[PATCH] D33135: [ASTMatchers] Add support for floatLiterals

[Aaron Ballman via Phabricator via cfe-commits]

aaron.ballman added a comment.

In https://reviews.llvm.org/D33135#754278, @Lekensteyn wrote:

> By the way, I think that `long double` is less common than long unsigned literals, so changing unsigned to uint64_t might be something more important?


I agree that it's likely a more common use case. There doesn't appear to be a suffix for __int128 (we talked about adding i128 once upon a time, but I don't believe it got in), so you may be okay using a `uint64_t` there rather than an `APInt`.



================
Comment at: include/clang/ASTMatchers/Dynamic/Parser.h:25
 /// <Boolean>           := true | false
+/// <Double>            := 1.0 | 2e-3 | 3.45e67
 /// <Unsigned>          := [0-9]+
----------------
Lekensteyn wrote:
> aaron.ballman wrote:
> > It'd be good to list the actual grammar rather than a few examples.
> I am concerned that listing a very precise grammar actually makes this less readable (see also the StringLiteral example).
> 
> If a grammar is used instead, how about this:
> 
>     <Double> := [0-9]+.[0-9]* | [0-9]+.[0-9]*[eE][-+]?[0-9]+
> 
That's reasonable enough.


================
Comment at: include/clang/ASTMatchers/Dynamic/VariantValue.h:335
     unsigned Unsigned;
+    double Double;
     bool Boolean;
----------------
Lekensteyn wrote:
> aaron.ballman wrote:
> > This may or may not be a good idea, but do we want to put the values into an APFloat rather than a double? My concern with double is that (0) it may be subtly different if the user wants a 16- or 32-bit float explicitly, (1) it won't be able to represent long double values, or quad double.
> > 
> > I'm thinking this value could be passed directly from the C++ API as an APFloat, float, or double, or provided using a StringRef for the dynamic API.
> (32-bit) double values are a superset of (16-bit) float values, that should be OK.
> Long doubles are possible in the AST (e.g. for `0.1L`), but neither C11 nor C++14 seem to define a quad double literal type (so that should be of a lesser concern).
> 
> Reasons why I chose for double instead of APFloat:
> - `strtod` is readily available and does not abort the program. By contrast, `APFloat(StringRef)` trips on assertions if the input is invalid.
> - I was not sure if the APFloat class can be used in an union.
The downside to using `strtod()` is that invalid input is silently accepted. However, assertions on invalid input is certainly not good either. It might be worth modifying `APFloat::convertFromString()` to accept invalid input and return an error.

I think instead of an `APFloat`, maybe using an `APValue` for both the `Unsigned` and `Double` fields might work. At the very least, it should give you implementation ideas.

There is a quad double literal suffix: `q`. It's only supported on some architectures, however. There are also imaginary numbers (`i`) and half (`h`).


================
Comment at: lib/ASTMatchers/Dynamic/Parser.cpp:180
   /// \brief Consume an unsigned literal.
   void consumeUnsignedLiteral(TokenInfo *Result) {
+    bool isFloatingLiteral = false;
----------------
Lekensteyn wrote:
> aaron.ballman wrote:
> > This function should be renamed and the comment updated.
> How does "consumeNumberLiteral" sound?
Sounds good to me.


================
Comment at: lib/ASTMatchers/Dynamic/Parser.cpp:209
+      double doubleValue = strtod(Result->Text.str().c_str(), &end);
+      if (*end == 0) {
+        Result->Kind = TokenInfo::TK_Literal;
----------------
You're failing to check errno here to ensure the value is actually valid.


https://reviews.llvm.org/D33135