[PATCH] D58292: Add support for importing ChooseExpr AST nodes.

Tom Roeder via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Thu Feb 21 15:44:39 PST 2019


tmroeder marked an inline comment as done.
tmroeder added inline comments.


================
Comment at: lib/AST/ASTImporter.cpp:6160
+  // condition-dependent.
+  bool CondIsTrue = false;
+  if (!E->isConditionDependent())
----------------
aaron.ballman wrote:
> tmroeder wrote:
> > a_sidorin wrote:
> > > aaron.ballman wrote:
> > > > tmroeder wrote:
> > > > > aaron.ballman wrote:
> > > > > > a_sidorin wrote:
> > > > > > > aaron.ballman wrote:
> > > > > > > > `bool CondIsTrue = E->isConditionDependent() ? false : E->isConditionTrue();`
> > > > > > > `bool CondIsTrue = E->isConditionDependent() && E->isConditionTrue();`?
> > > > > > I like that even better than my suggestion. :-)
> > > > > Wait, this doesn't have the same truth table as my original code.
> > > > > 
> > > > > let `CD = E->isConditionDependent()`
> > > > > let `CT = E->isConditionTrue()`
> > > > > 
> > > > > in
> > > > > 
> > > > > ```
> > > > > bool CondIsTrue = false;
> > > > > if (!CD)
> > > > >   CondIsTrue = CT;
> > > > > ```
> > > > > 
> > > > > has the table for `CondIsTrue`:
> > > > > 
> > > > > | `CD` | `CT` |  `CondIsTrue` |
> > > > > | T | T | F |
> > > > > | T | F | F |
> > > > > | F | T | T |
> > > > > | F | F | F |
> > > > > i.e., if CD is true, then CondIsTrue is always false. Otherwise, it's the value of CT.
> > > > > 
> > > > > The suggested line has the truth table
> > > > > 
> > > > > | `CD` | `CT` |  `CondIsTrue` |
> > > > > | T | T | T |
> > > > > | T | F | F |
> > > > > | F | T | F |
> > > > > | F | F | F |
> > > > > 
> > > > > That is, the effect of CD is swapped.
> > > > > 
> > > > > Aaron's suggestion matches my original table. I based my code on include/clang/AST/Expr.h line 4032, which asserts !isConditionDependent() in the implementation of isConditionTrue.
> > > > > 
> > > > > I realized this after I "fixed" my comment to match the implementation change. Am I missing something? Or is the assertion in Expr.h wrong? I think this should be
> > > > > 
> > > > > ```
> > > > > bool CondIsTrue = !E->isConditionDependent() && E->isConditionTrue();
> > > > > ```
> > > > > 
> > > > > I've changed my code to that and reverted the comment change.
> > > > Good catch -- I think my eyes just missed the change in logic. Perhaps we should add a test case that exercises this?
> > > Wow, that's a nice catch. Sorry for the misleading.
> > I started to look for a way to test this, then realized that while it's possible to test the code itself, it's impossible to make a ChooseExpr with isConditionDependent() that returns true for real code.
> > 
> > TL;DR: ChooseExpr is a C-only construct, and isConditionDependent is a C++-only condition; it's always false in C.
> > 
> > Details:
> > 
> > ChooseExpr only represents __builtin_choose_expr which is only valid in C, not C++. That means ChooseExpr::isConditionDependent() will always be false.
> > 
> > The definition is
> > 
> > ```
> > bool isConditionDependent() const {
> >   return getCond()->isTypeDependent() || getCond() ->isValueDependent();
> > }
> > ```
> > 
> > However Expr::isTypeDependent() (see Expr.h line 158) is a purely C++ property to do with templates ([temp.dep.expr]): it is true if the type of the expression could change from one template instantiation to another.
> > 
> > Similarly Expr::isValueDependent() (see Expr.h line 144) is a purely C++ property to do with templates ([temp.dep.expr]): it is true for value-dependent types in templates.
> > 
> > Both will always be false in all instantiations of ChooseExpr, due to the language difference.
> > 
> > So, I think ChooseExpr can be refactored to remove isConditionDependent and change its constructor to remove TypeDependent and ValueDependent.
> > 
> > If it's OK with you, I'll do that in a followup patch.
> > ChooseExpr only represents __builtin_choose_expr which is only valid in C, not C++. 
> 
> We allow it in C++, though: https://godbolt.org/z/_f1DPV
Hmm. Is that by design or chance? GCC doesn't allow it: https://godbolt.org/z/kvGCk1

Maybe it shouldn't be allowed?


Repository:
  rC Clang

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D58292/new/

https://reviews.llvm.org/D58292





More information about the cfe-commits mailing list