[llvm] [AMDGPU] Prefer lower total register usage in regions with spilling (PR #71882)

Jeffrey Byrnes via llvm-commits llvm-commits at lists.llvm.org
Thu Dec 14 13:40:39 PST 2023


https://github.com/jrbyrnes updated https://github.com/llvm/llvm-project/pull/71882

>From e8295773ac212824daa71875430cce70bdbbf8d0 Mon Sep 17 00:00:00 2001
From: Jeffrey Byrnes <Jeffrey.Byrnes at amd.com>
Date: Thu, 9 Nov 2023 12:03:06 -0800
Subject: [PATCH] [AMDGPU] Prefer lower total register usage in regions with
 spilling

Change-Id: Ib9872c0f675a60b098f73913bdc27d4bf52c7176
---
 .../Target/AMDGPU/GCNIterativeScheduler.cpp   |  12 +-
 llvm/lib/Target/AMDGPU/GCNRegPressure.cpp     |  75 +++-
 llvm/lib/Target/AMDGPU/GCNRegPressure.h       |  16 +-
 llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp   |   6 +-
 llvm/lib/Target/AMDGPU/GCNSubtarget.h         |   6 +
 .../Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp    |   4 +-
 llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h |   4 +
 .../CodeGen/AMDGPU/spill-regpressure-less.mir | 353 ++++++++++++++++++
 8 files changed, 459 insertions(+), 17 deletions(-)
 create mode 100644 llvm/test/CodeGen/AMDGPU/spill-regpressure-less.mir

diff --git a/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp b/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp
index d89c9b1febded0..34108dd5b330bc 100644
--- a/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp
@@ -409,9 +409,8 @@ void GCNIterativeScheduler::scheduleRegion(Region &R, Range &&Schedule,
 
 // Sort recorded regions by pressure - highest at the front
 void GCNIterativeScheduler::sortRegionsByPressure(unsigned TargetOcc) {
-  const auto &ST = MF.getSubtarget<GCNSubtarget>();
-  llvm::sort(Regions, [&ST, TargetOcc](const Region *R1, const Region *R2) {
-    return R2->MaxPressure.less(ST, R1->MaxPressure, TargetOcc);
+  llvm::sort(Regions, [this, TargetOcc](const Region *R1, const Region *R2) {
+    return R2->MaxPressure.less(MF, R1->MaxPressure, TargetOcc);
   });
 }
 
@@ -517,26 +516,25 @@ void GCNIterativeScheduler::scheduleLegacyMaxOccupancy(
 // Minimal Register Strategy
 
 void GCNIterativeScheduler::scheduleMinReg(bool force) {
-  const auto &ST = MF.getSubtarget<GCNSubtarget>();
   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
   const auto TgtOcc = MFI->getOccupancy();
   sortRegionsByPressure(TgtOcc);
 
   auto MaxPressure = Regions.front()->MaxPressure;
   for (auto *R : Regions) {
-    if (!force && R->MaxPressure.less(ST, MaxPressure, TgtOcc))
+    if (!force && R->MaxPressure.less(MF, MaxPressure, TgtOcc))
       break;
 
     BuildDAG DAG(*R, *this);
     const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this);
 
     const auto RP = getSchedulePressure(*R, MinSchedule);
-    LLVM_DEBUG(if (R->MaxPressure.less(ST, RP, TgtOcc)) {
+    LLVM_DEBUG(if (R->MaxPressure.less(MF, RP, TgtOcc)) {
       dbgs() << "\nWarning: Pressure becomes worse after minreg!";
       printSchedRP(dbgs(), R->MaxPressure, RP);
     });
 
-    if (!force && MaxPressure.less(ST, RP, TgtOcc))
+    if (!force && MaxPressure.less(MF, RP, TgtOcc))
       break;
 
     scheduleRegion(*R, MinSchedule, RP);
diff --git a/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp b/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
index a04c470b7b9762..f3b28031f0185d 100644
--- a/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
@@ -88,9 +88,10 @@ void GCNRegPressure::inc(unsigned Reg,
   }
 }
 
-bool GCNRegPressure::less(const GCNSubtarget &ST,
-                          const GCNRegPressure& O,
+bool GCNRegPressure::less(const MachineFunction &MF, const GCNRegPressure &O,
                           unsigned MaxOccupancy) const {
+  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
+
   const auto SGPROcc = std::min(MaxOccupancy,
                                 ST.getOccupancyWithNumSGPRs(getSGPRNum()));
   const auto VGPROcc =
@@ -104,18 +105,82 @@ bool GCNRegPressure::less(const GCNSubtarget &ST,
 
   const auto Occ = std::min(SGPROcc, VGPROcc);
   const auto OtherOcc = std::min(OtherSGPROcc, OtherVGPROcc);
+
+  // Give first precedence to the better occupancy.
   if (Occ != OtherOcc)
     return Occ > OtherOcc;
 
+  unsigned MaxVGPRs = ST.getMaxNumVGPRs(MF);
+  unsigned MaxSGPRs = ST.getMaxNumSGPRs(MF);
+
+  // SGPR excess pressure conditions
+  unsigned ExcessSGPR = std::max(static_cast<int>(getSGPRNum() - MaxSGPRs), 0);
+  unsigned OtherExcessSGPR =
+      std::max(static_cast<int>(O.getSGPRNum() - MaxSGPRs), 0);
+
+  auto WaveSize = ST.getWavefrontSize();
+  // The number of virtual VGPRs required to handle excess SGPR
+  unsigned SGPRSpills = (ExcessSGPR + (WaveSize - 1)) / WaveSize;
+  unsigned OtherSGPRSpills = (OtherExcessSGPR + (WaveSize - 1)) / WaveSize;
+
+  unsigned MaxArchVGPRs = ST.getAddressableNumArchVGPRs();
+
+  // Unified excess pressure conditions, accounting for VGPRs used for SGPR
+  // spills
+  unsigned ExcessVGPR = std::max(
+      static_cast<int>(getVGPRNum(ST.hasGFX90AInsts()) + SGPRSpills - MaxVGPRs),
+      0);
+  unsigned OtherExcessVGPR =
+      std::max(static_cast<int>(O.getVGPRNum(ST.hasGFX90AInsts()) +
+                                OtherSGPRSpills - MaxVGPRs),
+               0);
+  // Arch VGPR excess pressure conditions, accounting for VGPRs used for SGPR
+  // spills
+  unsigned ExcessArchVGPR = std::max(
+      static_cast<int>(getVGPRNum(false) + SGPRSpills - MaxArchVGPRs), 0);
+  unsigned OtherExcessArchVGPR = std::max(
+      static_cast<int>(O.getVGPRNum(false) + OtherSGPRSpills - MaxArchVGPRs),
+      0);
+  // AGPR excess pressure conditions
+  unsigned ExcessAGPR = std::max(
+      static_cast<int>(ST.hasGFX90AInsts() ? (getAGPRNum() - MaxArchVGPRs)
+                                           : (getAGPRNum() - MaxVGPRs)),
+      0);
+  unsigned OtherExcessAGPR = std::max(
+      static_cast<int>(ST.hasGFX90AInsts() ? (O.getAGPRNum() - MaxArchVGPRs)
+                                           : (O.getAGPRNum() - MaxVGPRs)),
+      0);
+
+  bool ExcessRP = ExcessSGPR || ExcessVGPR || ExcessArchVGPR || ExcessAGPR;
+  bool OtherExcessRP = OtherExcessSGPR || OtherExcessVGPR ||
+                       OtherExcessArchVGPR || OtherExcessAGPR;
+
+  // Give second precedence to the reduced number of spills to hold the register
+  // pressure.
+  if (ExcessRP || OtherExcessRP) {
+    // The difference in excess VGPR pressure, after including VGPRs used for
+    // SGPR spills
+    int VGPRDiff = ((OtherExcessVGPR + OtherExcessArchVGPR + OtherExcessAGPR) -
+                    (ExcessVGPR + ExcessArchVGPR + ExcessAGPR));
+
+    int SGPRDiff = OtherExcessSGPR - ExcessSGPR;
+
+    // If excess VGPR is the same, prefer the lower excess SGPR
+    if (VGPRDiff != 0)
+      return VGPRDiff > 0;
+    else if (SGPRDiff != 0)
+      return SGPRDiff > 0;
+  }
+
   bool SGPRImportant = SGPROcc < VGPROcc;
   const bool OtherSGPRImportant = OtherSGPROcc < OtherVGPROcc;
 
-  // if both pressures disagree on what is more important compare vgprs
+  // If both pressures disagree on what is more important compare vgprs.
   if (SGPRImportant != OtherSGPRImportant) {
     SGPRImportant = false;
   }
 
-  // compare large regs pressure
+  // Give thid precedence to lower register tuple pressure.
   bool SGPRFirst = SGPRImportant;
   for (int I = 2; I > 0; --I, SGPRFirst = !SGPRFirst) {
     if (SGPRFirst) {
@@ -130,6 +195,8 @@ bool GCNRegPressure::less(const GCNSubtarget &ST,
         return VW < OtherVW;
     }
   }
+
+  // Give final precedence to lower general RP.
   return SGPRImportant ? (getSGPRNum() < O.getSGPRNum()):
                          (getVGPRNum(ST.hasGFX90AInsts()) <
                           O.getVGPRNum(ST.hasGFX90AInsts()));
diff --git a/llvm/lib/Target/AMDGPU/GCNRegPressure.h b/llvm/lib/Target/AMDGPU/GCNRegPressure.h
index c750fe74749e2b..aa7457cd3f8f75 100644
--- a/llvm/lib/Target/AMDGPU/GCNRegPressure.h
+++ b/llvm/lib/Target/AMDGPU/GCNRegPressure.h
@@ -74,8 +74,20 @@ struct GCNRegPressure {
     return getOccupancy(ST) > O.getOccupancy(ST);
   }
 
-  bool less(const GCNSubtarget &ST, const GCNRegPressure& O,
-    unsigned MaxOccupancy = std::numeric_limits<unsigned>::max()) const;
+  /// Compares \p this GCNRegpressure to \p O, returning true if \p this is
+  /// less. Since GCNRegpressure contains different types of pressures, and due
+  /// to target-specific pecularities (e.g. we care about occupancy rather than
+  /// raw register usage), we determine if \p this GCNRegPressure is less than
+  /// \p O based on the following tiered comparisons (in order order of
+  /// precedence):
+  /// 1. Better occupancy
+  /// 2. Less spilling (first preference to VGPR spills, then to SGPR spills)
+  /// 3. Less tuple register pressure (first preference to VGPR tuples if we
+  /// determine that SGPR pressure is not important)
+  /// 4. Less raw register pressure (first preference to VGPR tuples if we
+  /// determine that SGPR pressure is not important)
+  bool less(const MachineFunction &MF, const GCNRegPressure &O,
+            unsigned MaxOccupancy = std::numeric_limits<unsigned>::max()) const;
 
   bool operator==(const GCNRegPressure &O) const {
     return std::equal(&Value[0], &Value[TOTAL_KINDS], O.Value);
diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
index c3d60b635d3240..f7d2b458151b9d 100644
--- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
@@ -961,6 +961,7 @@ void GCNSchedStage::checkScheduling() {
 
   unsigned MaxVGPRs = ST.getMaxNumVGPRs(MF);
   unsigned MaxSGPRs = ST.getMaxNumSGPRs(MF);
+
   if (PressureAfter.getVGPRNum(false) > MaxVGPRs ||
       PressureAfter.getAGPRNum() > MaxVGPRs ||
       PressureAfter.getSGPRNum() > MaxSGPRs) {
@@ -1183,9 +1184,8 @@ bool ILPInitialScheduleStage::shouldRevertScheduling(unsigned WavesAfter) {
 }
 
 bool GCNSchedStage::mayCauseSpilling(unsigned WavesAfter) {
-  if (WavesAfter <= MFI.getMinWavesPerEU() &&
-      !PressureAfter.less(ST, PressureBefore) &&
-      isRegionWithExcessRP()) {
+  if (WavesAfter <= MFI.getMinWavesPerEU() && isRegionWithExcessRP() &&
+      !PressureAfter.less(MF, PressureBefore)) {
     LLVM_DEBUG(dbgs() << "New pressure will result in more spilling.\n");
     return true;
   }
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 4c624ed59427c9..3b8edc001f372c 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -1292,6 +1292,12 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
     return AMDGPU::IsaInfo::getTotalNumVGPRs(this);
   }
 
+  /// \returns Addressable number of architectural VGPRs supported by the
+  /// subtarget.
+  unsigned getAddressableNumArchVGPRs() const {
+    return AMDGPU::IsaInfo::getAddressableNumArchVGPRs(this);
+  }
+
   /// \returns Addressable number of VGPRs supported by the subtarget.
   unsigned getAddressableNumVGPRs() const {
     return AMDGPU::IsaInfo::getAddressableNumVGPRs(this);
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 7b8a37532c9fad..c9c56b86187dc9 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -1026,10 +1026,12 @@ unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI) {
   return IsWave32 ? 1024 : 512;
 }
 
+unsigned getAddressableNumArchVGPRs(const MCSubtargetInfo *STI) { return 256; }
+
 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI) {
   if (STI->getFeatureBits().test(FeatureGFX90AInsts))
     return 512;
-  return 256;
+  return getAddressableNumArchVGPRs(STI);
 }
 
 unsigned getNumWavesPerEUWithNumVGPRs(const MCSubtargetInfo *STI,
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index 1e0994d0862cf5..ceee6915716db6 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -288,6 +288,10 @@ unsigned getVGPREncodingGranule(
 /// \returns Total number of VGPRs for given subtarget \p STI.
 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI);
 
+/// \returns Addressable number of architectural VGPRs for a given subtarget \p
+/// STI.
+unsigned getAddressableNumArchVGPRs(const MCSubtargetInfo *STI);
+
 /// \returns Addressable number of VGPRs for given subtarget \p STI.
 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI);
 
diff --git a/llvm/test/CodeGen/AMDGPU/spill-regpressure-less.mir b/llvm/test/CodeGen/AMDGPU/spill-regpressure-less.mir
new file mode 100644
index 00000000000000..f50688240fe8bd
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/spill-regpressure-less.mir
@@ -0,0 +1,353 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 3
+# RUN: llc -march=amdgcn -mcpu=gfx90a -run-pass=machine-scheduler -verify-misched -o - %s | FileCheck -check-prefix=GCN %s
+
+--- |
+  define amdgpu_kernel void @spill_regpressure_less() #0 {
+    ret void
+  }
+
+  attributes #0 = { "amdgpu-waves-per-eu"="8,8" }
+...
+
+---
+name:            spill_regpressure_less
+tracksRegLiveness: true
+machineFunctionInfo:
+  stackPtrOffsetReg: '$sgpr32'
+  occupancy:       8
+body:             |
+  bb.0:
+    ; GCN-LABEL: name: spill_regpressure_less
+    ; GCN: [[DEF:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF1:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF2:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF3:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF4:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF5:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF6:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF7:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF8:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF9:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF10:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF11:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF12:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF13:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF14:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF15:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF16:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF17:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF18:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF19:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF20:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF21:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF22:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF23:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF24:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF25:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF26:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF27:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF28:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF29:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF30:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF31:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF32:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF33:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF34:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF35:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF36:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF37:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF38:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF39:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF40:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF41:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF42:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF43:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF44:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF45:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF46:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF47:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF48:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF49:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF50:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF51:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF52:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF53:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF54:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF55:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF56:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF57:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF58:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF59:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF60:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF61:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF62:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF63:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF64:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF65:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF66:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: INLINEASM &"", 1 /* sideeffect attdialect */, implicit [[DEF]], implicit [[DEF1]], implicit [[DEF2]], implicit [[DEF3]], implicit [[DEF4]], implicit [[DEF5]], implicit [[DEF6]], implicit [[DEF7]], implicit [[DEF8]], implicit [[DEF9]], implicit [[DEF10]], implicit [[DEF11]], implicit [[DEF12]], implicit [[DEF13]], implicit [[DEF14]], implicit [[DEF15]], implicit [[DEF16]], implicit [[DEF17]], implicit [[DEF18]], implicit [[DEF19]], implicit [[DEF20]], implicit [[DEF21]], implicit [[DEF22]], implicit [[DEF23]], implicit [[DEF24]], implicit [[DEF25]], implicit [[DEF26]], implicit [[DEF27]], implicit [[DEF28]], implicit [[DEF29]], implicit [[DEF30]], implicit [[DEF31]], implicit [[DEF32]], implicit [[DEF33]], implicit [[DEF34]], implicit [[DEF35]], implicit [[DEF36]], implicit [[DEF37]], implicit [[DEF38]], implicit [[DEF39]], implicit [[DEF40]], implicit [[DEF41]], implicit [[DEF42]], implicit [[DEF43]], implicit [[DEF44]], implicit [[DEF45]], implicit [[DEF46]], implicit [[DEF47]], implicit [[DEF48]], implicit [[DEF49]], implicit [[DEF50]], implicit [[DEF51]], implicit [[DEF52]], implicit [[DEF53]], implicit [[DEF54]], implicit [[DEF55]], implicit [[DEF56]], implicit [[DEF57]], implicit [[DEF58]], implicit [[DEF59]], implicit [[DEF60]], implicit [[DEF61]], implicit [[DEF62]], implicit [[DEF63]], implicit [[DEF64]], implicit [[DEF65]], implicit [[DEF66]]
+    ; GCN-NEXT: KILL [[DEF]]
+    ; GCN-NEXT: KILL [[DEF1]]
+    ; GCN-NEXT: KILL [[DEF10]]
+    ; GCN-NEXT: KILL [[DEF12]]
+    ; GCN-NEXT: KILL [[DEF13]]
+    ; GCN-NEXT: KILL [[DEF14]]
+    ; GCN-NEXT: KILL [[DEF15]]
+    ; GCN-NEXT: KILL [[DEF16]]
+    ; GCN-NEXT: [[DEF67:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: KILL [[DEF17]]
+    ; GCN-NEXT: [[DEF68:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF69:%[0-9]+]]:sgpr_128 = IMPLICIT_DEF
+    ; GCN-NEXT: INLINEASM &"", 1 /* sideeffect attdialect */, implicit [[DEF69]], implicit [[DEF23]], implicit [[DEF24]], implicit [[DEF25]], implicit [[DEF26]], implicit [[DEF27]], implicit [[DEF28]]
+    ; GCN-NEXT: KILL [[DEF2]]
+    ; GCN-NEXT: KILL [[DEF3]]
+    ; GCN-NEXT: KILL [[DEF4]]
+    ; GCN-NEXT: KILL [[DEF5]]
+    ; GCN-NEXT: KILL [[DEF6]]
+    ; GCN-NEXT: KILL [[DEF7]]
+    ; GCN-NEXT: KILL [[DEF8]]
+    ; GCN-NEXT: KILL [[DEF9]]
+    ; GCN-NEXT: KILL [[DEF18]]
+    ; GCN-NEXT: KILL [[DEF19]]
+    ; GCN-NEXT: [[DEF70:%[0-9]+]]:sgpr_128 = IMPLICIT_DEF
+    ; GCN-NEXT: INLINEASM &"", 1 /* sideeffect attdialect */, implicit [[DEF70]], implicit [[DEF2]], implicit [[DEF3]], implicit [[DEF4]], implicit [[DEF5]], implicit [[DEF6]], implicit [[DEF7]], implicit [[DEF8]], implicit [[DEF9]]
+    ; GCN-NEXT: KILL [[DEF69]], implicit-def %70, implicit-def %71, implicit-def %72, implicit-def %73, implicit-def %74, implicit-def %75, implicit-def %76, implicit-def %77
+    ; GCN-NEXT: [[DEF71:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: [[DEF72:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: KILL [[DEF20]]
+    ; GCN-NEXT: [[DEF73:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: KILL [[DEF11]]
+    ; GCN-NEXT: [[DEF74:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: KILL [[DEF21]]
+    ; GCN-NEXT: [[DEF75:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: KILL [[DEF22]]
+    ; GCN-NEXT: [[DEF76:%[0-9]+]]:vgpr_32 = IMPLICIT_DEF
+    ; GCN-NEXT: KILL [[DEF23]]
+    ; GCN-NEXT: KILL [[DEF24]]
+    ; GCN-NEXT: KILL [[DEF25]]
+    ; GCN-NEXT: KILL [[DEF26]]
+    ; GCN-NEXT: KILL [[DEF27]]
+    ; GCN-NEXT: KILL [[DEF28]]
+    ; GCN-NEXT: KILL [[DEF29]]
+    ; GCN-NEXT: KILL [[DEF30]]
+    ; GCN-NEXT: KILL [[DEF31]]
+    ; GCN-NEXT: KILL [[DEF32]]
+    ; GCN-NEXT: KILL [[DEF33]]
+    ; GCN-NEXT: KILL [[DEF34]]
+    ; GCN-NEXT: KILL [[DEF35]]
+    ; GCN-NEXT: KILL [[DEF36]]
+    ; GCN-NEXT: KILL [[DEF37]]
+    ; GCN-NEXT: KILL [[DEF38]]
+    ; GCN-NEXT: KILL [[DEF39]]
+    ; GCN-NEXT: KILL [[DEF40]]
+    ; GCN-NEXT: KILL [[DEF41]]
+    ; GCN-NEXT: KILL [[DEF42]]
+    ; GCN-NEXT: KILL [[DEF43]]
+    ; GCN-NEXT: KILL [[DEF44]]
+    ; GCN-NEXT: KILL [[DEF45]]
+    ; GCN-NEXT: KILL [[DEF46]]
+    ; GCN-NEXT: KILL [[DEF47]]
+    ; GCN-NEXT: KILL [[DEF48]]
+    ; GCN-NEXT: KILL [[DEF49]]
+    ; GCN-NEXT: KILL [[DEF50]]
+    ; GCN-NEXT: KILL [[DEF51]]
+    ; GCN-NEXT: KILL [[DEF52]]
+    ; GCN-NEXT: KILL [[DEF53]]
+    ; GCN-NEXT: KILL [[DEF54]]
+    ; GCN-NEXT: KILL [[DEF55]]
+    ; GCN-NEXT: KILL [[DEF56]]
+    ; GCN-NEXT: KILL [[DEF57]]
+    ; GCN-NEXT: KILL [[DEF58]]
+    ; GCN-NEXT: KILL [[DEF59]]
+    ; GCN-NEXT: KILL [[DEF60]]
+    ; GCN-NEXT: KILL [[DEF61]]
+    ; GCN-NEXT: KILL [[DEF62]]
+    ; GCN-NEXT: KILL [[DEF63]]
+    ; GCN-NEXT: KILL [[DEF64]]
+    ; GCN-NEXT: KILL [[DEF65]]
+    ; GCN-NEXT: KILL [[DEF66]]
+    ; GCN-NEXT: KILL [[DEF67]]
+    ; GCN-NEXT: KILL [[DEF68]]
+    ; GCN-NEXT: KILL [[DEF71]]
+    ; GCN-NEXT: KILL [[DEF72]]
+    ; GCN-NEXT: KILL [[DEF73]]
+    ; GCN-NEXT: KILL [[DEF74]]
+    ; GCN-NEXT: KILL [[DEF75]]
+    ; GCN-NEXT: KILL [[DEF76]]
+    ; GCN-NEXT: KILL [[DEF70]]
+    ; GCN-NEXT: KILL %70
+    ; GCN-NEXT: KILL %71
+    ; GCN-NEXT: KILL %72
+    ; GCN-NEXT: KILL %73
+    ; GCN-NEXT: KILL %74
+    ; GCN-NEXT: KILL %75
+    ; GCN-NEXT: KILL %76
+    ; GCN-NEXT: KILL %77
+    %0:vgpr_32 = IMPLICIT_DEF
+    %1:vgpr_32 = IMPLICIT_DEF
+    %2:vgpr_32 = IMPLICIT_DEF
+    %3:vgpr_32 = IMPLICIT_DEF
+    %4:vgpr_32 = IMPLICIT_DEF
+    %5:vgpr_32 = IMPLICIT_DEF
+    %6:vgpr_32 = IMPLICIT_DEF
+    %7:vgpr_32 = IMPLICIT_DEF
+    %8:vgpr_32 = IMPLICIT_DEF
+    %9:vgpr_32 = IMPLICIT_DEF
+    %10:vgpr_32 = IMPLICIT_DEF
+    %11:vgpr_32 = IMPLICIT_DEF
+    %12:vgpr_32 = IMPLICIT_DEF
+    %13:vgpr_32 = IMPLICIT_DEF
+    %14:vgpr_32 = IMPLICIT_DEF
+    %15:vgpr_32 = IMPLICIT_DEF
+    %16:vgpr_32 = IMPLICIT_DEF
+    %17:vgpr_32 = IMPLICIT_DEF
+    %18:vgpr_32 = IMPLICIT_DEF
+    %19:vgpr_32 = IMPLICIT_DEF
+    %20:vgpr_32 = IMPLICIT_DEF
+    %21:vgpr_32 = IMPLICIT_DEF
+    %22:vgpr_32 = IMPLICIT_DEF
+    %23:vgpr_32 = IMPLICIT_DEF
+    %24:vgpr_32 = IMPLICIT_DEF
+    %25:vgpr_32 = IMPLICIT_DEF
+    %26:vgpr_32 = IMPLICIT_DEF
+    %27:vgpr_32 = IMPLICIT_DEF
+    %28:vgpr_32 = IMPLICIT_DEF
+    %29:vgpr_32 = IMPLICIT_DEF
+    %30:vgpr_32 = IMPLICIT_DEF
+    %31:vgpr_32 = IMPLICIT_DEF
+    %32:vgpr_32 = IMPLICIT_DEF
+    %33:vgpr_32 = IMPLICIT_DEF
+    %34:vgpr_32 = IMPLICIT_DEF
+    %35:vgpr_32 = IMPLICIT_DEF
+    %36:vgpr_32 = IMPLICIT_DEF
+    %37:vgpr_32 = IMPLICIT_DEF
+    %38:vgpr_32 = IMPLICIT_DEF
+    %39:vgpr_32 = IMPLICIT_DEF
+    %40:vgpr_32 = IMPLICIT_DEF
+    %41:vgpr_32 = IMPLICIT_DEF
+    %42:vgpr_32 = IMPLICIT_DEF
+    %43:vgpr_32 = IMPLICIT_DEF
+    %44:vgpr_32 = IMPLICIT_DEF
+    %45:vgpr_32 = IMPLICIT_DEF
+    %46:vgpr_32 = IMPLICIT_DEF
+    %47:vgpr_32 = IMPLICIT_DEF
+    %48:vgpr_32 = IMPLICIT_DEF
+    %49:vgpr_32 = IMPLICIT_DEF
+    %50:vgpr_32 = IMPLICIT_DEF
+    %51:vgpr_32 = IMPLICIT_DEF
+    %52:vgpr_32 = IMPLICIT_DEF
+    %53:vgpr_32 = IMPLICIT_DEF
+    %54:vgpr_32 = IMPLICIT_DEF
+    %55:vgpr_32 = IMPLICIT_DEF
+    %56:vgpr_32 = IMPLICIT_DEF
+    %57:vgpr_32 = IMPLICIT_DEF
+    %58:vgpr_32 = IMPLICIT_DEF
+    %59:vgpr_32 = IMPLICIT_DEF
+    %60:vgpr_32 = IMPLICIT_DEF
+    %61:vgpr_32 = IMPLICIT_DEF
+    %62:vgpr_32 = IMPLICIT_DEF
+    %63:vgpr_32 = IMPLICIT_DEF
+    %64:vgpr_32 = IMPLICIT_DEF
+    %65:vgpr_32 = IMPLICIT_DEF
+    %66:vgpr_32 = IMPLICIT_DEF
+    %67:vgpr_32 = IMPLICIT_DEF
+    %68:vgpr_32 = IMPLICIT_DEF
+    INLINEASM &"", 1, implicit %0, implicit %1, implicit %2, implicit %3, implicit %4, implicit %5, implicit %6, implicit %7, implicit %8, implicit %9, implicit %10, implicit %11, implicit %12, implicit %13, implicit %14, implicit %15, implicit %16, implicit %17, implicit %18, implicit %19, implicit %20, implicit %21, implicit %22, implicit %23, implicit %24, implicit %25, implicit %26, implicit %27, implicit %28, implicit %29, implicit %30, implicit %31, implicit %32, implicit %33, implicit %34, implicit %35, implicit %36, implicit %37, implicit %38, implicit %39, implicit %40, implicit %41, implicit %42, implicit %43, implicit %44, implicit %45, implicit %46, implicit %47, implicit %48, implicit %49, implicit %50, implicit %51, implicit %52, implicit %53, implicit %54, implicit %55, implicit %56, implicit %57, implicit %58, implicit %59, implicit %60, implicit %61, implicit %62, implicit %63, implicit %64, implicit %65, implicit %66
+    %69:sgpr_128 = IMPLICIT_DEF
+    INLINEASM &"", 1, implicit %69, implicit %23, implicit %24, implicit %25, implicit %26, implicit %27, implicit %28
+    KILL %0
+    KILL %1
+    KILL %2
+    KILL %3
+    KILL %4
+    KILL %5
+    KILL %6
+    KILL %7
+    KILL %8
+    KILL %9
+    KILL %10
+    KILL %12
+    KILL %13
+    KILL %14
+    KILL %15
+    KILL %16
+    KILL %17
+    KILL %18
+    KILL %19
+    KILL %69:sgpr_128, implicit-def %77:vgpr_32, implicit-def %78:vgpr_32, implicit-def %79:vgpr_32, implicit-def %80:vgpr_32, implicit-def %81:vgpr_32, implicit-def %82:vgpr_32, implicit-def %83:vgpr_32, implicit-def %84:vgpr_32
+    %70:vgpr_32 = IMPLICIT_DEF
+    %71:vgpr_32 = IMPLICIT_DEF
+    %72:vgpr_32 = IMPLICIT_DEF
+    %73:vgpr_32 = IMPLICIT_DEF
+    %74:vgpr_32 = IMPLICIT_DEF
+    %75:vgpr_32 = IMPLICIT_DEF
+    %76:sgpr_128 = IMPLICIT_DEF
+    INLINEASM &"", 1, implicit %76, implicit %2, implicit %3, implicit %4, implicit %5, implicit %6, implicit %7, implicit %8, implicit %9
+    KILL %20
+    KILL %11
+    KILL %21
+    KILL %22
+    KILL %23
+    KILL %24
+    KILL %25
+    KILL %26
+    KILL %27
+    KILL %28
+    KILL %29
+    KILL %30
+    KILL %31
+    KILL %32
+    KILL %33
+    KILL %34
+    KILL %35
+    KILL %36
+    KILL %37
+    KILL %38
+    KILL %39
+    KILL %40
+    KILL %41
+    KILL %42
+    KILL %43
+    KILL %44
+    KILL %45
+    KILL %46
+    KILL %47
+    KILL %48
+    KILL %49
+    KILL %50
+    KILL %51
+    KILL %52
+    KILL %53
+    KILL %54
+    KILL %55
+    KILL %56
+    KILL %57
+    KILL %58
+    KILL %59
+    KILL %60
+    KILL %61
+    KILL %62
+    KILL %63
+    KILL %64
+    KILL %65
+    KILL %66
+    KILL %67
+    KILL %68
+    KILL %70
+    KILL %71
+    KILL %72
+    KILL %73
+    KILL %74
+    KILL %75
+    KILL %76
+    KILL %77
+    KILL %78
+    KILL %79
+    KILL %80
+    KILL %81
+    KILL %82
+    KILL %83
+    KILL %84
+...
+## NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+# GCN: {{.*}}



More information about the llvm-commits mailing list