[llvm] [RISCV][P-ext] Improve the codegen for pzip (PR #208479)
Hongyu Chen via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 9 08:04:08 PDT 2026
https://github.com/XChy created https://github.com/llvm/llvm-project/pull/208479
This patch recognizes interleaved shuffle (as packed zip intrinsics use) and selects the corresponding instructions.
>From 1b47b03bc23ffbc1a55aa15614f97cb740a7135b Mon Sep 17 00:00:00 2001
From: XChy <xxs_chy at outlook.com>
Date: Thu, 9 Jul 2026 18:20:11 +0800
Subject: [PATCH] [RISCV][P-ext] Improve the codegen for pzip
---
llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 33 ++++++++++++++++++
llvm/lib/Target/RISCV/RISCVInstrInfoP.td | 21 ++++++++++++
llvm/test/CodeGen/RISCV/rvp-zip.ll | 38 +++------------------
3 files changed, 58 insertions(+), 34 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index df1692bdf9679..af2517a59d865 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -6314,6 +6314,36 @@ static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
return DAG.getBitcast(VT, DAG.getVectorShuffle(NewVT, DL, V0, V1, NewMask));
}
+// Match an interleave shuffle that forms a P-extension packed zip:
+// <a0, b0, a1, b1, ...> -> zip*p/wzip*p
+static SDValue lowerVECTOR_SHUFFLEAsPZip(ShuffleVectorSDNode *SVN,
+ SelectionDAG &DAG) {
+ SDValue V1 = SVN->getOperand(0);
+ SDValue V2 = SVN->getOperand(1);
+ SDLoc DL(SVN);
+ MVT VT = SVN->getSimpleValueType(0);
+ unsigned NumElts = VT.getVectorNumElements();
+ ArrayRef<int> Mask = SVN->getMask();
+
+ if ((VT != MVT::v8i8 && VT != MVT::v4i16) ||
+ V1.getSimpleValueType() != VT || V2.getSimpleValueType() != VT)
+ return SDValue();
+
+ SmallVector<unsigned, 2> StartIndexes;
+ if (!V2.isUndef() &&
+ ShuffleVectorInst::isInterleaveMask(Mask, 2, NumElts * 2,
+ StartIndexes)) {
+ int EvenSrc = StartIndexes[0];
+ int OddSrc = StartIndexes[1];
+ if (EvenSrc == 0 && OddSrc == (int)NumElts)
+ return DAG.getNode(RISCVISD::PZIP, DL, VT, V1, V2);
+ if (EvenSrc == (int)NumElts && OddSrc == 0)
+ return DAG.getNode(RISCVISD::PZIP, DL, VT, V2, V1);
+ }
+
+ return SDValue();
+}
+
SDValue RISCVTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
SelectionDAG &DAG) const {
SDValue V1 = Op.getOperand(0);
@@ -6348,6 +6378,9 @@ SDValue RISCVTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
DAG.getConstant(VT.getSizeInBits() / 2, DL, MVT::i64));
return DAG.getBitcast(VT, Srl);
}
+
+ if (SDValue V = lowerVECTOR_SHUFFLEAsPZip(SVN, DAG))
+ return V;
return SDValue();
}
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
index 1f547fe3ff2a2..7fa7c96ca99ca 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
@@ -1908,6 +1908,12 @@ def SDT_RISCVPackedUnary : SDTypeProfile<1, 1, [SDTCisVec<0>,
SDTCisSameAs<0, 1>]>;
def riscv_psabs : RVSDNode<"PSABS", SDT_RISCVPackedUnary>;
+// Interleave two packed vectors.
+def SDT_RISCVPZip : SDTypeProfile<1, 2, [SDTCisVec<0>,
+ SDTCisSameAs<0, 1>,
+ SDTCisSameAs<0, 2>]>;
+def riscv_pzip : RVSDNode<"PZIP", SDT_RISCVPZip>;
+
// Add one to the immediate. Used by RISCVISD::SATI.
def IncImm : SDNodeXForm<imm, [{
return CurDAG->getTargetConstant(N->getZExtValue() + 1, SDLoc(N),
@@ -2264,6 +2270,16 @@ let append Predicates = [IsRV32] in {
def : Pat<(v4i16 (zext (v4i8 GPR:$rs))), (WZIP8P GPR:$rs, (v4i8 X0))>;
def : Pat<(v2i32 (zext (v2i16 GPR:$rs))), (WZIP16P GPR:$rs, (v2i16 X0))>;
+ // Packed zip.
+ def : Pat<(v8i8 (riscv_pzip (v8i8 GPRPair:$rs1),
+ (v8i8 GPRPair:$rs2))),
+ (WZIP8P (i32 (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even)),
+ (i32 (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)))>;
+ def : Pat<(v4i16 (riscv_pzip (v4i16 GPRPair:$rs1),
+ (v4i16 GPRPair:$rs2))),
+ (WZIP16P (i32 (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even)),
+ (i32 (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)))>;
+
// Sign-extend patterns using pwadd with zero
def : Pat<(v4i16 (sext (v4i8 GPR:$rs))), (PWADD_B GPR:$rs, (v4i8 X0))>;
def : Pat<(v2i32 (sext (v2i16 GPR:$rs))), (PWADD_H GPR:$rs, (v2i16 X0))>;
@@ -2749,6 +2765,11 @@ let append Predicates = [IsRV64] in {
def : Pat<(v4i16 (zext_invec (v8i8 GPR:$rs))), (ZIP8P GPR:$rs, (v8i8 X0))>;
def : Pat<(v2i32 (zext_invec (v4i16 GPR:$rs))), (ZIP16P GPR:$rs, (v4i16 X0))>;
+ // Packed zip.
+ def : Pat<(v8i8 (riscv_pzip (v8i8 GPR:$rs1), (v8i8 GPR:$rs2))),
+ (ZIP8P GPR:$rs1, GPR:$rs2)>;
+ def : Pat<(v4i16 (riscv_pzip (v4i16 GPR:$rs1), (v4i16 GPR:$rs2))),
+ (ZIP16P GPR:$rs1, GPR:$rs2)>;
// Sign extend inreg patterns using psext. sext_invec is lowered to
// zext_invec+sext_inreg.
def : Pat<(v4i16 (sext_inreg GPR:$rs1, v4i8)), (PSEXT_H_B GPR:$rs1)>;
diff --git a/llvm/test/CodeGen/RISCV/rvp-zip.ll b/llvm/test/CodeGen/RISCV/rvp-zip.ll
index d639269c21eab..2f146afc23427 100644
--- a/llvm/test/CodeGen/RISCV/rvp-zip.ll
+++ b/llvm/test/CodeGen/RISCV/rvp-zip.ll
@@ -9,35 +9,12 @@
define <8 x i8> @test_pzip_v8i8(<4 x i8> %a, <4 x i8> %b) {
; RV32-LABEL: test_pzip_v8i8:
; RV32: # %bb.0:
-; RV32-NEXT: srli a2, a1, 24
-; RV32-NEXT: srli a3, a0, 24
-; RV32-NEXT: srli a4, a1, 16
-; RV32-NEXT: srli a5, a0, 16
-; RV32-NEXT: ppaire.b a2, a3, a2
-; RV32-NEXT: ppaire.b a3, a5, a4
-; RV32-NEXT: srli a4, a1, 8
-; RV32-NEXT: srli a5, a0, 8
-; RV32-NEXT: ppaire.b a0, a0, a1
-; RV32-NEXT: ppaire.b a4, a5, a4
-; RV32-NEXT: pack a1, a3, a2
-; RV32-NEXT: pack a0, a0, a4
+; RV32-NEXT: wzip8p a0, a0, a1
; RV32-NEXT: ret
;
; RV64-LABEL: test_pzip_v8i8:
; RV64: # %bb.0:
-; RV64-NEXT: srli a2, a1, 24
-; RV64-NEXT: srli a3, a0, 24
-; RV64-NEXT: srli a4, a1, 16
-; RV64-NEXT: srli a5, a0, 16
-; RV64-NEXT: ppaire.b a2, a3, a2
-; RV64-NEXT: ppaire.b a3, a5, a4
-; RV64-NEXT: srli a4, a1, 8
-; RV64-NEXT: srli a5, a0, 8
-; RV64-NEXT: ppaire.b a0, a0, a1
-; RV64-NEXT: ppaire.b a1, a5, a4
-; RV64-NEXT: ppaire.h a2, a3, a2
-; RV64-NEXT: ppaire.h a0, a0, a1
-; RV64-NEXT: pack a0, a0, a2
+; RV64-NEXT: zip8p a0, a0, a1
; RV64-NEXT: ret
%r = shufflevector <4 x i8> %a, <4 x i8> %b, <8 x i32> <i32 0, i32 4, i32 1, i32 5, i32 2, i32 6, i32 3, i32 7>
ret <8 x i8> %r
@@ -46,19 +23,12 @@ define <8 x i8> @test_pzip_v8i8(<4 x i8> %a, <4 x i8> %b) {
define <4 x i16> @test_pzip_v4i16(<2 x i16> %a, <2 x i16> %b) {
; RV32-LABEL: test_pzip_v4i16:
; RV32: # %bb.0:
-; RV32-NEXT: srli a2, a1, 16
-; RV32-NEXT: srli a3, a0, 16
-; RV32-NEXT: pack a0, a0, a1
-; RV32-NEXT: pack a1, a3, a2
+; RV32-NEXT: wzip16p a0, a0, a1
; RV32-NEXT: ret
;
; RV64-LABEL: test_pzip_v4i16:
; RV64: # %bb.0:
-; RV64-NEXT: srli a2, a1, 16
-; RV64-NEXT: srli a3, a0, 16
-; RV64-NEXT: ppaire.h a0, a0, a1
-; RV64-NEXT: ppaire.h a1, a3, a2
-; RV64-NEXT: pack a0, a0, a1
+; RV64-NEXT: zip16p a0, a0, a1
; RV64-NEXT: ret
%r = shufflevector <2 x i16> %a, <2 x i16> %b, <4 x i32> <i32 0, i32 2, i32 1, i32 3>
ret <4 x i16> %r
More information about the llvm-commits
mailing list