[WebAssembly] Bitselect and min/max builtins

[aheejin]

include/clang/Basic/BuiltinsWebAssembly.def

85

Do you think we need to provide this builtin also for other vector types?

[tlively]

include/clang/Basic/BuiltinsWebAssembly.def

85

This builtin is interesting because there is basically no reason for the mask to be the same type as the vectors. So we could have 36 versions of this, one for each pair of vector types. Or if we required the mask to be a vector of integers we could cut that down to just 24 versions. But no matter what types you use, this builtin does the exact same thing to the bits. So in the end it seems simpler to have just one version and have users convert between vector types. Without the -fno-lax-vector-conversions flag, clang will silently convert between vector types for you anyway, so there's no real ergonomic loss either (not that ergonomics really matter for builtins).

include/clang/Basic/BuiltinsWebAssembly.def

 BUILTIN(__builtin_wasm_trunc_saturate_u_i32_f32, "if", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_s_i32_f64, "id", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_u_i32_f64, "id", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_s_i64_f32, "LLif", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_u_i64_f32, "LLif", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_s_i64_f64, "LLid", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_u_i64_f64, "LLid", "nc")
 
+// Floating point min/max
+BUILTIN(__builtin_wasm_min_f32, "fff", "nc")
+BUILTIN(__builtin_wasm_max_f32, "fff", "nc")
+BUILTIN(__builtin_wasm_min_f64, "ddd", "nc")
+BUILTIN(__builtin_wasm_max_f64, "ddd", "nc")
+
 // SIMD builtins
 BUILTIN(__builtin_wasm_extract_lane_s_i8x16, "iV16cIi", "nc")
 BUILTIN(__builtin_wasm_extract_lane_u_i8x16, "iV16cIi", "nc")
 BUILTIN(__builtin_wasm_extract_lane_s_i16x8, "iV8sIi", "nc")
 BUILTIN(__builtin_wasm_extract_lane_u_i16x8, "iV8sIi", "nc")
 BUILTIN(__builtin_wasm_extract_lane_i32x4, "iV4iIi", "nc")
 BUILTIN(__builtin_wasm_extract_lane_i64x2, "LLiV2LLiIi", "nc")
 BUILTIN(__builtin_wasm_extract_lane_f32x4, "fV4fIi", "nc")

 BUILTIN(__builtin_wasm_add_saturate_s_i16x8, "V8sV8sV8s", "nc")
 BUILTIN(__builtin_wasm_add_saturate_u_i16x8, "V8sV8sV8s", "nc")
 
 BUILTIN(__builtin_wasm_sub_saturate_s_i8x16, "V16cV16cV16c", "nc")
 BUILTIN(__builtin_wasm_sub_saturate_u_i8x16, "V16cV16cV16c", "nc")
 BUILTIN(__builtin_wasm_sub_saturate_s_i16x8, "V8sV8sV8s", "nc")
 BUILTIN(__builtin_wasm_sub_saturate_u_i16x8, "V8sV8sV8s", "nc")
 
+BUILTIN(__builtin_wasm_bitselect, "V4iV4iV4iV4i", "nc")
+
 BUILTIN(__builtin_wasm_any_true_i8x16, "iV16c", "nc")
 BUILTIN(__builtin_wasm_any_true_i16x8, "iV8s", "nc")
 BUILTIN(__builtin_wasm_any_true_i32x4, "iV4i", "nc")
 BUILTIN(__builtin_wasm_any_true_i64x2, "iV2LLi", "nc")
 BUILTIN(__builtin_wasm_all_true_i8x16, "iV16c", "nc")
 BUILTIN(__builtin_wasm_all_true_i16x8, "iV8s", "nc")
 BUILTIN(__builtin_wasm_all_true_i32x4, "iV4i", "nc")
 BUILTIN(__builtin_wasm_all_true_i64x2, "iV2LLi", "nc")
 
 BUILTIN(__builtin_wasm_abs_f32x4, "V4fV4f", "nc")
 BUILTIN(__builtin_wasm_abs_f64x2, "V2dV2d", "nc")
 
+BUILTIN(__builtin_wasm_min_f32x4, "V4fV4fV4f", "nc")
+BUILTIN(__builtin_wasm_max_f32x4, "V4fV4fV4f", "nc")
+BUILTIN(__builtin_wasm_min_f64x2, "V2dV2dV2d", "nc")
+BUILTIN(__builtin_wasm_max_f64x2, "V2dV2dV2d", "nc")
+
 BUILTIN(__builtin_wasm_sqrt_f32x4, "V4fV4f", "nc")
 BUILTIN(__builtin_wasm_sqrt_f64x2, "V2dV2d", "nc")
 
 BUILTIN(__builtin_wasm_trunc_saturate_s_v4i32_v4f32, "V4iV4f", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_u_v4i32_v4f32, "V4iV4f", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_s_v2i64_v2f64, "V2LLiV2d", "nc")
 BUILTIN(__builtin_wasm_trunc_saturate_u_v2i64_v2f64, "V2LLiV2d", "nc")
 
 #undef BUILTIN

lib/CodeGen/CGBuiltin.cpp

Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_v4i32_v4f32:
   case WebAssembly::BI__builtin_wasm_trunc_saturate_u_v2i64_v2f64: {
     Value *Src = EmitScalarExpr(E->getArg(0));
     llvm::Type *ResT = ConvertType(E->getType());
     Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_saturate_unsigned,
                                      {ResT, Src->getType()});
     return Builder.CreateCall(Callee, {Src});
   }
+  case WebAssembly::BI__builtin_wasm_min_f32:
+  case WebAssembly::BI__builtin_wasm_min_f64:
+  case WebAssembly::BI__builtin_wasm_min_f32x4:
+  case WebAssembly::BI__builtin_wasm_min_f64x2: {
+    Value *LHS = EmitScalarExpr(E->getArg(0));
+    Value *RHS = EmitScalarExpr(E->getArg(1));
+    Value *Callee = CGM.getIntrinsic(Intrinsic::minimum,
+                                     ConvertType(E->getType()));
+    return Builder.CreateCall(Callee, {LHS, RHS});
+  }
+  case WebAssembly::BI__builtin_wasm_max_f32:
+  case WebAssembly::BI__builtin_wasm_max_f64:
+  case WebAssembly::BI__builtin_wasm_max_f32x4:
+  case WebAssembly::BI__builtin_wasm_max_f64x2: {
+    Value *LHS = EmitScalarExpr(E->getArg(0));
+    Value *RHS = EmitScalarExpr(E->getArg(1));
+    Value *Callee = CGM.getIntrinsic(Intrinsic::maximum,
+                                     ConvertType(E->getType()));
+    return Builder.CreateCall(Callee, {LHS, RHS});
+  }
   case WebAssembly::BI__builtin_wasm_extract_lane_s_i8x16:
   case WebAssembly::BI__builtin_wasm_extract_lane_u_i8x16:
   case WebAssembly::BI__builtin_wasm_extract_lane_s_i16x8:
   case WebAssembly::BI__builtin_wasm_extract_lane_u_i16x8:
   case WebAssembly::BI__builtin_wasm_extract_lane_i32x4:
   case WebAssembly::BI__builtin_wasm_extract_lane_i64x2:
   case WebAssembly::BI__builtin_wasm_extract_lane_f32x4:
   case WebAssembly::BI__builtin_wasm_extract_lane_f64x2: {
case WebAssembly::BI__builtin_wasm_sub_saturate_u_i16x8: {
     default:
       llvm_unreachable("unexpected builtin ID");
     }
     Value *LHS = EmitScalarExpr(E->getArg(0));
     Value *RHS = EmitScalarExpr(E->getArg(1));
     Value *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
     return Builder.CreateCall(Callee, {LHS, RHS});
   }
+  case WebAssembly::BI__builtin_wasm_bitselect: {
+    Value *V1 = EmitScalarExpr(E->getArg(0));
+    Value *V2 = EmitScalarExpr(E->getArg(1));
+    Value *C = EmitScalarExpr(E->getArg(2));
+    Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_bitselect,
+                                     ConvertType(E->getType()));
+    return Builder.CreateCall(Callee, {V1, V2, C});
+  }
   case WebAssembly::BI__builtin_wasm_any_true_i8x16:
   case WebAssembly::BI__builtin_wasm_any_true_i16x8:
   case WebAssembly::BI__builtin_wasm_any_true_i32x4:
   case WebAssembly::BI__builtin_wasm_any_true_i64x2:
   case WebAssembly::BI__builtin_wasm_all_true_i8x16:
   case WebAssembly::BI__builtin_wasm_all_true_i16x8:
   case WebAssembly::BI__builtin_wasm_all_true_i32x4:
   case WebAssembly::BI__builtin_wasm_all_true_i64x2: {

test/CodeGen/builtins-wasm.c

 }
 
 long long trunc_saturate_u_i64_f64(double f) {
   return __builtin_wasm_trunc_saturate_u_i64_f64(f);
   // WEBASSEMBLY: call i64 @llvm.wasm.trunc.saturate.unsigned.i64.f64(double %f)
   // WEBASSEMBLY-NEXT: ret
 }
 
+float min_f32(float x, float y) {
+  return __builtin_wasm_min_f32(x, y);
+  // WEBASSEMBLY: call float @llvm.minimum.f32(float %x, float %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+float max_f32(float x, float y) {
+  return __builtin_wasm_max_f32(x, y);
+  // WEBASSEMBLY: call float @llvm.maximum.f32(float %x, float %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+double min_f64(double x, double y) {
+  return __builtin_wasm_min_f64(x, y);
+  // WEBASSEMBLY: call double @llvm.minimum.f64(double %x, double %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+double max_f64(double x, double y) {
+  return __builtin_wasm_max_f64(x, y);
+  // WEBASSEMBLY: call double @llvm.maximum.f64(double %x, double %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+
 int extract_lane_s_i8x16(i8x16 v) {
   return __builtin_wasm_extract_lane_s_i8x16(v, 13);
   // WEBASSEMBLY: extractelement <16 x i8> %v, i32 13
   // WEBASSEMBLY-NEXT: sext
   // WEBASSEMBLY-NEXT: ret
 }
 
 int extract_lane_u_i8x16(i8x16 v) {

 
 i16x8 sub_saturate_u_i16x8(i16x8 x, i16x8 y) {
   return __builtin_wasm_sub_saturate_u_i16x8(x, y);
   // WEBASSEMBLY: call <8 x i16> @llvm.wasm.sub.saturate.unsigned.v8i16(
   // WEBASSEMBLY-SAME: <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(
+  // WEBASSEMBLY-SAME: <4 x i32> %x, <4 x i32> %y, <4 x i32> %c)
+  // WEBASSEMBLY-NEXT: ret
+}
+
 int any_true_i8x16(i8x16 x) {
   return __builtin_wasm_any_true_i8x16(x);
   // WEBASSEMBLY: call i32 @llvm.wasm.anytrue.v16i8(<16 x i8> %x)
   // WEBASSEMBLY: ret
 }
 
 int any_true_i16x8(i16x8 x) {
   return __builtin_wasm_any_true_i16x8(x);

 }
 
 f64x2 abs_f64x2(f64x2 x) {
   return __builtin_wasm_abs_f64x2(x);
   // WEBASSEMBLY: call <2 x double> @llvm.fabs.v2f64(<2 x double> %x)
   // WEBASSEMBLY: ret
 }
 
+f32x4 min_f32x4(f32x4 x, f32x4 y) {
+  return __builtin_wasm_min_f32x4(x, y);
+  // WEBASSEMBLY: call <4 x float> @llvm.minimum.v4f32(
+  // WEBASSEMBLY-SAME: <4 x float> %x, <4 x float> %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+f32x4 max_f32x4(f32x4 x, f32x4 y) {
+  return __builtin_wasm_max_f32x4(x, y);
+  // WEBASSEMBLY: call <4 x float> @llvm.maximum.v4f32(
+  // WEBASSEMBLY-SAME: <4 x float> %x, <4 x float> %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+f64x2 min_f64x2(f64x2 x, f64x2 y) {
+  return __builtin_wasm_min_f64x2(x, y);
+  // WEBASSEMBLY: call <2 x double> @llvm.minimum.v2f64(
+  // WEBASSEMBLY-SAME: <2 x double> %x, <2 x double> %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+f64x2 max_f64x2(f64x2 x, f64x2 y) {
+  return __builtin_wasm_max_f64x2(x, y);
+  // WEBASSEMBLY: call <2 x double> @llvm.maximum.v2f64(
+  // WEBASSEMBLY-SAME: <2 x double> %x, <2 x double> %y)
+  // WEBASSEMBLY-NEXT: ret
+}
+
 f32x4 sqrt_f32x4(f32x4 x) {
   return __builtin_wasm_sqrt_f32x4(x);
   // WEBASSEMBLY: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %x)
   // WEBASSEMBLY: ret
 }
 
 f64x2 sqrt_f64x2(f64x2 x) {
   return __builtin_wasm_sqrt_f64x2(x);