AMDGPU: Correct maximum possible private allocation size

We were assuming a much larger possible per-wave visible stack
allocation than is possible:

https://github.com/RadeonOpenCompute/ROCR-Runtime/blob/faa3ae51388517353afcdaf9c16621f879ef0a59/src/core/runtime/amd_gpu_agent.cpp#L70

Based on this, we can assume the high 15 bits of a frame index or sret
are 0. The frame index value is the per-lane offset, so the maximum
frame index value is MAX_WAVE_SCRATCH / wavesize.

Remove the corresponding subtarget feature and option that made
this configurable.

llvm-svn: 361541
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 341ef73..9938eea 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -458,13 +458,6 @@
 def FeatureMaxPrivateElementSize8 : FeatureMaxPrivateElementSize<8>;
 def FeatureMaxPrivateElementSize16 : FeatureMaxPrivateElementSize<16>;
 
-def FeatureEnableHugePrivateBuffer : SubtargetFeature<
-  "huge-private-buffer",
-  "EnableHugePrivateBuffer",
-  "true",
-  "Enable private/scratch buffer sizes greater than 128 GB"
->;
-
 def FeatureDumpCode : SubtargetFeature <"DumpCode",
   "DumpCode",
   "true",
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index a88218f..09b806b 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -190,7 +190,6 @@
     EnableCuMode(false),
     TrapHandler(false),
 
-    EnableHugePrivateBuffer(false),
     EnableLoadStoreOpt(false),
     EnableUnsafeDSOffsetFolding(false),
     EnableSIScheduler(false),
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 1ef7262..34166aa 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -299,7 +299,6 @@
   bool TrapHandler;
 
   // Used as options.
-  bool EnableHugePrivateBuffer;
   bool EnableLoadStoreOpt;
   bool EnableUnsafeDSOffsetFolding;
   bool EnableSIScheduler;
@@ -377,6 +376,9 @@
   SITargetLowering TLInfo;
   SIFrameLowering FrameLowering;
 
+  // See COMPUTE_TMPRING_SIZE.WAVESIZE, 13-bit field in units of 256-dword.
+  static const unsigned MaxWaveScratchSize = (256 * 4) * ((1 << 13) - 1);
+
 public:
   GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
                const GCNTargetMachine &TM);
@@ -436,6 +438,11 @@
     return Log2_32(WavefrontSize);
   }
 
+  /// Return the number of high bits known to be zero fror a frame index.
+  unsigned getKnownHighZeroBitsForFrameIndex() const {
+    return countLeadingZeros(MaxWaveScratchSize) + getWavefrontSizeLog2();
+  }
+
   int getLDSBankCount() const {
     return LDSBankCount;
   }
@@ -526,10 +533,6 @@
     return isAmdHsaOS() ? TrapHandlerAbiHsa : TrapHandlerAbiNone;
   }
 
-  bool enableHugePrivateBuffer() const {
-    return EnableHugePrivateBuffer;
-  }
-
   bool unsafeDSOffsetFoldingEnabled() const {
     return EnableUnsafeDSOffsetFolding;
   }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index c4c0e40..c2cda5e 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -93,12 +93,6 @@
   cl::desc("Use GPR indexing mode instead of movrel for vector indexing"),
   cl::init(false));
 
-static cl::opt<unsigned> AssumeFrameIndexHighZeroBits(
-  "amdgpu-frame-index-zero-bits",
-  cl::desc("High bits of frame index assumed to be zero"),
-  cl::init(5),
-  cl::ReallyHidden);
-
 static cl::opt<bool> DisableLoopAlignment(
   "amdgpu-disable-loop-alignment",
   cl::desc("Do not align and prefetch loops"),
@@ -2059,13 +2053,14 @@
     Reg = MF.addLiveIn(Reg, RC);
     SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
 
-    if (Arg.Flags.isSRet() && !getSubtarget()->enableHugePrivateBuffer()) {
+    if (Arg.Flags.isSRet()) {
       // The return object should be reasonably addressable.
 
       // FIXME: This helps when the return is a real sret. If it is a
       // automatically inserted sret (i.e. CanLowerReturn returns false), an
       // extra copy is inserted in SelectionDAGBuilder which obscures this.
-      unsigned NumBits = 32 - AssumeFrameIndexHighZeroBits;
+      unsigned NumBits
+        = 32 - getSubtarget()->getKnownHighZeroBitsForFrameIndex();
       Val = DAG.getNode(ISD::AssertZext, DL, VT, Val,
         DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), NumBits)));
     }
@@ -9970,14 +9965,10 @@
   TargetLowering::computeKnownBitsForFrameIndex(Op, Known, DemandedElts,
                                                 DAG, Depth);
 
-  if (getSubtarget()->enableHugePrivateBuffer())
-    return;
-
-  // Technically it may be possible to have a dispatch with a single workitem
-  // that uses the full private memory size, but that's not really useful. We
-  // can't use vaddr in MUBUF instructions if we don't know the address
+  // Set the high bits to zero based on the maximum allowed scratch size per
+  // wave. We can't use vaddr in MUBUF instructions if we don't know the address
   // calculation won't overflow, so assume the sign bit is never set.
-  Known.Zero.setHighBits(AssumeFrameIndexHighZeroBits);
+  Known.Zero.setHighBits(getSubtarget()->getKnownHighZeroBitsForFrameIndex());
 }
 
 unsigned SITargetLowering::getPrefLoopAlignment(MachineLoop *ML) const {