AMDGPU/SI: Disable unrolling in the loop vectorizer if the loop is not vectorized.

[cfang]

Summary

This additional unrolling (interleaving) will increase the register usage, and most likely hurts the performance.

[arsenm]

Can you add a test for this

[cfang]

Add a test case.

I was thinking that we should still be able to use the backend option: -force-target-max-scalar-interleave

But it seems the code in LoopVectorize.cpp prevents us doing this:

// Don't attempt if

// 1. the target claims to have no vector registers, and
// 2. interleaving won't help ILP.
//
// The second condition is necessary because, even if the target has no
// vector registers, loop vectorization may still enable scalar
// interleaving.
if (!TTI->getNumberOfRegisters(true) && TTI->getMaxInterleaveFactor(1) < 2) {
  return false;
}

[arsenm]

The test should go in test/Transforms/LoopVectorize/AMDGPU

[cfang]

Move test case to test/Transforms/LoopVectorize/AMDGPU

[arsenm]

You need to add an ew lit.local.cfg like the other target specific test directories

[cfang]

add lit.local.cfg file in the newly created test/Transforms/LoopVectorize/AMDGPU directory.

[arsenm]

LGTM

llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

 
 bool AMDGPUTTIImpl::isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
                                                  unsigned Alignment,
                                                  unsigned AddrSpace) const {
   return isLegalToVectorizeMemChain(ChainSizeInBytes, Alignment, AddrSpace);
 }
 
 unsigned AMDGPUTTIImpl::getMaxInterleaveFactor(unsigned VF) {
+  // Disable unrolling if the loop is not vectorized.
+  if (VF == 1)
+    return 1;
+
   // Semi-arbitrary large amount.
   return 64;
 }
 
 int AMDGPUTTIImpl::getArithmeticInstrCost(
     unsigned Opcode, Type *Ty, TTI::OperandValueKind Opd1Info,
     TTI::OperandValueKind Opd2Info, TTI::OperandValueProperties Opd1PropInfo,
     TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args ) {

llvm/trunk/test/Transforms/LoopVectorize/AMDGPU/lit.local.cfg

+if not 'AMDGPU' in config.root.targets:
+    config.unsupported = True

llvm/trunk/test/Transforms/LoopVectorize/AMDGPU/unroll-in-loop-vectorizer.ll

+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=fiji -loop-vectorize < %s | FileCheck %s
+
+
+; For AMDGPU, loop unroll in loop vectorizer is disabled when VF==1.
+;
+; CHECK-LABEL: @small_loop(
+; CHECK: store i32
+; CHECK-NOT: store i32
+; CHECK: ret
+define void @small_loop(i32* nocapture %inArray, i32 %size) nounwind {
+entry:
+  %0 = icmp sgt i32 %size, 0
+  br i1 %0, label %loop, label %exit
+
+loop:                                          ; preds = %entry, %loop
+  %iv = phi i32 [ %iv1, %loop ], [ 0, %entry ]
+  %1 = getelementptr inbounds i32, i32* %inArray, i32 %iv
+  %2 = load i32, i32* %1, align 4
+  %3 = add nsw i32 %2, 6
+  store i32 %3, i32* %1, align 4
+  %iv1 = add i32 %iv, 1
+;  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %cond = icmp eq i32 %iv1, %size
+  br i1 %cond, label %exit, label %loop
+
+exit:                                         ; preds = %loop, %entry
+  ret void
+}