https://github.com/lialan created https://github.com/llvm/llvm-project/pull/238174
Stacked on #237386. Use `RISCVISD::BuildPairGPRVec` directly when lowering RV32 packed byte build vectors. Constructing register pairs through `SCALAR_TO_VECTOR` and `CONCAT_VECTORS` lets the concat fold in #237386 turn them into sparse build vectors that unnecessarily pack undefined lanes. The direct construction restores `test_build_vector_i8` from 12 to 3 instructions and `test_purem_b` from 28 to 19. Validation: assertions-enabled `llc` rebuild; all 123 selected CodeGen tests pass; the complete affected RV32 file passes its pre-#237386 checks. Across 105 compared configurations, only the affected RV32 configuration changes relative to the bottom PR. >From 6654bca4957cea55a05c2f74043f594bc552257e Mon Sep 17 00:00:00 2001 From: Alan Li <[email protected]> Date: Mon, 2 Nov 2026 15:49:30 -0800 Subject: [PATCH] [RISCV] Build register pairs directly for byte vector lowering Use BuildPairGPRVec when lowering RV32 packed byte build vectors. The PPAIRE sequence only needs the low byte of each scalar register. Constructing pairs through scalar-to-vector concatenations lets DAG folding turn them into build vectors that unnecessarily pack undefined lanes. --- llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 22 ++- llvm/test/CodeGen/RISCV/rvp-simd-32.ll | 162 ++++++-------------- 2 files changed, 54 insertions(+), 130 deletions(-) diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp index 8072ba1b240c9d..f09eb11c7ed62b 100644 --- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp +++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp @@ -4798,20 +4798,16 @@ static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, if (VT == MVT::v4i8) { // <4 x i8> BUILD_VECTOR a, b, c, d -> PACK(PPACK.DH pair(a, c), pair(b, // d)) - SDValue Val0 = - DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i8, Op->getOperand(0)); - SDValue Val1 = - DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i8, Op->getOperand(1)); - SDValue Val2 = - DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i8, Op->getOperand(2)); - SDValue Val3 = - DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i8, Op->getOperand(3)); - SDValue Concat1 = - DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i8, Val0, Val2); - SDValue Concat2 = - DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i8, Val1, Val3); + SDValue Val0 = DAG.getAnyExtOrTrunc(Op->getOperand(0), DL, XLenVT); + SDValue Val1 = DAG.getAnyExtOrTrunc(Op->getOperand(1), DL, XLenVT); + SDValue Val2 = DAG.getAnyExtOrTrunc(Op->getOperand(2), DL, XLenVT); + SDValue Val3 = DAG.getAnyExtOrTrunc(Op->getOperand(3), DL, XLenVT); + SDValue Pair1 = + DAG.getNode(RISCVISD::BuildPairGPRVec, DL, MVT::v8i8, Val0, Val2); + SDValue Pair2 = + DAG.getNode(RISCVISD::BuildPairGPRVec, DL, MVT::v8i8, Val1, Val3); SDValue PPairE = - DAG.getNode(RISCVISD::PPAIRE, DL, MVT::v8i8, Concat1, Concat2); + DAG.getNode(RISCVISD::PPAIRE, DL, MVT::v8i8, Pair1, Pair2); SDValue Lo = DAG.getExtractSubvector(DL, MVT::v4i8, PPairE, 0); SDValue Hi = DAG.getExtractSubvector(DL, MVT::v4i8, PPairE, 4); diff --git a/llvm/test/CodeGen/RISCV/rvp-simd-32.ll b/llvm/test/CodeGen/RISCV/rvp-simd-32.ll index 48daae20325760..9d6e29422f2442 100644 --- a/llvm/test/CodeGen/RISCV/rvp-simd-32.ll +++ b/llvm/test/CodeGen/RISCV/rvp-simd-32.ll @@ -720,15 +720,6 @@ define <2 x i16> @test_padd_hs_splat_rhs(<2 x i16> %a, i16 %b) { define <4 x i8> @test_build_vector_i8(i8 %a, i8 %c, i8 %b, i8 %d) { ; RV32-LABEL: test_build_vector_i8: ; RV32: # %bb.0: -; RV32-NEXT: ppaire.b a3, a3, a0 -; RV32-NEXT: ppaire.b a4, a0, a0 -; RV32-NEXT: pack a3, a3, a4 -; RV32-NEXT: ppaire.b a2, a2, a0 -; RV32-NEXT: ppaire.b a1, a1, a0 -; RV32-NEXT: pack a1, a1, a4 -; RV32-NEXT: ppaire.b a0, a0, a0 -; RV32-NEXT: pack a2, a2, a4 -; RV32-NEXT: pack a0, a0, a4 ; RV32-NEXT: ppaire.db a0, a0, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret @@ -1000,24 +991,15 @@ define <4 x i8> @test_psll_bs_vec_shamt(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srli a2, a1, 24 ; RV32-NEXT: srli a3, a0, 24 -; RV32-NEXT: ppaire.b a4, a0, a0 -; RV32-NEXT: sll a2, a3, a2 -; RV32-NEXT: srli a3, a1, 8 -; RV32-NEXT: ppaire.b a2, a2, a0 -; RV32-NEXT: srli a5, a0, 8 -; RV32-NEXT: pack a7, a2, a4 -; RV32-NEXT: sll a2, a5, a3 -; RV32-NEXT: ppaire.b a2, a2, a0 -; RV32-NEXT: sll a3, a0, a1 +; RV32-NEXT: srli a4, a1, 8 +; RV32-NEXT: sll a3, a3, a2 +; RV32-NEXT: srli a2, a0, 8 +; RV32-NEXT: sll a6, a0, a1 ; RV32-NEXT: srli a1, a1, 16 ; RV32-NEXT: srli a0, a0, 16 -; RV32-NEXT: ppaire.b a3, a3, a0 -; RV32-NEXT: sll a0, a0, a1 -; RV32-NEXT: pack t1, a3, a4 -; RV32-NEXT: ppaire.b a0, a0, a0 -; RV32-NEXT: pack a6, a2, a4 -; RV32-NEXT: pack t2, a0, a4 -; RV32-NEXT: ppaire.db a0, t1, a6 +; RV32-NEXT: sll a2, a2, a4 +; RV32-NEXT: sll a7, a0, a1 +; RV32-NEXT: ppaire.db a0, a6, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret ; @@ -1173,26 +1155,17 @@ define <4 x i8> @test_psrl_bs_vec_shamt(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srli a2, a1, 24 ; RV32-NEXT: srli a3, a0, 24 -; RV32-NEXT: ppaire.b a4, a0, a0 -; RV32-NEXT: srl a2, a3, a2 -; RV32-NEXT: slli a3, a0, 16 -; RV32-NEXT: ppaire.b a2, a2, a0 -; RV32-NEXT: srli a5, a1, 8 -; RV32-NEXT: srli a6, a3, 24 -; RV32-NEXT: pack a3, a2, a4 -; RV32-NEXT: srl a2, a6, a5 -; RV32-NEXT: ppaire.b a2, a2, a0 +; RV32-NEXT: srli a4, a1, 8 +; RV32-NEXT: srl a3, a3, a2 +; RV32-NEXT: slli a2, a0, 16 +; RV32-NEXT: srli a2, a2, 24 ; RV32-NEXT: zext.b a5, a0 -; RV32-NEXT: srl a5, a5, a1 +; RV32-NEXT: srl a6, a5, a1 ; RV32-NEXT: slli a0, a0, 8 ; RV32-NEXT: srli a1, a1, 16 ; RV32-NEXT: srli a0, a0, 24 -; RV32-NEXT: ppaire.b a5, a5, a0 -; RV32-NEXT: srl a0, a0, a1 -; RV32-NEXT: pack a6, a5, a4 -; RV32-NEXT: ppaire.b a0, a0, a0 -; RV32-NEXT: pack a2, a2, a4 -; RV32-NEXT: pack a7, a0, a4 +; RV32-NEXT: srl a2, a2, a4 +; RV32-NEXT: srl a7, a0, a1 ; RV32-NEXT: ppaire.db a0, a6, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret @@ -1251,26 +1224,17 @@ define <4 x i8> @test_psra_bs_vec_shamt(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srli a2, a1, 24 ; RV32-NEXT: srai a3, a0, 24 -; RV32-NEXT: ppaire.b a4, a0, a0 -; RV32-NEXT: sra a2, a3, a2 -; RV32-NEXT: slli a3, a0, 16 -; RV32-NEXT: ppaire.b a2, a2, a0 -; RV32-NEXT: srli a5, a1, 8 -; RV32-NEXT: srai a6, a3, 24 -; RV32-NEXT: pack a3, a2, a4 -; RV32-NEXT: sra a2, a6, a5 -; RV32-NEXT: ppaire.b a2, a2, a0 +; RV32-NEXT: srli a4, a1, 8 +; RV32-NEXT: sra a3, a3, a2 +; RV32-NEXT: slli a2, a0, 16 +; RV32-NEXT: srai a2, a2, 24 ; RV32-NEXT: sext.b a5, a0 -; RV32-NEXT: sra a5, a5, a1 +; RV32-NEXT: sra a6, a5, a1 ; RV32-NEXT: slli a0, a0, 8 ; RV32-NEXT: srli a1, a1, 16 ; RV32-NEXT: srai a0, a0, 24 -; RV32-NEXT: ppaire.b a5, a5, a0 -; RV32-NEXT: sra a0, a0, a1 -; RV32-NEXT: pack a6, a5, a4 -; RV32-NEXT: ppaire.b a0, a0, a0 -; RV32-NEXT: pack a2, a2, a4 -; RV32-NEXT: pack a7, a0, a4 +; RV32-NEXT: sra a2, a2, a4 +; RV32-NEXT: sra a7, a0, a1 ; RV32-NEXT: ppaire.db a0, a6, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret @@ -1992,12 +1956,12 @@ define <4 x i8> @test_psdiv_b(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srai a2, a1, 24 ; RV32-NEXT: srai a3, a0, 24 -; RV32-NEXT: div a2, a3, a2 -; RV32-NEXT: slli a3, a1, 16 +; RV32-NEXT: div a3, a3, a2 +; RV32-NEXT: slli a2, a1, 16 ; RV32-NEXT: slli a4, a0, 16 -; RV32-NEXT: srai a3, a3, 24 +; RV32-NEXT: srai a2, a2, 24 ; RV32-NEXT: srai a4, a4, 24 -; RV32-NEXT: div a3, a4, a3 +; RV32-NEXT: div a2, a4, a2 ; RV32-NEXT: sext.b a4, a1 ; RV32-NEXT: sext.b a5, a0 ; RV32-NEXT: div a4, a5, a4 @@ -2005,17 +1969,8 @@ define <4 x i8> @test_psdiv_b(<4 x i8> %a, <4 x i8> %b) { ; RV32-NEXT: slli a0, a0, 8 ; RV32-NEXT: srai a1, a1, 24 ; RV32-NEXT: srai a0, a0, 24 -; RV32-NEXT: div a0, a0, a1 -; RV32-NEXT: ppaire.b a1, a2, a0 -; RV32-NEXT: ppaire.b a2, a0, a0 -; RV32-NEXT: pack a1, a1, a2 -; RV32-NEXT: ppaire.b a3, a3, a0 -; RV32-NEXT: ppaire.b a4, a4, a0 -; RV32-NEXT: pack a4, a4, a2 -; RV32-NEXT: ppaire.b a5, a0, a0 -; RV32-NEXT: pack a0, a3, a2 -; RV32-NEXT: pack a5, a5, a2 -; RV32-NEXT: ppaire.db a0, a4, a0 +; RV32-NEXT: div a5, a0, a1 +; RV32-NEXT: ppaire.db a0, a4, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret ; @@ -2102,12 +2057,12 @@ define <4 x i8> @test_pudiv_b(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srli a2, a1, 24 ; RV32-NEXT: srli a3, a0, 24 -; RV32-NEXT: divu a2, a3, a2 -; RV32-NEXT: slli a3, a1, 16 +; RV32-NEXT: divu a3, a3, a2 +; RV32-NEXT: slli a2, a1, 16 ; RV32-NEXT: slli a4, a0, 16 -; RV32-NEXT: srli a3, a3, 24 +; RV32-NEXT: srli a2, a2, 24 ; RV32-NEXT: srli a4, a4, 24 -; RV32-NEXT: divu a3, a4, a3 +; RV32-NEXT: divu a2, a4, a2 ; RV32-NEXT: zext.b a4, a1 ; RV32-NEXT: zext.b a5, a0 ; RV32-NEXT: divu a4, a5, a4 @@ -2115,17 +2070,8 @@ define <4 x i8> @test_pudiv_b(<4 x i8> %a, <4 x i8> %b) { ; RV32-NEXT: slli a0, a0, 8 ; RV32-NEXT: srli a1, a1, 24 ; RV32-NEXT: srli a0, a0, 24 -; RV32-NEXT: divu a0, a0, a1 -; RV32-NEXT: ppaire.b a1, a2, a0 -; RV32-NEXT: ppaire.b a2, a0, a0 -; RV32-NEXT: pack a1, a1, a2 -; RV32-NEXT: ppaire.b a3, a3, a0 -; RV32-NEXT: ppaire.b a4, a4, a0 -; RV32-NEXT: pack a4, a4, a2 -; RV32-NEXT: ppaire.b a5, a0, a0 -; RV32-NEXT: pack a0, a3, a2 -; RV32-NEXT: pack a5, a5, a2 -; RV32-NEXT: ppaire.db a0, a4, a0 +; RV32-NEXT: divu a5, a0, a1 +; RV32-NEXT: ppaire.db a0, a4, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret ; @@ -2208,12 +2154,12 @@ define <4 x i8> @test_psrem_b(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srai a2, a1, 24 ; RV32-NEXT: srai a3, a0, 24 -; RV32-NEXT: rem a2, a3, a2 -; RV32-NEXT: slli a3, a1, 16 +; RV32-NEXT: rem a3, a3, a2 +; RV32-NEXT: slli a2, a1, 16 ; RV32-NEXT: slli a4, a0, 16 -; RV32-NEXT: srai a3, a3, 24 +; RV32-NEXT: srai a2, a2, 24 ; RV32-NEXT: srai a4, a4, 24 -; RV32-NEXT: rem a3, a4, a3 +; RV32-NEXT: rem a2, a4, a2 ; RV32-NEXT: sext.b a4, a1 ; RV32-NEXT: sext.b a5, a0 ; RV32-NEXT: rem a4, a5, a4 @@ -2221,17 +2167,8 @@ define <4 x i8> @test_psrem_b(<4 x i8> %a, <4 x i8> %b) { ; RV32-NEXT: slli a0, a0, 8 ; RV32-NEXT: srai a1, a1, 24 ; RV32-NEXT: srai a0, a0, 24 -; RV32-NEXT: rem a0, a0, a1 -; RV32-NEXT: ppaire.b a1, a2, a0 -; RV32-NEXT: ppaire.b a2, a0, a0 -; RV32-NEXT: pack a1, a1, a2 -; RV32-NEXT: ppaire.b a3, a3, a0 -; RV32-NEXT: ppaire.b a4, a4, a0 -; RV32-NEXT: pack a4, a4, a2 -; RV32-NEXT: ppaire.b a5, a0, a0 -; RV32-NEXT: pack a0, a3, a2 -; RV32-NEXT: pack a5, a5, a2 -; RV32-NEXT: ppaire.db a0, a4, a0 +; RV32-NEXT: rem a5, a0, a1 +; RV32-NEXT: ppaire.db a0, a4, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret ; @@ -2294,12 +2231,12 @@ define <4 x i8> @test_purem_b(<4 x i8> %a, <4 x i8> %b) { ; RV32: # %bb.0: ; RV32-NEXT: srli a2, a1, 24 ; RV32-NEXT: srli a3, a0, 24 -; RV32-NEXT: remu a2, a3, a2 -; RV32-NEXT: slli a3, a1, 16 +; RV32-NEXT: remu a3, a3, a2 +; RV32-NEXT: slli a2, a1, 16 ; RV32-NEXT: slli a4, a0, 16 -; RV32-NEXT: srli a3, a3, 24 +; RV32-NEXT: srli a2, a2, 24 ; RV32-NEXT: srli a4, a4, 24 -; RV32-NEXT: remu a3, a4, a3 +; RV32-NEXT: remu a2, a4, a2 ; RV32-NEXT: zext.b a4, a1 ; RV32-NEXT: zext.b a5, a0 ; RV32-NEXT: remu a4, a5, a4 @@ -2307,17 +2244,8 @@ define <4 x i8> @test_purem_b(<4 x i8> %a, <4 x i8> %b) { ; RV32-NEXT: slli a0, a0, 8 ; RV32-NEXT: srli a1, a1, 24 ; RV32-NEXT: srli a0, a0, 24 -; RV32-NEXT: remu a0, a0, a1 -; RV32-NEXT: ppaire.b a1, a2, a0 -; RV32-NEXT: ppaire.b a2, a0, a0 -; RV32-NEXT: pack a1, a1, a2 -; RV32-NEXT: ppaire.b a3, a3, a0 -; RV32-NEXT: ppaire.b a4, a4, a0 -; RV32-NEXT: pack a4, a4, a2 -; RV32-NEXT: ppaire.b a5, a0, a0 -; RV32-NEXT: pack a0, a3, a2 -; RV32-NEXT: pack a5, a5, a2 -; RV32-NEXT: ppaire.db a0, a4, a0 +; RV32-NEXT: remu a5, a0, a1 +; RV32-NEXT: ppaire.db a0, a4, a2 ; RV32-NEXT: pack a0, a0, a1 ; RV32-NEXT: ret ; _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
