Fix pthread_attr_t having 9 elems instead of 10

[aardappel]

Regression introduced here: #13006
We are relying on them having 10 elems thru the _a_prio field (in wasm64) and also offsets in library_pthread.js

[sbc100]

I would expect this to match the code in upstream musl:

https://github.com/emscripten-core/musl/blob/85e0e3519655220688e757b9d5bfd314923548bd/include/alltypes.h.in#L85

i.e. it would expect it to with be 14 or 9 for each other.. what am I missing here?

[aardappel]

Not sure.. we have:

#define __SU (sizeof(size_t)/sizeof(int))
#define _a_prio __u.__i[3*__SU+3]

so accessing that field causes:

system\lib\libc\musl\src\thread\pthread_attr_get.c:22:26: error: array index 9 is past the end of the array (which contains 9 elements)
      [-Werror,-Warray-bounds]
        param->sched_priority = a->_a_prio;

[sbc100]

Can you copy what musl does and do sizeof(long)==8?14:9 .. for consistency with upstream?

[aardappel]

that code was never in there and would give us suddenly 14 fields in wasm32, which also would require our JS bindings to further become conditional on wasm32/64, see e.g. https://github.com/emscripten-core/emscripten/pull/15229/files#diff-31ce84d9e9748b123befe225a55c67718b58011f9b0a099c821e5c5f0019b1fa

[sbc100]

OK fair enough. This change lgtm for now then.

[sbc100]

Wait, but don't the JS bindings already need to have different offsets for wasm64 since _a_prio is different for wasm64 (for example)

[aardappel]

not sure if we access _a_prio in JS (we access it in C), but we access the canvas names field which is currently at a single offset in both. See link above.

[sbc100]

Hmm, it does look like we access _a_stacksize, _a_stackaddr and _a_detach using hardcoded constants in the JS code:

emscripten/src/library_pthread.js

Lines 778 to 780 in ec42ea2

	stackSize = {{{ makeGetValue('attr', 0/*_a_stacksize*/, 'i32') }}};
	stackBase = {{{ makeGetValue('attr', 8/*_a_stackaddr*/, 'i32') }}};
	if ({{{ makeGetValue('attr', 12/*_a_detach*/, 'i32') }}}) {

These probably need to be fixes (although not necessarily as part of this PR).

The reference to the canvas is not longer hardcoded and uses the struct info so it should be fine, even if the struct layout varies between 32 and 64 (which it does):

emscripten/src/library_pthread.js

Line 667 in ec42ea2

var transferredCanvasNames = attr ? {{{ makeGetValue('attr', C_STRUCTS.pthread_attr_t._a_transferredcanvases, POINTER_TYPE) }}} : 0;

[sbc100]

lgtm for now since it fixes the wasm64 libc build.