[clang] [llvm] [RISCV][P-ext] Add scalar ssh1sadd intrinsic (PR #239234)
via cfe-commits
cfe-commits at lists.llvm.org
Wed Nov 11 20:18:47 PST 2026
https://github.com/sihuan updated https://github.com/llvm/llvm-project/pull/239234
>From 01897dc9ca2182430fbf3e2b530624a4112bb5f5 Mon Sep 17 00:00:00 2001
From: SiHuaN <liyongtai at iscas.ac.cn>
Date: Thu, 5 Nov 2026 13:02:36 +0000
Subject: [PATCH] [RISCV][P-ext] Add scalar ssh1sadd intrinsic
ssh1sadd computes sat_add(sat_shl1(rs1), rs2). RV32 selects ssh1sadd
directly; RV64 has no scalar form and uses pssh1sadd.w on the low 32-bit
lane.
---
clang/include/clang/Basic/BuiltinsRISCV.td | 3 +++
clang/lib/CodeGen/TargetBuiltins/RISCV.cpp | 5 +++++
clang/lib/Headers/riscv_packed_simd.h | 3 +++
clang/test/CodeGen/RISCV/rvp-intrinsics.c | 18 ++++++++++++++++++
.../intrinsic-header-tests/riscv_packed_simd.c | 7 +++++++
llvm/include/llvm/IR/IntrinsicsRISCV.td | 3 +++
llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 10 +++++++---
llvm/lib/Target/RISCV/RISCVInstrInfoP.td | 7 +++++++
llvm/test/CodeGen/RISCV/rvp-scalar-rv64.ll | 12 ++++++++++++
llvm/test/CodeGen/RISCV/rvp-scalar.ll | 16 ++++++++++++++++
10 files changed, 81 insertions(+), 3 deletions(-)
diff --git a/clang/include/clang/Basic/BuiltinsRISCV.td b/clang/include/clang/Basic/BuiltinsRISCV.td
index 19438ec8d6b0bb..fa4cb4808f7cfc 100644
--- a/clang/include/clang/Basic/BuiltinsRISCV.td
+++ b/clang/include/clang/Basic/BuiltinsRISCV.td
@@ -271,6 +271,9 @@ def aaddu_u32 : RISCVBuiltin<"uint32_t(uint32_t, uint32_t)">;
def asub_i32 : RISCVBuiltin<"int32_t(int32_t, int32_t)">;
def asubu_u32 : RISCVBuiltin<"uint32_t(uint32_t, uint32_t)">;
+// Scalar Shift-Add
+def ssh1sadd_i32 : RISCVBuiltin<"int32_t(int32_t, int32_t)">;
+
// Scalar Saturation
def sati_i32 : RISCVBuiltin<"int32_t(int32_t, _Constant unsigned int)">;
def usati_u32 : RISCVBuiltin<"uint32_t(int32_t, _Constant unsigned int)">;
diff --git a/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp b/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp
index 19885655bfba0d..423a15b5167864 100644
--- a/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp
@@ -1700,6 +1700,11 @@ Value *CodeGenFunction::EmitRISCVBuiltinExpr(unsigned BuiltinID,
ID = Intrinsic::riscv_asubu_u32;
break;
+ // Scalar Shift-Add
+ case RISCV::BI__builtin_riscv_ssh1sadd_i32:
+ ID = Intrinsic::riscv_ssh1sadd_i32;
+ break;
+
// Scalar Multiply High
case RISCV::BI__builtin_riscv_mulhr_i32:
case RISCV::BI__builtin_riscv_mulhru_u32:
diff --git a/clang/lib/Headers/riscv_packed_simd.h b/clang/lib/Headers/riscv_packed_simd.h
index 4459bd7e715fc0..116da9e42512aa 100644
--- a/clang/lib/Headers/riscv_packed_simd.h
+++ b/clang/lib/Headers/riscv_packed_simd.h
@@ -470,6 +470,9 @@ __packed_binary_builtin(aaddu_u32, uint32_t, __builtin_riscv_aaddu_u32)
__packed_binary_builtin(asub_i32, int32_t, __builtin_riscv_asub_i32)
__packed_binary_builtin(asubu_u32, uint32_t, __builtin_riscv_asubu_u32)
+/* Scalar Shift-Add */
+__packed_binary_builtin(ssh1sadd_i32, int32_t, __builtin_riscv_ssh1sadd_i32)
+
/* Scalar Saturating Addition and Subtraction */
__packed_binary_builtin(sadd_i32, int32_t, __builtin_elementwise_add_sat)
__packed_binary_builtin(saddu_u32, uint32_t, __builtin_elementwise_add_sat)
diff --git a/clang/test/CodeGen/RISCV/rvp-intrinsics.c b/clang/test/CodeGen/RISCV/rvp-intrinsics.c
index c91f3b674ab407..2357234a280170 100644
--- a/clang/test/CodeGen/RISCV/rvp-intrinsics.c
+++ b/clang/test/CodeGen/RISCV/rvp-intrinsics.c
@@ -204,6 +204,24 @@ uint32_t test_asubu_u32(uint32_t a, uint32_t b) {
return __riscv_asubu_u32(a, b);
}
+/* Scalar Shift-Add */
+
+// RV32-LABEL: define dso_local i32 @test_ssh1sadd_i32(
+// RV32-SAME: i32 noundef [[A:%.*]], i32 noundef [[B:%.*]]) #[[ATTR0]] {
+// RV32-NEXT: [[ENTRY:.*:]]
+// RV32-NEXT: [[TMP0:%.*]] = call i32 @llvm.riscv.ssh1sadd.i32(i32 [[A]], i32 [[B]])
+// RV32-NEXT: ret i32 [[TMP0]]
+//
+// RV64-LABEL: define dso_local signext i32 @test_ssh1sadd_i32(
+// RV64-SAME: i32 noundef signext [[A:%.*]], i32 noundef signext [[B:%.*]]) #[[ATTR0]] {
+// RV64-NEXT: [[ENTRY:.*:]]
+// RV64-NEXT: [[TMP0:%.*]] = call i32 @llvm.riscv.ssh1sadd.i32(i32 [[A]], i32 [[B]])
+// RV64-NEXT: ret i32 [[TMP0]]
+//
+int32_t test_ssh1sadd_i32(int32_t a, int32_t b) {
+ return __riscv_ssh1sadd_i32(a, b);
+}
+
/* Scalar Widening Zip */
// RV32-LABEL: define dso_local i64 @test_wzip8p_64(
diff --git a/cross-project-tests/intrinsic-header-tests/riscv_packed_simd.c b/cross-project-tests/intrinsic-header-tests/riscv_packed_simd.c
index 2357a1ff1ca5f6..4e24833d6b018c 100644
--- a/cross-project-tests/intrinsic-header-tests/riscv_packed_simd.c
+++ b/cross-project-tests/intrinsic-header-tests/riscv_packed_simd.c
@@ -6600,3 +6600,10 @@ int32x2_t test_pslidedownx_i32x2(int32x2_t rd, int32x2_t rs1, unsigned rs2) {
uint32x2_t test_pslidedownx_u32x2(uint32x2_t rd, uint32x2_t rs1, unsigned rs2) {
return __riscv_pslidedownx_u32x2(rd, rs1, rs2);
}
+
+// CHECK-LABEL: test_ssh1sadd_i32:
+// RV32: ssh1sadd{{[[:space:]]}}
+// RV64: pssh1sadd.w{{[[:space:]]}}
+int32_t test_ssh1sadd_i32(int32_t rs1, int32_t rs2) {
+ return __riscv_ssh1sadd_i32(rs1, rs2);
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsRISCV.td b/llvm/include/llvm/IR/IntrinsicsRISCV.td
index c5a5d4b4a61c11..d077c8e124498f 100644
--- a/llvm/include/llvm/IR/IntrinsicsRISCV.td
+++ b/llvm/include/llvm/IR/IntrinsicsRISCV.td
@@ -2184,6 +2184,9 @@ class RVPBinaryIntrinsic
def int_riscv_asub_i32 : RVPScalarBinaryIntrinsic;
def int_riscv_asubu_u32 : RVPScalarBinaryIntrinsic;
+ // Scalar Shift-Add. On RV64 this maps to pssh1sadd.w on the low words.
+ def int_riscv_ssh1sadd_i32 : RVPScalarBinaryIntrinsic;
+
// Scalar Multiply High. On RV32 these map to the M-extension mulh
// instructions and the P-extension mulhr instructions.
class RVPScalarMulHighIntrinsic
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 0d7be09555ed6e..d35d9f045e5475 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -12506,6 +12506,8 @@ static unsigned getRVPScalarOpcode(unsigned IntNo) {
return ISD::AVGSUBFLOORS;
case Intrinsic::riscv_asubu_u32:
return ISD::AVGSUBFLOORU;
+ case Intrinsic::riscv_ssh1sadd_i32:
+ return RISCVISD::SSH1SADD;
case Intrinsic::riscv_mulhr_i32:
return RISCVISD::MULHR;
case Intrinsic::riscv_mulhru_u32:
@@ -13493,6 +13495,7 @@ SDValue RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
case Intrinsic::riscv_aaddu_u32:
case Intrinsic::riscv_asub_i32:
case Intrinsic::riscv_asubu_u32:
+ case Intrinsic::riscv_ssh1sadd_i32:
case Intrinsic::riscv_mulhr_i32:
case Intrinsic::riscv_mulhru_u32:
case Intrinsic::riscv_mulhsu_i32:
@@ -18583,6 +18586,7 @@ void RISCVTargetLowering::ReplaceNodeResults(SDNode *N,
case Intrinsic::riscv_aaddu_u32:
case Intrinsic::riscv_asub_i32:
case Intrinsic::riscv_asubu_u32:
+ case Intrinsic::riscv_ssh1sadd_i32:
case Intrinsic::riscv_mulhr_i32:
case Intrinsic::riscv_mulhru_u32:
case Intrinsic::riscv_mulhsu_i32:
@@ -18595,9 +18599,9 @@ void RISCVTargetLowering::ReplaceNodeResults(SDNode *N,
case Intrinsic::riscv_mhraccu_u32:
case Intrinsic::riscv_mhaccsu_i32:
case Intrinsic::riscv_mhraccsu_i32: {
- // RV64 has no scalar averaging, mulh, mulq or mhacc instructions; reuse
- // the packed .w family on the low words, whose element 0 is the scalar
- // result.
+ // RV64 has no scalar averaging, shift-add, mulh, mulq or mhacc
+ // instructions; reuse the packed .w family on the low words, whose
+ // element 0 is the scalar result.
assert(Subtarget.is64Bit() && N->getValueType(0) == MVT::i32 &&
"Unexpected subtarget or result type");
unsigned Opc = getRVPScalarOpcode(IntNo);
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
index 2ecec70322234a..ae91e457815b23 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
@@ -2035,6 +2035,9 @@ def riscv_mulhr : RVSDNode<"MULHR", SDTIntBinOp, [SDNPCommutative]>;
def riscv_mulhru : RVSDNode<"MULHRU", SDTIntBinOp, [SDNPCommutative]>;
def riscv_mulhrsu : RVSDNode<"MULHRSU", SDTIntBinOp>;
+// Shift-add.
+def riscv_ssh1sadd : RVSDNode<"SSH1SADD", SDTIntBinOp>;
+
def SDT_RISCVSameTernary : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>,
SDTCisSameAs<0, 2>,
SDTCisSameAs<0, 3>]>;
@@ -2613,6 +2616,9 @@ let append Predicates = [IsRV32] in {
def : Pat<(XLenVT (saddsat (saddsat GPR:$rs1, GPR:$rs1), GPR:$rs2)),
(SSH1SADD GPR:$rs1, GPR:$rs2)>;
+ // 32-bit shift-add pattern
+ def : PatGprGpr<riscv_ssh1sadd, SSH1SADD, i32>;
+
// 32-bit averaging patterns
def : PatGprGpr<avgfloors, AADD, i32>;
def : PatGprGpr<avgflooru, AADDU, i32>;
@@ -3465,6 +3471,7 @@ let append Predicates = [IsRV64] in {
// 32-bit shift-add patterns
def : PatPSh1Add<PSH1ADD_W, v2i32>;
def : PatPSSh1SAdd<PSSH1SADD_W, v2i32>;
+ def : PatGprGpr<riscv_ssh1sadd, PSSH1SADD_W, v2i32>;
// 32-bit exchanged add/sub patterns
def : PatGprGpr<int_riscv_pas, PAS_WX, v2i32>;
diff --git a/llvm/test/CodeGen/RISCV/rvp-scalar-rv64.ll b/llvm/test/CodeGen/RISCV/rvp-scalar-rv64.ll
index ff4d49db1aba7e..bc36f912f77859 100644
--- a/llvm/test/CodeGen/RISCV/rvp-scalar-rv64.ll
+++ b/llvm/test/CodeGen/RISCV/rvp-scalar-rv64.ll
@@ -1676,3 +1676,15 @@ define void @merge_wusers_v8i8(<8 x i8> %a, i8 %x, ptr %p) nounwind {
store i32 %t, ptr %p
ret void
}
+
+declare i32 @llvm.riscv.ssh1sadd.i32(i32, i32)
+
+define signext i32 @ssh1sadd_intrinsic_i32(i32 signext %a, i32 signext %b) nounwind {
+; CHECK-LABEL: ssh1sadd_intrinsic_i32:
+; CHECK: # %bb.0:
+; CHECK-NEXT: pssh1sadd.w a0, a0, a1
+; CHECK-NEXT: sext.w a0, a0
+; CHECK-NEXT: ret
+ %res = call i32 @llvm.riscv.ssh1sadd.i32(i32 %a, i32 %b)
+ ret i32 %res
+}
diff --git a/llvm/test/CodeGen/RISCV/rvp-scalar.ll b/llvm/test/CodeGen/RISCV/rvp-scalar.ll
index a68592f7ed7ae9..13579bf2b9a11a 100644
--- a/llvm/test/CodeGen/RISCV/rvp-scalar.ll
+++ b/llvm/test/CodeGen/RISCV/rvp-scalar.ll
@@ -4628,3 +4628,19 @@ define i64 @srxi_i64_103(i64 %a, i64 %b) {
%1 = tail call i64 @llvm.fshr.i64(i64 %a, i64 %b, i64 103)
ret i64 %1
}
+
+declare i32 @llvm.riscv.ssh1sadd.i32(i32, i32)
+
+define i32 @ssh1sadd_intrinsic_i32(i32 %a, i32 %b) nounwind {
+; RV32-LABEL: ssh1sadd_intrinsic_i32:
+; RV32: # %bb.0:
+; RV32-NEXT: ssh1sadd a0, a0, a1
+; RV32-NEXT: ret
+;
+; RV64-LABEL: ssh1sadd_intrinsic_i32:
+; RV64: # %bb.0:
+; RV64-NEXT: pssh1sadd.w a0, a0, a1
+; RV64-NEXT: ret
+ %res = call i32 @llvm.riscv.ssh1sadd.i32(i32 %a, i32 %b)
+ ret i32 %res
+}
More information about the cfe-commits
mailing list