`clang-format` is not idempotent on certain files

[winterqt]

Given this file, running the following commands:

clang-format phmap.h > phmap-1.h
clang-format phmap-1.h > phmap-2.h
diff -u phmap-1.h phmap-2.h

shows that clang-format is not idempotent:

--- phmap-1.h	2024-09-06 13:18:49.621712646 -0400
+++ phmap-2.h	2024-09-06 13:18:54.169121591 -0400
@@ -2734,7 +2734,7 @@
   parallel_hash_set() noexcept(
       std::is_nothrow_default_constructible<hasher>::value &&
       std::is_nothrow_default_constructible<key_equal>::value &&
-      std::is_nothrow_default_constructible<allocator_type>::value) {}
+      std::is_nothrow_default_constructible<allocator_type>::value){}
 
 #if (__cplusplus >= 201703L || _MSVC_LANG >= 201402) &&                        \
     (defined(_MSC_VER) || defined(__clang__) ||                                \
@@ -2745,7 +2745,8 @@
                              const allocator_type &alloc = allocator_type())
       : parallel_hash_set(
             typename Inner::Params{bucket_cnt, hash_param, eq, alloc},
-            phmap::make_index_sequence<num_tables>{}) {}
+            phmap::make_index_sequence<num_tables>{}) {
+  }
 
   template <std::size_t... i>
   parallel_hash_set(typename Inner::Params const &p,

However, running clang-format on phmap-2.h does not produce any further changes.

I am able to reproduce this on both macOS and Ubuntu, using clang-format version 18.1.8 (from Nixpkgs and the LLVM APT repo respectively).

[winterqt]

Might be a dupe of #58202; though that was for Objective-C, and this is for C++.

[llvmbot]

@llvm/issue-subscribers-bug

Author: Winter (winterqt)

Given [this file](https://github.com/greg7mdp/parallel-hashmap/blob/be6a2c79857c9ea76760ca6ce782e1609713428e/parallel_hashmap/phmap.h), running the following commands:

clang-format phmap.h > phmap-1.h
clang-format phmap-1.h > phmap-2.h
diff -u phmap-1.h phmap-2.h

shows that clang-format is not idempotent:

--- phmap-1.h	2024-09-06 13:18:49.621712646 -0400
+++ phmap-2.h	2024-09-06 13:18:54.169121591 -0400
@@ -2734,7 +2734,7 @@
   parallel_hash_set() noexcept(
       std::is_nothrow_default_constructible<hasher>::value &&
       std::is_nothrow_default_constructible<key_equal>::value &&
-      std::is_nothrow_default_constructible<allocator_type>::value) {}
+      std::is_nothrow_default_constructible<allocator_type>::value){}
 
 #if (__cplusplus >= 201703L || _MSVC_LANG >= 201402) &&                        \
     (defined(_MSC_VER) || defined(__clang__) ||                                \
@@ -2745,7 +2745,8 @@
                              const allocator_type &alloc = allocator_type())
       : parallel_hash_set(
             typename Inner::Params{bucket_cnt, hash_param, eq, alloc},
-            phmap::make_index_sequence<num_tables>{}) {}
+            phmap::make_index_sequence<num_tables>{}) {
+  }
 
   template <std::size_t... i>
   parallel_hash_set(typename Inner::Params const &p,

However, running clang-format on phmap-2.h does not produce any further changes.

I am able to reproduce this on both macOS and Ubuntu, using clang-format version 18.1.8 (from Nixpkgs and the LLVM APT repo respectively).

[aij]

I can confirm the same problem with the nightly build from apt.llvm.org

$ clang-format-20 --version
Ubuntu clang-format version 20.0.0 (++20241023031243+629d9809ab14-1~exp1~20241023151404.2006)

It did reach a fixpoint after the second iteration with default settings.

FWIW, I tried formatting a C++ codebase at work (which unfortunately I can't share) where clang-format was non-idempotent on 47 files, and some files took 6 iterations to converge. One file never converged but started cycling through 3 different options. I'll see if I can produce a minimal repro of the last one as I suspect it may be a slightly different bug.

[aij]

Here's an example where repeated formatting fails to converge: nonconvergence.cc.txt (Sorry had to add a .txt extension to make GitHub happy)

Repro with

clang-format-20 --style='{BasedOnStyle: llvm, ReflowComments: false}' nonconvergence.cc -i

It just keeps flipping back and forth with these two diffs

diff --git nonconvergence.cc nonconvergence.cc
index 5aa29b1..10efce4 100644
--- nonconvergence.cc
+++ nonconvergence.cc
@@ -8,8 +8,8 @@ int window_type =
 #ifdef _WIN32
         do_the_thing()
 #else // _WIN32
-// No need for this on other platforms because we need extra text here to make
-// more lines.
+        // No need for this on other platforms because we need extra text here to make
+        // more lines.
         0
 #endif      
             ));

diff --git nonconvergence.cc nonconvergence.cc
index 10efce4..5aa29b1 100644
--- nonconvergence.cc
+++ nonconvergence.cc
@@ -8,8 +8,8 @@ int window_type =
 #ifdef _WIN32
         do_the_thing()
 #else // _WIN32
-        // No need for this on other platforms because we need extra text here to make
-        // more lines.
+// No need for this on other platforms because we need extra text here to make
+// more lines.
         0
 #endif      
             ));

I even tried adding the --fail-on-incomplete-format option and it still return exit code 0.

[owenca]

Here's an example where repeated formatting fails to converge: nonconvergence.cc.txt (Sorry had to add a .txt extension to make GitHub happy)

This is unrelated. Please split it to another issue.