https://github.com/XChy created https://github.com/llvm/llvm-project/pull/218430

See also 
https://github.com/riscv/riscv-p-spec/blob/master/P-ext-intrinsics.adoc#packed-multiplication-with-horizontal-addition.

>From 43bcc1fcfb9201599983b71aaa21b3d9bbb06dc3 Mon Sep 17 00:00:00 2001
From: XChy <[email protected]>
Date: Mon, 24 Aug 2026 22:10:31 +0800
Subject: [PATCH] [RISCV] Support Packed Multiplication with Horizontal
 Addition

---
 clang/include/clang/Basic/BuiltinsRISCV.td    |  41 ++
 clang/lib/CodeGen/TargetBuiltins/RISCV.cpp    | 110 +++++
 clang/lib/Headers/riscv_packed_simd.h         |  41 ++
 .../riscv_packed_simd.c                       | 285 +++++++++++++
 llvm/include/llvm/IR/IntrinsicsRISCV.td       |  19 +
 llvm/lib/Target/RISCV/RISCVISelLowering.cpp   | 100 +++++
 llvm/lib/Target/RISCV/RISCVInstrInfoP.td      | 140 +++++++
 llvm/test/CodeGen/RISCV/rvp-simd-32.ll        | 117 ++++++
 llvm/test/CodeGen/RISCV/rvp-simd-64.ll        | 375 ++++++++++++++++++
 9 files changed, 1228 insertions(+)

diff --git a/clang/include/clang/Basic/BuiltinsRISCV.td 
b/clang/include/clang/Basic/BuiltinsRISCV.td
index 89dbf736f4435..9cddafd659af1 100644
--- a/clang/include/clang/Basic/BuiltinsRISCV.td
+++ b/clang/include/clang/Basic/BuiltinsRISCV.td
@@ -291,6 +291,47 @@ def pmhaccsu_i32x2 : RISCVBuiltin<"_Vector<2, 
int>(_Vector<2, int>, _Vector<2, i
 def pmhraccsu_i16x4 : RISCVBuiltin<"_Vector<4, short>(_Vector<4, short>, 
_Vector<4, short>, _Vector<4, unsigned short>)">;
 def pmhraccsu_i32x2 : RISCVBuiltin<"_Vector<2, int>(_Vector<2, int>, 
_Vector<2, int>, _Vector<2, unsigned int>)">;
 
+// Packed Multiplication with Horizontal Addition (32-bit)
+def pm4add_i8x4 : RISCVBuiltin<"int(_Vector<4, signed char>, _Vector<4, signed 
char>)">;
+def pm2add_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pm2add_x_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pm4addu_u8x4 : RISCVBuiltin<"unsigned int(_Vector<4, unsigned char>, 
_Vector<4, unsigned char>)">;
+def pm2addu_u16x2 : RISCVBuiltin<"unsigned int(_Vector<2, unsigned short>, 
_Vector<2, unsigned short>)">;
+def pmq2add_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pmqr2add_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pm2sadd_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pm2sadd_x_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, 
short>)">;
+def pm2sub_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pm2sub_x_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, short>)">;
+def pm4addsu_i8x4 : RISCVBuiltin<"int(_Vector<4, signed char>, _Vector<4, 
unsigned char>)">;
+def pm2addsu_i16x2 : RISCVBuiltin<"int(_Vector<2, short>, _Vector<2, unsigned 
short>)">;
+
+// Packed Multiplication with Horizontal Addition (64-bit)
+def pm4add_i8x8 : RISCVBuiltin<"_Vector<2, int>(_Vector<8, signed char>, 
_Vector<8, signed char>)">;
+def pm2add_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, _Vector<4, 
short>)">;
+def pm2add_x_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, short>)">;
+def pm4addu_u8x8 : RISCVBuiltin<"_Vector<2, unsigned int>(_Vector<8, unsigned 
char>, _Vector<8, unsigned char>)">;
+def pm2addu_u16x4 : RISCVBuiltin<"_Vector<2, unsigned int>(_Vector<4, unsigned 
short>, _Vector<4, unsigned short>)">;
+def pmq2add_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, short>)">;
+def pmqr2add_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, short>)">;
+def pm2sadd_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, short>)">;
+def pm2sadd_x_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, short>)">;
+def pm2sub_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, _Vector<4, 
short>)">;
+def pm2sub_x_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, short>)">;
+def pm4addsu_i8x8 : RISCVBuiltin<"_Vector<2, int>(_Vector<8, signed char>, 
_Vector<8, unsigned char>)">;
+def pm2addsu_i16x4 : RISCVBuiltin<"_Vector<2, int>(_Vector<4, short>, 
_Vector<4, unsigned short>)">;
+def pm2add_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, int>)">;
+def pm2add_x_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, int>)">;
+def pm2addu_u32x2 : RISCVBuiltin<"uint64_t(_Vector<2, unsigned int>, 
_Vector<2, unsigned int>)">;
+def pmq2add_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, int>)">;
+def pm2sub_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, int>)">;
+def pm2sub_x_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, int>)">;
+def pm2addsu_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, 
unsigned int>)">;
+def pmqr2add_i32x2 : RISCVBuiltin<"int64_t(_Vector<2, int>, _Vector<2, int>)">;
+def pm4add_i16x4 : RISCVBuiltin<"int64_t(_Vector<4, short>, _Vector<4, 
short>)">;
+def pm4addu_u16x4 : RISCVBuiltin<"uint64_t(_Vector<4, unsigned short>, 
_Vector<4, unsigned short>)">;
+def pm4addsu_i16x4 : RISCVBuiltin<"int64_t(_Vector<4, short>, _Vector<4, 
unsigned short>)">;
+
 // Packed Absolute Difference Sum (32-bit)
 def pabdsumu_u8x4_u32 : RISCVBuiltin<"unsigned int(_Vector<4, unsigned char>, 
_Vector<4, unsigned char>)">;
 def pabdsumau_u8x4_u32 : RISCVBuiltin<"unsigned int(unsigned int, _Vector<4, 
unsigned char>, _Vector<4, unsigned char>)">;
diff --git a/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp 
b/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp
index 88fedecbee504..5c5ed9083057f 100644
--- a/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/RISCV.cpp
@@ -1531,6 +1531,116 @@ Value *CodeGenFunction::EmitRISCVBuiltinExpr(unsigned 
BuiltinID,
     break;
   }
 
+  // Packed Multiplication with Horizontal Addition
+  case RISCV::BI__builtin_riscv_pm4add_i8x4:
+  case RISCV::BI__builtin_riscv_pm4add_i8x8:
+  case RISCV::BI__builtin_riscv_pm4add_i16x4:
+  case RISCV::BI__builtin_riscv_pm2add_i16x2:
+  case RISCV::BI__builtin_riscv_pm2add_i16x4:
+  case RISCV::BI__builtin_riscv_pm2add_i32x2:
+  case RISCV::BI__builtin_riscv_pm2add_x_i16x2:
+  case RISCV::BI__builtin_riscv_pm2add_x_i16x4:
+  case RISCV::BI__builtin_riscv_pm2add_x_i32x2:
+  case RISCV::BI__builtin_riscv_pm4addu_u8x4:
+  case RISCV::BI__builtin_riscv_pm4addu_u8x8:
+  case RISCV::BI__builtin_riscv_pm4addu_u16x4:
+  case RISCV::BI__builtin_riscv_pm2addu_u16x2:
+  case RISCV::BI__builtin_riscv_pm2addu_u16x4:
+  case RISCV::BI__builtin_riscv_pm2addu_u32x2:
+  case RISCV::BI__builtin_riscv_pmq2add_i16x2:
+  case RISCV::BI__builtin_riscv_pmq2add_i16x4:
+  case RISCV::BI__builtin_riscv_pmq2add_i32x2:
+  case RISCV::BI__builtin_riscv_pmqr2add_i16x2:
+  case RISCV::BI__builtin_riscv_pmqr2add_i16x4:
+  case RISCV::BI__builtin_riscv_pmqr2add_i32x2:
+  case RISCV::BI__builtin_riscv_pm2sadd_i16x2:
+  case RISCV::BI__builtin_riscv_pm2sadd_i16x4:
+  case RISCV::BI__builtin_riscv_pm2sadd_x_i16x2:
+  case RISCV::BI__builtin_riscv_pm2sadd_x_i16x4:
+  case RISCV::BI__builtin_riscv_pm2sub_i16x2:
+  case RISCV::BI__builtin_riscv_pm2sub_i16x4:
+  case RISCV::BI__builtin_riscv_pm2sub_i32x2:
+  case RISCV::BI__builtin_riscv_pm2sub_x_i16x2:
+  case RISCV::BI__builtin_riscv_pm2sub_x_i16x4:
+  case RISCV::BI__builtin_riscv_pm2sub_x_i32x2:
+  case RISCV::BI__builtin_riscv_pm4addsu_i8x4:
+  case RISCV::BI__builtin_riscv_pm4addsu_i8x8:
+  case RISCV::BI__builtin_riscv_pm4addsu_i16x4:
+  case RISCV::BI__builtin_riscv_pm2addsu_i16x2:
+  case RISCV::BI__builtin_riscv_pm2addsu_i16x4:
+  case RISCV::BI__builtin_riscv_pm2addsu_i32x2: {
+    switch (BuiltinID) {
+    default:
+      llvm_unreachable("unexpected builtin ID");
+    case RISCV::BI__builtin_riscv_pm4add_i8x4:
+    case RISCV::BI__builtin_riscv_pm4add_i8x8:
+    case RISCV::BI__builtin_riscv_pm4add_i16x4:
+      ID = Intrinsic::riscv_pm4add;
+      break;
+    case RISCV::BI__builtin_riscv_pm2add_i16x2:
+    case RISCV::BI__builtin_riscv_pm2add_i16x4:
+    case RISCV::BI__builtin_riscv_pm2add_i32x2:
+      ID = Intrinsic::riscv_pm2add;
+      break;
+    case RISCV::BI__builtin_riscv_pm2add_x_i16x2:
+    case RISCV::BI__builtin_riscv_pm2add_x_i16x4:
+    case RISCV::BI__builtin_riscv_pm2add_x_i32x2:
+      ID = Intrinsic::riscv_pm2add_x;
+      break;
+    case RISCV::BI__builtin_riscv_pm4addu_u8x4:
+    case RISCV::BI__builtin_riscv_pm4addu_u8x8:
+    case RISCV::BI__builtin_riscv_pm4addu_u16x4:
+      ID = Intrinsic::riscv_pm4addu;
+      break;
+    case RISCV::BI__builtin_riscv_pm2addu_u16x2:
+    case RISCV::BI__builtin_riscv_pm2addu_u16x4:
+    case RISCV::BI__builtin_riscv_pm2addu_u32x2:
+      ID = Intrinsic::riscv_pm2addu;
+      break;
+    case RISCV::BI__builtin_riscv_pmq2add_i16x2:
+    case RISCV::BI__builtin_riscv_pmq2add_i16x4:
+    case RISCV::BI__builtin_riscv_pmq2add_i32x2:
+      ID = Intrinsic::riscv_pmq2add;
+      break;
+    case RISCV::BI__builtin_riscv_pmqr2add_i16x2:
+    case RISCV::BI__builtin_riscv_pmqr2add_i16x4:
+    case RISCV::BI__builtin_riscv_pmqr2add_i32x2:
+      ID = Intrinsic::riscv_pmqr2add;
+      break;
+    case RISCV::BI__builtin_riscv_pm2sadd_i16x2:
+    case RISCV::BI__builtin_riscv_pm2sadd_i16x4:
+      ID = Intrinsic::riscv_pm2sadd;
+      break;
+    case RISCV::BI__builtin_riscv_pm2sadd_x_i16x2:
+    case RISCV::BI__builtin_riscv_pm2sadd_x_i16x4:
+      ID = Intrinsic::riscv_pm2sadd_x;
+      break;
+    case RISCV::BI__builtin_riscv_pm2sub_i16x2:
+    case RISCV::BI__builtin_riscv_pm2sub_i16x4:
+    case RISCV::BI__builtin_riscv_pm2sub_i32x2:
+      ID = Intrinsic::riscv_pm2sub;
+      break;
+    case RISCV::BI__builtin_riscv_pm2sub_x_i16x2:
+    case RISCV::BI__builtin_riscv_pm2sub_x_i16x4:
+    case RISCV::BI__builtin_riscv_pm2sub_x_i32x2:
+      ID = Intrinsic::riscv_pm2sub_x;
+      break;
+    case RISCV::BI__builtin_riscv_pm4addsu_i8x4:
+    case RISCV::BI__builtin_riscv_pm4addsu_i8x8:
+    case RISCV::BI__builtin_riscv_pm4addsu_i16x4:
+      ID = Intrinsic::riscv_pm4addsu;
+      break;
+    case RISCV::BI__builtin_riscv_pm2addsu_i16x2:
+    case RISCV::BI__builtin_riscv_pm2addsu_i16x4:
+    case RISCV::BI__builtin_riscv_pm2addsu_i32x2:
+      ID = Intrinsic::riscv_pm2addsu;
+      break;
+    }
+
+    IntrinsicTypes = {ResultType, Ops[0]->getType()};
+    break;
+  }
+
   // Packed Reduction Sum
   case RISCV::BI__builtin_riscv_predsum_i8x4_i32:
   case RISCV::BI__builtin_riscv_predsum_i16x2_i32:
diff --git a/clang/lib/Headers/riscv_packed_simd.h 
b/clang/lib/Headers/riscv_packed_simd.h
index 318900805043c..92d01e0868c1c 100644
--- a/clang/lib/Headers/riscv_packed_simd.h
+++ b/clang/lib/Headers/riscv_packed_simd.h
@@ -813,6 +813,47 @@ __packed_ternary_builtin_mixed(pmhaccsu_i32x2, int32x2_t, 
int32x2_t, uint32x2_t,
 __packed_ternary_builtin_mixed(pmhraccsu_i16x4, int16x4_t, int16x4_t, 
uint16x4_t, __builtin_riscv_pmhraccsu_i16x4)
 __packed_ternary_builtin_mixed(pmhraccsu_i32x2, int32x2_t, int32x2_t, 
uint32x2_t, __builtin_riscv_pmhraccsu_i32x2)
 
+/* Packed Multiplication with Horizontal Addition (32-bit) */
+__packed_binary_builtin_mixed(pm4add_i8x4, int32_t, int8x4_t, int8x4_t, 
__builtin_riscv_pm4add_i8x4)
+__packed_binary_builtin_mixed(pm2add_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pm2add_i16x2)
+__packed_binary_builtin_mixed(pm2add_x_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pm2add_x_i16x2)
+__packed_binary_builtin_mixed(pm4addu_u8x4, uint32_t, uint8x4_t, uint8x4_t, 
__builtin_riscv_pm4addu_u8x4)
+__packed_binary_builtin_mixed(pm2addu_u16x2, uint32_t, uint16x2_t, uint16x2_t, 
__builtin_riscv_pm2addu_u16x2)
+__packed_binary_builtin_mixed(pmq2add_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pmq2add_i16x2)
+__packed_binary_builtin_mixed(pmqr2add_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pmqr2add_i16x2)
+__packed_binary_builtin_mixed(pm2sadd_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pm2sadd_i16x2)
+__packed_binary_builtin_mixed(pm2sadd_x_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pm2sadd_x_i16x2)
+__packed_binary_builtin_mixed(pm2sub_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pm2sub_i16x2)
+__packed_binary_builtin_mixed(pm2sub_x_i16x2, int32_t, int16x2_t, int16x2_t, 
__builtin_riscv_pm2sub_x_i16x2)
+__packed_binary_builtin_mixed(pm4addsu_i8x4, int32_t, int8x4_t, uint8x4_t, 
__builtin_riscv_pm4addsu_i8x4)
+__packed_binary_builtin_mixed(pm2addsu_i16x2, int32_t, int16x2_t, uint16x2_t, 
__builtin_riscv_pm2addsu_i16x2)
+
+/* Packed Multiplication with Horizontal Addition (64-bit) */
+__packed_binary_builtin_mixed(pm4add_i8x8, int32x2_t, int8x8_t, int8x8_t, 
__builtin_riscv_pm4add_i8x8)
+__packed_binary_builtin_mixed(pm2add_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pm2add_i16x4)
+__packed_binary_builtin_mixed(pm2add_x_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pm2add_x_i16x4)
+__packed_binary_builtin_mixed(pm4addu_u8x8, uint32x2_t, uint8x8_t, uint8x8_t, 
__builtin_riscv_pm4addu_u8x8)
+__packed_binary_builtin_mixed(pm2addu_u16x4, uint32x2_t, uint16x4_t, 
uint16x4_t, __builtin_riscv_pm2addu_u16x4)
+__packed_binary_builtin_mixed(pmq2add_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pmq2add_i16x4)
+__packed_binary_builtin_mixed(pmqr2add_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pmqr2add_i16x4)
+__packed_binary_builtin_mixed(pm2sadd_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pm2sadd_i16x4)
+__packed_binary_builtin_mixed(pm2sadd_x_i16x4, int32x2_t, int16x4_t, 
int16x4_t, __builtin_riscv_pm2sadd_x_i16x4)
+__packed_binary_builtin_mixed(pm2sub_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pm2sub_i16x4)
+__packed_binary_builtin_mixed(pm2sub_x_i16x4, int32x2_t, int16x4_t, int16x4_t, 
__builtin_riscv_pm2sub_x_i16x4)
+__packed_binary_builtin_mixed(pm4addsu_i8x8, int32x2_t, int8x8_t, uint8x8_t, 
__builtin_riscv_pm4addsu_i8x8)
+__packed_binary_builtin_mixed(pm2addsu_i16x4, int32x2_t, int16x4_t, 
uint16x4_t, __builtin_riscv_pm2addsu_i16x4)
+__packed_binary_builtin_mixed(pm2add_i32x2, int64_t, int32x2_t, int32x2_t, 
__builtin_riscv_pm2add_i32x2)
+__packed_binary_builtin_mixed(pm2add_x_i32x2, int64_t, int32x2_t, int32x2_t, 
__builtin_riscv_pm2add_x_i32x2)
+__packed_binary_builtin_mixed(pm2addu_u32x2, uint64_t, uint32x2_t, uint32x2_t, 
__builtin_riscv_pm2addu_u32x2)
+__packed_binary_builtin_mixed(pmq2add_i32x2, int64_t, int32x2_t, int32x2_t, 
__builtin_riscv_pmq2add_i32x2)
+__packed_binary_builtin_mixed(pm2sub_i32x2, int64_t, int32x2_t, int32x2_t, 
__builtin_riscv_pm2sub_i32x2)
+__packed_binary_builtin_mixed(pm2sub_x_i32x2, int64_t, int32x2_t, int32x2_t, 
__builtin_riscv_pm2sub_x_i32x2)
+__packed_binary_builtin_mixed(pm2addsu_i32x2, int64_t, int32x2_t, uint32x2_t, 
__builtin_riscv_pm2addsu_i32x2)
+__packed_binary_builtin_mixed(pmqr2add_i32x2, int64_t, int32x2_t, int32x2_t, 
__builtin_riscv_pmqr2add_i32x2)
+__packed_binary_builtin_mixed(pm4add_i16x4, int64_t, int16x4_t, int16x4_t, 
__builtin_riscv_pm4add_i16x4)
+__packed_binary_builtin_mixed(pm4addu_u16x4, uint64_t, uint16x4_t, uint16x4_t, 
__builtin_riscv_pm4addu_u16x4)
+__packed_binary_builtin_mixed(pm4addsu_i16x4, int64_t, int16x4_t, uint16x4_t, 
__builtin_riscv_pm4addsu_i16x4)
+
 /* Packed Absolute Difference Sum (32-bit) */
 __packed_abdsum(pabdsumu_u8x4_u32, uint32_t, uint8x4_t, 
__builtin_riscv_pabdsumu_u8x4_u32)
 __packed_abdsum_acc(pabdsumau_u8x4_u32, uint32_t, uint8x4_t, 
__builtin_riscv_pabdsumau_u8x4_u32)
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 75dcb796b66cd..7a4d909968b23 100644
--- a/cross-project-tests/intrinsic-header-tests/riscv_packed_simd.c
+++ b/cross-project-tests/intrinsic-header-tests/riscv_packed_simd.c
@@ -3406,3 +3406,288 @@ int32x2_t test_pnclipp_i32x2(int64_t a, int64_t b) {
 uint32x2_t test_pnclipup_u32x2(uint64_t a, uint64_t b) {
   return __riscv_pnclipup_u32x2(a, b);
 }
+
+// CHECK-LABEL: test_pm4add_i8x4:
+// RV32:        pm4add.b
+// RV64:        pm4add.b
+int32_t test_pm4add_i8x4(int8x4_t rs1, int8x4_t rs2) {
+  return __riscv_pm4add_i8x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2add_i16x2:
+// RV32:        pm2add.h
+// RV64:        pm2add.h
+int32_t test_pm2add_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pm2add_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2add_x_i16x2:
+// RV32:        pm2add.hx
+// RV64:        pm2add.hx
+int32_t test_pm2add_x_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pm2add_x_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4addu_u8x4:
+// RV32:        pm4addu.b
+// RV64:        pm4addu.b
+uint32_t test_pm4addu_u8x4(uint8x4_t rs1, uint8x4_t rs2) {
+  return __riscv_pm4addu_u8x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2addu_u16x2:
+// RV32:        pm2addu.h
+// RV64:        pm2addu.h
+uint32_t test_pm2addu_u16x2(uint16x2_t rs1, uint16x2_t rs2) {
+  return __riscv_pm2addu_u16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pmq2add_i16x2:
+// RV32:        pmq2add.h
+// RV64:        pmq2add.h
+int32_t test_pmq2add_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pmq2add_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pmqr2add_i16x2:
+// RV32:        pmqr2add.h
+// RV64:        pmqr2add.h
+int32_t test_pmqr2add_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pmqr2add_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sadd_i16x2:
+// RV32:        pm2sadd.h
+// RV64:        pm2sadd.h
+int32_t test_pm2sadd_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pm2sadd_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sadd_x_i16x2:
+// RV32:        pm2sadd.hx
+// RV64:        pm2sadd.hx
+int32_t test_pm2sadd_x_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pm2sadd_x_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sub_i16x2:
+// RV32:        pm2sub.h
+// RV64:        pm2sub.h
+int32_t test_pm2sub_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pm2sub_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sub_x_i16x2:
+// RV32:        pm2sub.hx
+// RV64:        pm2sub.hx
+int32_t test_pm2sub_x_i16x2(int16x2_t rs1, int16x2_t rs2) {
+  return __riscv_pm2sub_x_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4addsu_i8x4:
+// RV32:        pm4addsu.b
+// RV64:        pm4addsu.b
+int32_t test_pm4addsu_i8x4(int8x4_t rs1, uint8x4_t rs2) {
+  return __riscv_pm4addsu_i8x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2addsu_i16x2:
+// RV32:        pm2addsu.h
+// RV64:        pm2addsu.h
+int32_t test_pm2addsu_i16x2(int16x2_t rs1, uint16x2_t rs2) {
+  return __riscv_pm2addsu_i16x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4add_i8x8:
+// RV32:        pm4add.b
+// RV32:        pm4add.b
+// RV64:        pm4add.b
+int32x2_t test_pm4add_i8x8(int8x8_t rs1, int8x8_t rs2) {
+  return __riscv_pm4add_i8x8(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2add_i16x4:
+// RV32:        pm2add.h
+// RV32:        pm2add.h
+// RV64:        pm2add.h
+int32x2_t test_pm2add_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm2add_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2add_x_i16x4:
+// RV32:        pm2add.hx
+// RV32:        pm2add.hx
+// RV64:        pm2add.hx
+int32x2_t test_pm2add_x_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm2add_x_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4addu_u8x8:
+// RV32:        pm4addu.b
+// RV32:        pm4addu.b
+// RV64:        pm4addu.b
+uint32x2_t test_pm4addu_u8x8(uint8x8_t rs1, uint8x8_t rs2) {
+  return __riscv_pm4addu_u8x8(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2addu_u16x4:
+// RV32:        pm2addu.h
+// RV32:        pm2addu.h
+// RV64:        pm2addu.h
+uint32x2_t test_pm2addu_u16x4(uint16x4_t rs1, uint16x4_t rs2) {
+  return __riscv_pm2addu_u16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pmq2add_i16x4:
+// RV32:        pmq2add.h
+// RV32:        pmq2add.h
+// RV64:        pmq2add.h
+int32x2_t test_pmq2add_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pmq2add_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pmqr2add_i16x4:
+// RV32:        pmqr2add.h
+// RV32:        pmqr2add.h
+// RV64:        pmqr2add.h
+int32x2_t test_pmqr2add_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pmqr2add_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sadd_i16x4:
+// RV32:        pm2sadd.h
+// RV32:        pm2sadd.h
+// RV64:        pm2sadd.h
+int32x2_t test_pm2sadd_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm2sadd_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sadd_x_i16x4:
+// RV32:        pm2sadd.hx
+// RV32:        pm2sadd.hx
+// RV64:        pm2sadd.hx
+int32x2_t test_pm2sadd_x_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm2sadd_x_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sub_i16x4:
+// RV32:        pm2sub.h
+// RV32:        pm2sub.h
+// RV64:        pm2sub.h
+int32x2_t test_pm2sub_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm2sub_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sub_x_i16x4:
+// RV32:        pm2sub.hx
+// RV32:        pm2sub.hx
+// RV64:        pm2sub.hx
+int32x2_t test_pm2sub_x_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm2sub_x_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4addsu_i8x8:
+// RV32:        pm4addsu.b
+// RV32:        pm4addsu.b
+// RV64:        pm4addsu.b
+int32x2_t test_pm4addsu_i8x8(int8x8_t rs1, uint8x8_t rs2) {
+  return __riscv_pm4addsu_i8x8(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2addsu_i16x4:
+// RV32:        pm2addsu.h
+// RV32:        pm2addsu.h
+// RV64:        pm2addsu.h
+int32x2_t test_pm2addsu_i16x4(int16x4_t rs1, uint16x4_t rs2) {
+  return __riscv_pm2addsu_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2add_i32x2:
+// RV32:        wmul
+// RV32:        wmacc
+// RV64:        pm2add.w
+int64_t test_pm2add_i32x2(int32x2_t rs1, int32x2_t rs2) {
+  return __riscv_pm2add_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2add_x_i32x2:
+// RV32:        wmul
+// RV32:        wmacc
+// RV64:        pm2add.wx
+int64_t test_pm2add_x_i32x2(int32x2_t rs1, int32x2_t rs2) {
+  return __riscv_pm2add_x_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2addu_u32x2:
+// RV32:        wmulu
+// RV32:        wmaccu
+// RV64:        pm2addu.w
+uint64_t test_pm2addu_u32x2(uint32x2_t rs1, uint32x2_t rs2) {
+  return __riscv_pm2addu_u32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pmq2add_i32x2:
+// RV32:        mqwacc
+// RV32:        mqwacc
+// RV64:        pmq2add.w
+int64_t test_pmq2add_i32x2(int32x2_t rs1, int32x2_t rs2) {
+  return __riscv_pmq2add_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pmqr2add_i32x2:
+// RV32:        mqrwacc
+// RV32:        mqrwacc
+// RV64:        pmqr2add.w
+int64_t test_pmqr2add_i32x2(int32x2_t rs1, int32x2_t rs2) {
+  return __riscv_pmqr2add_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sub_i32x2:
+// RV32:        wmul
+// RV32:        wmul
+// RV32:        subd
+// RV64:        pm2sub.w
+int64_t test_pm2sub_i32x2(int32x2_t rs1, int32x2_t rs2) {
+  return __riscv_pm2sub_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2sub_x_i32x2:
+// RV32:        wmul
+// RV32:        wmul
+// RV32:        subd
+// RV64:        pm2sub.wx
+int64_t test_pm2sub_x_i32x2(int32x2_t rs1, int32x2_t rs2) {
+  return __riscv_pm2sub_x_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm2addsu_i32x2:
+// RV32:        wmulsu
+// RV32:        wmaccsu
+// RV64:        pm2addsu.w
+int64_t test_pm2addsu_i32x2(int32x2_t rs1, uint32x2_t rs2) {
+  return __riscv_pm2addsu_i32x2(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4add_i16x4:
+// RV32:        pm2wadd.h
+// RV32:        pm2wadda.h
+// RV64:        pm4add.h
+int64_t test_pm4add_i16x4(int16x4_t rs1, int16x4_t rs2) {
+  return __riscv_pm4add_i16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4addu_u16x4:
+// RV32:        pm2waddu.h
+// RV32:        pm2waddau.h
+// RV64:        pm4addu.h
+uint64_t test_pm4addu_u16x4(uint16x4_t rs1, uint16x4_t rs2) {
+  return __riscv_pm4addu_u16x4(rs1, rs2);
+}
+
+// CHECK-LABEL: test_pm4addsu_i16x4:
+// RV32:        pm2waddsu.h
+// RV32:        pm2waddasu.h
+// RV64:        pm4addsu.h
+int64_t test_pm4addsu_i16x4(int16x4_t rs1, uint16x4_t rs2) {
+  return __riscv_pm4addsu_i16x4(rs1, rs2);
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsRISCV.td 
b/llvm/include/llvm/IR/IntrinsicsRISCV.td
index 8fa0ee78ca2a5..68fc04ed60176 100644
--- a/llvm/include/llvm/IR/IntrinsicsRISCV.td
+++ b/llvm/include/llvm/IR/IntrinsicsRISCV.td
@@ -2125,6 +2125,25 @@ class RVPBinaryIntrinsic
   def int_riscv_pmhaccsu  : RVPTernaryIntrinsic;
   def int_riscv_pmhraccsu : RVPTernaryIntrinsic;
 
+  // Packed Multiplication with Horizontal Addition.
+  class RVPHorizontalMulIntrinsic
+      : DefaultAttrsIntrinsic<[llvm_any_ty],
+                              [llvm_anyvector_ty, LLVMMatchType<1>],
+                              [IntrNoMem, IntrSpeculatable]>;
+  def int_riscv_pm4add     : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2add     : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2add_x   : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm4addu    : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2addu    : RVPHorizontalMulIntrinsic;
+  def int_riscv_pmq2add    : RVPHorizontalMulIntrinsic;
+  def int_riscv_pmqr2add   : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2sadd    : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2sadd_x  : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2sub     : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2sub_x   : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm4addsu   : RVPHorizontalMulIntrinsic;
+  def int_riscv_pm2addsu   : RVPHorizontalMulIntrinsic;
+
   // Packed Absolute Difference Sum.
   def int_riscv_pabdsumu
       : DefaultAttrsIntrinsic<[llvm_anyint_ty],
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp 
b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 194aad22775cf..8f80fbfa5c338 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -12254,6 +12254,40 @@ static unsigned getRVPMulHighAccumulateOpcode(unsigned 
IntNo) {
   }
 }
 
+static unsigned getRVPHorizontalMulOpcode(unsigned IntNo) {
+  switch (IntNo) {
+  default:
+    llvm_unreachable(
+        "Unexpected RISC-V packed horizontal multiply intrinsic");
+  case Intrinsic::riscv_pm4add:
+    return RISCVISD::PM4ADD;
+  case Intrinsic::riscv_pm2add:
+    return RISCVISD::PM2ADD;
+  case Intrinsic::riscv_pm2add_x:
+    return RISCVISD::PM2ADD_X;
+  case Intrinsic::riscv_pm4addu:
+    return RISCVISD::PM4ADDU;
+  case Intrinsic::riscv_pm2addu:
+    return RISCVISD::PM2ADDU;
+  case Intrinsic::riscv_pmq2add:
+    return RISCVISD::PMQ2ADD;
+  case Intrinsic::riscv_pmqr2add:
+    return RISCVISD::PMQR2ADD;
+  case Intrinsic::riscv_pm2sadd:
+    return RISCVISD::PM2SADD;
+  case Intrinsic::riscv_pm2sadd_x:
+    return RISCVISD::PM2SADD_X;
+  case Intrinsic::riscv_pm2sub:
+    return RISCVISD::PM2SUB;
+  case Intrinsic::riscv_pm2sub_x:
+    return RISCVISD::PM2SUB_X;
+  case Intrinsic::riscv_pm4addsu:
+    return RISCVISD::PM4ADDSU;
+  case Intrinsic::riscv_pm2addsu:
+    return RISCVISD::PM2ADDSU;
+  }
+}
+
 SDValue RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                      SelectionDAG &DAG) const {
   unsigned IntNo = Op.getConstantOperandVal(0);
@@ -12643,6 +12677,36 @@ SDValue 
RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
 
     return DAG.getNode(MulOpc, DL, VT, Rd, Rs1, Rs2);
   }
+  case Intrinsic::riscv_pm4add:
+  case Intrinsic::riscv_pm2add:
+  case Intrinsic::riscv_pm2add_x:
+  case Intrinsic::riscv_pm4addu:
+  case Intrinsic::riscv_pm2addu:
+  case Intrinsic::riscv_pmq2add:
+  case Intrinsic::riscv_pmqr2add:
+  case Intrinsic::riscv_pm2sadd:
+  case Intrinsic::riscv_pm2sadd_x:
+  case Intrinsic::riscv_pm2sub:
+  case Intrinsic::riscv_pm2sub_x:
+  case Intrinsic::riscv_pm4addsu:
+  case Intrinsic::riscv_pm2addsu: {
+    EVT VT = Op.getValueType();
+    unsigned Opc = getRVPHorizontalMulOpcode(IntNo);
+    SDValue Rs1 = Op.getOperand(1);
+    SDValue Rs2 = Op.getOperand(2);
+
+    // RV32 applies the 32-bit instruction independently to both halves of a
+    // 64-bit packed input.
+    if (!Subtarget.is64Bit() && VT == MVT::v2i32) {
+      auto [Rs1Lo, Rs1Hi] = DAG.SplitVector(Rs1, DL);
+      auto [Rs2Lo, Rs2Hi] = DAG.SplitVector(Rs2, DL);
+      SDValue Lo = DAG.getNode(Opc, DL, MVT::i32, Rs1Lo, Rs2Lo);
+      SDValue Hi = DAG.getNode(Opc, DL, MVT::i32, Rs1Hi, Rs2Hi);
+      return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Lo, Hi);
+    }
+
+    return DAG.getNode(Opc, DL, VT, Rs1, Rs2);
+  }
   case Intrinsic::riscv_pmerge: {
     EVT VT = Op.getValueType();
     auto buildMerge = [&](SDValue Rs1, SDValue Rs2, SDValue Mask,
@@ -16734,6 +16798,42 @@ void RISCVTargetLowering::ReplaceNodeResults(SDNode *N,
       Results.push_back(DAG.getExtractSubvector(DL, VT, Res, 0));
       return;
     }
+    case Intrinsic::riscv_pm4add:
+    case Intrinsic::riscv_pm2add:
+    case Intrinsic::riscv_pm2add_x:
+    case Intrinsic::riscv_pm4addu:
+    case Intrinsic::riscv_pm2addu:
+    case Intrinsic::riscv_pmq2add:
+    case Intrinsic::riscv_pmqr2add:
+    case Intrinsic::riscv_pm2sadd:
+    case Intrinsic::riscv_pm2sadd_x:
+    case Intrinsic::riscv_pm2sub:
+    case Intrinsic::riscv_pm2sub_x:
+    case Intrinsic::riscv_pm4addsu:
+    case Intrinsic::riscv_pm2addsu: {
+      MVT VT = N->getSimpleValueType(0);
+      unsigned Opc = getRVPHorizontalMulOpcode(IntNo);
+      SDValue Rs1 = N->getOperand(1);
+      SDValue Rs2 = N->getOperand(2);
+
+      if (!Subtarget.is64Bit() && VT == MVT::i64) {
+        SDValue Pair = DAG.getNode(Opc, DL, MVT::v2i32, Rs1, Rs2);
+        Results.push_back(DAG.getBitcast(MVT::i64, Pair));
+        return;
+      }
+
+      assert(Subtarget.is64Bit() && VT == MVT::i32 &&
+             "Unexpected horizontal multiply legalization");
+      MVT SrcVT = Rs1.getSimpleValueType();
+      MVT WideSrcVT = SrcVT == MVT::v4i8 ? MVT::v8i8 : MVT::v4i16;
+      Rs1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, WideSrcVT, Rs1,
+                        DAG.getUNDEF(SrcVT));
+      Rs2 = DAG.getNode(ISD::CONCAT_VECTORS, DL, WideSrcVT, Rs2,
+                        DAG.getUNDEF(SrcVT));
+      SDValue Wide = DAG.getNode(Opc, DL, MVT::v2i32, Rs1, Rs2);
+      Results.push_back(DAG.getExtractVectorElt(DL, MVT::i32, Wide, 0));
+      return;
+    }
     case Intrinsic::riscv_paadd:
     case Intrinsic::riscv_paaddu:
     case Intrinsic::riscv_pasub:
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td 
b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
index 9b72141ef9bcd..ee7b78fcfa9cc 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoP.td
@@ -1923,6 +1923,23 @@ def riscv_mhraccu  : RVSDNode<"MHRACCU", 
SDT_RISCVSameTernary>;
 def riscv_mhaccsu  : RVSDNode<"MHACCSU", SDT_RISCVSameTernary>;
 def riscv_mhraccsu : RVSDNode<"MHRACCSU", SDT_RISCVSameTernary>;
 
+// Multiply packed elements and horizontally combine each adjacent group.
+def SDT_RISCVHorizontalMul
+    : SDTypeProfile<1, 2, [SDTCisVec<1>, SDTCisSameAs<1, 2>]>;
+def riscv_pm4add    : RVSDNode<"PM4ADD", SDT_RISCVHorizontalMul>;
+def riscv_pm2add    : RVSDNode<"PM2ADD", SDT_RISCVHorizontalMul>;
+def riscv_pm2add_x  : RVSDNode<"PM2ADD_X", SDT_RISCVHorizontalMul>;
+def riscv_pm4addu   : RVSDNode<"PM4ADDU", SDT_RISCVHorizontalMul>;
+def riscv_pm2addu   : RVSDNode<"PM2ADDU", SDT_RISCVHorizontalMul>;
+def riscv_pmq2add   : RVSDNode<"PMQ2ADD", SDT_RISCVHorizontalMul>;
+def riscv_pmqr2add  : RVSDNode<"PMQR2ADD", SDT_RISCVHorizontalMul>;
+def riscv_pm2sadd   : RVSDNode<"PM2SADD", SDT_RISCVHorizontalMul>;
+def riscv_pm2sadd_x : RVSDNode<"PM2SADD_X", SDT_RISCVHorizontalMul>;
+def riscv_pm2sub    : RVSDNode<"PM2SUB", SDT_RISCVHorizontalMul>;
+def riscv_pm2sub_x  : RVSDNode<"PM2SUB_X", SDT_RISCVHorizontalMul>;
+def riscv_pm4addsu  : RVSDNode<"PM4ADDSU", SDT_RISCVHorizontalMul>;
+def riscv_pm2addsu  : RVSDNode<"PM2ADDSU", SDT_RISCVHorizontalMul>;
+
 // "Q-format" multiplication
 def riscv_mulq  : RVSDNode<"MULQ",  SDTIntBinOp, [SDNPCommutative]>;
 def riscv_mulqr : RVSDNode<"MULQR", SDTIntBinOp, [SDNPCommutative]>;
@@ -2023,6 +2040,11 @@ def RISCVBuildPairGPRVec : RVSDNode<"BuildPairGPRVec",
                                     SDT_RISCVBuildPairGPRVec>;
 def RISCVSplitGPRVec     : RVSDNode<"SplitGPRVec", SDT_RISCVSplitGPRVec>;
 
+class PatHorizontalMul<SDPatternOperator OpNode, RVInst Inst,
+                       ValueType ResultVT, ValueType SrcVT>
+    : Pat<(ResultVT (OpNode (SrcVT GPR:$rs1), (SrcVT GPR:$rs2))),
+          (Inst GPR:$rs1, GPR:$rs2)>;
+
 // (rs1 << 1) + rs2
 class PatPSh1Add<RVInst Inst, ValueType vt>
     : Pat<(vt (add (riscv_pshl (vt GPR:$rs1), (XLenVT 1)), (vt GPR:$rs2))),
@@ -2300,6 +2322,97 @@ let append Predicates = [IsRV32] in {
   def : PatGprGpr<riscv_mulq, MULQ, i32>;
   def : PatGprGpr<riscv_mulqr, MULQR, i32>;
 
+  // Packed multiplication with horizontal addition patterns.
+  def : PatHorizontalMul<riscv_pm4add, PM4ADD_B, i32, v4i8>;
+  def : PatHorizontalMul<riscv_pm2add, PM2ADD_H, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm2add_x, PM2ADD_HX, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm4addu, PM4ADDU_B, i32, v4i8>;
+  def : PatHorizontalMul<riscv_pm2addu, PM2ADDU_H, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pmq2add, PMQ2ADD_H, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pmqr2add, PMQR2ADD_H, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm2sadd, PM2SADD_H, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm2sadd_x, PM2SADD_HX, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm2sub, PM2SUB_H, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm2sub_x, PM2SUB_HX, i32, v2i16>;
+  def : PatHorizontalMul<riscv_pm4addsu, PM4ADDSU_B, i32, v4i8>;
+  def : PatHorizontalMul<riscv_pm2addsu, PM2ADDSU_H, i32, v2i16>;
+
+  // Select the RV32 v2i32 result produced by type legalization into the
+  // instruction sequences specified for the 64-bit intrinsics.
+  def : Pat<(v2i32 (riscv_pm2add (v2i32 GPRPair:$rs1),
+                                 (v2i32 GPRPair:$rs2))),
+            (WMACC (WMUL (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                          (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                   (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                   (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pm2add_x (v2i32 GPRPair:$rs1),
+                                   (v2i32 GPRPair:$rs2))),
+            (WMACC (WMUL (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                          (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd)),
+                   (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                   (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even))>;
+  def : Pat<(v2i32 (riscv_pm2addu (v2i32 GPRPair:$rs1),
+                                  (v2i32 GPRPair:$rs2))),
+            (WMACCU (WMULU (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                            (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                    (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                    (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pmq2add (v2i32 GPRPair:$rs1),
+                                  (v2i32 GPRPair:$rs2))),
+            (MQWACC (v2i32 (MQWACC
+                       (v2i32 (BuildGPRPair (i32 X0), (i32 X0))),
+                       (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                       (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even))),
+                    (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                    (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pmqr2add (v2i32 GPRPair:$rs1),
+                                   (v2i32 GPRPair:$rs2))),
+            (MQRWACC (v2i32 (MQRWACC
+                        (v2i32 (BuildGPRPair (i32 X0), (i32 X0))),
+                        (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                        (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even))),
+                     (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                     (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pm2sub (v2i32 GPRPair:$rs1),
+                                 (v2i32 GPRPair:$rs2))),
+            (SUBD (WMUL (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                         (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                  (WMUL (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                         (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd)))>;
+  def : Pat<(v2i32 (riscv_pm2sub_x (v2i32 GPRPair:$rs1),
+                                   (v2i32 GPRPair:$rs2))),
+            (SUBD (WMUL (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                         (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd)),
+                  (WMUL (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                         (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)))>;
+  def : Pat<(v2i32 (riscv_pm2addsu (v2i32 GPRPair:$rs1),
+                                   (v2i32 GPRPair:$rs2))),
+            (WMACCSU (WMULSU (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                              (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                      (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                      (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pm4add (v4i16 GPRPair:$rs1),
+                                 (v4i16 GPRPair:$rs2))),
+            (PM2WADDA_H
+                (PM2WADD_H (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                            (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pm4addu (v4i16 GPRPair:$rs1),
+                                  (v4i16 GPRPair:$rs2))),
+            (PM2WADDAU_H
+                (PM2WADDU_H (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                             (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+  def : Pat<(v2i32 (riscv_pm4addsu (v4i16 GPRPair:$rs1),
+                                   (v4i16 GPRPair:$rs2))),
+            (PM2WADDASU_H
+                (PM2WADDSU_H (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_even),
+                              (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_even)),
+                (EXTRACT_SUBREG GPRPair:$rs1, sub_gpr_odd),
+                (EXTRACT_SUBREG GPRPair:$rs2, sub_gpr_odd))>;
+
   // Halfword multiply patterns where one operand is a sext.h or zext.h and
   // the other is a sext.h or zext.h or is known to be sign/zero-extended. We
   // prefer plain mul when both operands are known to be sign/zero-extended.
@@ -2849,6 +2962,33 @@ let append Predicates = [IsRV64] in {
   def : PatGprGpr<riscv_mulq, PMULQ_W, v2i32>;
   def : PatGprGpr<riscv_mulqr, PMULQR_W, v2i32>;
 
+  // Packed multiplication with horizontal addition patterns.
+  def : PatHorizontalMul<riscv_pm4add, PM4ADD_B, v2i32, v8i8>;
+  def : PatHorizontalMul<riscv_pm2add, PM2ADD_H, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm2add_x, PM2ADD_HX, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm4addu, PM4ADDU_B, v2i32, v8i8>;
+  def : PatHorizontalMul<riscv_pm2addu, PM2ADDU_H, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pmq2add, PMQ2ADD_H, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pmqr2add, PMQR2ADD_H, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm2sadd, PM2SADD_H, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm2sadd_x, PM2SADD_HX, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm2sub, PM2SUB_H, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm2sub_x, PM2SUB_HX, v2i32, v4i16>;
+  def : PatHorizontalMul<riscv_pm4addsu, PM4ADDSU_B, v2i32, v8i8>;
+  def : PatHorizontalMul<riscv_pm2addsu, PM2ADDSU_H, v2i32, v4i16>;
+
+  def : PatHorizontalMul<riscv_pm2add, PM2ADD_W, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pm2add_x, PM2ADD_WX, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pm2addu, PM2ADDU_W, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pmq2add, PMQ2ADD_W, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pmqr2add, PMQR2ADD_W, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pm2sub, PM2SUB_W, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pm2sub_x, PM2SUB_WX, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pm2addsu, PM2ADDSU_W, i64, v2i32>;
+  def : PatHorizontalMul<riscv_pm4add, PM4ADD_H, i64, v4i16>;
+  def : PatHorizontalMul<riscv_pm4addu, PM4ADDU_H, i64, v4i16>;
+  def : PatHorizontalMul<riscv_pm4addsu, PM4ADDSU_H, i64, v4i16>;
+
   // 8/16/32-bit multiply low patterns
   // FIXME custom lower
   def : Pat<(v8i8 (mul GPR:$rs1, GPR:$rs2)),
diff --git a/llvm/test/CodeGen/RISCV/rvp-simd-32.ll 
b/llvm/test/CodeGen/RISCV/rvp-simd-32.ll
index d38012912872d..103d91eba595f 100644
--- a/llvm/test/CodeGen/RISCV/rvp-simd-32.ll
+++ b/llvm/test/CodeGen/RISCV/rvp-simd-32.ll
@@ -3156,3 +3156,120 @@ define <2 x i16> @test_pnclipup_v2i16(i32 %a, i32 %b) {
   %r = call <2 x i16> @llvm.riscv.pnclipup.v2i16.i32(i32 %a, i32 %b)
   ret <2 x i16> %r
 }
+
+define i32 @test_pm4add_v4i8_i32(<4 x i8> %a, <4 x i8> %b) {
+; CHECK-LABEL: test_pm4add_v4i8_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm4add.b a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm4add.i32.v4i8(<4 x i8> %a, <4 x i8> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2add_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2add_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2add.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2add.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2add_x_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2add_x_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2add.hx a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2add.x.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm4addu_v4i8_i32(<4 x i8> %a, <4 x i8> %b) {
+; CHECK-LABEL: test_pm4addu_v4i8_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm4addu.b a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm4addu.i32.v4i8(<4 x i8> %a, <4 x i8> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2addu_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2addu_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2addu.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2addu.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pmq2add_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pmq2add_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pmq2add.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pmq2add.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pmqr2add_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pmqr2add_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pmqr2add.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pmqr2add.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2sadd_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2sadd_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2sadd.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2sadd.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2sadd_x_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2sadd_x_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2sadd.hx a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2sadd.x.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2sub_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2sub_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2sub.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2sub.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2sub_x_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2sub_x_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2sub.hx a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2sub.x.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm4addsu_v4i8_i32(<4 x i8> %a, <4 x i8> %b) {
+; CHECK-LABEL: test_pm4addsu_v4i8_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm4addsu.b a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm4addsu.i32.v4i8(<4 x i8> %a, <4 x i8> %b)
+  ret i32 %r
+}
+
+define i32 @test_pm2addsu_v2i16_i32(<2 x i16> %a, <2 x i16> %b) {
+; CHECK-LABEL: test_pm2addsu_v2i16_i32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pm2addsu.h a0, a0, a1
+; CHECK-NEXT:    ret
+  %r = call i32 @llvm.riscv.pm2addsu.i32.v2i16(<2 x i16> %a, <2 x i16> %b)
+  ret i32 %r
+}
diff --git a/llvm/test/CodeGen/RISCV/rvp-simd-64.ll 
b/llvm/test/CodeGen/RISCV/rvp-simd-64.ll
index 96f16c3d52103..705718e63ba4f 100644
--- a/llvm/test/CodeGen/RISCV/rvp-simd-64.ll
+++ b/llvm/test/CodeGen/RISCV/rvp-simd-64.ll
@@ -6936,3 +6936,378 @@ define i64 @test_pnclipup_v2i32(i64 %a, i64 %b) {
   %s = bitcast <2 x i32> %r to i64
   ret i64 %s
 }
+
+define <2 x i32> @test_pm4add_v8i8_v2i32(<8 x i8> %a, <8 x i8> %b) {
+; RV32-LABEL: test_pm4add_v8i8_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm4add.b a1, a1, a3
+; RV32-NEXT:    pm4add.b a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm4add_v8i8_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm4add.b a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm4add.v2i32.v8i8(<8 x i8> %a, <8 x i8> %b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2add_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2add_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2add.h a1, a1, a3
+; RV32-NEXT:    pm2add.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2add_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2add.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2add.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2add_x_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2add_x_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2add.hx a1, a1, a3
+; RV32-NEXT:    pm2add.hx a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2add_x_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2add.hx a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2add.x.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm4addu_v8i8_v2i32(<8 x i8> %a, <8 x i8> %b) {
+; RV32-LABEL: test_pm4addu_v8i8_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm4addu.b a1, a1, a3
+; RV32-NEXT:    pm4addu.b a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm4addu_v8i8_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm4addu.b a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm4addu.v2i32.v8i8(<8 x i8> %a, <8 x i8> %b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2addu_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2addu_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2addu.h a1, a1, a3
+; RV32-NEXT:    pm2addu.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2addu_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2addu.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2addu.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pmq2add_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pmq2add_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pmq2add.h a1, a1, a3
+; RV32-NEXT:    pmq2add.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pmq2add_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pmq2add.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pmq2add.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pmqr2add_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pmqr2add_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pmqr2add.h a1, a1, a3
+; RV32-NEXT:    pmqr2add.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pmqr2add_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pmqr2add.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pmqr2add.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2sadd_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2sadd_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2sadd.h a1, a1, a3
+; RV32-NEXT:    pm2sadd.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2sadd_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2sadd.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2sadd.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2sadd_x_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2sadd_x_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2sadd.hx a1, a1, a3
+; RV32-NEXT:    pm2sadd.hx a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2sadd_x_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2sadd.hx a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2sadd.x.v2i32.v4i16(<4 x i16> %a, <4 x 
i16> %b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2sub_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2sub_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2sub.h a1, a1, a3
+; RV32-NEXT:    pm2sub.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2sub_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2sub.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2sub.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2sub_x_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2sub_x_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2sub.hx a1, a1, a3
+; RV32-NEXT:    pm2sub.hx a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2sub_x_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2sub.hx a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2sub.x.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm4addsu_v8i8_v2i32(<8 x i8> %a, <8 x i8> %b) {
+; RV32-LABEL: test_pm4addsu_v8i8_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm4addsu.b a1, a1, a3
+; RV32-NEXT:    pm4addsu.b a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm4addsu_v8i8_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm4addsu.b a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm4addsu.v2i32.v8i8(<8 x i8> %a, <8 x i8> %b)
+  ret <2 x i32> %r
+}
+
+define <2 x i32> @test_pm2addsu_v4i16_v2i32(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm2addsu_v4i16_v2i32:
+; RV32:       # %bb.0:
+; RV32-NEXT:    pm2addsu.h a1, a1, a3
+; RV32-NEXT:    pm2addsu.h a0, a0, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2addsu_v4i16_v2i32:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2addsu.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call <2 x i32> @llvm.riscv.pm2addsu.v2i32.v4i16(<4 x i16> %a, <4 x i16> 
%b)
+  ret <2 x i32> %r
+}
+
+define i64 @test_pm2add_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pm2add_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    wmul a0, a0, a2
+; RV32-NEXT:    wmacc a0, a4, a3
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2add_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2add.w a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm2add.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm2add_x_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pm2add_x_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    wmul a0, a0, a3
+; RV32-NEXT:    wmacc a0, a4, a2
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2add_x_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2add.wx a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm2add.x.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm2addu_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pm2addu_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    wmulu a0, a0, a2
+; RV32-NEXT:    wmaccu a0, a4, a3
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2addu_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2addu.w a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm2addu.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pmq2add_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pmq2add_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    li a4, 0
+; RV32-NEXT:    li a5, 0
+; RV32-NEXT:    mqwacc a4, a0, a2
+; RV32-NEXT:    mqwacc a4, a1, a3
+; RV32-NEXT:    mvd a0, a4
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pmq2add_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pmq2add.w a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pmq2add.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pmqr2add_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pmqr2add_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    li a4, 0
+; RV32-NEXT:    li a5, 0
+; RV32-NEXT:    mqrwacc a4, a0, a2
+; RV32-NEXT:    mqrwacc a4, a1, a3
+; RV32-NEXT:    mvd a0, a4
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pmqr2add_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pmqr2add.w a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pmqr2add.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm2sub_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pm2sub_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    wmul a4, a1, a3
+; RV32-NEXT:    wmul a0, a0, a2
+; RV32-NEXT:    subd a0, a0, a4
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2sub_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2sub.w a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm2sub.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm2sub_x_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pm2sub_x_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    wmul a4, a1, a2
+; RV32-NEXT:    wmul a0, a0, a3
+; RV32-NEXT:    subd a0, a0, a4
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2sub_x_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2sub.wx a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm2sub.x.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm2addsu_v2i32_i64(<2 x i32> %a, <2 x i32> %b) {
+; RV32-LABEL: test_pm2addsu_v2i32_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    wmulsu a0, a0, a2
+; RV32-NEXT:    wmaccsu a0, a4, a3
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm2addsu_v2i32_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm2addsu.w a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm2addsu.i64.v2i32(<2 x i32> %a, <2 x i32> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm4add_v4i16_i64(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm4add_v4i16_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    pm2wadd.h a0, a0, a2
+; RV32-NEXT:    pm2wadda.h a0, a4, a3
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm4add_v4i16_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm4add.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm4add.i64.v4i16(<4 x i16> %a, <4 x i16> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm4addu_v4i16_i64(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm4addu_v4i16_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    pm2waddu.h a0, a0, a2
+; RV32-NEXT:    pm2waddau.h a0, a4, a3
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm4addu_v4i16_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm4addu.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm4addu.i64.v4i16(<4 x i16> %a, <4 x i16> %b)
+  ret i64 %r
+}
+
+define i64 @test_pm4addsu_v4i16_i64(<4 x i16> %a, <4 x i16> %b) {
+; RV32-LABEL: test_pm4addsu_v4i16_i64:
+; RV32:       # %bb.0:
+; RV32-NEXT:    mv a4, a1
+; RV32-NEXT:    pm2waddsu.h a0, a0, a2
+; RV32-NEXT:    pm2waddasu.h a0, a4, a3
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: test_pm4addsu_v4i16_i64:
+; RV64:       # %bb.0:
+; RV64-NEXT:    pm4addsu.h a0, a0, a1
+; RV64-NEXT:    ret
+  %r = call i64 @llvm.riscv.pm4addsu.i64.v4i16(<4 x i16> %a, <4 x i16> %b)
+  ret i64 %r
+}

_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to