[WebAssembly] __builtin_wasm_replace_lane_* builtins

[tlively]

Summary

Depends on D52852.

[craig.topper]

include/clang/Basic/BuiltinsWebAssembly.def

55

I think maybe you want LL instead of W, but I'm not sure.

[aheejin]

include/clang/Basic/BuiltinsWebAssembly.def

55

Not sure what the differences are either. I've used LL for other builtins in this file though.

[tlively]

• Force constant folding for lane

[craig.topper]

include/clang/Basic/BuiltinsWebAssembly.def

55

W changes based on what's returned from getInt64Type() for the target. I think it will be either L or LL. It looks to be almost exclusively used by MSVC compatibility intrinsics except for two builtins in NVPTX.

[tlively]

include/clang/Basic/BuiltinsWebAssembly.def

55

Ok, I'll switch to using LL.

[tlively]

• Replace Wi with LLi and remove V:128:

[dschuff]

W is int64_t (which is long long for wasm32 and would probably be long for wasm64 since that would probably LP64 (like Linux) rather than LLP64 (Like Win64). So if we want it to be the same for both wasm32 and wasm64, I guess we want LL.

[tlively]

In D52856#1255905, @dschuff wrote:

W is int64_t (which is long long for wasm32 and would probably be long for wasm64 since that would probably LP64 (like Linux) rather than LLP64 (Like Win64). So if we want it to be the same for both wasm32 and wasm64, I guess we want LL.

Ideally I would be able to exactly specify int8_t, int32_t, etc, but c/s/i/LLi works too.

include/clang/Basic/BuiltinsWebAssembly.def

 BUILTIN(__builtin_wasm_extract_lane_u_i8x16, "iV16cIi", "ncV:128:")
 BUILTIN(__builtin_wasm_extract_lane_s_i16x8, "iV8sIi", "ncV:128:")
 BUILTIN(__builtin_wasm_extract_lane_u_i16x8, "iV8sIi", "ncV:128:")
 BUILTIN(__builtin_wasm_extract_lane_i32x4, "iV4iIi", "ncV:128:")
 BUILTIN(__builtin_wasm_extract_lane_i64x2, "WiV2WiIi", "ncV:128:")
 BUILTIN(__builtin_wasm_extract_lane_f32x4, "fV4fIi", "ncV:128:")
 BUILTIN(__builtin_wasm_extract_lane_f64x2, "dV2dIi", "ncV:128:")
 
+BUILTIN(__builtin_wasm_replace_lane_i8x16, "V16cV16cIii", "ncV:128:")
+BUILTIN(__builtin_wasm_replace_lane_i16x8, "V8sV8sIii", "ncV:128:")
+BUILTIN(__builtin_wasm_replace_lane_i32x4, "V4iV4iIii", "ncV:128:")
+BUILTIN(__builtin_wasm_replace_lane_i64x2, "V2WiV2WiIiWi", "ncV:128:")
+BUILTIN(__builtin_wasm_replace_lane_f32x4, "V4fV4fIif", "ncV:128:")
+BUILTIN(__builtin_wasm_replace_lane_f64x2, "V2dV2dIid", "ncV:128:")
+
 #undef BUILTIN

lib/CodeGen/CGBuiltin.cpp

Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
   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: {
     Value *Vec = EmitScalarExpr(E->getArg(0));
     Value *Lane = EmitScalarExpr(E->getArg(1));
     return Builder.CreateExtractElement(Vec, Lane);
   }
+  case WebAssembly::BI__builtin_wasm_replace_lane_i8x16:
+  case WebAssembly::BI__builtin_wasm_replace_lane_i16x8: {
+    Value *Vec = EmitScalarExpr(E->getArg(0));
+    Value *Lane = EmitScalarExpr(E->getArg(1));
+    Value *Val = EmitScalarExpr(E->getArg(2));
+    llvm::Type *ElemType = ConvertType(E->getType())->getVectorElementType();
+    Value *Trunc = Builder.CreateTrunc(Val, ElemType);
+    return Builder.CreateInsertElement(Vec, Trunc, Lane);
+  }
+  case WebAssembly::BI__builtin_wasm_replace_lane_i32x4:
+  case WebAssembly::BI__builtin_wasm_replace_lane_i64x2:
+  case WebAssembly::BI__builtin_wasm_replace_lane_f32x4:
+  case WebAssembly::BI__builtin_wasm_replace_lane_f64x2: {
+    Value *Vec = EmitScalarExpr(E->getArg(0));
+    Value *Lane = EmitScalarExpr(E->getArg(1));
+    Value *Val = EmitScalarExpr(E->getArg(2));
+    return Builder.CreateInsertElement(Vec, Val, Lane);
+  }
 
   default:
     return nullptr;
   }
 }
 
 Value *CodeGenFunction::EmitHexagonBuiltinExpr(unsigned BuiltinID,
                                                const CallExpr *E) {

test/CodeGen/builtins-wasm.c

float f17(f32x4 v) {
   // WEBASSEMBLY-NEXT: ret
 }
 
 double f18(f64x2 v) {
   return __builtin_wasm_extract_lane_f64x2(v, 1);
   // WEBASSEMBLY: extractelement <2 x double> %v, i32 1
   // WEBASSEMBLY-NEXT: ret
 }
+
+i8x16 f19(i8x16 v, int x) {
+  return __builtin_wasm_replace_lane_i8x16(v, 13, x);
+  // WEBASSEMBLY: trunc i32 %x to i8
+  // WEBASSEMBLY-NEXT: insertelement <16 x i8> %v, i8 %{{.*}}, i32 13
+  // WEBASSEMBLY-NEXT: ret
+}
+
+i16x8 f20(i16x8 v, int x) {
+  return __builtin_wasm_replace_lane_i16x8(v, 7, x);
+  // WEBASSEMBLY: trunc i32 %x to i16
+  // WEBASSEMBLY-NEXT: insertelement <8 x i16> %v, i16 %{{.*}}, i32 7
+  // WEBASSEMBLY-NEXT: ret
+}
+
+i32x4 f21(i32x4 v, int x) {
+  return __builtin_wasm_replace_lane_i32x4(v, 3, x);
+  // WEBASSEMBLY: insertelement <4 x i32> %v, i32 %x, i32 3
+  // WEBASSEMBLY-NEXT: ret
+}
+
+i64x2 f22(i64x2 v, long long x) {
+  return __builtin_wasm_replace_lane_i64x2(v, 1, x);
+  // WEBASSEMBLY: insertelement <2 x i64> %v, i64 %x, i32 1
+  // WEBASSEMBLY-NEXT: ret
+}
+
+f32x4 f23(f32x4 v, float x) {
+  return __builtin_wasm_replace_lane_f32x4(v, 3, x);
+  // WEBASSEMBLY: insertelement <4 x float> %v, float %x, i32 3
+  // WEBASSEMBLY-NEXT: ret
+}
+
+f64x2 f24(f64x2 v, double x) {
+  return __builtin_wasm_replace_lane_f64x2(v, 1, x);
+  // WEBASSEMBLY: insertelement <2 x double> %v, double %x, i32 1
+  // WEBASSEMBLY-NEXT: ret
+}