[SelectionDAG][RISCV] Add support for splitting vp.splice #145184
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Use a stack based expansion similar to the non-VP splice. This code has been in our downstream for a while. I don't know how often is exercised though. Our downstream was missing clipping for the immediate value to keep it in range of the stack object so I've added it. As I'm writing this, I realize we might need to clip EVL1 when we create the pointer to store the second vector.
|
@llvm/pr-subscribers-llvm-selectiondag Author: Craig Topper (topperc) ChangesUse a stack based expansion similar to the non-VP splice. This code has been in our downstream for a while. I don't know how often is exercised though. Our downstream was missing clipping for the immediate value to keep it in range of the stack object so I've added it. As I'm writing this, I realize we might need to clip EVL1 when we create the pointer to store the second vector. Full diff: https://github.com/llvm/llvm-project/pull/145184.diff 3 Files Affected:
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h b/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
index a541833684f38..8643ae9d78159 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h
@@ -985,6 +985,7 @@ class LLVM_LIBRARY_VISIBILITY DAGTypeLegalizer {
void SplitVecRes_VECTOR_INTERLEAVE(SDNode *N);
void SplitVecRes_VAARG(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_FP_TO_XINT_SAT(SDNode *N, SDValue &Lo, SDValue &Hi);
+ void SplitVecRes_VP_SPLICE(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_VP_REVERSE(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_PARTIAL_REDUCE_MLA(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_GET_ACTIVE_LANE_MASK(SDNode *N, SDValue &Lo, SDValue &Hi);
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index c56cfec81acdd..745f386e5a30f 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -1382,6 +1382,9 @@ void DAGTypeLegalizer::SplitVectorResult(SDNode *N, unsigned ResNo) {
case ISD::UDIVFIXSAT:
SplitVecRes_FIX(N, Lo, Hi);
break;
+ case ISD::EXPERIMENTAL_VP_SPLICE:
+ SplitVecRes_VP_SPLICE(N, Lo, Hi);
+ break;
case ISD::EXPERIMENTAL_VP_REVERSE:
SplitVecRes_VP_REVERSE(N, Lo, Hi);
break;
@@ -3209,6 +3212,79 @@ void DAGTypeLegalizer::SplitVecRes_VP_REVERSE(SDNode *N, SDValue &Lo,
std::tie(Lo, Hi) = DAG.SplitVector(Load, DL);
}
+void DAGTypeLegalizer::SplitVecRes_VP_SPLICE(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ EVT VT = N->getValueType(0);
+ SDValue V1 = N->getOperand(0);
+ SDValue V2 = N->getOperand(1);
+ int64_t Imm = cast<ConstantSDNode>(N->getOperand(2))->getSExtValue();
+ SDValue Mask = N->getOperand(3);
+ SDValue EVL1 = N->getOperand(4);
+ SDValue EVL2 = N->getOperand(5);
+ SDLoc DL(N);
+
+ // Since EVL2 is considered the real VL it gets promoted during
+ // SelectionDAGBuilder. Promote EVL1 here if needed.
+ if (getTypeAction(EVL1.getValueType()) == TargetLowering::TypePromoteInteger)
+ EVL1 = ZExtPromotedInteger(EVL1);
+
+ Align Alignment = DAG.getReducedAlign(VT, /*UseABI=*/false);
+
+ EVT MemVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+ VT.getVectorElementCount() * 2);
+ SDValue StackPtr = DAG.CreateStackTemporary(MemVT.getStoreSize(), Alignment);
+ EVT PtrVT = StackPtr.getValueType();
+ auto &MF = DAG.getMachineFunction();
+ auto FrameIndex = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
+ auto PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIndex);
+
+ MachineMemOperand *StoreMMO = DAG.getMachineFunction().getMachineMemOperand(
+ PtrInfo, MachineMemOperand::MOStore, LocationSize::beforeOrAfterPointer(),
+ Alignment);
+ MachineMemOperand *LoadMMO = DAG.getMachineFunction().getMachineMemOperand(
+ PtrInfo, MachineMemOperand::MOLoad, LocationSize::beforeOrAfterPointer(),
+ Alignment);
+
+ unsigned EltWidth = VT.getScalarSizeInBits() / 8;
+ SDValue OffsetToV2 =
+ DAG.getNode(ISD::MUL, DL, PtrVT, DAG.getZExtOrTrunc(EVL1, DL, PtrVT),
+ DAG.getConstant(EltWidth, DL, PtrVT));
+ SDValue StackPtr2 = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, OffsetToV2);
+
+ SDValue TrueMask = DAG.getBoolConstant(true, DL, Mask.getValueType(), VT);
+ SDValue StoreV1 = DAG.getStoreVP(DAG.getEntryNode(), DL, V1, StackPtr,
+ DAG.getUNDEF(PtrVT), TrueMask, EVL1,
+ V1.getValueType(), StoreMMO, ISD::UNINDEXED);
+
+ SDValue StoreV2 =
+ DAG.getStoreVP(StoreV1, DL, V2, StackPtr2, DAG.getUNDEF(PtrVT), TrueMask,
+ EVL2, V2.getValueType(), StoreMMO, ISD::UNINDEXED);
+
+ SDValue Load;
+ if (Imm >= 0) {
+ StackPtr = TLI.getVectorElementPointer(DAG, StackPtr, VT, N->getOperand(2));
+ Load = DAG.getLoadVP(VT, DL, StoreV2, StackPtr, Mask, EVL2, LoadMMO);
+ } else {
+ uint64_t TrailingElts = -Imm;
+ SDValue TrailingBytes = DAG.getConstant(TrailingElts * EltWidth, DL, PtrVT);
+
+ // Make sure TrailingBytes doesn't exceed the size of vec1.
+ TrailingBytes = DAG.getNode(ISD::UMIN, DL, PtrVT, TrailingBytes, OffsetToV2);
+
+ // Calculate the start address of the spliced result.
+ StackPtr2 = DAG.getNode(ISD::SUB, DL, PtrVT, StackPtr2, TrailingBytes);
+ Load = DAG.getLoadVP(VT, DL, StoreV2, StackPtr2, Mask, EVL2, LoadMMO);
+ }
+
+ EVT LoVT, HiVT;
+ std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
+ Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, LoVT, Load,
+ DAG.getVectorIdxConstant(0, DL));
+ Hi =
+ DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HiVT, Load,
+ DAG.getVectorIdxConstant(LoVT.getVectorMinNumElements(), DL));
+}
+
void DAGTypeLegalizer::SplitVecRes_PARTIAL_REDUCE_MLA(SDNode *N, SDValue &Lo,
SDValue &Hi) {
SDLoc DL(N);
diff --git a/llvm/test/CodeGen/RISCV/rvv/vp-splice.ll b/llvm/test/CodeGen/RISCV/rvv/vp-splice.ll
index a4f91c3e7c99e..a90de98cdcb81 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vp-splice.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vp-splice.ll
@@ -286,3 +286,130 @@ define <vscale x 2 x float> @test_vp_splice_nxv2f32_masked(<vscale x 2 x float>
%v = call <vscale x 2 x float> @llvm.experimental.vp.splice.nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, i32 5, <vscale x 2 x i1> %mask, i32 %evla, i32 %evlb)
ret <vscale x 2 x float> %v
}
+
+define <vscale x 16 x i64> @test_vp_splice_nxv16i64(<vscale x 16 x i64> %va, <vscale x 16 x i64> %vb, i32 zeroext %evla, i32 zeroext %evlb) nounwind {
+; CHECK-LABEL: test_vp_splice_nxv16i64:
+; CHECK: # %bb.0:
+; CHECK-NEXT: csrr a4, vlenb
+; CHECK-NEXT: slli a1, a4, 3
+; CHECK-NEXT: add a5, a0, a1
+; CHECK-NEXT: vl8re64.v v24, (a5)
+; CHECK-NEXT: mv a5, a2
+; CHECK-NEXT: bltu a2, a4, .LBB21_2
+; CHECK-NEXT: # %bb.1:
+; CHECK-NEXT: mv a5, a4
+; CHECK-NEXT: .LBB21_2:
+; CHECK-NEXT: addi sp, sp, -80
+; CHECK-NEXT: sd ra, 72(sp) # 8-byte Folded Spill
+; CHECK-NEXT: sd s0, 64(sp) # 8-byte Folded Spill
+; CHECK-NEXT: addi s0, sp, 80
+; CHECK-NEXT: csrr a6, vlenb
+; CHECK-NEXT: slli a6, a6, 5
+; CHECK-NEXT: sub sp, sp, a6
+; CHECK-NEXT: andi sp, sp, -64
+; CHECK-NEXT: vl8re64.v v0, (a0)
+; CHECK-NEXT: addi a0, sp, 64
+; CHECK-NEXT: sub a6, a2, a4
+; CHECK-NEXT: vsetvli zero, a5, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v8, (a0)
+; CHECK-NEXT: sltu a5, a2, a6
+; CHECK-NEXT: addi a5, a5, -1
+; CHECK-NEXT: and a5, a5, a6
+; CHECK-NEXT: add a6, a0, a1
+; CHECK-NEXT: vsetvli zero, a5, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v16, (a6)
+; CHECK-NEXT: sub a5, a3, a4
+; CHECK-NEXT: slli a2, a2, 3
+; CHECK-NEXT: sltu a6, a3, a5
+; CHECK-NEXT: add a2, a0, a2
+; CHECK-NEXT: addi a0, a6, -1
+; CHECK-NEXT: add a6, a2, a1
+; CHECK-NEXT: and a0, a0, a5
+; CHECK-NEXT: vsetvli zero, a0, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v24, (a6)
+; CHECK-NEXT: bltu a3, a4, .LBB21_4
+; CHECK-NEXT: # %bb.3:
+; CHECK-NEXT: mv a3, a4
+; CHECK-NEXT: .LBB21_4:
+; CHECK-NEXT: vsetvli zero, a3, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v0, (a2)
+; CHECK-NEXT: addi a2, sp, 104
+; CHECK-NEXT: add a1, a2, a1
+; CHECK-NEXT: vsetvli zero, a0, e64, m8, ta, ma
+; CHECK-NEXT: vle64.v v16, (a1)
+; CHECK-NEXT: vsetvli zero, a3, e64, m8, ta, ma
+; CHECK-NEXT: vle64.v v8, (a2)
+; CHECK-NEXT: addi sp, s0, -80
+; CHECK-NEXT: ld ra, 72(sp) # 8-byte Folded Reload
+; CHECK-NEXT: ld s0, 64(sp) # 8-byte Folded Reload
+; CHECK-NEXT: addi sp, sp, 80
+; CHECK-NEXT: ret
+ %v = call <vscale x 16 x i64> @llvm.experimental.vp.splice.nxv16i64(<vscale x 16 x i64> %va, <vscale x 16 x i64> %vb, i32 5, <vscale x 16 x i1> splat (i1 1), i32 %evla, i32 %evlb)
+ ret <vscale x 16 x i64> %v
+}
+
+define <vscale x 16 x i64> @test_vp_splice_nxv16i64_negative_offset(<vscale x 16 x i64> %va, <vscale x 16 x i64> %vb, i32 zeroext %evla, i32 zeroext %evlb) nounwind {
+; CHECK-LABEL: test_vp_splice_nxv16i64_negative_offset:
+; CHECK: # %bb.0:
+; CHECK-NEXT: csrr a4, vlenb
+; CHECK-NEXT: slli a1, a4, 3
+; CHECK-NEXT: add a5, a0, a1
+; CHECK-NEXT: vl8re64.v v0, (a5)
+; CHECK-NEXT: mv a5, a2
+; CHECK-NEXT: bltu a2, a4, .LBB22_2
+; CHECK-NEXT: # %bb.1:
+; CHECK-NEXT: mv a5, a4
+; CHECK-NEXT: .LBB22_2:
+; CHECK-NEXT: addi sp, sp, -80
+; CHECK-NEXT: sd ra, 72(sp) # 8-byte Folded Spill
+; CHECK-NEXT: sd s0, 64(sp) # 8-byte Folded Spill
+; CHECK-NEXT: addi s0, sp, 80
+; CHECK-NEXT: csrr a6, vlenb
+; CHECK-NEXT: slli a6, a6, 5
+; CHECK-NEXT: sub sp, sp, a6
+; CHECK-NEXT: andi sp, sp, -64
+; CHECK-NEXT: vl8re64.v v24, (a0)
+; CHECK-NEXT: addi a0, sp, 64
+; CHECK-NEXT: sub a6, a2, a4
+; CHECK-NEXT: vsetvli zero, a5, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v8, (a0)
+; CHECK-NEXT: sltu a5, a2, a6
+; CHECK-NEXT: addi a5, a5, -1
+; CHECK-NEXT: and a5, a5, a6
+; CHECK-NEXT: add a6, a0, a1
+; CHECK-NEXT: vsetvli zero, a5, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v16, (a6)
+; CHECK-NEXT: sub a6, a3, a4
+; CHECK-NEXT: slli a2, a2, 3
+; CHECK-NEXT: sltu a7, a3, a6
+; CHECK-NEXT: add a5, a0, a2
+; CHECK-NEXT: addi a7, a7, -1
+; CHECK-NEXT: and a0, a7, a6
+; CHECK-NEXT: add a6, a5, a1
+; CHECK-NEXT: vsetvli zero, a0, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v0, (a6)
+; CHECK-NEXT: bltu a3, a4, .LBB22_4
+; CHECK-NEXT: # %bb.3:
+; CHECK-NEXT: mv a3, a4
+; CHECK-NEXT: .LBB22_4:
+; CHECK-NEXT: li a4, 8
+; CHECK-NEXT: vsetvli zero, a3, e64, m8, ta, ma
+; CHECK-NEXT: vse64.v v24, (a5)
+; CHECK-NEXT: bltu a2, a4, .LBB22_6
+; CHECK-NEXT: # %bb.5:
+; CHECK-NEXT: li a2, 8
+; CHECK-NEXT: .LBB22_6:
+; CHECK-NEXT: sub a5, a5, a2
+; CHECK-NEXT: add a1, a5, a1
+; CHECK-NEXT: vsetvli zero, a0, e64, m8, ta, ma
+; CHECK-NEXT: vle64.v v16, (a1)
+; CHECK-NEXT: vsetvli zero, a3, e64, m8, ta, ma
+; CHECK-NEXT: vle64.v v8, (a5)
+; CHECK-NEXT: addi sp, s0, -80
+; CHECK-NEXT: ld ra, 72(sp) # 8-byte Folded Reload
+; CHECK-NEXT: ld s0, 64(sp) # 8-byte Folded Reload
+; CHECK-NEXT: addi sp, sp, 80
+; CHECK-NEXT: ret
+ %v = call <vscale x 16 x i64> @llvm.experimental.vp.splice.nxv16i64(<vscale x 16 x i64> %va, <vscale x 16 x i64> %vb, i32 -1, <vscale x 16 x i1> splat (i1 1), i32 %evla, i32 %evlb)
+ ret <vscale x 16 x i64> %v
+}
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Use a stack based expansion similar to the non-VP splice.
This code has been in our downstream for a while. I don't know how often is exercised though. Our downstream was missing clipping for the immediate value to keep it in range of the stack object so I've added it.
As I'm writing this, I realize we might need to clip EVL1 when we create the pointer to store the second vector.