[WebAssembly] Add experimental SIMD dot product instruction

Summary:
This instruction is not merged to the spec proposal, but we need it to
be implemented in the toolchain to experiment with it. It is available
only on an opt-in basis through a clang builtin.

Defined in https://github.com/WebAssembly/simd/pull/127.

Depends on D69696.

Reviewers: aheejin

Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, sunfish, cfe-commits, llvm-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D69697

diff --git a/clang/include/clang/Basic/BuiltinsWebAssembly.def b/clang/include/clang/Basic/BuiltinsWebAssembly.def
index 2233c20..2c57c5b 100644 (file)
--- a/clang/include/clang/Basic/BuiltinsWebAssembly.def
+++ b/clang/include/clang/Basic/BuiltinsWebAssembly.def
@@ -132,6 +132,8 @@ TARGET_BUILTIN(__builtin_wasm_max_f32x4, "V4fV4fV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_min_f64x2, "V2dV2dV2d", "nc", "unimplemented-simd128")
 TARGET_BUILTIN(__builtin_wasm_max_f64x2, "V2dV2dV2d", "nc", "unimplemented-simd128")
 
+TARGET_BUILTIN(__builtin_wasm_dot_s_i32x4_i16x8, "V4iV8sV8s", "nc", "simd128")
+
 TARGET_BUILTIN(__builtin_wasm_sqrt_f32x4, "V4fV4f", "nc", "unimplemented-simd128")
 TARGET_BUILTIN(__builtin_wasm_sqrt_f64x2, "V2dV2d", "nc", "unimplemented-simd128")
 

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index fef5cab..cc2cbb9 100644 (file)
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -14360,6 +14360,12 @@ Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
     Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
     return Builder.CreateCall(Callee, {LHS, RHS});
   }
+  case WebAssembly::BI__builtin_wasm_dot_s_i32x4_i16x8: {
+    Value *LHS = EmitScalarExpr(E->getArg(0));
+    Value *RHS = EmitScalarExpr(E->getArg(1));
+    Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_dot);
+    return Builder.CreateCall(Callee, {LHS, RHS});
+  }
   case WebAssembly::BI__builtin_wasm_any_true_i8x16:
   case WebAssembly::BI__builtin_wasm_any_true_i16x8:
   case WebAssembly::BI__builtin_wasm_any_true_i32x4:

diff --git a/clang/test/CodeGen/builtins-wasm.c b/clang/test/CodeGen/builtins-wasm.c
index c07e73f..eea51eb 100644 (file)
--- a/clang/test/CodeGen/builtins-wasm.c
+++ b/clang/test/CodeGen/builtins-wasm.c
@@ -436,6 +436,12 @@ i32x4 max_u_i32x4(i32x4 x, i32x4 y) {
   // 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);
   // WEBASSEMBLY: call <4 x i32> @llvm.wasm.bitselect.v4i32(

diff --git a/llvm/include/llvm/IR/IntrinsicsWebAssembly.td b/llvm/include/llvm/IR/IntrinsicsWebAssembly.td
index b99abb4..422334a 100644 (file)
--- a/llvm/include/llvm/IR/IntrinsicsWebAssembly.td
+++ b/llvm/include/llvm/IR/IntrinsicsWebAssembly.td
@@ -152,6 +152,10 @@ def int_wasm_qfms :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_dot :
+  Intrinsic<[llvm_v4i32_ty],
+            [llvm_v8i16_ty, llvm_v8i16_ty],
+            [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_narrow_signed :
   Intrinsic<[llvm_anyvector_ty],
             [llvm_anyvector_ty, LLVMMatchType<1>],

diff --git a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
index 2a423ac..137234f 100644 (file)
--- a/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
+++ b/llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td
@@ -738,6 +738,13 @@ defm MAX_S : SIMDBinaryIntNoI64x2<int_wasm_max_signed, "max_s", 96>;
 defm MAX_U : SIMDBinaryIntNoI64x2<int_wasm_max_unsigned, "max_u", 97>;
 } // isCommutable = 1
 
+// Widening dot product: i32x4.dot_i16x8_s
+let isCommutable = 1 in
+defm DOT : SIMD_I<(outs V128:$dst), (ins V128:$lhs, V128:$rhs), (outs), (ins),
+                  [(set V128:$dst, (int_wasm_dot V128:$lhs, V128:$rhs))],
+                  "i32x4.dot_i16x8_s\t$dst, $lhs, $rhs", "i32x4.dot_i16x8_s",
+                  217>;
+
 //===----------------------------------------------------------------------===//
 // Floating-point unary arithmetic
 //===----------------------------------------------------------------------===//

diff --git a/llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll b/llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll
index d404e0b..7571524 100644 (file)
--- a/llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll
+++ b/llvm/test/CodeGen/WebAssembly/simd-intrinsics.ll
@@ -387,6 +387,16 @@ define <4 x i32> @max_u_v4i32(<4 x i32> %x, <4 x i32> %y) {
   ret <4 x i32> %a
 }
 
+; CHECK-LABEL: dot:
+; SIMD128-NEXT: .functype dot (v128, v128) -> (v128){{$}}
+; SIMD128-NEXT: i32x4.dot_i16x8_s $push[[R:[0-9]+]]=, $0, $1{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x i32> @llvm.wasm.dot(<8 x i16>, <8 x i16>)
+define <4 x i32> @dot(<8 x i16> %x, <8 x i16> %y) {
+  %a = call <4 x i32> @llvm.wasm.dot(<8 x i16> %x, <8 x i16> %y)
+  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]+]]=, $0{{$}}

diff --git a/llvm/test/MC/WebAssembly/simd-encodings.s b/llvm/test/MC/WebAssembly/simd-encodings.s
index adddcf5..e1cb5fe 100644 (file)
--- a/llvm/test/MC/WebAssembly/simd-encodings.s
+++ b/llvm/test/MC/WebAssembly/simd-encodings.s
@@ -571,4 +571,7 @@ main:
     # CHECK: v128.andnot # encoding: [0xfd,0xd8,0x01]
     v128.andnot
 
+    # CHECK: i32x4.dot_i16x8_s # encoding: [0xfd,0xd9,0x01]
+    i32x4.dot_i16x8_s
+
     end_function