From 65f94b33808d7d69539961a6f5a2168f0a1eef41 Mon Sep 17 00:00:00 2001
From: Piotr Sobczak <Piotr.Sobczak@amd.com>
Date: Fri, 15 Nov 2019 15:07:41 +0100
Subject: [InstCombine][AMDGPU] Trim more components of *buffer_load

Summary:
Add trimming of unused components of s_buffer_load.

Extend trimming of *buffer_load to also include
unused components at the beginning of vectors and update offset.

Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D70315
---
 .../InstCombine/InstCombineSimplifyDemanded.cpp    | 77 +++++++++++++++++-----
 1 file changed, 60 insertions(+), 17 deletions(-)

(limited to 'llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp')

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
index 359fc55035f..4e509816716 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -984,13 +984,65 @@ Value *InstCombiner::simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II,
   if (VWidth == 1)
     return nullptr;
 
-  ConstantInt *NewDMask = nullptr;
+  IRBuilderBase::InsertPointGuard Guard(Builder);
+  Builder.SetInsertPoint(II);
+
+  // Assume the arguments are unchanged and later override them, if needed.
+  SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end());
 
   if (DMaskIdx < 0) {
-    // Pretend that a prefix of elements is demanded to simplify the code
-    // below.
-    DemandedElts = (1 << DemandedElts.getActiveBits()) - 1;
+    // Buffer case.
+
+    const unsigned ActiveBits = DemandedElts.getActiveBits();
+    const unsigned UnusedComponentsAtFront = DemandedElts.countTrailingZeros();
+
+    // Start assuming the prefix of elements is demanded, but possibly clear some other bits if
+    // there are trailing zeros (unused components at front) and update offset.
+    DemandedElts = (1 << ActiveBits) - 1;
+
+    if (UnusedComponentsAtFront > 0) {
+      static const unsigned InvalidOffsetIdx = 0xf;
+
+      unsigned OffsetIdx;
+      switch (II->getIntrinsicID()) {
+      case Intrinsic::amdgcn_raw_buffer_load:
+      case Intrinsic::amdgcn_raw_buffer_load_format:
+        OffsetIdx = 1;
+        break;
+      case Intrinsic::amdgcn_s_buffer_load:
+        // If resulting type is vec3, there is no point in trimming the
+        // load with updated offset, as the vec3 would most likely be widened to
+        // vec4 anyway during lowering.
+        if (ActiveBits == 4 && UnusedComponentsAtFront == 1)
+          OffsetIdx = InvalidOffsetIdx;
+        else
+          OffsetIdx = 1;
+        break;
+      case Intrinsic::amdgcn_buffer_load:
+      case Intrinsic::amdgcn_buffer_load_format:
+      case Intrinsic::amdgcn_struct_buffer_load:
+      case Intrinsic::amdgcn_struct_buffer_load_format:
+        OffsetIdx = 2;
+        break;
+      default:
+        // TODO: handle *tbuffer* intrinsics.
+        OffsetIdx = InvalidOffsetIdx;
+        break;
+      }
+
+      if (OffsetIdx != InvalidOffsetIdx) {
+        // Clear demanded bits and update the offset.
+        DemandedElts &= ~((1 << UnusedComponentsAtFront) - 1);
+        auto Offset = II->getArgOperand(OffsetIdx);
+        unsigned SingleComponentSizeInBits = getDataLayout().getTypeSizeInBits(II->getType()->getScalarType());
+        unsigned OffsetAdd = UnusedComponentsAtFront * SingleComponentSizeInBits / 8;
+        auto OffsetAddVal = ConstantInt::get(Offset->getType(), OffsetAdd);
+        Args[OffsetIdx] = Builder.CreateAdd(Offset, OffsetAddVal);
+      }
+    }
   } else {
+    // Image case.
+
     ConstantInt *DMask = cast<ConstantInt>(II->getArgOperand(DMaskIdx));
     unsigned DMaskVal = DMask->getZExtValue() & 0xf;
 
@@ -1009,7 +1061,7 @@ Value *InstCombiner::simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II,
     }
 
     if (DMaskVal != NewDMaskVal)
-      NewDMask = ConstantInt::get(DMask->getType(), NewDMaskVal);
+      Args[DMaskIdx] = ConstantInt::get(DMask->getType(), NewDMaskVal);
   }
 
   unsigned NewNumElts = DemandedElts.countPopulation();
@@ -1017,8 +1069,8 @@ Value *InstCombiner::simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II,
     return UndefValue::get(II->getType());
 
   if (NewNumElts >= VWidth && DemandedElts.isMask()) {
-    if (NewDMask)
-      II->setArgOperand(DMaskIdx, NewDMask);
+    if (DMaskIdx >= 0)
+      II->setArgOperand(DMaskIdx, Args[DMaskIdx]);
     return nullptr;
   }
 
@@ -1041,16 +1093,6 @@ Value *InstCombiner::simplifyAMDGCNMemoryIntrinsicDemanded(IntrinsicInst *II,
   OverloadTys[0] = NewTy;
   Function *NewIntrin = Intrinsic::getDeclaration(M, IID, OverloadTys);
 
-  SmallVector<Value *, 16> Args;
-  for (unsigned I = 0, E = II->getNumArgOperands(); I != E; ++I)
-    Args.push_back(II->getArgOperand(I));
-
-  if (NewDMask)
-    Args[DMaskIdx] = NewDMask;
-
-  IRBuilderBase::InsertPointGuard Guard(Builder);
-  Builder.SetInsertPoint(II);
-
   CallInst *NewCall = Builder.CreateCall(NewIntrin, Args);
   NewCall->takeName(II);
   NewCall->copyMetadata(*II);
@@ -1719,6 +1761,7 @@ Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
     case Intrinsic::amdgcn_raw_buffer_load:
     case Intrinsic::amdgcn_raw_buffer_load_format:
     case Intrinsic::amdgcn_raw_tbuffer_load:
+    case Intrinsic::amdgcn_s_buffer_load:
     case Intrinsic::amdgcn_struct_buffer_load:
     case Intrinsic::amdgcn_struct_buffer_load_format:
     case Intrinsic::amdgcn_struct_tbuffer_load:
-- 
cgit v1.2.3