[WebAssembly] Implement prototype SIMD rounding instructions

[tlively]

Summary

As specified in https://github.com/WebAssembly/simd/pull/232. These
instructions are implemented as LLVM intrinsics for now rather than
normal ISel patterns to make these instructions opt-in. Once the
instructions are merged to the spec proposal, the intrinsics will be
replaced with proper ISel patterns.

clang/include/clang/Basic/BuiltinsWebAssembly.def

 TARGET_BUILTIN(__builtin_wasm_max_f32x4, "V4fV4fV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_pmin_f32x4, "V4fV4fV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_pmax_f32x4, "V4fV4fV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_min_f64x2, "V2dV2dV2d", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_max_f64x2, "V2dV2dV2d", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_pmin_f64x2, "V2dV2dV2d", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_pmax_f64x2, "V2dV2dV2d", "nc", "simd128")
 
+TARGET_BUILTIN(__builtin_wasm_ceil_f32x4, "V4fV4f", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_floor_f32x4, "V4fV4f", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_trunc_f32x4, "V4fV4f", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_nearest_f32x4, "V4fV4f", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_ceil_f64x2, "V2dV2d", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_floor_f64x2, "V2dV2d", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_trunc_f64x2, "V2dV2d", "nc", "simd128")
+TARGET_BUILTIN(__builtin_wasm_nearest_f64x2, "V2dV2d", "nc", "simd128")
+
 TARGET_BUILTIN(__builtin_wasm_dot_s_i32x4_i16x8, "V4iV8sV8s", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_sqrt_f32x4, "V4fV4f", "nc", "simd128")
 TARGET_BUILTIN(__builtin_wasm_sqrt_f64x2, "V2dV2d", "nc", "simd128")
 
 TARGET_BUILTIN(__builtin_wasm_qfma_f32x4, "V4fV4fV4fV4f", "nc", "unimplemented-simd128")
 TARGET_BUILTIN(__builtin_wasm_qfms_f32x4, "V4fV4fV4fV4f", "nc", "unimplemented-simd128")
 TARGET_BUILTIN(__builtin_wasm_qfma_f64x2, "V2dV2dV2dV2d", "nc", "unimplemented-simd128")

clang/lib/CodeGen/CGBuiltin.cpp

Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
   case WebAssembly::BI__builtin_wasm_pmax_f32x4:
   case WebAssembly::BI__builtin_wasm_pmax_f64x2: {
     Value *LHS = EmitScalarExpr(E->getArg(0));
     Value *RHS = EmitScalarExpr(E->getArg(1));
     Function *Callee =
         CGM.getIntrinsic(Intrinsic::wasm_pmax, ConvertType(E->getType()));
     return Builder.CreateCall(Callee, {LHS, RHS});
   }
+  case WebAssembly::BI__builtin_wasm_ceil_f32x4:
+  case WebAssembly::BI__builtin_wasm_floor_f32x4:
+  case WebAssembly::BI__builtin_wasm_trunc_f32x4:
+  case WebAssembly::BI__builtin_wasm_nearest_f32x4:
+  case WebAssembly::BI__builtin_wasm_ceil_f64x2:
+  case WebAssembly::BI__builtin_wasm_floor_f64x2:
+  case WebAssembly::BI__builtin_wasm_trunc_f64x2:
+  case WebAssembly::BI__builtin_wasm_nearest_f64x2: {
+    unsigned IntNo;
+    switch (BuiltinID) {
+    case WebAssembly::BI__builtin_wasm_ceil_f32x4:
+    case WebAssembly::BI__builtin_wasm_ceil_f64x2:
+      IntNo = Intrinsic::wasm_ceil;
+      break;
+    case WebAssembly::BI__builtin_wasm_floor_f32x4:
+    case WebAssembly::BI__builtin_wasm_floor_f64x2:
+      IntNo = Intrinsic::wasm_floor;
+      break;
+    case WebAssembly::BI__builtin_wasm_trunc_f32x4:
+    case WebAssembly::BI__builtin_wasm_trunc_f64x2:
+      IntNo = Intrinsic::wasm_trunc;
+      break;
+    case WebAssembly::BI__builtin_wasm_nearest_f32x4:
+    case WebAssembly::BI__builtin_wasm_nearest_f64x2:
+      IntNo = Intrinsic::wasm_nearest;
+      break;
+    default:
+      llvm_unreachable("unexpected builtin ID");
+    }
+    Value *Value = EmitScalarExpr(E->getArg(0));
+    Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
+    return Builder.CreateCall(Callee, Value);
+  }
   case WebAssembly::BI__builtin_wasm_swizzle_v8x16: {
     Value *Src = EmitScalarExpr(E->getArg(0));
     Value *Indices = EmitScalarExpr(E->getArg(1));
     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_swizzle);
     return Builder.CreateCall(Callee, {Src, Indices});
   }
   case WebAssembly::BI__builtin_wasm_extract_lane_s_i8x16:
   case WebAssembly::BI__builtin_wasm_extract_lane_u_i8x16:

clang/test/CodeGen/builtins-wasm.c

 
 f64x2 pmax_f64x2(f64x2 x, f64x2 y) {
   return __builtin_wasm_pmax_f64x2(x, y);
   // WEBASSEMBLY: call <2 x double> @llvm.wasm.pmax.v2f64(
   // WEBASSEMBLY-SAME: <2 x double> %x, <2 x double> %y)
   // WEBASSEMBLY-NEXT: ret
 }
 
+f32x4 ceil_f32x4(f32x4 x) {
+  return __builtin_wasm_ceil_f32x4(x);
+  // WEBASSEMBLY: call <4 x float> @llvm.wasm.ceil.v4f32(<4 x float> %x)
+  // WEBASSEMBLY: ret
+}
+
+f32x4 floor_f32x4(f32x4 x) {
+  return __builtin_wasm_floor_f32x4(x);
+  // WEBASSEMBLY: call <4 x float> @llvm.wasm.floor.v4f32(<4 x float> %x)
+  // WEBASSEMBLY: ret
+}
+
+f32x4 trunc_f32x4(f32x4 x) {
+  return __builtin_wasm_trunc_f32x4(x);
+  // WEBASSEMBLY: call <4 x float> @llvm.wasm.trunc.v4f32(<4 x float> %x)
+  // WEBASSEMBLY: ret
+}
+
+f32x4 nearest_f32x4(f32x4 x) {
+  return __builtin_wasm_nearest_f32x4(x);
+  // WEBASSEMBLY: call <4 x float> @llvm.wasm.nearest.v4f32(<4 x float> %x)
+  // WEBASSEMBLY: ret
+}
+
+f64x2 ceil_f64x2(f64x2 x) {
+  return __builtin_wasm_ceil_f64x2(x);
+  // WEBASSEMBLY: call <2 x double> @llvm.wasm.ceil.v2f64(<2 x double> %x)
+  // WEBASSEMBLY: ret
+}
+
+f64x2 floor_f64x2(f64x2 x) {
+  return __builtin_wasm_floor_f64x2(x);
+  // WEBASSEMBLY: call <2 x double> @llvm.wasm.floor.v2f64(<2 x double> %x)
+  // WEBASSEMBLY: ret
+}
+
+f64x2 trunc_f64x2(f64x2 x) {
+  return __builtin_wasm_trunc_f64x2(x);
+  // WEBASSEMBLY: call <2 x double> @llvm.wasm.trunc.v2f64(<2 x double> %x)
+  // WEBASSEMBLY: ret
+}
+
+f64x2 nearest_f64x2(f64x2 x) {
+  return __builtin_wasm_nearest_f64x2(x);
+  // WEBASSEMBLY: call <2 x double> @llvm.wasm.nearest.v2f64(<2 x double> %x)
+  // WEBASSEMBLY: 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);

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

def int_wasm_widen_low_unsigned :
   Intrinsic<[llvm_anyvector_ty],
             [llvm_anyvector_ty],
             [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_widen_high_unsigned :
   Intrinsic<[llvm_anyvector_ty],
             [llvm_anyvector_ty],
             [IntrNoMem, IntrSpeculatable]>;
 
-// TODO: Replace these intrinsics with normal ISel patterns once the
-// pmin/pmax instructions are merged to the spec proposal.
+// TODO: Replace these intrinsics with normal ISel patterns
 def int_wasm_pmin :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_pmax :
   Intrinsic<[llvm_anyvector_ty],
             [LLVMMatchType<0>, LLVMMatchType<0>],
             [IntrNoMem, IntrSpeculatable]>;
 
+// TODO: Replace these instrinsics with normal ISel patterns once the
+// rounding instructions are merged to the proposal
+// (https://github.com/WebAssembly/simd/pull/232).
+def int_wasm_ceil :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMMatchType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_floor :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMMatchType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_trunc :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMMatchType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+def int_wasm_nearest :
+  Intrinsic<[llvm_anyvector_ty],
+            [LLVMMatchType<0>],
+            [IntrNoMem, IntrSpeculatable]>;
+
 //===----------------------------------------------------------------------===//
 // Bulk memory intrinsics
 //===----------------------------------------------------------------------===//
 
 def int_wasm_memory_init :
   Intrinsic<[],
             [llvm_i32_ty, llvm_i32_ty, llvm_ptr_ty, llvm_i32_ty, llvm_i32_ty],
             [IntrWriteMem, IntrInaccessibleMemOrArgMemOnly, WriteOnly<ArgIndex<2>>,

llvm/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td

 defm ABS : SIMDUnaryFP<fabs, "abs", 224>;
 
 // Negation: neg
 defm NEG : SIMDUnaryFP<fneg, "neg", 225>;
 
 // Square root: sqrt
 defm SQRT : SIMDUnaryFP<fsqrt, "sqrt", 227>;
 
+// Rounding: ceil, floor, trunc, nearest
+defm CEIL : SIMDUnary<v4f32, "f32x4", int_wasm_ceil, "ceil", 216>;
+defm FLOOR : SIMDUnary<v4f32, "f32x4", int_wasm_floor, "floor", 217>;
+defm TRUNC: SIMDUnary<v4f32, "f32x4", int_wasm_trunc, "trunc", 218>;
+defm NEAREST: SIMDUnary<v4f32, "f32x4", int_wasm_nearest, "nearest", 219>;
+defm CEIL : SIMDUnary<v2f64, "f64x2", int_wasm_ceil, "ceil", 220>;
+defm FLOOR : SIMDUnary<v2f64, "f64x2", int_wasm_floor, "floor", 221>;
+defm TRUNC: SIMDUnary<v2f64, "f64x2", int_wasm_trunc, "trunc", 222>;
+defm NEAREST: SIMDUnary<v2f64, "f64x2", int_wasm_nearest, "nearest", 223>;
+
 //===----------------------------------------------------------------------===//
 // Floating-point binary arithmetic
 //===----------------------------------------------------------------------===//
 
 multiclass SIMDBinaryFP<SDNode node, string name, bits<32> baseInst> {
   defm "" : SIMDBinary<v4f32, "f32x4", node, name, baseInst>;
   defm "" : SIMDBinary<v2f64, "f64x2", node, name, !add(baseInst, 12)>;
 }

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

 ; SIMD128-NEXT: f32x4.pmax $push[[R:[0-9]+]]=, $0, $1{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare <4 x float> @llvm.wasm.pmax.v4f32(<4 x float>, <4 x float>)
 define <4 x float> @pmax_v4f32(<4 x float> %a, <4 x float> %b) {
   %v = call <4 x float> @llvm.wasm.pmax.v4f32(<4 x float> %a, <4 x float> %b)
   ret <4 x float> %v
 }
 
+; CHECK-LABEL: ceil_v4f32:
+; SIMD128-NEXT: .functype ceil_v4f32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.ceil $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.ceil.v4f32(<4 x float>)
+define <4 x float> @ceil_v4f32(<4 x float> %a) {
+  %v = call <4 x float> @llvm.wasm.ceil.v4f32(<4 x float> %a)
+  ret <4 x float> %v
+}
+
+; CHECK-LABEL: floor_v4f32:
+; SIMD128-NEXT: .functype floor_v4f32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.floor $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.floor.v4f32(<4 x float>)
+define <4 x float> @floor_v4f32(<4 x float> %a) {
+  %v = call <4 x float> @llvm.wasm.floor.v4f32(<4 x float> %a)
+  ret <4 x float> %v
+}
+
+; CHECK-LABEL: trunc_v4f32:
+; SIMD128-NEXT: .functype trunc_v4f32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.trunc $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.trunc.v4f32(<4 x float>)
+define <4 x float> @trunc_v4f32(<4 x float> %a) {
+  %v = call <4 x float> @llvm.wasm.trunc.v4f32(<4 x float> %a)
+  ret <4 x float> %v
+}
+
+; CHECK-LABEL: nearest_v4f32:
+; SIMD128-NEXT: .functype nearest_v4f32 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f32x4.nearest $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <4 x float> @llvm.wasm.nearest.v4f32(<4 x float>)
+define <4 x float> @nearest_v4f32(<4 x float> %a) {
+  %v = call <4 x float> @llvm.wasm.nearest.v4f32(<4 x float> %a)
+  ret <4 x float> %v
+}
+
 ; CHECK-LABEL: qfma_v4f32:
 ; SIMD128-NEXT: .functype qfma_v4f32 (v128, v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: f32x4.qfma $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare <4 x float> @llvm.wasm.qfma.v4f32(<4 x float>, <4 x float>, <4 x float>)
 define <4 x float> @qfma_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c) {
   %v = call <4 x float> @llvm.wasm.qfma.v4f32(
     <4 x float> %a, <4 x float> %b, <4 x float> %c

 ; SIMD128-NEXT: f64x2.pmax $push[[R:[0-9]+]]=, $0, $1{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare <2 x double> @llvm.wasm.pmax.v2f64(<2 x double>, <2 x double>)
 define <2 x double> @pmax_v2f64(<2 x double> %a, <2 x double> %b) {
   %v = call <2 x double> @llvm.wasm.pmax.v2f64(<2 x double> %a, <2 x double> %b)
   ret <2 x double> %v
 }
 
+; CHECK-LABEL: ceil_v2f64:
+; SIMD128-NEXT: .functype ceil_v2f64 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f64x2.ceil $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <2 x double> @llvm.wasm.ceil.v2f64(<2 x double>)
+define <2 x double> @ceil_v2f64(<2 x double> %a) {
+  %v = call <2 x double> @llvm.wasm.ceil.v2f64(<2 x double> %a)
+  ret <2 x double> %v
+}
+
+; CHECK-LABEL: floor_v2f64:
+; SIMD128-NEXT: .functype floor_v2f64 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f64x2.floor $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <2 x double> @llvm.wasm.floor.v2f64(<2 x double>)
+define <2 x double> @floor_v2f64(<2 x double> %a) {
+  %v = call <2 x double> @llvm.wasm.floor.v2f64(<2 x double> %a)
+  ret <2 x double> %v
+}
+
+; CHECK-LABEL: trunc_v2f64:
+; SIMD128-NEXT: .functype trunc_v2f64 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f64x2.trunc $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <2 x double> @llvm.wasm.trunc.v2f64(<2 x double>)
+define <2 x double> @trunc_v2f64(<2 x double> %a) {
+  %v = call <2 x double> @llvm.wasm.trunc.v2f64(<2 x double> %a)
+  ret <2 x double> %v
+}
+
+; CHECK-LABEL: nearest_v2f64:
+; SIMD128-NEXT: .functype nearest_v2f64 (v128) -> (v128){{$}}
+; SIMD128-NEXT: f64x2.nearest $push[[R:[0-9]+]]=, $0{{$}}
+; SIMD128-NEXT: return $pop[[R]]{{$}}
+declare <2 x double> @llvm.wasm.nearest.v2f64(<2 x double>)
+define <2 x double> @nearest_v2f64(<2 x double> %a) {
+  %v = call <2 x double> @llvm.wasm.nearest.v2f64(<2 x double> %a)
+  ret <2 x double> %v
+}
+
 ; CHECK-LABEL: qfma_v2f64:
 ; SIMD128-NEXT: .functype qfma_v2f64 (v128, v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: f64x2.qfma $push[[R:[0-9]+]]=, $0, $1, $2{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 declare <2 x double> @llvm.wasm.qfma.v2f64(<2 x double>, <2 x double>, <2 x double>)
 define <2 x double> @qfma_v2f64(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
   %v = call <2 x double> @llvm.wasm.qfma.v2f64(
     <2 x double> %a, <2 x double> %b, <2 x double> %c

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

main:
     i64x2.add
 
     # CHECK: i64x2.sub # encoding: [0xfd,0xd1,0x01]
     i64x2.sub
 
     # CHECK: i64x2.mul # encoding: [0xfd,0xd5,0x01]
     i64x2.mul
 
+    # CHECK: f32x4.ceil # encoding: [0xfd,0xd8,0x01]
+    f32x4.ceil
+
+    # CHECK: f32x4.floor # encoding: [0xfd,0xd9,0x01]
+    f32x4.floor
+
+    # CHECK: f32x4.trunc # encoding: [0xfd,0xda,0x01]
+    f32x4.trunc
+
+    # CHECK: f32x4.nearest # encoding: [0xfd,0xdb,0x01]
+    f32x4.nearest
+
+    # CHECK: f64x2.ceil # encoding: [0xfd,0xdc,0x01]
+    f64x2.ceil
+
+    # CHECK: f64x2.floor # encoding: [0xfd,0xdd,0x01]
+    f64x2.floor
+
+    # CHECK: f64x2.trunc # encoding: [0xfd,0xde,0x01]
+    f64x2.trunc
+
+    # CHECK: f64x2.nearest # encoding: [0xfd,0xdf,0x01]
+    f64x2.nearest
+
     # CHECK: f32x4.abs # encoding: [0xfd,0xe0,0x01]
     f32x4.abs
 
     # CHECK: f32x4.neg # encoding: [0xfd,0xe1,0x01]
     f32x4.neg
 
     # CHECK: f32x4.sqrt # encoding: [0xfd,0xe3,0x01]
     f32x4.sqrt