From 367664725157cef23135854cb48e7358399b24c9 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Tue, 25 Apr 2017 23:40:57 +0000
Subject: [PATCH] AMDGPU: Shift down reserved SP register like scratch wave
 offset

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@301367 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AMDGPU/SIFrameLowering.cpp | 74 +++++++++++++++++++++------
 lib/Target/AMDGPU/SIFrameLowering.h   |  2 +-
 2 files changed, 59 insertions(+), 17 deletions(-)

diff --git a/lib/Target/AMDGPU/SIFrameLowering.cpp b/lib/Target/AMDGPU/SIFrameLowering.cpp
index e03a416bd6c..86e3b37b09e 100644
--- a/lib/Target/AMDGPU/SIFrameLowering.cpp
+++ b/lib/Target/AMDGPU/SIFrameLowering.cpp
@@ -143,7 +143,10 @@ unsigned SIFrameLowering::getReservedPrivateSegmentBufferReg(
   return ScratchRsrcReg;
 }
 
-unsigned SIFrameLowering::getReservedPrivateSegmentWaveByteOffsetReg(
+// Shift down registers reserved for the scratch wave offset and stack pointer
+// SGPRs.
+std::pair<unsigned, unsigned>
+SIFrameLowering::getReservedPrivateSegmentWaveByteOffsetReg(
   const SISubtarget &ST,
   const SIInstrInfo *TII,
   const SIRegisterInfo *TRI,
@@ -154,18 +157,20 @@ unsigned SIFrameLowering::getReservedPrivateSegmentWaveByteOffsetReg(
 
   // No replacement necessary.
   if (ScratchWaveOffsetReg == AMDGPU::NoRegister ||
-      !MRI.isPhysRegUsed(ScratchWaveOffsetReg))
-    return AMDGPU::NoRegister;
+      !MRI.isPhysRegUsed(ScratchWaveOffsetReg)) {
+    assert(MFI->getStackPtrOffsetReg() == AMDGPU::NoRegister);
+    return std::make_pair(AMDGPU::NoRegister, AMDGPU::NoRegister);
+  }
 
-  if (ST.hasSGPRInitBug() ||
-      ScratchWaveOffsetReg != TRI->reservedPrivateSegmentWaveByteOffsetReg(MF))
-    return ScratchWaveOffsetReg;
+  unsigned SPReg = MFI->getStackPtrOffsetReg();
+  if (ST.hasSGPRInitBug())
+    return std::make_pair(ScratchWaveOffsetReg, SPReg);
 
   unsigned NumPreloaded = MFI->getNumPreloadedSGPRs();
 
   ArrayRef<MCPhysReg> AllSGPRs = getAllSGPRs(ST, MF);
   if (NumPreloaded > AllSGPRs.size())
-    return ScratchWaveOffsetReg;
+    return std::make_pair(ScratchWaveOffsetReg, SPReg);
 
   AllSGPRs = AllSGPRs.slice(NumPreloaded);
 
@@ -180,22 +185,42 @@ unsigned SIFrameLowering::getReservedPrivateSegmentWaveByteOffsetReg(
   //     register from the list to consider, it means that when this
   //     register is being used for the scratch wave offset and there
   //     are no other free SGPRs, then the value will stay in this register.
+  // + 1 if stack pointer is used.
   // ----
-  //  13
-  if (AllSGPRs.size() < 13)
-    return ScratchWaveOffsetReg;
+  //  13 (+1)
+  unsigned ReservedRegCount = 13;
+  if (SPReg != AMDGPU::NoRegister)
+    ++ReservedRegCount;
+
+  if (AllSGPRs.size() < ReservedRegCount)
+    return std::make_pair(ScratchWaveOffsetReg, SPReg);
 
-  for (MCPhysReg Reg : AllSGPRs.drop_back(13)) {
+  bool HandledScratchWaveOffsetReg =
+    ScratchWaveOffsetReg != TRI->reservedPrivateSegmentWaveByteOffsetReg(MF);
+
+  for (MCPhysReg Reg : AllSGPRs.drop_back(ReservedRegCount)) {
     // Pick the first unallocated SGPR. Be careful not to pick an alias of the
     // scratch descriptor, since we haven’t added its uses yet.
     if (!MRI.isPhysRegUsed(Reg) && MRI.isAllocatable(Reg)) {
-      MRI.replaceRegWith(ScratchWaveOffsetReg, Reg);
-      MFI->setScratchWaveOffsetReg(Reg);
-      return Reg;
+      if (!HandledScratchWaveOffsetReg) {
+        HandledScratchWaveOffsetReg = true;
+
+        MRI.replaceRegWith(ScratchWaveOffsetReg, Reg);
+        MFI->setScratchWaveOffsetReg(Reg);
+        ScratchWaveOffsetReg = Reg;
+      } else {
+        if (SPReg == AMDGPU::NoRegister)
+          break;
+
+        MRI.replaceRegWith(SPReg, Reg);
+        MFI->setStackPtrOffsetReg(Reg);
+        SPReg = Reg;
+        break;
+      }
     }
   }
 
-  return ScratchWaveOffsetReg;
+  return std::make_pair(ScratchWaveOffsetReg, SPReg);
 }
 
 void SIFrameLowering::emitPrologue(MachineFunction &MF,
@@ -233,9 +258,26 @@ void SIFrameLowering::emitPrologue(MachineFunction &MF,
   if (MF.getFrameInfo().hasStackObjects() && MFI->hasFlatScratchInit())
     emitFlatScratchInit(ST, MF, MBB);
 
+  unsigned SPReg = MFI->getStackPtrOffsetReg();
+  if (SPReg != AMDGPU::NoRegister) {
+    DebugLoc DL;
+    int64_t StackSize = MF.getFrameInfo().getStackSize();
+
+    if (StackSize == 0) {
+      BuildMI(MBB, MBB.begin(), DL, TII->get(AMDGPU::COPY), SPReg)
+        .addReg(MFI->getScratchWaveOffsetReg());
+    } else {
+      BuildMI(MBB, MBB.begin(), DL, TII->get(AMDGPU::S_ADD_U32), SPReg)
+        .addReg(MFI->getScratchWaveOffsetReg())
+        .addImm(StackSize * ST.getWavefrontSize());
+    }
+  }
+
   unsigned ScratchRsrcReg
     = getReservedPrivateSegmentBufferReg(ST, TII, TRI, MFI, MF);
-  unsigned ScratchWaveOffsetReg
+
+  unsigned ScratchWaveOffsetReg;
+  std::tie(ScratchWaveOffsetReg, SPReg)
     = getReservedPrivateSegmentWaveByteOffsetReg(ST, TII, TRI, MFI, MF);
 
   // It's possible to have uses of only ScratchWaveOffsetReg without
diff --git a/lib/Target/AMDGPU/SIFrameLowering.h b/lib/Target/AMDGPU/SIFrameLowering.h
index 1bfc08093da..7ccd02b3c86 100644
--- a/lib/Target/AMDGPU/SIFrameLowering.h
+++ b/lib/Target/AMDGPU/SIFrameLowering.h
@@ -49,7 +49,7 @@ private:
     SIMachineFunctionInfo *MFI,
     MachineFunction &MF) const;
 
-  unsigned getReservedPrivateSegmentWaveByteOffsetReg(
+  std::pair<unsigned, unsigned> getReservedPrivateSegmentWaveByteOffsetReg(
     const SISubtarget &ST,
     const SIInstrInfo *TII,
     const SIRegisterInfo *TRI,
-- 
2.50.1