuristics::tryCriticalResourceDependency(
     GenericSchedulerBase::SchedCandidate &TryCand,
     GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary *Zone) const {
@@ -806,8 +859,15 @@ bool AMDGPUCoExecSchedStrategy::tryCandidateCoexec(SchedCandidate &Cand,
   if (SameBoundary) {
     // Compare candidates by the stall they would introduce if
     // scheduled in the current cycle.
-    if (Heurs.tryEffectiveStall(Cand, TryCand, *Zone))
+    if (Heurs.tryEffectiveStall(TryCand, Cand, *Zone)) {
+      LastAMDGPUReason = AMDGPUSchedReason::Stall;
       return TryCand.Reason != NoCand;
+    }
+
+    if (Heurs.tryMemoryPipeline(TryCand, Cand)) {
+      LastAMDGPUReason = AMDGPUSchedReason::MemoryPipeline;
+      return TryCand.Reason != NoCand;
+    }
 
     Heurs.sortHWUIResources();
     if (Heurs.tryCriticalResource(TryCand, Cand, Zone)) {
@@ -852,12 +912,6 @@ bool AMDGPUCoExecSchedStrategy::tryCandidateCoexec(SchedCandidate &Cand,
     return TryCand.Reason != NoCand;
 
   if (SameBoundary) {
-    // Avoid serializing long latency dependence chains.
-    // For acyclic path limited loops, latency was already checked above.
-    if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency &&
-        !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, *Zone))
-      return TryCand.Reason != NoCand;
-
     // Fall through to original instruction order.
     if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) ||
         (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h
index 177f77dde7562..4b1568c674300 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h
@@ -127,6 +127,8 @@ inline FlavorGroup all() {
 /// than the generic CandReason enum for debugging purposes.
 enum class AMDGPUSchedReason : uint8_t {
   None,
+  Stall,
+  MemoryPipeline,
   CritResourceBalance, // tryCriticalResource chose based on resource pressure
   CritResourceDep,     // tryCriticalResourceDependency chose based on enabling
   NUM_REASONS
@@ -136,6 +138,10 @@ constexpr StringRef getReasonName(AMDGPUSchedReason R) {
   switch (R) {
   case AMDGPUSchedReason::None:
     return "None";
+  case AMDGPUSchedReason::Stall:
+    return "Stall";
+  case AMDGPUSchedReason::MemoryPipeline:
+    return "MemoryPipeline";
   case AMDGPUSchedReason::CritResourceBalance:
     return "CritResource";
   case AMDGPUSchedReason::CritResourceDep:
@@ -236,6 +242,15 @@ class HardwareUnitInfo {
            ScheduledSUs[ScheduledSUs.size() - BufferSize]->TopReadyCycle;
   }
 
+  /// \returns the most recently scheduled SU for this HardwareUnit.
+  SUnit *getLastScheduledSU() {
+    unsigned ScheduledCount = ScheduledSUs.size();
+    if (!ScheduledCount)
+      return nullptr;
+
+    return ScheduledSUs[ScheduledCount - 1];
+  }
+
   /// \returns the SUnit with higher priority or nullptr if they are the same.
   /// This method looks through the PrioritySUs to determine if one SU is more
   /// prioritized than the other. If neither are in the PrioritySUs list, then
@@ -349,6 +364,9 @@ class CandidateHeuristics {
                          GenericSchedulerBase::SchedCandidate &Cand,
                          SchedBoundary &Zone);
 
+  bool tryMemoryPipeline(GenericSchedulerBase::SchedCandidate &TryCand,
+                         GenericSchedulerBase::SchedCandidate &Cand);
+
   /// Check for critical resource consumption. Prefer the candidate that uses
   /// the most prioritized HardwareUnit. If both candidates use the same
   /// HarwareUnit, prefer the candidate with higher priority on that
diff --git a/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir b/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir
index 0a6f2fe9375d5..7868c7dcf88a3 100644
--- a/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir
+++ b/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir
@@ -5,6 +5,9 @@
 --- |
   define void @test-sched-effective-stall() #0 { ret void }
   define void @test-sched-pending-structural-stall() #0 { ret void }
+  define void @test-fence-stall() #0 { ret void }
+  define void @test-tensorcnt-stall() #0 { ret void }
+  define void @test-dscnt-stall() #0 { ret void }
 
   attributes #0 = { "amdgpu-waves-per-eu"="1,1" }
 ...
@@ -121,3 +124,323 @@ body: |
     S_NOP 0
     S_ENDPGM 0, implicit %10, implicit %11
 ...
+
+
+---
+name: test-fence-stall
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; DEFAULT-LABEL: name: test-fence-stall
+    ; DEFAULT: [[DEF:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF1:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF2:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF3:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 64, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 128, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 192, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 256, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 320, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 384, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 448, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 512, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 576, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 640, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 704, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 768, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 832, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 896, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 960, 0, implicit $exec
+    ; DEFAULT-NEXT: dead [[DEF5:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead [[DEF6:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF7:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF8]], [[DEF7]].sub1, implicit-def dead $scc
+    ; DEFAULT-NEXT: ATOMIC_FENCE 4, 2
+    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: S_ENDPGM 0, amdgpu_allvgprs
+    ;
+    ; COEXEC-LABEL: name: test-fence-stall
+    ; COEXEC: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 64, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 128, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 192, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 256, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 320, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 384, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 448, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 512, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 576, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 640, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 704, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 768, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 832, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 896, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 960, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DEF1:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: dead [[DEF2:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF3:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF5:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF6:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF7:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF8:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
+    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF4]], [[DEF3]].sub1, implicit-def dead $scc
+    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: ATOMIC_FENCE 4, 2
+    ; COEXEC-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_ENDPGM 0, amdgpu_allvgprs
+    %95:vgpr_32 = IMPLICIT_DEF
+    %90:vreg_256_align2 = IMPLICIT_DEF
+    %91:vreg_256_align2 = IMPLICIT_DEF
+    %85:sreg_64_xexec = IMPLICIT_DEF
+    %249:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %2373:vreg_512_align2 = IMPLICIT_DEF
+    %2369:vreg_512_align2 = IMPLICIT_DEF
+    %2353:vreg_128_lo256_align2 = IMPLICIT_DEF
+    %8:vgpr_32_lo256 = IMPLICIT_DEF
+    undef %134.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 0, 0, implicit $exec
+    %134.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 64, 0, implicit $exec
+    undef %143.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 128, 0, implicit $exec
+    %143.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 192, 0, implicit $exec
+    undef %152.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 256, 0, implicit $exec
+    %152.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 320, 0, implicit $exec
+    undef %161.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 384, 0, implicit $exec
+    %161.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 448, 0, implicit $exec
+    undef %170.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 512, 0, implicit $exec
+    %170.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 576, 0, implicit $exec
+    undef %179.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 640, 0, implicit $exec
+    %179.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 704, 0, implicit $exec
+    undef %188.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 768, 0, implicit $exec
+    %188.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 832, 0, implicit $exec
+    %249:sreg_32_xm0_xexec = S_ADD_I32 %249:sreg_32_xm0_xexec, %85.sub1:sreg_64_xexec, implicit-def dead $scc
+    undef %197.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 896, 0, implicit $exec
+    %197.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 960, 0, implicit $exec
+    ATOMIC_FENCE 4, 2
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    S_ENDPGM 0, amdgpu_allvgprs
+...
+
+
+---
+name: test-tensorcnt-stall
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; DEFAULT-LABEL: name: test-tensorcnt-stall
+    ; DEFAULT: [[DEF:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF1:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF2:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF3:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 64, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 128, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 192, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 256, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 320, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 384, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 448, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 512, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 576, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 640, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 704, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 768, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 832, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 896, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 960, 0, implicit $exec
+    ; DEFAULT-NEXT: dead [[DEF5:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead [[DEF6:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF7:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF8]], [[DEF7]].sub1, implicit-def dead $scc
+    ; DEFAULT-NEXT: S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: S_ENDPGM 0, amdgpu_allvgprs
+    ;
+    ; COEXEC-LABEL: name: test-tensorcnt-stall
+    ; COEXEC: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 64, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 128, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 192, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 256, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 320, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 384, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 448, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 512, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 576, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 640, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 704, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 768, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 832, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 896, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 960, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DEF1:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: dead [[DEF2:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF3:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF5:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF6:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF7:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF8:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
+    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF4]], [[DEF3]].sub1, implicit-def dead $scc
+    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    ; COEXEC-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_ENDPGM 0, amdgpu_allvgprs
+    %95:vgpr_32 = IMPLICIT_DEF
+    %90:vreg_256_align2 = IMPLICIT_DEF
+    %91:vreg_256_align2 = IMPLICIT_DEF
+    %85:sreg_64_xexec = IMPLICIT_DEF
+    %249:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %2373:vreg_512_align2 = IMPLICIT_DEF
+    %2369:vreg_512_align2 = IMPLICIT_DEF
+    %2353:vreg_128_lo256_align2 = IMPLICIT_DEF
+    %8:vgpr_32_lo256 = IMPLICIT_DEF
+    undef %134.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 0, 0, implicit $exec
+    %134.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 64, 0, implicit $exec
+    undef %143.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 128, 0, implicit $exec
+    %143.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 192, 0, implicit $exec
+    undef %152.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 256, 0, implicit $exec
+    %152.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 320, 0, implicit $exec
+    undef %161.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 384, 0, implicit $exec
+    %161.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 448, 0, implicit $exec
+    undef %170.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 512, 0, implicit $exec
+    %170.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 576, 0, implicit $exec
+    undef %179.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 640, 0, implicit $exec
+    %179.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 704, 0, implicit $exec
+    undef %188.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 768, 0, implicit $exec
+    %188.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 832, 0, implicit $exec
+    %249:sreg_32_xm0_xexec = S_ADD_I32 %249:sreg_32_xm0_xexec, %85.sub1:sreg_64_xexec, implicit-def dead $scc
+    undef %197.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 896, 0, implicit $exec
+    %197.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 960, 0, implicit $exec
+    S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    S_ENDPGM 0, amdgpu_allvgprs
+...
+
+---
+name: test-dscnt-stall
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; DEFAULT-LABEL: name: test-dscnt-stall
+    ; DEFAULT: [[DEF:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF1:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF2:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF3:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 64, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 128, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 192, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 256, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 320, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 384, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 448, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 512, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 576, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 640, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 704, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 768, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 832, 0, implicit $exec
+    ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 896, 0, implicit $exec
+    ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 960, 0, implicit $exec
+    ; DEFAULT-NEXT: dead [[DEF5:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead [[DEF6:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF7:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; DEFAULT-NEXT: [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; DEFAULT-NEXT: dead [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF8]], [[DEF7]].sub1, implicit-def dead $scc
+    ; DEFAULT-NEXT: S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: S_ENDPGM 0, amdgpu_allvgprs
+    ;
+    ; COEXEC-LABEL: name: test-dscnt-stall
+    ; COEXEC: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 64, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 128, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 192, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 256, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 320, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 384, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 448, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 512, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_4:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 576, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 640, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_5:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 704, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 768, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_6:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 832, 0, implicit $exec
+    ; COEXEC-NEXT: undef [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 896, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DS_LOAD_TR16_B128_7:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF]], 960, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DEF1:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: dead [[DEF2:%[0-9]+]]:vreg_256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF3:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF5:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF6:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF7:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
+    ; COEXEC-NEXT: [[DEF8:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
+    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF4]], [[DEF3]].sub1, implicit-def dead $scc
+    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    ; COEXEC-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_ENDPGM 0, amdgpu_allvgprs
+    %95:vgpr_32 = IMPLICIT_DEF
+    %90:vreg_256_align2 = IMPLICIT_DEF
+    %91:vreg_256_align2 = IMPLICIT_DEF
+    %85:sreg_64_xexec = IMPLICIT_DEF
+    %249:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %2373:vreg_512_align2 = IMPLICIT_DEF
+    %2369:vreg_512_align2 = IMPLICIT_DEF
+    %2353:vreg_128_lo256_align2 = IMPLICIT_DEF
+    %8:vgpr_32_lo256 = IMPLICIT_DEF
+    undef %134.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 0, 0, implicit $exec
+    %134.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 64, 0, implicit $exec
+    undef %143.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 128, 0, implicit $exec
+    %143.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 192, 0, implicit $exec
+    undef %152.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 256, 0, implicit $exec
+    %152.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 320, 0, implicit $exec
+    undef %161.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 384, 0, implicit $exec
+    %161.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 448, 0, implicit $exec
+    undef %170.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 512, 0, implicit $exec
+    %170.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 576, 0, implicit $exec
+    undef %179.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 640, 0, implicit $exec
+    %179.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 704, 0, implicit $exec
+    undef %188.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 768, 0, implicit $exec
+    %188.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 832, 0, implicit $exec
+    %249:sreg_32_xm0_xexec = S_ADD_I32 %249:sreg_32_xm0_xexec, %85.sub1:sreg_64_xexec, implicit-def dead $scc
+    undef %197.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 896, 0, implicit $exec
+    %197.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 960, 0, implicit $exec
+    S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    S_ENDPGM 0, amdgpu_allvgprs
+...

>From 8203e7624d0df463aa76282e9ab1418e31e39fde Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Wed, 25 Mar 2026 18:22:24 -0700
Subject: [PATCH 2/7] Add a comment

Change-Id: I447f7f1fb185b18924cfd98249b5a0a05fef2484
---
 llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h
index 4b1568c674300..6e056a9554afa 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.h
@@ -364,6 +364,13 @@ class CandidateHeuristics {
                          GenericSchedulerBase::SchedCandidate &Cand,
                          SchedBoundary &Zone);
 
+  /// Prioritize instructions involved the memory pipeline. Currently we don't have
+  /// any modelling of pipelined loads, so we control the layout of the pipeline
+  /// per iteration by giving the user some control over the stalls (e.g. between
+  /// s_barrier_signal and s_barrier_wait) and scheduling the pipeline instructions
+  /// as soon as they are ready.
+  ///
+  /// TODO -- add better modelling and heuristics for pipelining based scheduling.
   bool tryMemoryPipeline(GenericSchedulerBase::SchedCandidate &TryCand,
                          GenericSchedulerBase::SchedCandidate &Cand);
 

>From 2bfdfe27b509bfde351dea64571971d919f8a06c Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Tue, 7 Apr 2026 16:04:56 -0700
Subject: [PATCH 3/7] Make fence heuristic work bottom-up

Change-Id: I629cbc8905b87a962e8b123287e5f60a3154df6b
---
 .../AMDGPU/AMDGPUCoExecSchedStrategy.cpp      | 36 ++++++++++---------
 1 file changed, 19 insertions(+), 17 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
index 0292deee4ae9c..1db3f6201d1b0 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
@@ -375,8 +375,7 @@ void CandidateHeuristics::sortHWUIResources() {
 
 unsigned CandidateHeuristics::getStructuralStallCycles(SchedBoundary &Zone,
                                                        SUnit *SU) {
-  // Only implemented for top-down scheduling currently.
-  if (!Zone.isTop() || !SU)
+  if (!SU)
     return 0;
 
   MachineInstr *MI = SU->getInstr();
@@ -411,6 +410,10 @@ unsigned CandidateHeuristics::getStructuralStallCycles(SchedBoundary &Zone,
 bool CandidateHeuristics::tryEffectiveStall(
     GenericSchedulerBase::SchedCandidate &TryCand,
     GenericSchedulerBase::SchedCandidate &Cand, SchedBoundary &Zone) {
+  // Only implemented for top-down scheduling
+  if (!Zone.isTop())
+    return 0;
+
   auto getBufferFullStalls = [this, &Zone](SUnit *SU) -> unsigned {
     InstructionFlavor Flavor = classifyFlavor(
         *SU->getInstr(), *static_cast<const SIInstrInfo *>(DAG->TII));
@@ -420,8 +423,6 @@ bool CandidateHeuristics::tryEffectiveStall(
     if (HWUI->getBufferSize() == 0)
       return 0;
 
-    // getBufferAvailableCycle assumes top-down scheduling.
-    assert(Zone.isTop());
     unsigned CurrCycle = Zone.getCurrCycle();
     unsigned BufferReadyCycle = HWUI->getBufferAvailableCycle(CurrCycle);
     if (BufferReadyCycle <= CurrCycle)
@@ -432,29 +433,30 @@ bool CandidateHeuristics::tryEffectiveStall(
 
   unsigned CurrCycle = Zone.getCurrCycle();
 
-  auto getFenceStalls = [this, &CurrCycle](SUnit *SU) -> unsigned {
+  auto getFenceStalls = [this, &CurrCycle, &Zone](SUnit *SU) -> unsigned {
     InstructionFlavor Flavor = classifyFlavor(
         *SU->getInstr(), *static_cast<const SIInstrInfo *>(DAG->TII));
 
-    if (Flavor != InstructionFlavor::Fence)
+    bool IsTop = Zone.isTop();
+    if ((Flavor != InstructionFlavor::Fence && IsTop) || (Flavor != InstructionFlavor::DS && !IsTop))
       return 0;
 
-    HardwareUnitInfo *FenceHWUI = getHWUIFromFlavor(Flavor);
-    HardwareUnitInfo *DSHWUI = getHWUIFromFlavor(InstructionFlavor::DS);
+    HardwareUnitInfo *ConsumerHWUI = getHWUIFromFlavor(Flavor);
+    HardwareUnitInfo *ProducerHWUI = getHWUIFromFlavor(IsTop ? InstructionFlavor::DS : InstructionFlavor::Fence);
 
-    SUnit *LastDS = DSHWUI->getLastScheduledSU();
-    if (!LastDS)
+    SUnit *LastProducer = ProducerHWUI->getLastScheduledSU();
+    if (!LastProducer)
       return 0;
 
-    SUnit *LastFence = FenceHWUI->getLastScheduledSU();
-    unsigned LastFenceCycle = LastFence ? LastFence->TopReadyCycle : 0;
-    unsigned LastDSCycle = LastDS->TopReadyCycle;
+    SUnit *LastConsumer = ConsumerHWUI->getLastScheduledSU();
+    unsigned LastConsumerCycle = LastConsumer ? LastConsumer->TopReadyCycle : 0;
+    unsigned LastProducerCycle = LastProducer->TopReadyCycle;
 
-    if (LastDSCycle < LastFenceCycle)
+    if (LastProducerCycle < LastConsumerCycle)
       return 0;
 
-    unsigned LastDSFinish = LastDSCycle + getHWUICyclesForSU(LastDS);
-    return LastDSFinish <= CurrCycle ? 0 : LastDSFinish - CurrCycle;
+    unsigned FenceStallFinish = LastProducerCycle + getHWUICyclesForSU(IsTop ? LastProducer : SU);
+    return FenceStallFinish <= CurrCycle ? 0 : FenceStallFinish - CurrCycle;
   };
 
   // Treat stalls as a single scheduling cost for the current cycle.
@@ -911,7 +913,7 @@ bool AMDGPUCoExecSchedStrategy::tryCandidateCoexec(SchedCandidate &Cand,
                   Cand, RegMax, TRI, DAG->MF))
     return TryCand.Reason != NoCand;
 
-  if (SameBoundary) {
+  if (SameBoundary) {  
     // Fall through to original instruction order.
     if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) ||
         (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {

>From 466f439973fe56e3ed34e14ca73dd5f809e3e1af Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Tue, 7 Apr 2026 16:08:36 -0700
Subject: [PATCH 4/7] Add back tryLatency

Change-Id: I12d4f255c48ed77ba927eb3b192e5903f1f5e24f
---
 llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
index 1db3f6201d1b0..b4751027960b4 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
@@ -913,7 +913,13 @@ bool AMDGPUCoExecSchedStrategy::tryCandidateCoexec(SchedCandidate &Cand,
                   Cand, RegMax, TRI, DAG->MF))
     return TryCand.Reason != NoCand;
 
-  if (SameBoundary) {  
+  if (SameBoundary) {
+    // Avoid serializing long latency dependence chains.
+    // For acyclic path limited loops, latency was already checked above.
+    if (!RegionPolicy.DisableLatencyHeuristic && TryCand.Policy.ReduceLatency &&
+        !Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, *Zone))
+      return TryCand.Reason != NoCand;
+
     // Fall through to original instruction order.
     if ((Zone->isTop() && TryCand.SU->NodeNum < Cand.SU->NodeNum) ||
         (!Zone->isTop() && TryCand.SU->NodeNum > Cand.SU->NodeNum)) {

>From 64b74c414ba4e7675b5fe18aec622b330d217129 Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Tue, 7 Apr 2026 16:09:51 -0700
Subject: [PATCH 5/7] clang-format

Change-Id: I534b1a979f55339a814ef3416c2492252845add5
---
 llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
index b4751027960b4..e95ee4d0d7092 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
@@ -438,11 +438,13 @@ bool CandidateHeuristics::tryEffectiveStall(
         *SU->getInstr(), *static_cast<const SIInstrInfo *>(DAG->TII));
 
     bool IsTop = Zone.isTop();
-    if ((Flavor != InstructionFlavor::Fence && IsTop) || (Flavor != InstructionFlavor::DS && !IsTop))
+    if ((Flavor != InstructionFlavor::Fence && IsTop) ||
+        (Flavor != InstructionFlavor::DS && !IsTop))
       return 0;
 
     HardwareUnitInfo *ConsumerHWUI = getHWUIFromFlavor(Flavor);
-    HardwareUnitInfo *ProducerHWUI = getHWUIFromFlavor(IsTop ? InstructionFlavor::DS : InstructionFlavor::Fence);
+    HardwareUnitInfo *ProducerHWUI = getHWUIFromFlavor(
+        IsTop ? InstructionFlavor::DS : InstructionFlavor::Fence);
 
     SUnit *LastProducer = ProducerHWUI->getLastScheduledSU();
     if (!LastProducer)
@@ -455,7 +457,8 @@ bool CandidateHeuristics::tryEffectiveStall(
     if (LastProducerCycle < LastConsumerCycle)
       return 0;
 
-    unsigned FenceStallFinish = LastProducerCycle + getHWUICyclesForSU(IsTop ? LastProducer : SU);
+    unsigned FenceStallFinish =
+        LastProducerCycle + getHWUICyclesForSU(IsTop ? LastProducer : SU);
     return FenceStallFinish <= CurrCycle ? 0 : FenceStallFinish - CurrCycle;
   };
 

>From 6744fde8019a9e47f28d5e51ca442482fc71983e Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Tue, 7 Apr 2026 16:11:50 -0700
Subject: [PATCH 6/7] Add comment

Change-Id: I2180bba631fe4a01ed3c3fbcfa8c19cbefa84133
---
 llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
index e95ee4d0d7092..eb259e2d1a09d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
@@ -457,6 +457,7 @@ bool CandidateHeuristics::tryEffectiveStall(
     if (LastProducerCycle < LastConsumerCycle)
       return 0;
 
+    // Latency comes from DS regardless of bottom-up / top-down.
     unsigned FenceStallFinish =
         LastProducerCycle + getHWUICyclesForSU(IsTop ? LastProducer : SU);
     return FenceStallFinish <= CurrCycle ? 0 : FenceStallFinish - CurrCycle;

>From 3e26f22711c122ee2883c2c981763f7b9c22b639 Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Wed, 15 Apr 2026 16:14:27 -0700
Subject: [PATCH 7/7] Adrress comments from
 https://github.com/llvm/llvm-project/pull/188658

Change-Id: Ia94c567a753168c1ffa16dc5d91195e7dd0ba044
---
 .../AMDGPU/AMDGPUCoExecSchedStrategy.cpp      |   6 +-
 .../AMDGPU/coexec-sched-effective-stall.mir   | 222 +++++++++---------
 2 files changed, 114 insertions(+), 114 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
index eb259e2d1a09d..d710d40a7ea98 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCoExecSchedStrategy.cpp
@@ -680,9 +680,9 @@ void AMDGPUCoExecSchedStrategy::initPolicy(MachineBasicBlock::iterator Begin,
                                            MachineBasicBlock::iterator End,
                                            unsigned NumRegionInstrs) {
   GCNSchedStrategy::initPolicy(Begin, End, NumRegionInstrs);
-  assert((PreRADirection == MISched::Unspecified ||
-          PreRADirection == MISched::TopDown) &&
-         "coexec scheduler only supports top-down scheduling");
+  if (PreRADirection == MISched::BottomUp ||
+      PreRADirection == MISched::Bidirectional)
+    report_fatal_error("CoExecSchedStrategy only support TopDown scheduling.");
   RegionPolicy.OnlyTopDown = true;
   RegionPolicy.OnlyBottomUp = false;
 }
diff --git a/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir b/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir
index 7868c7dcf88a3..772eb91d44c59 100644
--- a/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir
+++ b/llvm/test/CodeGen/AMDGPU/coexec-sched-effective-stall.mir
@@ -136,14 +136,14 @@ body: |
     ; DEFAULT-NEXT: [[DEF1:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
     ; DEFAULT-NEXT: [[DEF2:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
     ; DEFAULT-NEXT: [[DEF3:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
-    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 0, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 64, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 128, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 192, 0, implicit $exec
-    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 256, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 320, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 384, 0, implicit $exec
@@ -162,7 +162,7 @@ body: |
     ; DEFAULT-NEXT: [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
     ; DEFAULT-NEXT: dead [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF8]], [[DEF7]].sub1, implicit-def dead $scc
     ; DEFAULT-NEXT: ATOMIC_FENCE 4, 2
-    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: S_ENDPGM 0, amdgpu_allvgprs
     ;
     ; COEXEC-LABEL: name: test-fence-stall
@@ -191,44 +191,44 @@ body: |
     ; COEXEC-NEXT: [[DEF6:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
     ; COEXEC-NEXT: [[DEF7:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
     ; COEXEC-NEXT: [[DEF8:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
-    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; COEXEC-NEXT: dead [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF4]], [[DEF3]].sub1, implicit-def dead $scc
-    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; COEXEC-NEXT: ATOMIC_FENCE 4, 2
     ; COEXEC-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: ATOMIC_FENCE 4, 2
+    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; COEXEC-NEXT: S_ENDPGM 0, amdgpu_allvgprs
-    %95:vgpr_32 = IMPLICIT_DEF
-    %90:vreg_256_align2 = IMPLICIT_DEF
-    %91:vreg_256_align2 = IMPLICIT_DEF
-    %85:sreg_64_xexec = IMPLICIT_DEF
-    %249:sreg_32_xm0_xexec = IMPLICIT_DEF
-    %2373:vreg_512_align2 = IMPLICIT_DEF
-    %2369:vreg_512_align2 = IMPLICIT_DEF
-    %2353:vreg_128_lo256_align2 = IMPLICIT_DEF
+    %0:vgpr_32 = IMPLICIT_DEF
+    %1:vreg_256_align2 = IMPLICIT_DEF
+    %2:vreg_256_align2 = IMPLICIT_DEF
+    %3:sreg_64_xexec = IMPLICIT_DEF
+    %4:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %5:vreg_512_align2 = IMPLICIT_DEF
+    %6:vreg_512_align2 = IMPLICIT_DEF
+    %7:vreg_128_lo256_align2 = IMPLICIT_DEF
     %8:vgpr_32_lo256 = IMPLICIT_DEF
-    undef %134.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 0, 0, implicit $exec
-    %134.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 64, 0, implicit $exec
-    undef %143.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 128, 0, implicit $exec
-    %143.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 192, 0, implicit $exec
-    undef %152.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 256, 0, implicit $exec
-    %152.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 320, 0, implicit $exec
-    undef %161.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 384, 0, implicit $exec
-    %161.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 448, 0, implicit $exec
-    undef %170.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 512, 0, implicit $exec
-    %170.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 576, 0, implicit $exec
-    undef %179.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 640, 0, implicit $exec
-    %179.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 704, 0, implicit $exec
-    undef %188.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 768, 0, implicit $exec
-    %188.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 832, 0, implicit $exec
-    %249:sreg_32_xm0_xexec = S_ADD_I32 %249:sreg_32_xm0_xexec, %85.sub1:sreg_64_xexec, implicit-def dead $scc
-    undef %197.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 896, 0, implicit $exec
-    %197.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 960, 0, implicit $exec
+    undef %9.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 0, 0, implicit $exec
+    %9.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 64, 0, implicit $exec
+    undef %10.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 128, 0, implicit $exec
+    %10.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 192, 0, implicit $exec
+    undef %11.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 256, 0, implicit $exec
+    %11.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 320, 0, implicit $exec
+    undef %12.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 384, 0, implicit $exec
+    %12.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 448, 0, implicit $exec
+    undef %13.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 512, 0, implicit $exec
+    %13.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 576, 0, implicit $exec
+    undef %14.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 640, 0, implicit $exec
+    %14.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 704, 0, implicit $exec
+    undef %15.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 768, 0, implicit $exec
+    %15.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 832, 0, implicit $exec
+    %4:sreg_32_xm0_xexec = S_ADD_I32 %4:sreg_32_xm0_xexec, %3.sub1:sreg_64_xexec, implicit-def dead $scc
+    undef %16.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 896, 0, implicit $exec
+    %16.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 960, 0, implicit $exec
     ATOMIC_FENCE 4, 2
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     S_ENDPGM 0, amdgpu_allvgprs
 ...
 
@@ -243,14 +243,14 @@ body: |
     ; DEFAULT-NEXT: [[DEF1:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
     ; DEFAULT-NEXT: [[DEF2:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
     ; DEFAULT-NEXT: [[DEF3:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
-    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 0, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 64, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 128, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 192, 0, implicit $exec
-    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 256, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 320, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 384, 0, implicit $exec
@@ -269,7 +269,7 @@ body: |
     ; DEFAULT-NEXT: [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
     ; DEFAULT-NEXT: dead [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF8]], [[DEF7]].sub1, implicit-def dead $scc
     ; DEFAULT-NEXT: S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
-    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: S_ENDPGM 0, amdgpu_allvgprs
     ;
     ; COEXEC-LABEL: name: test-tensorcnt-stall
@@ -298,44 +298,44 @@ body: |
     ; COEXEC-NEXT: [[DEF6:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
     ; COEXEC-NEXT: [[DEF7:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
     ; COEXEC-NEXT: [[DEF8:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
-    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; COEXEC-NEXT: dead [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF4]], [[DEF3]].sub1, implicit-def dead $scc
-    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; COEXEC-NEXT: S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
     ; COEXEC-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; COEXEC-NEXT: S_ENDPGM 0, amdgpu_allvgprs
-    %95:vgpr_32 = IMPLICIT_DEF
-    %90:vreg_256_align2 = IMPLICIT_DEF
-    %91:vreg_256_align2 = IMPLICIT_DEF
-    %85:sreg_64_xexec = IMPLICIT_DEF
-    %249:sreg_32_xm0_xexec = IMPLICIT_DEF
-    %2373:vreg_512_align2 = IMPLICIT_DEF
-    %2369:vreg_512_align2 = IMPLICIT_DEF
-    %2353:vreg_128_lo256_align2 = IMPLICIT_DEF
+    %0:vgpr_32 = IMPLICIT_DEF
+    %1:vreg_256_align2 = IMPLICIT_DEF
+    %2:vreg_256_align2 = IMPLICIT_DEF
+    %3:sreg_64_xexec = IMPLICIT_DEF
+    %4:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %5:vreg_512_align2 = IMPLICIT_DEF
+    %6:vreg_512_align2 = IMPLICIT_DEF
+    %7:vreg_128_lo256_align2 = IMPLICIT_DEF
     %8:vgpr_32_lo256 = IMPLICIT_DEF
-    undef %134.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 0, 0, implicit $exec
-    %134.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 64, 0, implicit $exec
-    undef %143.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 128, 0, implicit $exec
-    %143.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 192, 0, implicit $exec
-    undef %152.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 256, 0, implicit $exec
-    %152.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 320, 0, implicit $exec
-    undef %161.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 384, 0, implicit $exec
-    %161.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 448, 0, implicit $exec
-    undef %170.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 512, 0, implicit $exec
-    %170.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 576, 0, implicit $exec
-    undef %179.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 640, 0, implicit $exec
-    %179.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 704, 0, implicit $exec
-    undef %188.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 768, 0, implicit $exec
-    %188.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 832, 0, implicit $exec
-    %249:sreg_32_xm0_xexec = S_ADD_I32 %249:sreg_32_xm0_xexec, %85.sub1:sreg_64_xexec, implicit-def dead $scc
-    undef %197.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 896, 0, implicit $exec
-    %197.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 960, 0, implicit $exec
+    undef %9.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 0, 0, implicit $exec
+    %9.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 64, 0, implicit $exec
+    undef %10.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 128, 0, implicit $exec
+    %10.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 192, 0, implicit $exec
+    undef %11.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 256, 0, implicit $exec
+    %11.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 320, 0, implicit $exec
+    undef %12.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 384, 0, implicit $exec
+    %12.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 448, 0, implicit $exec
+    undef %13.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 512, 0, implicit $exec
+    %13.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 576, 0, implicit $exec
+    undef %14.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 640, 0, implicit $exec
+    %14.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 704, 0, implicit $exec
+    undef %15.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 768, 0, implicit $exec
+    %15.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 832, 0, implicit $exec
+    %4:sreg_32_xm0_xexec = S_ADD_I32 %4:sreg_32_xm0_xexec, %3.sub1:sreg_64_xexec, implicit-def dead $scc
+    undef %16.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 896, 0, implicit $exec
+    %16.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 960, 0, implicit $exec
     S_WAIT_TENSORCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     S_ENDPGM 0, amdgpu_allvgprs
 ...
 
@@ -349,14 +349,14 @@ body: |
     ; DEFAULT-NEXT: [[DEF1:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
     ; DEFAULT-NEXT: [[DEF2:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
     ; DEFAULT-NEXT: [[DEF3:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
-    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 0, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 64, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 128, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_1:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 192, 0, implicit $exec
-    ; DEFAULT-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 256, 0, implicit $exec
     ; DEFAULT-NEXT: dead undef [[DS_LOAD_TR16_B128_2:%[0-9]+]].sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 320, 0, implicit $exec
     ; DEFAULT-NEXT: undef [[DS_LOAD_TR16_B128_3:%[0-9]+]].sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 [[DEF4]], 384, 0, implicit $exec
@@ -375,7 +375,7 @@ body: |
     ; DEFAULT-NEXT: [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = IMPLICIT_DEF
     ; DEFAULT-NEXT: dead [[DEF8:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF8]], [[DEF7]].sub1, implicit-def dead $scc
     ; DEFAULT-NEXT: S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
-    ; DEFAULT-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; DEFAULT-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF]], [[DEF1]], 8, 0, [[DEF2]].sub0, [[DEF3]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; DEFAULT-NEXT: S_ENDPGM 0, amdgpu_allvgprs
     ;
     ; COEXEC-LABEL: name: test-dscnt-stall
@@ -404,43 +404,43 @@ body: |
     ; COEXEC-NEXT: [[DEF6:%[0-9]+]]:vreg_512_align2 = IMPLICIT_DEF
     ; COEXEC-NEXT: [[DEF7:%[0-9]+]]:vreg_128_lo256_align2 = IMPLICIT_DEF
     ; COEXEC-NEXT: [[DEF8:%[0-9]+]]:vgpr_32_lo256 = IMPLICIT_DEF
-    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; COEXEC-NEXT: dead [[DEF4:%[0-9]+]]:sreg_32_xm0_xexec = S_ADD_I32 [[DEF4]], [[DEF3]].sub1, implicit-def dead $scc
-    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; COEXEC-NEXT: S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
     ; COEXEC-NEXT: dead early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    ; COEXEC-NEXT: dead early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
+    ; COEXEC-NEXT: dead early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    ; COEXEC-NEXT: dead early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr [[DEF5]], [[DEF6]], 8, 0, [[DEF7]].sub0, [[DEF8]], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     ; COEXEC-NEXT: S_ENDPGM 0, amdgpu_allvgprs
-    %95:vgpr_32 = IMPLICIT_DEF
-    %90:vreg_256_align2 = IMPLICIT_DEF
-    %91:vreg_256_align2 = IMPLICIT_DEF
-    %85:sreg_64_xexec = IMPLICIT_DEF
-    %249:sreg_32_xm0_xexec = IMPLICIT_DEF
-    %2373:vreg_512_align2 = IMPLICIT_DEF
-    %2369:vreg_512_align2 = IMPLICIT_DEF
-    %2353:vreg_128_lo256_align2 = IMPLICIT_DEF
+    %0:vgpr_32 = IMPLICIT_DEF
+    %1:vreg_256_align2 = IMPLICIT_DEF
+    %2:vreg_256_align2 = IMPLICIT_DEF
+    %3:sreg_64_xexec = IMPLICIT_DEF
+    %4:sreg_32_xm0_xexec = IMPLICIT_DEF
+    %5:vreg_512_align2 = IMPLICIT_DEF
+    %6:vreg_512_align2 = IMPLICIT_DEF
+    %7:vreg_128_lo256_align2 = IMPLICIT_DEF
     %8:vgpr_32_lo256 = IMPLICIT_DEF
-    undef %134.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 0, 0, implicit $exec
-    %134.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 64, 0, implicit $exec
-    undef %143.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 128, 0, implicit $exec
-    %143.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 192, 0, implicit $exec
-    undef %152.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 256, 0, implicit $exec
-    %152.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 320, 0, implicit $exec
-    undef %161.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 384, 0, implicit $exec
-    %161.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 448, 0, implicit $exec
-    undef %170.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 512, 0, implicit $exec
-    %170.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 576, 0, implicit $exec
-    undef %179.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 640, 0, implicit $exec
-    %179.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 704, 0, implicit $exec
-    undef %188.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 768, 0, implicit $exec
-    %188.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 832, 0, implicit $exec
-    %249:sreg_32_xm0_xexec = S_ADD_I32 %249:sreg_32_xm0_xexec, %85.sub1:sreg_64_xexec, implicit-def dead $scc
-    undef %197.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 896, 0, implicit $exec
-    %197.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %95:vgpr_32, 960, 0, implicit $exec
+    undef %9.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 0, 0, implicit $exec
+    %9.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 64, 0, implicit $exec
+    undef %10.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 128, 0, implicit $exec
+    %10.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 192, 0, implicit $exec
+    undef %11.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 256, 0, implicit $exec
+    %11.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 320, 0, implicit $exec
+    undef %12.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 384, 0, implicit $exec
+    %12.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 448, 0, implicit $exec
+    undef %13.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 512, 0, implicit $exec
+    %13.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 576, 0, implicit $exec
+    undef %14.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 640, 0, implicit $exec
+    %14.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 704, 0, implicit $exec
+    undef %15.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 768, 0, implicit $exec
+    %15.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 832, 0, implicit $exec
+    %4:sreg_32_xm0_xexec = S_ADD_I32 %4:sreg_32_xm0_xexec, %3.sub1:sreg_64_xexec, implicit-def dead $scc
+    undef %16.sub0_sub1_sub2_sub3:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 896, 0, implicit $exec
+    %16.sub4_sub5_sub6_sub7:vreg_256_align2 = DS_LOAD_TR16_B128 %0:vgpr_32, 960, 0, implicit $exec
     S_WAIT_DSCNT 5, implicit-def dead $tensorcnt, implicit $tensorcnt
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
-    early-clobber %2379:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %2373:vreg_512_align2, %2369:vreg_512_align2, 8, 0, %2353.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %17:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %18:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %19:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    early-clobber %20:vreg_256_align2 = V_WMMA_SCALE_F32_16X16X128_F8F6F4_f8_f8_w32_threeaddr %5:vreg_512_align2, %6:vreg_512_align2, 8, 0, %7.sub0:vreg_128_lo256_align2, %8:vgpr_32_lo256, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec
     S_ENDPGM 0, amdgpu_allvgprs
 ...



More information about the llvm-branch-commits mailing list