[RISCV][IA] Use strided load for one active deinterleaveN(load)

[preames]

This adds the analogous handling we use for the shuffle lowering to the deinterleaveN intrinsic path.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Philip Reames (preames)

Changes

This adds the analogous handling we use for the shuffle lowering to the deinterleaveN intrinsic path.

---

Full diff: https://github.com/llvm/llvm-project/pull/148892.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp (+29-5)
• (modified) llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll (+7-4)

diff --git a/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp b/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp
index 39603b92cc2f7..7dac87b07e990 100644
--- a/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInterleavedAccess.cpp
@@ -243,20 +243,44 @@ bool RISCVTargetLowering::lowerDeinterleaveIntrinsicToLoad(
   assert(LI->isSimple());
   IRBuilder<> Builder(LI);
 
-  Value *FirstActive =
-      *llvm::find_if(DeinterleaveValues, [](Value *V) { return V != nullptr; });
-  VectorType *ResVTy = cast<VectorType>(FirstActive->getType());
+  auto FirstActiveItr =
+      llvm::find_if(DeinterleaveValues, [](Value *V) { return V != nullptr; });
+  VectorType *ResVTy = cast<VectorType>((*FirstActiveItr)->getType());
 
   const DataLayout &DL = LI->getDataLayout();
-
   if (!isLegalInterleavedAccessType(ResVTy, Factor, LI->getAlign(),
                                     LI->getPointerAddressSpace(), DL))
     return false;
 
-  Value *Return;
   Type *PtrTy = LI->getPointerOperandType();
   Type *XLenTy = Type::getIntNTy(LI->getContext(), Subtarget.getXLen());
 
+  // If the segment load is going to be performed segment at a time anyways
+  // and there's only one element used, use a strided load instead.  This
+  // will be equally fast, and create less vector register pressure.
+  if (!Subtarget.hasOptimizedSegmentLoadStore(Factor) &&
+      1 == llvm::count_if(DeinterleaveValues,
+                          [](Value *V) { return V != nullptr; })) {
+    unsigned Idx = std::distance(DeinterleaveValues.begin(), FirstActiveItr);
+    unsigned ScalarSizeInBytes = DL.getTypeStoreSize(ResVTy->getElementType());
+    Value *Stride = ConstantInt::get(XLenTy, Factor * ScalarSizeInBytes);
+    Value *Offset = ConstantInt::get(XLenTy, Idx * ScalarSizeInBytes);
+    Value *BasePtr = Builder.CreatePtrAdd(LI->getPointerOperand(), Offset);
+    Value *Mask = Builder.getAllOnesMask(ResVTy->getElementCount());
+    Type *I32 = Type::getIntNTy(LI->getContext(), 32);
+    Value *VL = Builder.CreateElementCount(I32, ResVTy->getElementCount());
+
+    CallInst *CI =
+        Builder.CreateIntrinsic(Intrinsic::experimental_vp_strided_load,
+                                {ResVTy, BasePtr->getType(), Stride->getType()},
+                                {BasePtr, Stride, Mask, VL});
+    Align A = commonAlignment(LI->getAlign(), Idx * ScalarSizeInBytes);
+    CI->addParamAttr(0, Attribute::getWithAlignment(CI->getContext(), A));
+    (*FirstActiveItr)->replaceAllUsesWith(CI);
+    return true;
+  }
+
+  Value *Return;
   if (isa<FixedVectorType>(ResVTy)) {
     Value *VL = Builder.CreateElementCount(XLenTy, ResVTy->getElementCount());
     Value *Mask = Builder.getAllOnesMask(ResVTy->getElementCount());
diff --git a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll
index 9af92aa995f1f..e28428224c2ec 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll
@@ -407,8 +407,9 @@ define { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x
 define <vscale x 8 x i8> @vector_deinterleave_load_factor4_oneactive(ptr %p) {
 ; CHECK-LABEL: vector_deinterleave_load_factor4_oneactive:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
-; CHECK-NEXT:    vlseg4e8.v v8, (a0)
+; CHECK-NEXT:    li a1, 4
+; CHECK-NEXT:    vsetvli a2, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vlse8.v v8, (a0), a1
 ; CHECK-NEXT:    ret
   %vec = load <vscale x 32 x i8>, ptr %p
   %d0 = call { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8> } @llvm.vector.deinterleave4(<vscale x 32 x i8> %vec)
@@ -419,8 +420,10 @@ define <vscale x 8 x i8> @vector_deinterleave_load_factor4_oneactive(ptr %p) {
 define <vscale x 8 x i8> @vector_deinterleave_load_factor4_oneactive2(ptr %p) {
 ; CHECK-LABEL: vector_deinterleave_load_factor4_oneactive2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a1, zero, e8, m1, ta, ma
-; CHECK-NEXT:    vlseg4e8.v v5, (a0)
+; CHECK-NEXT:    addi a0, a0, 3
+; CHECK-NEXT:    li a1, 4
+; CHECK-NEXT:    vsetvli a2, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vlse8.v v8, (a0), a1
 ; CHECK-NEXT:    ret
   %vec = load <vscale x 32 x i8>, ptr %p
   %d0 = call { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8> } @llvm.vector.deinterleave4(<vscale x 32 x i8> %vec)

[topperc]

Suggested change

	// If the segment load is going to be performed segment at a time anyways
	// If the segment load is going to be performed one segment at a time anyways

[topperc]

element -> field?

[topperc]

Suggested change

	// will be equally fast, and create less vector register pressure.
	// will be equally fast, and creates less vector register pressure.

[topperc]

Type::getInt32Ty?

[lukel97]

LGTM

[lukel97]

Nit

Suggested change

	1 == llvm::count_if(DeinterleaveValues,
	[](Value *V) { return V != nullptr; })) {
	llvm::count_if(DeinterleaveValues,
	[](Value *V) { return V != nullptr; }) == 1) {

[preames]

This patch conflicts with #148716. As I was already planning on reworking this to be a combine eventually, I'm going to close this for the moment, and post something new once the API has stabilized a bit.