[llvm] r266244 - AMDGPU/SI: Add support for spilling VGPRs without having to scavenge registers

[Tom Stellard via llvm-commits]

Author: tstellar
Date: Wed Apr 13 15:44:16 2016
New Revision: 266244

URL: http://llvm.org/viewvc/llvm-project?rev=266244&view=rev
Log:
AMDGPU/SI: Add support for spilling VGPRs without having to scavenge registers

Summary:
When we are spilling SGPRs to scratch memory, we usually don't have
free SGPRs to do the address calculation, so we need to re-use the
ScratchOffset register for the calculation.

Reviewers: arsenm

Subscribers: arsenm, llvm-commits

Differential Revision: http://reviews.llvm.org/D18917

Modified:
    llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp
    llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.h

Modified: llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp?rev=266244&r1=266243&r2=266244&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.cpp Wed Apr 13 15:44:16 2016
@@ -246,29 +246,42 @@ void SIRegisterInfo::buildScratchLoadSto
                                            unsigned Value,
                                            unsigned ScratchRsrcReg,
                                            unsigned ScratchOffset,
-                                           int64_t Offset) const {
+                                           int64_t Offset,
+                                           RegScavenger *RS) const {
 
   MachineBasicBlock *MBB = MI->getParent();
   MachineFunction *MF = MI->getParent()->getParent();
-  MachineRegisterInfo &MRI = MF->getRegInfo();
   const SIInstrInfo *TII =
       static_cast<const SIInstrInfo *>(MF->getSubtarget().getInstrInfo());
-  LLVMContext &Ctx = MF->getFunction()->getContext();
   DebugLoc DL = MI->getDebugLoc();
   bool IsStore = TII->get(LoadStoreOp).mayStore();
 
   bool RanOutOfSGPRs = false;
   bool Scavenged = false;
   unsigned SOffset = ScratchOffset;
+  unsigned OriginalImmOffset = Offset;
 
   unsigned NumSubRegs = getNumSubRegsForSpillOp(MI->getOpcode());
   unsigned Size = NumSubRegs * 4;
 
   if (!isUInt<12>(Offset + Size)) {
-    SOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
+    SOffset = AMDGPU::NoRegister;
+
+    // We don't have access to the register scavenger if this function is called
+    // during  PEI::scavengeFrameVirtualRegs().
+    if (RS)
+      SOffset = RS->FindUnusedReg(&AMDGPU::SGPR_32RegClass);
+
     if (SOffset == AMDGPU::NoRegister) {
+      // There are no free SGPRs, and since we are in the process of spilling
+      // VGPRs too.  Since we need a VGPR in order to spill SGPRs (this is true
+      // on SI/CI and on VI it is true until we implement spilling using scalar
+      // stores), we have no way to free up an SGPR.  Our solution here is to
+      // add the offset directly to the ScratchOffset register, and then
+      // subtract the offset after the spill to return ScratchOffset to it's
+      // original value.
       RanOutOfSGPRs = true;
-      SOffset = AMDGPU::SGPR0;
+      SOffset = ScratchOffset;
     } else {
       Scavenged = true;
     }
@@ -278,9 +291,6 @@ void SIRegisterInfo::buildScratchLoadSto
     Offset = 0;
   }
 
-  if (RanOutOfSGPRs)
-    Ctx.emitError("Ran out of SGPRs for spilling VGPRS");
-
   for (unsigned i = 0, e = NumSubRegs; i != e; ++i, Offset += 4) {
     unsigned SubReg = NumSubRegs > 1 ?
         getPhysRegSubReg(Value, &AMDGPU::VGPR_32RegClass, i) :
@@ -301,6 +311,13 @@ void SIRegisterInfo::buildScratchLoadSto
       .addReg(Value, RegState::Implicit | getDefRegState(!IsStore))
       .setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
   }
+
+  if (RanOutOfSGPRs) {
+    // Subtract the offset we added to the ScratchOffset register.
+    BuildMI(*MBB, MI, DL, TII->get(AMDGPU::S_SUB_U32), ScratchOffset)
+            .addReg(ScratchOffset)
+            .addImm(OriginalImmOffset);
+  }
 }
 
 void SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
@@ -451,7 +468,7 @@ void SIRegisterInfo::eliminateFrameIndex
             TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_rsrc)->getReg(),
             TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_offset)->getReg(),
             FrameInfo->getObjectOffset(Index) +
-            TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm());
+            TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm(), RS);
       MI->eraseFromParent();
       break;
     case AMDGPU::SI_SPILL_V32_RESTORE:
@@ -465,7 +482,7 @@ void SIRegisterInfo::eliminateFrameIndex
             TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_rsrc)->getReg(),
             TII->getNamedOperand(*MI, AMDGPU::OpName::scratch_offset)->getReg(),
             FrameInfo->getObjectOffset(Index) +
-            TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm());
+            TII->getNamedOperand(*MI, AMDGPU::OpName::offset)->getImm(), RS);
       MI->eraseFromParent();
       break;
     }

Modified: llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.h?rev=266244&r1=266243&r2=266244&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.h (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIRegisterInfo.h Wed Apr 13 15:44:16 2016
@@ -171,7 +171,8 @@ private:
   void buildScratchLoadStore(MachineBasicBlock::iterator MI,
                              unsigned LoadStoreOp, unsigned Value,
                              unsigned ScratchRsrcReg, unsigned ScratchOffset,
-                             int64_t Offset) const;
+                             int64_t Offset,
+                             RegScavenger *RS) const;
 };
 
 } // End namespace llvm