AMDGPU: Cache information about register pressure sets

[• tstellarAMD]

Summary

We can statically decide whether or not a register pressure set is for
SGPRs or VGPRs, so we don't need to re-compute this information in
SIRegisterInfo::getRegPressureSetLimit().

[• tstellarAMD]

Ping.

[arsenm]

lib/Target/AMDGPU/SIRegisterInfo.cpp

118–121 ↗

(On Diff #40597)

Are both SGPRPressureSets and VGPRPressureSets necessary? Can you just have one and assume if it isn't it's the other?

[• tstellarAMD]

lib/Target/AMDGPU/SIRegisterInfo.cpp

118–121 ↗

(On Diff #40597)

It's because of the VS_32 register class has both SGPRs. and VGPRs.

[arsenm]

LGTM

lib/Target/AMDGPU/SIRegisterInfo.cpp

118–121 ↗

(On Diff #40597)

I should try to reapply my patch to try to get VS_32 to not influence pressure

llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.h

 #include "llvm/Support/Debug.h"
 
 namespace llvm {
 
 struct SIRegisterInfo final : public AMDGPURegisterInfo {
 private:
   unsigned SGPR32SetID;
   unsigned VGPR32SetID;
+  BitVector SGPRPressureSets;
+  BitVector VGPRPressureSets;
 
   void reserveRegisterTuples(BitVector &, unsigned Reg) const;
+  void classifyPressureSet(unsigned PSetID, unsigned Reg,
+                           BitVector &PressureSets) const;
 
 public:
   SIRegisterInfo();
 
   /// Return the end register initially reserved for the scratch buffer in case
   /// spilling is needed.
   unsigned reservedPrivateSegmentBufferReg(const MachineFunction &MF) const;
 

llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp

 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/RegisterScavenging.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/LLVMContext.h"
 
 using namespace llvm;
 
-SIRegisterInfo::SIRegisterInfo() : AMDGPURegisterInfo() {
+static bool hasPressureSet(const int *PSets, unsigned PSetID) {
+  for (unsigned i = 0; PSets[i] != -1; ++i) {
+    if (PSets[i] == (int)PSetID)
+      return true;
+  }
+  return false;
+}
+
+void SIRegisterInfo::classifyPressureSet(unsigned PSetID, unsigned Reg,
+                                         BitVector &PressureSets) const {
+  for (MCRegUnitIterator U(Reg, this); U.isValid(); ++U) {
+    const int *PSets = getRegUnitPressureSets(*U);
+    if (hasPressureSet(PSets, PSetID)) {
+      PressureSets.set(PSetID);
+      break;
+    }
+  }
+}
+
+SIRegisterInfo::SIRegisterInfo() : AMDGPURegisterInfo(),
+                                   SGPRPressureSets(getNumRegPressureSets()),
+                                   VGPRPressureSets(getNumRegPressureSets()) {
   unsigned NumRegPressureSets = getNumRegPressureSets();
 
   SGPR32SetID = NumRegPressureSets;
   VGPR32SetID = NumRegPressureSets;
   for (unsigned i = 0; i < NumRegPressureSets; ++i) {
     if (strncmp("SGPR_32", getRegPressureSetName(i), 7) == 0)
       SGPR32SetID = i;
     else if (strncmp("VGPR_32", getRegPressureSetName(i), 7) == 0)
       VGPR32SetID = i;
+
+    classifyPressureSet(i, AMDGPU::SGPR0, SGPRPressureSets);
+    classifyPressureSet(i, AMDGPU::VGPR0, VGPRPressureSets);
   }
   assert(SGPR32SetID < NumRegPressureSets &&
          VGPR32SetID < NumRegPressureSets);
 }
 
 void SIRegisterInfo::reserveRegisterTuples(BitVector &Reserved, unsigned Reg) const {
   MCRegAliasIterator R(Reg, this, true);
 
unsigned SIRegisterInfo::getRegPressureSetLimit(const MachineFunction &MF,
   const AMDGPUSubtarget &STI = MF.getSubtarget<AMDGPUSubtarget>();
   // FIXME: We should adjust the max number of waves based on LDS size.
   unsigned SGPRLimit = getNumSGPRsAllowed(STI.getGeneration(),
                                           STI.getMaxWavesPerCU());
   unsigned VGPRLimit = getNumVGPRsAllowed(STI.getMaxWavesPerCU());
 
   unsigned VSLimit = SGPRLimit + VGPRLimit;
 
-  for (regclass_iterator I = regclass_begin(), E = regclass_end();
-       I != E; ++I) {
-    const TargetRegisterClass *RC = *I;
-
-    unsigned NumSubRegs = std::max((int)RC->getSize() / 4, 1);
-    unsigned Limit;
-
-    if (isPseudoRegClass(RC)) {
-      // FIXME: This is a hack. We should never be considering the pressure of
-      // these since no virtual register should ever have this class.
-      Limit = VSLimit;
-    } else if (isSGPRClass(RC)) {
-      Limit = SGPRLimit / NumSubRegs;
-    } else {
-      Limit = VGPRLimit / NumSubRegs;
-    }
+  if (SGPRPressureSets.test(Idx) && VGPRPressureSets.test(Idx)) {
+    return VSLimit;
 
-    const int *Sets = getRegClassPressureSets(RC);
-    assert(Sets);
-    for (unsigned i = 0; Sets[i] != -1; ++i) {
-      if (Sets[i] == (int)Idx)
-        return Limit;
-    }
-  }
-  return 256;
+  if (SGPRPressureSets.test(Idx))
+    return SGPRLimit;
+
+  return VGPRLimit;
 }
 
 bool SIRegisterInfo::requiresRegisterScavenging(const MachineFunction &Fn) const {
   return Fn.getFrameInfo()->hasStackObjects();
 }
 
 bool
 SIRegisterInfo::requiresFrameIndexScavenging(const MachineFunction &MF) const {