Recursive call stack makes release build 4x slower

[JohelEGP]

In analyzing cppfront with Clang Build Analyzer, I noticed a very odd item:

**** Functions that took longest to compile:
  8037 ms: cpp2::primary_expression_node::is_fold_expression() const (/home/johel/Documents/C++/Forks/hsutter/waarudo/source/cppfront.cpp)
   417 ms: void cpp2::iteration_statement_node::visit<cpp2::sema>(cpp2::sema&, ... (/home/johel/Documents/C++/Forks/hsutter/waarudo/source/cppfront.cpp)

These are only the first two items.
primary_expression_node::is_fold_expression is exorbitantly expensive to compile.

I have a matrix of compilers (Clang, GCC) × configurations (Debug, Release).
Only Clang Release is affected, its compilation time increasing by 145 s.

Normally:

Configuration	Clang 18	GCC 14
Debug	25 s	47 s
Release	199 s	95 s

With the patch below:

Configuration	Clang 18	GCC 14
Debug	25 s	47 s
Release	54 s	95 s

This git patch removes the offending call that makes it expensive to optimize.

diff --git a/source/parse.h b/source/parse.h
index 103f9ce..bae5e7f 100644
--- a/source/parse.h
+++ b/source/parse.h
@@ -1556,7 +1556,7 @@ auto primary_expression_node::is_fold_expression() const
     break;case identifier:
         return *std::get<identifier>(expr) == "...";
     break;case expression_list:
-        return std::get<expression_list>(expr)->is_fold_expression();
+        return false; // std::get<expression_list>(expr)->is_fold_expression();
     break;case id_expression:
         return std::get<id_expression>(expr)->is_fold_expression();
     break;default: ; // the others can't contain folds

Analyzing with this patch drops the offending function from the default output.

I suspect this is due to the recursive nature of primary_expression_node::is_fold_expression.
This is the stack when first breaking at the offending call while debugging regression-tests/pure2-print.cpp2.

From level 18 to 1, those are all calls to is_fold_expression members.
And if primary_expression_node contains an expression_list_node, that can repeat, ad infinitum.

[JohelEGP]

This is a flame graph of the build with a Release configuration of Clang.
Unfortunately, I was unable to build a Debug configuration of Clang.

[hsutter]

Thanks @JohelEGP
That's strange... FWIW, on my laptop, a full compile of cppfront using GCC 10 and Clang 12 takes 17 seconds (each) with default optimizations, 25 and 33 seconds (respectively) with -O2.
Update: And MSVC 2022 takes 6 seconds in debug and 16 seconds with -O2.

[JohelEGP]

Maybe it's a regression by a more recent combination of LLVM version and C++ standard version.