[BUG]: Cannot compile emscripten/pyodide project

[mducle]

Required prerequisites

• Make sure you've read the documentation. Your issue may be addressed there.
• Search the issue tracker and Discussions to verify that this hasn't already been reported. +1 or comment there if it has.
• Consider asking first in the Gitter chat room or in a Discussion.

What version (or hash if on master) of pybind11 are you using?

3.0.4

Problem description

I'm working on a project using pybind11 and CMake for pyodide/webassembly. With newer versions of CMake (>=4.1.0) and pybind11 (>=3.0.2) the cross-compilation gives an error. Using either older CMake (<4.1.0) and newer pybind (>=3.0.2) or new CMake and old pybind gives successful compilation.

I think this is caused by PR #5829 which only sets the internal variable _PYBIND11_CROSSCOMPILING if the CMAKE_CROSSCOMPILING_EMULATOR is not defined (code) - however, policy CMP0190 introduced by CMake 4.1 actually requires this to be defined and set to the program used to execute the cross-compiler. As such when running CMake with the emscripten cross-compiler, CMAKE_CROSSCOMPILING_EMULATOR is always defined which means the check fails and _PYBIND11_CROSSCOMPILING is always set to OFF and the compile then fails when the pointer size does not match (webassembly is 32-bit, the host system is 64-bit.

This can be reproduced with the cmake_example project with some modifications in a fork:

git clone --recurse-submodule https://github.com/mducle/cmake_example
cd cmake_example/pybind11
git checkout v3.0.4
cd ..
python -m pip install cibuildwheel cmake
CIBW_BUILD="cp312*" python -m cibuildwheel --platform pyodide

This will give an error Cannot run the interpreter but this is spurious - it's because the cross-compiling flag is not set correctly.

Now, I can get the compilation to work by:

• Changing cmake version to 4.0.0 or earlier in the pip install
• Changing pybind version to 3.0.1 or earlier in the checkout
• Keep both versions up-to-date and editing tools/pybind11Common.cmake to remove the NOT in line 55

Reproducible example code

Is this a regression? Put the last known working version here if it is.

3.0.1

[rwgk]

@henryiii Could you help triaging this issue?

[rwgk]

Cursor GPT-5.4 Extra High Fast (no manual edits)

---

From a quick triage, this does look like a plausible real regression, and the report already points to a fairly specific suspicious code path.

What seems solid already:

• the version window is reasonably clean: the report says this fails with pybind11 >= 3.0.2 together with CMake >= 4.1, while older CMake with newer pybind11 and newer CMake with pybind11 <= 3.0.1 both work
• the suspected change is localized to PR #5829, specifically the new cross-compiling logic in tools/pybind11Common.cmake
• in current pybind11, PYBIND11_USE_CROSSCOMPILING is enabled for CMake >= 4.1 when CMP0190 is NEW, but _PYBIND11_CROSSCOMPILING is then set to ON only when CMAKE_CROSSCOMPILING_EMULATOR is not defined
• under CMP0190 NEW, having CMAKE_CROSSCOMPILING_EMULATOR defined is actually expected in the kind of cross-compiling setup described here, so _PYBIND11_CROSSCOMPILING ending up OFF in that case does look suspicious

That would also fit the reported symptom pattern: if pybind11 incorrectly thinks this is not a cross-compiling case, it can fall back into host/interpreter/pointer-size checks that do not make sense for an Emscripten/Pyodide target, which would explain the spurious Cannot run the interpreter / bitness-mismatch style failure.

I have not drilled deeper than that yet, so I am not claiming the exact fix is proven already. But based on the report as written, this looks concrete enough to be worth a closer look, not just a vague environment problem.

[rwgk]

@mducle

Keep both versions up-to-date and editing tools/pybind11Common.cmake to remove the NOT in line 55

Do you think that could be a good permanent fix? Could you send a PR?

[mducle]

@mhsmith - could please you check that the logic in line 55 of tools/pybind11Common.cmake is correct as is, or if the NOT is incorrect?

Although removing the NOT works for me cross-compiling to emscripten, I don't know if it works with cmake >= 4.1 with Android which is what your PR was about, and I don't have an Android cross-compiling toolchain setup to check.

[mhsmith]

Sorry for the slow reply, I'll look into this tomorrow.

[mhsmith]

I agree this is a bug. My intention was to default to enabling _PYBIND11_CROSSCOMPILING in some situations, not to override anything which was explicitly set by the project. Sorry about that.

I think the correct fix is something like this (untested):

--- a/tools/pybind11Common.cmake
+++ b/tools/pybind11Common.cmake
@@ -43,7 +43,9 @@ set(pybind11_INCLUDE_DIRS
 
 # CMP0190 prohibits calling FindPython with both Interpreter and Development components
 # when cross-compiling, unless the CMAKE_CROSSCOMPILING_EMULATOR variable is defined.
-if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.1")
+if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.1"
+   AND NOT DEFINED PYBIND11_USE_CROSSCOMPILING
+   AND NOT DEFINED CMAKE_CROSSCOMPILING_EMULATOR)
   cmake_policy(GET CMP0190 _pybind11_cmp0190)
   if(_pybind11_cmp0190 STREQUAL "NEW")
     set(PYBIND11_USE_CROSSCOMPILING "ON")
@@ -51,8 +53,7 @@ if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.1")
 endif()
 
 if(CMAKE_CROSSCOMPILING
-   AND PYBIND11_USE_CROSSCOMPILING
-   AND NOT DEFINED CMAKE_CROSSCOMPILING_EMULATOR)
+   AND PYBIND11_USE_CROSSCOMPILING)
   set(_PYBIND11_CROSSCOMPILING
       ON
       CACHE INTERNAL "")