[llvm] [RISCV] Add intrinsics for strided segment stores with fixed vectors (PR #152038)
Min-Yih Hsu via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 8 14:07:08 PDT 2025
https://github.com/mshockwave updated https://github.com/llvm/llvm-project/pull/152038
>From bf4b7f17b1eb3fe1706f89d710c89cddd3433e81 Mon Sep 17 00:00:00 2001
From: Min-Yih Hsu <min.hsu at sifive.com>
Date: Mon, 4 Aug 2025 14:32:46 -0700
Subject: [PATCH 1/3] [RISCV] Add intrinsics for strided segment stores with
fixed vectors
---
llvm/include/llvm/IR/IntrinsicsRISCV.td | 11 ++
llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 159 ++++++++++++------
.../RISCV/rvv/fixed-vectors-ssegN-store.ll | 72 ++++++++
3 files changed, 192 insertions(+), 50 deletions(-)
create mode 100644 llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ssegN-store.ll
diff --git a/llvm/include/llvm/IR/IntrinsicsRISCV.td b/llvm/include/llvm/IR/IntrinsicsRISCV.td
index 99f975faeb85e..c56e390d578a4 100644
--- a/llvm/include/llvm/IR/IntrinsicsRISCV.td
+++ b/llvm/include/llvm/IR/IntrinsicsRISCV.td
@@ -1736,6 +1736,17 @@ let TargetPrefix = "riscv" in {
[llvm_anyptr_ty, LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>,
llvm_anyint_ty]),
[NoCapture<ArgIndex<nf>>, IntrWriteMem]>;
+
+ // Input: (<stored values>..., pointer, offset, mask, vl)
+ def int_riscv_sseg # nf # _store_mask
+ : DefaultAttrsIntrinsic<[],
+ !listconcat([llvm_anyvector_ty],
+ !listsplat(LLVMMatchType<0>,
+ !add(nf, -1)),
+ [llvm_anyptr_ty, llvm_anyint_ty,
+ LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>,
+ llvm_anyint_ty]),
+ [NoCapture<ArgIndex<nf>>, IntrWriteMem]>;
}
} // TargetPrefix = "riscv"
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index e09e6fb5b26ee..671a107cc8f6b 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -1839,6 +1839,17 @@ bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 3,
/*IsStore*/ true,
/*IsUnitStrided*/ false, /*UsePtrVal*/ true);
+ case Intrinsic::riscv_sseg2_store_mask:
+ case Intrinsic::riscv_sseg3_store_mask:
+ case Intrinsic::riscv_sseg4_store_mask:
+ case Intrinsic::riscv_sseg5_store_mask:
+ case Intrinsic::riscv_sseg6_store_mask:
+ case Intrinsic::riscv_sseg7_store_mask:
+ case Intrinsic::riscv_sseg8_store_mask:
+ // Operands are (vec, ..., vec, ptr, offset, mask, vl)
+ return SetRVVLoadStoreInfo(/*PtrOp*/ I.arg_size() - 4,
+ /*IsStore*/ true,
+ /*IsUnitStrided*/ false, /*UsePtrVal*/ true);
case Intrinsic::riscv_vlm:
return SetRVVLoadStoreInfo(/*PtrOp*/ 0,
/*IsStore*/ false,
@@ -11077,69 +11088,117 @@ SDValue RISCVTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
return lowerVectorIntrinsicScalars(Op, DAG, Subtarget);
}
-SDValue RISCVTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
- SelectionDAG &DAG) const {
- unsigned IntNo = Op.getConstantOperandVal(1);
+static SDValue
+lowerFixedVectorSegStoreIntrinsics(unsigned IntNo, SDValue Op,
+ const RISCVSubtarget &Subtarget,
+ SelectionDAG &DAG) {
+ bool IsStrided;
switch (IntNo) {
- default:
- break;
case Intrinsic::riscv_seg2_store_mask:
case Intrinsic::riscv_seg3_store_mask:
case Intrinsic::riscv_seg4_store_mask:
case Intrinsic::riscv_seg5_store_mask:
case Intrinsic::riscv_seg6_store_mask:
case Intrinsic::riscv_seg7_store_mask:
- case Intrinsic::riscv_seg8_store_mask: {
- SDLoc DL(Op);
- static const Intrinsic::ID VssegInts[] = {
- Intrinsic::riscv_vsseg2_mask, Intrinsic::riscv_vsseg3_mask,
- Intrinsic::riscv_vsseg4_mask, Intrinsic::riscv_vsseg5_mask,
- Intrinsic::riscv_vsseg6_mask, Intrinsic::riscv_vsseg7_mask,
- Intrinsic::riscv_vsseg8_mask};
+ case Intrinsic::riscv_seg8_store_mask:
+ IsStrided = false;
+ break;
+ case Intrinsic::riscv_sseg2_store_mask:
+ case Intrinsic::riscv_sseg3_store_mask:
+ case Intrinsic::riscv_sseg4_store_mask:
+ case Intrinsic::riscv_sseg5_store_mask:
+ case Intrinsic::riscv_sseg6_store_mask:
+ case Intrinsic::riscv_sseg7_store_mask:
+ case Intrinsic::riscv_sseg8_store_mask:
+ IsStrided = true;
+ break;
+ default:
+ llvm_unreachable("unexpected intrinsic ID");
+ }
- // Operands: (chain, int_id, vec*, ptr, mask, vl)
- unsigned NF = Op->getNumOperands() - 5;
- assert(NF >= 2 && NF <= 8 && "Unexpected seg number");
- MVT XLenVT = Subtarget.getXLenVT();
- MVT VT = Op->getOperand(2).getSimpleValueType();
- MVT ContainerVT = getContainerForFixedLengthVector(VT);
- unsigned Sz = NF * ContainerVT.getVectorMinNumElements() *
- ContainerVT.getScalarSizeInBits();
- EVT VecTupTy = MVT::getRISCVVectorTupleVT(Sz, NF);
+ SDLoc DL(Op);
+ static const Intrinsic::ID VssegInts[] = {
+ Intrinsic::riscv_vsseg2_mask, Intrinsic::riscv_vsseg3_mask,
+ Intrinsic::riscv_vsseg4_mask, Intrinsic::riscv_vsseg5_mask,
+ Intrinsic::riscv_vsseg6_mask, Intrinsic::riscv_vsseg7_mask,
+ Intrinsic::riscv_vsseg8_mask};
+ static const Intrinsic::ID VsssegInts[] = {
+ Intrinsic::riscv_vssseg2_mask, Intrinsic::riscv_vssseg3_mask,
+ Intrinsic::riscv_vssseg4_mask, Intrinsic::riscv_vssseg5_mask,
+ Intrinsic::riscv_vssseg6_mask, Intrinsic::riscv_vssseg7_mask,
+ Intrinsic::riscv_vssseg8_mask};
+
+ // Operands: (chain, int_id, vec*, ptr, mask, vl) or
+ // (chain, int_id, vec*, ptr, stride, mask, vl)
+ unsigned NF = Op->getNumOperands() - (IsStrided ? 6 : 5);
+ assert(NF >= 2 && NF <= 8 && "Unexpected seg number");
+ MVT XLenVT = Subtarget.getXLenVT();
+ MVT VT = Op->getOperand(2).getSimpleValueType();
+ MVT ContainerVT = ::getContainerForFixedLengthVector(DAG, VT, Subtarget);
+ unsigned Sz = NF * ContainerVT.getVectorMinNumElements() *
+ ContainerVT.getScalarSizeInBits();
+ EVT VecTupTy = MVT::getRISCVVectorTupleVT(Sz, NF);
+
+ SDValue VL = Op.getOperand(Op.getNumOperands() - 1);
+ SDValue Mask = Op.getOperand(Op.getNumOperands() - 2);
+ MVT MaskVT = Mask.getSimpleValueType();
+ MVT MaskContainerVT =
+ ::getContainerForFixedLengthVector(DAG, MaskVT, Subtarget);
+ Mask = convertToScalableVector(MaskContainerVT, Mask, DAG, Subtarget);
- SDValue VL = Op.getOperand(Op.getNumOperands() - 1);
- SDValue Mask = Op.getOperand(Op.getNumOperands() - 2);
- MVT MaskVT = Mask.getSimpleValueType();
- MVT MaskContainerVT =
- ::getContainerForFixedLengthVector(DAG, MaskVT, Subtarget);
- Mask = convertToScalableVector(MaskContainerVT, Mask, DAG, Subtarget);
+ SDValue IntID = DAG.getTargetConstant(
+ IsStrided ? VsssegInts[NF - 2] : VssegInts[NF - 2], DL, XLenVT);
+ SDValue Ptr = Op->getOperand(NF + 2);
- SDValue IntID = DAG.getTargetConstant(VssegInts[NF - 2], DL, XLenVT);
- SDValue Ptr = Op->getOperand(NF + 2);
+ auto *FixedIntrinsic = cast<MemIntrinsicSDNode>(Op);
- auto *FixedIntrinsic = cast<MemIntrinsicSDNode>(Op);
+ SDValue StoredVal = DAG.getUNDEF(VecTupTy);
+ for (unsigned i = 0; i < NF; i++)
+ StoredVal = DAG.getNode(
+ RISCVISD::TUPLE_INSERT, DL, VecTupTy, StoredVal,
+ convertToScalableVector(ContainerVT, FixedIntrinsic->getOperand(2 + i),
+ DAG, Subtarget),
+ DAG.getTargetConstant(i, DL, MVT::i32));
- SDValue StoredVal = DAG.getUNDEF(VecTupTy);
- for (unsigned i = 0; i < NF; i++)
- StoredVal = DAG.getNode(
- RISCVISD::TUPLE_INSERT, DL, VecTupTy, StoredVal,
- convertToScalableVector(
- ContainerVT, FixedIntrinsic->getOperand(2 + i), DAG, Subtarget),
- DAG.getTargetConstant(i, DL, MVT::i32));
+ SmallVector<SDValue, 10> Ops = {
+ FixedIntrinsic->getChain(),
+ IntID,
+ StoredVal,
+ Ptr,
+ Mask,
+ VL,
+ DAG.getTargetConstant(Log2_64(VT.getScalarSizeInBits()), DL, XLenVT)};
+ if (IsStrided)
+ Ops.insert(std::next(Ops.begin(), 4),
+ Op.getOperand(Op.getNumOperands() - 3));
+
+ return DAG.getMemIntrinsicNode(
+ ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other), Ops,
+ FixedIntrinsic->getMemoryVT(), FixedIntrinsic->getMemOperand());
+}
+
+SDValue RISCVTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
+ SelectionDAG &DAG) const {
+ unsigned IntNo = Op.getConstantOperandVal(1);
+ switch (IntNo) {
+ default:
+ break;
+ case Intrinsic::riscv_seg2_store_mask:
+ case Intrinsic::riscv_seg3_store_mask:
+ case Intrinsic::riscv_seg4_store_mask:
+ case Intrinsic::riscv_seg5_store_mask:
+ case Intrinsic::riscv_seg6_store_mask:
+ case Intrinsic::riscv_seg7_store_mask:
+ case Intrinsic::riscv_seg8_store_mask:
+ case Intrinsic::riscv_sseg2_store_mask:
+ case Intrinsic::riscv_sseg3_store_mask:
+ case Intrinsic::riscv_sseg4_store_mask:
+ case Intrinsic::riscv_sseg5_store_mask:
+ case Intrinsic::riscv_sseg6_store_mask:
+ case Intrinsic::riscv_sseg7_store_mask:
+ case Intrinsic::riscv_sseg8_store_mask:
+ return lowerFixedVectorSegStoreIntrinsics(IntNo, Op, Subtarget, DAG);
- SDValue Ops[] = {
- FixedIntrinsic->getChain(),
- IntID,
- StoredVal,
- Ptr,
- Mask,
- VL,
- DAG.getTargetConstant(Log2_64(VT.getScalarSizeInBits()), DL, XLenVT)};
-
- return DAG.getMemIntrinsicNode(
- ISD::INTRINSIC_VOID, DL, DAG.getVTList(MVT::Other), Ops,
- FixedIntrinsic->getMemoryVT(), FixedIntrinsic->getMemOperand());
- }
case Intrinsic::riscv_sf_vc_xv_se:
return getVCIXISDNodeVOID(Op, DAG, RISCVISD::SF_VC_XV_SE);
case Intrinsic::riscv_sf_vc_iv_se:
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ssegN-store.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ssegN-store.ll
new file mode 100644
index 0000000000000..abf289419ec16
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-ssegN-store.ll
@@ -0,0 +1,72 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=riscv64 -mattr=+v -verify-machineinstrs < %s | FileCheck %s
+
+define void @store_factor2(<8 x i8> %v0, <8 x i8> %v1, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor2:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg2e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg2.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
+
+define void @store_factor3(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor3:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg3e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg3.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
+
+define void @store_factor4(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor4:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg4e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg4.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
+
+define void @store_factor5(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor5:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg5e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg5.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
+
+define void @store_factor6(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, <8 x i8> %v5, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor6:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg6e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg6.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, <8 x i8> %v5, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
+
+define void @store_factor7(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, <8 x i8> %v5, <8 x i8> %v6, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor7:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg7e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg7.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, <8 x i8> %v5, <8 x i8> %v6, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
+
+define void @store_factor8(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, <8 x i8> %v5, <8 x i8> %v6, <8 x i8> %v7, ptr %ptr, i64 %stride) {
+; CHECK-LABEL: store_factor8:
+; CHECK: # %bb.0:
+; CHECK-NEXT: vsetivli zero, 8, e8, mf2, ta, ma
+; CHECK-NEXT: vssseg8e8.v v8, (a0), a1
+; CHECK-NEXT: ret
+ call void @llvm.riscv.sseg8.store.mask.v8i8.i64.i64(<8 x i8> %v0, <8 x i8> %v1, <8 x i8> %v2, <8 x i8> %v3, <8 x i8> %v4, <8 x i8> %v5, <8 x i8> %v6, <8 x i8> %v7, ptr %ptr, i64 %stride, <8 x i1> splat (i1 true), i64 8)
+ ret void
+}
>From 1e07d8447845ceba957a4c78774c97d521ce3d4f Mon Sep 17 00:00:00 2001
From: Min-Yih Hsu <min.hsu at sifive.com>
Date: Mon, 4 Aug 2025 14:47:05 -0700
Subject: [PATCH 2/3] fixup! [RISCV] Add intrinsics for strided segment stores
with fixed vectors
---
llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 1 +
1 file changed, 1 insertion(+)
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 671a107cc8f6b..e3ad6af4ad024 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -11168,6 +11168,7 @@ lowerFixedVectorSegStoreIntrinsics(unsigned IntNo, SDValue Op,
Mask,
VL,
DAG.getTargetConstant(Log2_64(VT.getScalarSizeInBits()), DL, XLenVT)};
+ // Insert the stride operand.
if (IsStrided)
Ops.insert(std::next(Ops.begin(), 4),
Op.getOperand(Op.getNumOperands() - 3));
>From d94082dbc044db49f64665225ce2dd97b2dd4cf2 Mon Sep 17 00:00:00 2001
From: Min-Yih Hsu <min.hsu at sifive.com>
Date: Fri, 8 Aug 2025 14:05:24 -0700
Subject: [PATCH 3/3] fixup! Rename offset to stride
---
llvm/include/llvm/IR/IntrinsicsRISCV.td | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/include/llvm/IR/IntrinsicsRISCV.td b/llvm/include/llvm/IR/IntrinsicsRISCV.td
index c56e390d578a4..2d44c8e0d6e25 100644
--- a/llvm/include/llvm/IR/IntrinsicsRISCV.td
+++ b/llvm/include/llvm/IR/IntrinsicsRISCV.td
@@ -1737,7 +1737,7 @@ let TargetPrefix = "riscv" in {
llvm_anyint_ty]),
[NoCapture<ArgIndex<nf>>, IntrWriteMem]>;
- // Input: (<stored values>..., pointer, offset, mask, vl)
+ // Input: (<stored values>..., pointer, stride, mask, vl)
def int_riscv_sseg # nf # _store_mask
: DefaultAttrsIntrinsic<[],
!listconcat([llvm_anyvector_ty],
More information about the llvm-commits
mailing list