[WebAssembly] Prototype extending pairwise add instructions

[tlively]

Summary

As proposed in https://github.com/WebAssembly/simd/pull/380. This commit makes
the new instructions available only via clang builtins and LLVM intrinsics to
make their use opt-in while they are still being evaluated for inclusion in the
SIMD proposal.

Depends on D93771.

clang/include/clang/Basic/BuiltinsWebAssembly.def

 TARGET_BUILTIN(__builtin_wasm_extmul_low_i16x8_u_i32x4, "V4UiV8UsV8Us", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_extmul_high_i16x8_u_i32x4, "V4UiV8UsV8Us", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_extmul_low_i32x4_s_i64x2, "V2LLiV4iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_extmul_high_i32x4_s_i64x2, "V2LLiV4iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_extmul_low_i32x4_u_i64x2, "V2ULLiV4UiV4Ui", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_extmul_high_i32x4_u_i64x2, "V2ULLiV4UiV4Ui", "nc", "simd128")
 
+TARGET_BUILTIN(__builtin_wasm_extadd_pairwise_i8x16_s_i16x8, "V8sV16Sc", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_extadd_pairwise_i8x16_u_i16x8, "V8UsV16Uc", "nc", "simd128")
+
+TARGET_BUILTIN(__builtin_wasm_extadd_pairwise_i16x8_s_i32x4, "V4iV8s", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_extadd_pairwise_i16x8_u_i32x4, "V4UiV8Us", "nc", "simd128")
+
 TARGET_BUILTIN(__builtin_wasm_bitselect, "V4iV4iV4iV4i", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_signselect_i8x16, "V16ScV16ScV16ScV16Sc", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_signselect_i16x8, "V8sV8sV8sV8s", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_signselect_i32x4, "V4iV4iV4iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_signselect_i64x2, "V2LLiV2LLiV2LLiV2LLi", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_shuffle_v8x16, "V16ScV16ScV16ScIiIiIiIiIiIiIiIiIiIiIiIiIiIiIiIi", "nc", "simd128")

clang/lib/CodeGen/CGBuiltin.cpp

case WebAssembly::BI__builtin_wasm_extmul_high_i32x4_u_i64x2:
       break;
     default:
       llvm_unreachable("unexptected builtin ID");
     }
 
     Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
     return Builder.CreateCall(Callee, {LHS, RHS});
   }
+  case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8:
+  case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8:
+  case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4:
+  case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4: {
+    Value *Vec = EmitScalarExpr(E->getArg(0));
+    unsigned IntNo;
+    switch (BuiltinID) {
+    case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8:
+    case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4:
+      IntNo = Intrinsic::wasm_extadd_pairwise_signed;
+      break;
+    case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8:
+    case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4:
+      IntNo = Intrinsic::wasm_extadd_pairwise_unsigned;
+      break;
+    default:
+      llvm_unreachable("unexptected builtin ID");
+    }
+
+    Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
+    return Builder.CreateCall(Callee, Vec);
+  }
   case WebAssembly::BI__builtin_wasm_bitselect: {
     Value *V1 = EmitScalarExpr(E->getArg(0));
     Value *V2 = EmitScalarExpr(E->getArg(1));
     Value *C = EmitScalarExpr(E->getArg(2));
     Function *Callee =
         CGM.getIntrinsic(Intrinsic::wasm_bitselect, ConvertType(E->getType()));
     return Builder.CreateCall(Callee, {V1, V2, C});
   }

clang/test/CodeGen/builtins-wasm.c

 
 u64x2 extmul_high_i32x4_u_i64x2(u32x4 x, u32x4 y) {
   return __builtin_wasm_extmul_high_i32x4_u_i64x2(x, y);
   // WEBASSEMBLY: call <2 x i64> @llvm.wasm.extmul.high.unsigned.v2i64(
   // WEBASSEMBLY-SAME: <4 x i32> %x, <4 x i32> %y)
   // WEBASSEMBLY-NEXT: ret
 }
 
+i16x8 extadd_pairwise_i8x16_s_i16x8(i8x16 v) {
+  return __builtin_wasm_extadd_pairwise_i8x16_s_i16x8(v);
+  // WEBASSEMBLY: call <8 x i16> @llvm.wasm.extadd.pairwise.signed.v8i16(
+  // WEBASSEMBLY-SAME: <16 x i8> %v)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+u16x8 extadd_pairwise_i8x16_u_i16x8(u8x16 v) {
+  return __builtin_wasm_extadd_pairwise_i8x16_u_i16x8(v);
+  // WEBASSEMBLY: call <8 x i16> @llvm.wasm.extadd.pairwise.unsigned.v8i16(
+  // WEBASSEMBLY-SAME: <16 x i8> %v)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+i32x4 extadd_pairwise_i16x8_s_i32x4(i16x8 v) {
+  return __builtin_wasm_extadd_pairwise_i16x8_s_i32x4(v);
+  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.extadd.pairwise.signed.v4i32(
+  // WEBASSEMBLY-SAME: <8 x i16> %v)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+u32x4 extadd_pairwise_i16x8_u_i32x4(u16x8 v) {
+  return __builtin_wasm_extadd_pairwise_i16x8_u_i32x4(v);
+  // WEBASSEMBLY: call <4 x i32> @llvm.wasm.extadd.pairwise.unsigned.v4i32(
+  // WEBASSEMBLY-SAME: <8 x i16> %v)
+  // WEBASSEMBLY-NEXT: ret
+}
+
 i32x4 dot_i16x8_s(i16x8 x, i16x8 y) {
   return __builtin_wasm_dot_s_i32x4_i16x8(x, y);
   // WEBASSEMBLY: call <4 x i32> @llvm.wasm.dot(<8 x i16> %x, <8 x i16> %y)
   // WEBASSEMBLY-NEXT: ret
 }
 
 i32x4 bitselect(i32x4 x, i32x4 y, i32x4 c) {
   return __builtin_wasm_bitselect(x, y, c);

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

def int_wasm_extmul_low_unsigned :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMSubdivide2VectorType<0>, LLVMSubdivide2VectorType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_extmul_high_unsigned :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMSubdivide2VectorType<0>, LLVMSubdivide2VectorType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 
+def int_wasm_extadd_pairwise_signed :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMSubdivide2VectorType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_extadd_pairwise_unsigned :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMSubdivide2VectorType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+
 def int_wasm_signselect :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 
 // TODO: Remove this intrinsic and the associated builtin if i64x2.eq gets
 // merged to the proposal.
 def int_wasm_eq :

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

 foreach t2 = !foldl(
   []<ValueType>, [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64],
   acc, cur, !if(!eq(!cast<string>(t1), !cast<string>(cur)),
     acc, !listconcat(acc, [cur])
   )
 ) in
 def : Pat<(t1 (bitconvert (t2 V128:$v))), (t1 V128:$v)>;
 
+// Extended pairwise addition
+defm "" : SIMDConvert<I16x8, I8x16, int_wasm_extadd_pairwise_signed,
+                      "extadd_pairwise_i8x16_s", 0xc2>;
+defm "" : SIMDConvert<I16x8, I8x16, int_wasm_extadd_pairwise_unsigned,
+                      "extadd_pairwise_i8x16_u", 0xc3>;
+defm "" : SIMDConvert<I32x4, I16x8, int_wasm_extadd_pairwise_signed,
+                      "extadd_pairwise_i16x8_s", 0xa5>;
+defm "" : SIMDConvert<I32x4, I16x8, int_wasm_extadd_pairwise_unsigned,
+                      "extadd_pairwise_i16x8_u", 0xa6>;
+
+
 //===----------------------------------------------------------------------===//
 // Quasi-Fused Multiply- Add and Subtract (QFMA/QFMS)
 //===----------------------------------------------------------------------===//
 
 multiclass SIMDQFM<Vec vec, bits<32> simdopA, bits<32> simdopS> {
   defm QFMA_#vec :
     SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c),
            (outs), (ins),

llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll

 declare <8 x i16> @llvm.wasm.extmul.high.unsigned.v8i16(<16 x i8>, <16 x i8>)
 define <8 x i16> @extmul_high_u_v8i16(<16 x i8> %x, <16 x i8> %y) {
   %a = call <8 x i16> @llvm.wasm.extmul.high.unsigned.v8i16(
     <16 x i8> %x, <16 x i8> %y
   )
   ret <8 x i16> %a
 }
 
+; CHECK-LABEL: extadd_pairwise_s_v8i16:
+; SIMD128-NEXT: .functype extadd_pairwise_s_v8i16 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.extadd_pairwise_i8x16_s $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.extadd.pairwise.signed.v8i16(<16 x i8>)
+define <8 x i16> @extadd_pairwise_s_v8i16(<16 x i8> %x) {
+  %a = call <8 x i16> @llvm.wasm.extadd.pairwise.signed.v8i16(<16 x i8> %x)
+  ret <8 x i16> %a
+}
+
+; CHECK-LABEL: extadd_pairwise_u_v8i16:
+; SIMD128-NEXT: .functype extadd_pairwise_u_v8i16 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i16x8.extadd_pairwise_i8x16_u $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <8 x i16> @llvm.wasm.extadd.pairwise.unsigned.v8i16(<16 x i8>)
+define <8 x i16> @extadd_pairwise_u_v8i16(<16 x i8> %x) {
+  %a = call <8 x i16> @llvm.wasm.extadd.pairwise.unsigned.v8i16(<16 x i8> %x)
+  ret <8 x i16> %a
+}
+
 ; CHECK-LABEL: any_v8i16:
 ; SIMD128-NEXT: .functype any_v8i16 (v128) -> (i32){{$}}
 ; SIMD128-NEXT: i16x8.any_true $push[[R:[0-9]+]]=, $0{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare i32 @llvm.wasm.anytrue.v8i16(<8 x i16>)
 define i32 @any_v8i16(<8 x i16> %x) {
   %a = call i32 @llvm.wasm.anytrue.v8i16(<8 x i16> %x)
   ret i32 %a

 declare <4 x i32> @llvm.wasm.extmul.high.unsigned.v4i32(<8 x i16>, <8 x i16>)
 define <4 x i32> @extmul_high_u_v4i32(<8 x i16> %x, <8 x i16> %y) {
   %a = call <4 x i32> @llvm.wasm.extmul.high.unsigned.v4i32(
     <8 x i16> %x, <8 x i16> %y
   )
   ret <4 x i32> %a
 }
 
+; CHECK-LABEL: extadd_pairwise_s_v4i32:
+; SIMD128-NEXT: .functype extadd_pairwise_s_v4i32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.extadd_pairwise_i16x8_s $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.extadd.pairwise.signed.v4i32(<8 x i16>)
+define <4 x i32> @extadd_pairwise_s_v4i32(<8 x i16> %x) {
+  %a = call <4 x i32> @llvm.wasm.extadd.pairwise.signed.v4i32(<8 x i16> %x)
+  ret <4 x i32> %a
+}
+
+; CHECK-LABEL: extadd_pairwise_u_v4i32:
+; SIMD128-NEXT: .functype extadd_pairwise_u_v4i32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.extadd_pairwise_i16x8_u $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.extadd.pairwise.unsigned.v4i32(<8 x i16>)
+define <4 x i32> @extadd_pairwise_u_v4i32(<8 x i16> %x) {
+  %a = call <4 x i32> @llvm.wasm.extadd.pairwise.unsigned.v4i32(<8 x i16> %x)
+  ret <4 x i32> %a
+}
+
+
 ; CHECK-LABEL: any_v4i32:
 ; SIMD128-NEXT: .functype any_v4i32 (v128) -> (i32){{$}}
 ; SIMD128-NEXT: i32x4.any_true $push[[R:[0-9]+]]=, $0{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare i32 @llvm.wasm.anytrue.v4i32(<4 x i32>)
 define i32 @any_v4i32(<4 x i32> %x) {
   %a = call i32 @llvm.wasm.anytrue.v4i32(<4 x i32> %x)
   ret i32 %a

llvm/test/MC/WebAssembly/simd-encodings.s

main:
     i16x8.signselect
 
     # CHECK: i32x4.signselect # encoding: [0xfd,0x7f]
     i32x4.signselect
 
     # CHECK: i64x2.signselect # encoding: [0xfd,0x94,0x01]
     i64x2.signselect
 
+    # CHECK: i16x8.extadd_pairwise_i8x16_s # encoding: [0xfd,0xc2,0x01]
+    i16x8.extadd_pairwise_i8x16_s
+
+    # CHECK: i16x8.extadd_pairwise_i8x16_u # encoding: [0xfd,0xc3,0x01]
+    i16x8.extadd_pairwise_i8x16_u
+
+    # CHECK: i32x4.extadd_pairwise_i16x8_s # encoding: [0xfd,0xa5,0x01]
+    i32x4.extadd_pairwise_i16x8_s
+
+    # CHECK: i32x4.extadd_pairwise_i16x8_u # encoding: [0xfd,0xa6,0x01]
+    i32x4.extadd_pairwise_i16x8_u
+
     end_function