[WebAssembly] SIMD bitmask intrinsics and builtin functions

Summary:
These experimental new instructions are proposed in
WebAssembly/simd#201.

Reviewers: aheejin

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

Tags: #clang

Differential Revision: https://reviews.llvm.org/D76397

Author: tlively

clang/include/clang/Basic/BuiltinsWebAssembly.def

@@ -125,6 +125,10 @@ TARGET_BUILTIN(__builtin_wasm_all_true_i16x8, "iV8s", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_all_true_i32x4, "iV4i", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_all_true_i64x2, "iV2LLi", "nc", "unimplemented-simd128")
 
+TARGET_BUILTIN(__builtin_wasm_bitmask_i8x16, "iV16c", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_bitmask_i16x8, "iV8s", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_bitmask_i32x4, "iV4i", "nc", "simd128")
+
 TARGET_BUILTIN(__builtin_wasm_abs_f32x4, "V4fV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_abs_f64x2, "V2dV2d", "nc", "simd128")
 

clang/lib/CodeGen/CGBuiltin.cpp

@@ -15142,6 +15142,14 @@ Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
     Function *Callee = CGM.getIntrinsic(IntNo, Vec->getType());
     return Builder.CreateCall(Callee, {Vec});
   }
+  case WebAssembly::BI__builtin_wasm_bitmask_i8x16:
+  case WebAssembly::BI__builtin_wasm_bitmask_i16x8:
+  case WebAssembly::BI__builtin_wasm_bitmask_i32x4: {
+    Value *Vec = EmitScalarExpr(E->getArg(0));
+    Function *Callee =
+        CGM.getIntrinsic(Intrinsic::wasm_bitmask, Vec->getType());
+    return Builder.CreateCall(Callee, {Vec});
+  }
   case WebAssembly::BI__builtin_wasm_abs_f32x4:
   case WebAssembly::BI__builtin_wasm_abs_f64x2: {
     Value *Vec = EmitScalarExpr(E->getArg(0));

clang/test/CodeGen/builtins-wasm.c

@@ -511,6 +511,24 @@ int all_true_i64x2(i64x2 x) {
   // WEBASSEMBLY: ret
 }
 
+int bitmask_i8x16(i8x16 x) {
+  return __builtin_wasm_bitmask_i8x16(x);
+  // WEBASSEMBLY: call i32 @llvm.wasm.bitmask.v16i8(<16 x i8> %x)
+  // WEBASSEMBLY: ret
+}
+
+int bitmask_i16x8(i16x8 x) {
+  return __builtin_wasm_bitmask_i16x8(x);
+  // WEBASSEMBLY: call i32 @llvm.wasm.bitmask.v8i16(<8 x i16> %x)
+  // WEBASSEMBLY: ret
+}
+
+int bitmask_i32x4(i32x4 x) {
+  return __builtin_wasm_bitmask_i32x4(x);
+  // WEBASSEMBLY: call i32 @llvm.wasm.bitmask.v4i32(<4 x i32> %x)
+  // WEBASSEMBLY: ret
+}
+
 f32x4 abs_f32x4(f32x4 x) {
   return __builtin_wasm_abs_f32x4(x);
   // WEBASSEMBLY: call <4 x float> @llvm.fabs.v4f32(<4 x float> %x)

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

@@ -129,6 +129,10 @@ def int_wasm_alltrue :
   Intrinsic<[llvm_i32_ty],
             [llvm_anyvector_ty],
             [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_bitmask :
+  Intrinsic<[llvm_i32_ty],
+            [llvm_anyvector_ty],
+            [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_qfma :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

@@ -606,6 +606,18 @@ def : Pat<(i32 (seteq
           (i32 (!cast<NI>(reduction[1]#"_"#ty) (ty V128:$x)))>;
 }
 
+multiclass SIMDBitmask<ValueType vec_t, string vec, bits<32> simdop> {
+  defm _#vec_t : SIMD_I<(outs I32:$dst), (ins V128:$vec), (outs), (ins),
+                         [(set I32:$dst,
+                           (i32 (int_wasm_bitmask (vec_t V128:$vec)))
+                         )],
+                         vec#".bitmask\t$dst, $vec", vec#".bitmask", simdop>;
+}
+
+defm BITMASK : SIMDBitmask<v16i8, "i8x16", 228>;
+defm BITMASK : SIMDBitmask<v8i16, "i16x8", 229>;
+defm BITMASK : SIMDBitmask<v4i32, "i32x4", 230>;
+
 //===----------------------------------------------------------------------===//
 // Bit shifts
 //===----------------------------------------------------------------------===//

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

@@ -95,6 +95,16 @@ define i32 @all_v16i8(<16 x i8> %x) {
   ret i32 %a
 }
 
+; CHECK-LABEL: bitmask_v16i8:
+; SIMD128-NEXT: .functype bitmask_v16i8 (v128) -> (i32){{$}}
+; SIMD128-NEXT: i8x16.bitmask $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare i32 @llvm.wasm.bitmask.v16i8(<16 x i8>)
+define i32 @bitmask_v16i8(<16 x i8> %x) {
+  %a = call i32 @llvm.wasm.bitmask.v16i8(<16 x i8> %x)
+  ret i32 %a
+}
+
 ; CHECK-LABEL: bitselect_v16i8:
 ; SIMD128-NEXT: .functype bitselect_v16i8 (v128, v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: v128.bitselect $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
@@ -208,6 +218,16 @@ define i32 @all_v8i16(<8 x i16> %x) {
   ret i32 %a
 }
 
+; CHECK-LABEL: bitmask_v8i16:
+; SIMD128-NEXT: .functype bitmask_v8i16 (v128) -> (i32){{$}}
+; SIMD128-NEXT: i16x8.bitmask $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare i32 @llvm.wasm.bitmask.v8i16(<8 x i16>)
+define i32 @bitmask_v8i16(<8 x i16> %x) {
+  %a = call i32 @llvm.wasm.bitmask.v8i16(<8 x i16> %x)
+  ret i32 %a
+}
+
 ; CHECK-LABEL: bitselect_v8i16:
 ; SIMD128-NEXT: .functype bitselect_v8i16 (v128, v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: v128.bitselect $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
@@ -317,6 +337,16 @@ define i32 @all_v4i32(<4 x i32> %x) {
   ret i32 %a
 }
 
+; CHECK-LABEL: bitmask_v4i32:
+; SIMD128-NEXT: .functype bitmask_v4i32 (v128) -> (i32){{$}}
+; SIMD128-NEXT: i32x4.bitmask $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare i32 @llvm.wasm.bitmask.v4i32(<4 x i32>)
+define i32 @bitmask_v4i32(<4 x i32> %x) {
+  %a = call i32 @llvm.wasm.bitmask.v4i32(<4 x i32> %x)
+  ret i32 %a
+}
+
 ; CHECK-LABEL: bitselect_v4i32:
 ; SIMD128-NEXT: .functype bitselect_v4i32 (v128, v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: v128.bitselect $push[[R:[0-9]+]]=, $0, $1, $2{{$}}

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

@@ -580,4 +580,13 @@ main:
     # CHECK: i32x4.dot_i16x8_s # encoding: [0xfd,0xdb,0x01]
     i32x4.dot_i16x8_s
 
+    # CHECK: i8x16.bitmask # encoding: [0xfd,0xe4,0x01]
+    i8x16.bitmask
+
+    # CHECK: i16x8.bitmask # encoding: [0xfd,0xe5,0x01]
+    i16x8.bitmask
+
+    # CHECK: i32x4.bitmask # encoding: [0xfd,0xe6,0x01]
+    i32x4.bitmask
+
     end_function