[WebAssembly] Trapping fptoint builtins and intrinsics

[tlively]

Summary

The WebAssembly backend lowers fptoint instructions to a code sequence
that checks for overflow to avoid traps because fptoint is supposed to
be speculatable. These new builtins and intrinsics give users a way to
depend on the trapping semantics of the underlying instructions and
avoid the extra code generated normally.

Patch by coffee and tlively.

clang/include/clang/Basic/BuiltinsWebAssembly.def

 TARGET_BUILTIN(__builtin_wasm_throw, "vIUiv*", "r", "exception-handling")
 TARGET_BUILTIN(__builtin_wasm_rethrow_in_catch, "v", "r", "exception-handling")
 
 // Atomic wait and notify.
 BUILTIN(__builtin_wasm_atomic_wait_i32, "ii*iLLi", "n")
 BUILTIN(__builtin_wasm_atomic_wait_i64, "iLLi*LLiLLi", "n")
 BUILTIN(__builtin_wasm_atomic_notify, "Uii*Ui", "n")
 
+// Trapping fp-to-int conversions
+BUILTIN(__builtin_wasm_trunc_s_i32_f32, "if", "nc")
+BUILTIN(__builtin_wasm_trunc_u_i32_f32, "if", "nc")
+BUILTIN(__builtin_wasm_trunc_s_i32_f64, "id", "nc")
+BUILTIN(__builtin_wasm_trunc_u_i32_f64, "id", "nc")
+BUILTIN(__builtin_wasm_trunc_s_i64_f32, "LLif", "nc")
+BUILTIN(__builtin_wasm_trunc_u_i64_f32, "LLif", "nc")
+BUILTIN(__builtin_wasm_trunc_s_i64_f64, "LLid", "nc")
+BUILTIN(__builtin_wasm_trunc_u_i64_f64, "LLid", "nc")
+
 // Saturating fp-to-int conversions
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_s_i32_f32, "if", "nc", "nontrapping-fptoint")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_u_i32_f32, "if", "nc", "nontrapping-fptoint")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_s_i32_f64, "id", "nc", "nontrapping-fptoint")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_u_i32_f64, "id", "nc", "nontrapping-fptoint")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_s_i64_f32, "LLif", "nc", "nontrapping-fptoint")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_u_i64_f32, "LLif", "nc", "nontrapping-fptoint")
 TARGET_BUILTIN(__builtin_wasm_trunc_saturate_s_i64_f64, "LLid", "nc", "nontrapping-fptoint")

clang/lib/CodeGen/CGBuiltin.cpp

case WebAssembly::BI__builtin_wasm_atomic_wait_i64: {
     return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
   }
   case WebAssembly::BI__builtin_wasm_atomic_notify: {
     Value *Addr = EmitScalarExpr(E->getArg(0));
     Value *Count = EmitScalarExpr(E->getArg(1));
     Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_atomic_notify);
     return Builder.CreateCall(Callee, {Addr, Count});
   }
+  case WebAssembly::BI__builtin_wasm_trunc_s_i32_f32:
+  case WebAssembly::BI__builtin_wasm_trunc_s_i32_f64:
+  case WebAssembly::BI__builtin_wasm_trunc_s_i64_f32:
+  case WebAssembly::BI__builtin_wasm_trunc_s_i64_f64: {
+    Value *Src = EmitScalarExpr(E->getArg(0));
+    llvm::Type *ResT = ConvertType(E->getType());
+    Function *Callee =
+        CGM.getIntrinsic(Intrinsic::wasm_trunc_signed, {ResT, Src->getType()});
+    return Builder.CreateCall(Callee, {Src});
+  }
+  case WebAssembly::BI__builtin_wasm_trunc_u_i32_f32:
+  case WebAssembly::BI__builtin_wasm_trunc_u_i32_f64:
+  case WebAssembly::BI__builtin_wasm_trunc_u_i64_f32:
+  case WebAssembly::BI__builtin_wasm_trunc_u_i64_f64: {
+    Value *Src = EmitScalarExpr(E->getArg(0));
+    llvm::Type *ResT = ConvertType(E->getType());
+    Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_unsigned,
+                                        {ResT, Src->getType()});
+    return Builder.CreateCall(Callee, {Src});
+  }
   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f32:
   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f64:
   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f32:
   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f64:
   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32x4_f32x4:
   case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64x2_f64x2: {
     Value *Src = EmitScalarExpr(E->getArg(0));
     llvm::Type *ResT = ConvertType(E->getType());

clang/test/CodeGen/builtins-wasm.c

 }
 
 unsigned int atomic_notify(int *addr, unsigned int count) {
   return __builtin_wasm_atomic_notify(addr, count);
   // WEBASSEMBLY32: call i32 @llvm.wasm.atomic.notify(i32* %{{.*}}, i32 %{{.*}})
   // WEBASSEMBLY64: call i32 @llvm.wasm.atomic.notify(i32* %{{.*}}, i32 %{{.*}})
 }
 
+int trunc_s_i32_f32(float f) {
+  return __builtin_wasm_trunc_s_i32_f32(f);
+  // WEBASSEMBLY: call i32 @llvm.wasm.trunc.signed.i32.f32(float %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+int trunc_u_i32_f32(float f) {
+  return __builtin_wasm_trunc_u_i32_f32(f);
+  // WEBASSEMBLY: call i32 @llvm.wasm.trunc.unsigned.i32.f32(float %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+int trunc_s_i32_f64(double f) {
+  return __builtin_wasm_trunc_s_i32_f64(f);
+  // WEBASSEMBLY: call i32 @llvm.wasm.trunc.signed.i32.f64(double %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+int trunc_u_i32_f64(double f) {
+  return __builtin_wasm_trunc_u_i32_f64(f);
+  // WEBASSEMBLY: call i32 @llvm.wasm.trunc.unsigned.i32.f64(double %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+long long trunc_s_i64_f32(float f) {
+  return __builtin_wasm_trunc_s_i64_f32(f);
+  // WEBASSEMBLY: call i64 @llvm.wasm.trunc.signed.i64.f32(float %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+long long trunc_u_i64_f32(float f) {
+  return __builtin_wasm_trunc_u_i64_f32(f);
+  // WEBASSEMBLY: call i64 @llvm.wasm.trunc.unsigned.i64.f32(float %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+long long trunc_s_i64_f64(double f) {
+  return __builtin_wasm_trunc_s_i64_f64(f);
+  // WEBASSEMBLY: call i64 @llvm.wasm.trunc.signed.i64.f64(double %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
+long long trunc_u_i64_f64(double f) {
+  return __builtin_wasm_trunc_u_i64_f64(f);
+  // WEBASSEMBLY: call i64 @llvm.wasm.trunc.unsigned.i64.f64(double %f)
+  // WEBASSEMBLY-NEXT: ret
+}
+
 int trunc_saturate_s_i32_f32(float f) {
   return __builtin_wasm_trunc_saturate_s_i32_f32(f);
   // WEBASSEMBLY: call i32 @llvm.wasm.trunc.saturate.signed.i32.f32(float %f)
   // WEBASSEMBLY-NEXT: ret
 }
 
 int trunc_saturate_u_i32_f32(float f) {
   return __builtin_wasm_trunc_saturate_u_i32_f32(f);

llvm/include/llvm/IR/IntrinsicsWebAssembly.td

 def int_wasm_memory_size : Intrinsic<[llvm_anyint_ty],
                                      [llvm_i32_ty],
                                      [IntrReadMem]>;
 def int_wasm_memory_grow : Intrinsic<[llvm_anyint_ty],
                                      [llvm_i32_ty, LLVMMatchType<0>],
                                      []>;
 
 //===----------------------------------------------------------------------===//
+// Trapping float-to-int conversions
+//===----------------------------------------------------------------------===//
+
+def int_wasm_trunc_signed : Intrinsic<[llvm_anyint_ty],
+                                      [llvm_anyfloat_ty],
+                                      [IntrNoMem]>;
+def int_wasm_trunc_unsigned : Intrinsic<[llvm_anyint_ty],
+                                        [llvm_anyfloat_ty],
+                                        [IntrNoMem]>;
+
+//===----------------------------------------------------------------------===//
 // Saturating float-to-int conversions
 //===----------------------------------------------------------------------===//
 
 def int_wasm_trunc_saturate_signed : Intrinsic<[llvm_anyint_ty],
                                                [llvm_anyfloat_ty],
                                                [IntrNoMem, IntrSpeculatable]>;
 def int_wasm_trunc_saturate_unsigned : Intrinsic<[llvm_anyint_ty],
                                                  [llvm_anyfloat_ty],

llvm/lib/Target/WebAssembly/WebAssemblyInstrConv.td

 defm I64_TRUNC_S_F64 : I<(outs I64:$dst), (ins F64:$src), (outs), (ins),
                          [], "i64.trunc_f64_s\t$dst, $src", "i64.trunc_f64_s",
                          0xb0>;
 defm I64_TRUNC_U_F64 : I<(outs I64:$dst), (ins F64:$src), (outs), (ins),
                          [], "i64.trunc_f64_u\t$dst, $src", "i64.trunc_f64_u",
                          0xb1>;
 } // hasSideEffects = 1
 
+def : Pat<(int_wasm_trunc_signed F32:$src),
+          (I32_TRUNC_S_F32 F32:$src)>;
+def : Pat<(int_wasm_trunc_unsigned F32:$src),
+          (I32_TRUNC_U_F32 F32:$src)>;
+def : Pat<(int_wasm_trunc_signed F64:$src),
+          (I32_TRUNC_S_F64 F64:$src)>;
+def : Pat<(int_wasm_trunc_unsigned F64:$src),
+          (I32_TRUNC_U_F64 F64:$src)>;
+def : Pat<(int_wasm_trunc_signed F32:$src),
+          (I64_TRUNC_S_F32 F32:$src)>;
+def : Pat<(int_wasm_trunc_unsigned F32:$src),
+          (I64_TRUNC_U_F32 F32:$src)>;
+def : Pat<(int_wasm_trunc_signed F64:$src),
+          (I64_TRUNC_S_F64 F64:$src)>;
+def : Pat<(int_wasm_trunc_unsigned F64:$src),
+          (I64_TRUNC_U_F64 F64:$src)>;
+
 defm F32_CONVERT_S_I32 : I<(outs F32:$dst), (ins I32:$src), (outs), (ins),
                            [(set F32:$dst, (sint_to_fp I32:$src))],
                            "f32.convert_i32_s\t$dst, $src", "f32.convert_i32_s",
                            0xb2>;
 defm F32_CONVERT_U_I32 : I<(outs F32:$dst), (ins I32:$src), (outs), (ins),
                            [(set F32:$dst, (uint_to_fp I32:$src))],
                            "f32.convert_i32_u\t$dst, $src", "f32.convert_i32_u",
                            0xb3>;

llvm/test/CodeGen/WebAssembly/conv-trap.ll

 ; CHECK-NEXT: BB
 ; CHECK-NEXT: end_block
 ; CHECK-NEXT: i64.const $push[[ALT:[0-9]+]]=, 0{{$}}
 ; CHECK-NEXT: return $pop[[ALT]]{{$}}
 define i64 @i64_trunc_u_f64(double %x) {
   %a = fptoui double %x to i64
   ret i64 %a
 }
+
+; CHECK-LABEL: llvm_wasm_trunc_signed_i32_f32:
+; CHECK-NEXT: .functype llvm_wasm_trunc_signed_i32_f32 (f32) -> (i32)
+; CHECK-NEXT: i32.trunc_f32_s $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i32 @llvm.wasm.trunc.signed.i32.f32(float)
+define i32 @llvm_wasm_trunc_signed_i32_f32(float %f) {
+  %a = call i32 @llvm.wasm.trunc.signed.i32.f32(float %f)
+  ret i32 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_unsigned_i32_f32:
+; CHECK-NEXT: .functype llvm_wasm_trunc_unsigned_i32_f32 (f32) -> (i32)
+; CHECK-NEXT: i32.trunc_f32_u $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i32 @llvm.wasm.trunc.unsigned.i32.f32(float)
+define i32 @llvm_wasm_trunc_unsigned_i32_f32(float %f) {
+  %a = call i32 @llvm.wasm.trunc.unsigned.i32.f32(float %f)
+  ret i32 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_signed_i32_f64:
+; CHECK-NEXT: .functype llvm_wasm_trunc_signed_i32_f64 (f64) -> (i32)
+; CHECK-NEXT: i32.trunc_f64_s $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i32 @llvm.wasm.trunc.signed.i32.f64(double)
+define i32 @llvm_wasm_trunc_signed_i32_f64(double %f) {
+  %a = call i32 @llvm.wasm.trunc.signed.i32.f64(double %f)
+  ret i32 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_unsigned_i32_f64:
+; CHECK-NEXT: .functype llvm_wasm_trunc_unsigned_i32_f64 (f64) -> (i32)
+; CHECK-NEXT: i32.trunc_f64_u $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i32 @llvm.wasm.trunc.unsigned.i32.f64(double)
+define i32 @llvm_wasm_trunc_unsigned_i32_f64(double %f) {
+  %a = call i32 @llvm.wasm.trunc.unsigned.i32.f64(double %f)
+  ret i32 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_signed_i64_f32:
+; CHECK-NEXT: .functype llvm_wasm_trunc_signed_i64_f32 (f32) -> (i64)
+; CHECK-NEXT: i64.trunc_f32_s $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i64 @llvm.wasm.trunc.signed.i64.f32(float)
+define i64 @llvm_wasm_trunc_signed_i64_f32(float %f) {
+  %a = call i64 @llvm.wasm.trunc.signed.i64.f32(float %f)
+  ret i64 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_unsigned_i64_f32:
+; CHECK-NEXT: .functype llvm_wasm_trunc_unsigned_i64_f32 (f32) -> (i64)
+; CHECK-NEXT: i64.trunc_f32_u $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i64 @llvm.wasm.trunc.unsigned.i64.f32(float)
+define i64 @llvm_wasm_trunc_unsigned_i64_f32(float %f) {
+  %a = call i64 @llvm.wasm.trunc.unsigned.i64.f32(float %f)
+  ret i64 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_signed_i64_f64:
+; CHECK-NEXT: .functype llvm_wasm_trunc_signed_i64_f64 (f64) -> (i64)
+; CHECK-NEXT: i64.trunc_f64_s $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i64 @llvm.wasm.trunc.signed.i64.f64(double)
+define i64 @llvm_wasm_trunc_signed_i64_f64(double %f) {
+  %a = call i64 @llvm.wasm.trunc.signed.i64.f64(double %f)
+  ret i64 %a
+}
+
+; CHECK-LABEL: llvm_wasm_trunc_unsigned_i64_f64:
+; CHECK-NEXT: .functype llvm_wasm_trunc_unsigned_i64_f64 (f64) -> (i64)
+; CHECK-NEXT: i64.trunc_f64_u $push[[L0:[0-9]+]]=, $0{{$}}
+; CHECK-NEXT: return $pop[[L0]]{{$}}
+declare i64 @llvm.wasm.trunc.unsigned.i64.f64(double)
+define i64 @llvm_wasm_trunc_unsigned_i64_f64(double %f) {
+  %a = call i64 @llvm.wasm.trunc.unsigned.i64.f64(double %f)
+  ret i64 %a
+}