[AMDGPU] gfx1010 loop alignment

Differential Revision: https://reviews.llvm.org/D61529

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@359935 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index 7de4dd9..37d4c83 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -99,6 +99,11 @@ static cl::opt<unsigned> AssumeFrameIndexHighZeroBits(
   cl::init(5),
   cl::ReallyHidden);
 
+static cl::opt<bool> DisableLoopAlignment(
+  "amdgpu-disable-loop-alignment",
+  cl::desc("Do not align and prefetch loops"),
+  cl::init(false));
+
 static unsigned findFirstFreeSGPR(CCState &CCInfo) {
   unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
   for (unsigned Reg = 0; Reg < NumSGPRs; ++Reg) {
@@ -9966,6 +9971,77 @@ void SITargetLowering::computeKnownBitsForFrameIndex(const SDValue Op,
   Known.Zero.setHighBits(AssumeFrameIndexHighZeroBits);
 }
 
+unsigned SITargetLowering::getPrefLoopAlignment(MachineLoop *ML) const {
+  const unsigned PrefAlign = TargetLowering::getPrefLoopAlignment(ML);
+  const unsigned CacheLineAlign = 6; // log2(64)
+
+  // Pre-GFX10 target did not benefit from loop alignment
+  if (!ML || DisableLoopAlignment ||
+      (getSubtarget()->getGeneration() < AMDGPUSubtarget::GFX10) ||
+      getSubtarget()->hasInstFwdPrefetchBug())
+    return PrefAlign;
+
+  // On GFX10 I$ is 4 x 64 bytes cache lines.
+  // By default prefetcher keeps one cache line behind and reads two ahead.
+  // We can modify it with S_INST_PREFETCH for larger loops to have two lines
+  // behind and one ahead.
+  // Therefor we can benefit from aligning loop headers if loop fits 192 bytes.
+  // If loop fits 64 bytes it always spans no more than two cache lines and
+  // does not need an alignment.
+  // Else if loop is less or equal 128 bytes we do not need to modify prefetch,
+  // Else if loop is less or equal 192 bytes we need two lines behind.
+
+  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
+  const MachineBasicBlock *Header = ML->getHeader();
+  if (Header->getAlignment() != PrefAlign)
+    return Header->getAlignment(); // Already processed.
+
+  unsigned LoopSize = 0;
+  for (const MachineBasicBlock *MBB : ML->blocks()) {
+    // If inner loop block is aligned assume in average half of the alignment
+    // size to be added as nops.
+    if (MBB != Header)
+      LoopSize += (1 << MBB->getAlignment()) / 2;
+
+    for (const MachineInstr &MI : *MBB) {
+      LoopSize += TII->getInstSizeInBytes(MI);
+      if (LoopSize > 192)
+        return PrefAlign;
+    }
+  }
+
+  if (LoopSize <= 64)
+    return PrefAlign;
+
+  if (LoopSize <= 128)
+    return CacheLineAlign;
+
+  // If any of parent loops is surrounded by prefetch instructions do not
+  // insert new for inner loop, which would reset parent's settings.
+  for (MachineLoop *P = ML->getParentLoop(); P; P = P->getParentLoop()) {
+    if (MachineBasicBlock *Exit = P->getExitBlock()) {
+      auto I = Exit->getFirstNonDebugInstr();
+      if (I != Exit->end() && I->getOpcode() == AMDGPU::S_INST_PREFETCH)
+        return CacheLineAlign;
+    }
+  }
+
+  MachineBasicBlock *Pre = ML->getLoopPreheader();
+  MachineBasicBlock *Exit = ML->getExitBlock();
+
+  if (Pre && Exit) {
+    BuildMI(*Pre, Pre->getFirstTerminator(), DebugLoc(),
+            TII->get(AMDGPU::S_INST_PREFETCH))
+      .addImm(1); // prefetch 2 lines behind PC
+
+    BuildMI(*Exit, Exit->getFirstNonDebugInstr(), DebugLoc(),
+            TII->get(AMDGPU::S_INST_PREFETCH))
+      .addImm(2); // prefetch 1 line behind PC
+  }
+
+  return CacheLineAlign;
+}
+
 LLVM_ATTRIBUTE_UNUSED
 static bool isCopyFromRegOfInlineAsm(const SDNode *N) {
   assert(N->getOpcode() == ISD::CopyFromReg);

diff --git a/lib/Target/AMDGPU/SIISelLowering.h b/lib/Target/AMDGPU/SIISelLowering.h
index 045439f..60a474f 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.h
+++ b/lib/Target/AMDGPU/SIISelLowering.h
@@ -367,6 +367,8 @@ public:
                                     bool SNaN = false,
                                     unsigned Depth = 0) const override;
   AtomicExpansionKind shouldExpandAtomicRMWInIR(AtomicRMWInst *) const override;
+
+  unsigned getPrefLoopAlignment(MachineLoop *ML) const override;
 };
 
 } // End namespace llvm