Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Replace usage of EM_BOOL with C/C++ bool type. NFC #22155

Merged
merged 2 commits into from
Oct 2, 2024

Conversation

sbc100
Copy link
Collaborator

@sbc100 sbc100 commented Jun 27, 2024

These macros were already updated to expand to the C/C++ bool type
back in #22157. This is the natural NFC followup to that change.

@sbc100 sbc100 requested review from dschuff and kripken June 27, 2024 17:42
@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 27, 2024

This change was created using sed to substitutions

@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 27, 2024

Updated to include struct size saves and code size savings.

sbc100 added a commit to sbc100/emscripten-glfw that referenced this pull request Jun 27, 2024
I'm trying to change the type of EM_BOOL and this is currently
blocking that:
  emscripten-core/emscripten#22155
@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 27, 2024

I could do this in two phases perhaps:

  • Narrow the type of EM_BOOL
  • Remove the use of EM_BOOL

@kripken
Copy link
Member

kripken commented Jun 27, 2024

Why do struct sizes change? I'd expect an int and an unpacked bool to both take 32 bits. Is that wrong?

@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 27, 2024

Why do struct sizes change? I'd expect an int and an unpacked bool to both take 32 bits. Is that wrong?

I guess bool is just a single byte in C/C++?

Indeed this program printf 1 1 on my desktop system:

$ cat test.c 
#include <stdbool.h>
#include <stdio.h>

int main() {
  printf("%ld %ld\n", sizeof(bool), _Alignof(bool));
  return 0;
}

@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 27, 2024

Why do struct sizes change? I'd expect an int and an unpacked bool to both take 32 bits. Is that wrong?

I guess bool is just a single byte in C/C++?

Indeed this program printf 1 1 on my desktop system:

$ cat test.c 
#include <stdbool.h>
#include <stdio.h>

int main() {
  printf("%ld %ld\n", sizeof(bool), _Alignof(bool));
  return 0;
}

(same for C++ BTW)

@kripken
Copy link
Member

kripken commented Jun 27, 2024

By itself it's 1 byte, I guess, but inside a struct the field after it might be aligned:

#include <stdbool.h>
#include <stdio.h>

class C {
  bool x;
  int y;
};

int main() {
  printf("%ld\n", sizeof(C));
  return 0;
}

That prints 8 for me, so 4 bytes for each field. I'd expect that to be the common case, I'm surprised we save anything actually...

@kripken
Copy link
Member

kripken commented Jun 27, 2024

Anyhow, yeah, if this causes struct layout changes then splitting the PR as much as possible sgtm.

@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 27, 2024

By itself it's 1 byte, I guess, but inside a struct the field after it might be aligned:

#include <stdbool.h>
#include <stdio.h>

class C {
  bool x;
  int y;
};

int main() {
  printf("%ld\n", sizeof(C));
  return 0;
}

That prints 8 for me, so 4 bytes for each field. I'd expect that to be the common case, I'm surprised we save anything actually...

Right, that is just normal struct layout rules. If you put the bool last you will see difference. Or if you have several booleans toegther in the struct.

Lowering the alignment of a type can for sure shrink struct sizes. That is the primary use of the alignment of a given type. Higher alignment == worse for struct size.

sbc100 added a commit to sbc100/emscripten that referenced this pull request Jun 27, 2024
This reduces the size of several structs and results in code size
savings in some cases.

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: emscripten-core#22155.
sbc100 added a commit to sbc100/emscripten that referenced this pull request Jun 27, 2024
This reduces the size of several structs and results in code size
savings in some cases.

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: emscripten-core#22155.
sbc100 added a commit to sbc100/emscripten that referenced this pull request Jun 27, 2024
This reduces the size of several structs and results in code size
savings in some cases.

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: emscripten-core#22155.
sbc100 added a commit to sbc100/emscripten that referenced this pull request Jun 27, 2024
This reduces the size of several structs and results in code size
savings in some cases.

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: emscripten-core#22155.
sbc100 added a commit to sbc100/emscripten that referenced this pull request Jun 27, 2024
This reduces the size of several structs and results in code size
savings in some cases.

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: emscripten-core#22155.
sbc100 added a commit to sbc100/emscripten that referenced this pull request Jun 28, 2024
This reduces the size of several structs and results in code size
savings in some cases.

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: emscripten-core#22155.
sbc100 added a commit that referenced this pull request Jun 28, 2024
This reduces the size of several structs and can result in code size
savings in some cases. The reason
the code size savings don't show up in trivial examples is (I believe)
because this change also increases
the use of HEAP8 (where previously some examples only depended on
HEAP32).

This change is split of from a larger change I have planned to remove
the use of EM_BOOL completely: #22155.
@sbc100
Copy link
Collaborator Author

sbc100 commented Jun 29, 2024

Rebased now that #22157 has landed. I'm planning on waiting for a release or two before landing this larger change.

Copy link
Member

@kripken kripken left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lgtm in a separate release as you said.

Also, do we have a test that EM_BOOL, EM_TRUE, EM_FALSE still work as legacy?

@sbc100
Copy link
Collaborator Author

sbc100 commented Jul 1, 2024

Added a test.

@kripken
Copy link
Member

kripken commented Jul 1, 2024

Thanks - which is it though? (this is a pretty big diff and I don't see a separate commit for it, this would have been a nice case to not squash actually I think?)

@kripken
Copy link
Member

kripken commented Jul 1, 2024

Thanks, lgtm.

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 3, 2024

This change is now ready to land I think

@sbc100 sbc100 changed the title Replace EM_BOOL with normal C/C++ bool type. NFC Replace usage of EM_BOOL with C/C++ bool type. NFC Sep 3, 2024
@sbc100 sbc100 requested a review from kripken September 3, 2024 21:34
@sbc100 sbc100 enabled auto-merge (squash) September 3, 2024 21:42
ChangeLog.md Outdated Show resolved Hide resolved
ChangeLog.md Outdated Show resolved Hide resolved
@juj
Copy link
Collaborator

juj commented Sep 3, 2024

Just to note, this looks like a massive user-facing breaking ABI change. The previous sizeof(EM_BOOL) change from 4 bytes to 1 byte also was a breaking ABI change, and it regressed my Emscripten downstream library at juj/wasm_webgpu#46 (I was lucky to have added #pragma clang diagnostic error "-Wpadded" in my code, otherwise it would have invited silent failure). There could be projects out there that have used EM_BOOL, and their pre-compiled code will break on this change.

@juj
Copy link
Collaborator

juj commented Sep 3, 2024

Looking at stdbool.h, it has

#ifndef _STDBOOL_H
#define _STDBOOL_H

#ifndef __cplusplus

#define true 1
#define false 0
#define bool _Bool

#endif

#define __bool_true_false_are_defined 1

#endif

What happens if C code is compiled with -std=c89 or -std=c90 that does not yet have the _Bool data type? Is there a #define somewhere else that lets that code compile?

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 3, 2024

Looking at stdbool.h, it has

#ifndef _STDBOOL_H
#define _STDBOOL_H

#ifndef __cplusplus

#define true 1
#define false 0
#define bool _Bool

#endif

#define __bool_true_false_are_defined 1

#endif

What happens if C code is compiled with -std=c89 or -std=c90 that does not yet have the _Bool data type? Is there a #define somewhere else that lets that code compile?

Musl's <stdbool.h> continues to work regardless of which C standard you use:

$ cat test.c 
#include <stdbool.h>
bool x = true; 
$ ./emcc -std=c89 test.c

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 3, 2024

Just to note, this looks like a massive user-facing breaking ABI change. The previous sizeof(EM_BOOL) change from 4 bytes to 1 byte also was a breaking ABI change, and it regressed my Emscripten downstream library at juj/wasm_webgpu#46 (I was lucky to have added #pragma clang diagnostic error "-Wpadded" in my code, otherwise it would have invited silent failure). There could be projects out there that have used EM_BOOL, and their pre-compiled code will break on this change.

I think this particular change is an NFC. That change that could be potentially breaking is the one that already landed and was already released in 3.1.62: #22157. Right?

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 3, 2024

Just to note, this looks like a massive user-facing breaking ABI change. The previous sizeof(EM_BOOL) change from 4 bytes to 1 byte also was a breaking ABI change, and it regressed my Emscripten downstream library at juj/wasm_webgpu#46 (I was lucky to have added #pragma clang diagnostic error "-Wpadded" in my code, otherwise it would have invited silent failure). There could be projects out there that have used EM_BOOL, and their pre-compiled code will break on this change.

I think this particular change is an NFC. That change that could be potentially breaking is the one that already landed and was already released in 3.1.62: #22157. Right?

If you like we can wait a few more releases to see if any other fallout is reported.

So far I have only heard of one regression which was someone who was using the int type for a callback return type rather then EM_BOOL.

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 3, 2024

Just to note, this looks like a massive user-facing breaking ABI change. The previous sizeof(EM_BOOL) change from 4 bytes to 1 byte also was a breaking ABI change, and it regressed my Emscripten downstream library at juj/wasm_webgpu#46 (I was lucky to have added #pragma clang diagnostic error "-Wpadded" in my code, otherwise it would have invited silent failure). There could be projects out there that have used EM_BOOL, and their pre-compiled code will break on this change.

Just in case I'm missing something, here are the 3 types of possible fallout that I can think of that might result from this type of change:

  1. Somebody linking object files compiled with different versions of emscripten (struct sizes disagree)
  2. Somebody hardcoding struct offsets (for example in JS code that doesn't use macros for this).
  3. Somebody passing a callback function using int in the signature where EM_BOOL should be.

Are there any others that we should also consider?

Note: Only (1) and (2) above can result in silent failure whereas (3) is a compiler error.

@juj
Copy link
Collaborator

juj commented Sep 4, 2024

here are the 3 types of possible fallout that I can think of that might result

I think these scenarios are probably accurate. In (1) I would note that it is not necessarily only struct sizes that disagree, but also C++ name mangling that changes, since EM_BOOL used to be an int, but becomes a bool.. so I presume that LLVM will give the C++ symbol name a different mangling.

So if someone has a codebase where they define a function with EM_BOOL in the signature in a .cpp file, then they would not have ABI compatibility to link against code that uses those functions compiled with previous version of Emscripten.

2. Somebody hardcoding struct offsets (for example in JS code that doesn't use macros for this).

Unless I have missed a development(?), Emscripten end users cannot use these {{{ C_STRUCTS.emscripten_fetch_t.numBytes }}} types of directives. These are an Emscripten authors only thing. So this "somebody" would be all Emscripten users?

That change that could be potentially breaking is the one that already landed and was already released in 3.1.62: #22157. Right?

Yeah, the actual breakage occurred in #22157. After that, this renaming of EM_BOOL uses to bool looks cosmetic.

This pair of changes looks like first #22157 broke ABI compatibility of EM_BOOL, and then this PR is moving Emscripten away from using EM_BOOL at all. I wonder if the first change was needed for this second change to occur?

If I were to rewind back, I wonder if it might be less disrupting to keep EM_BOOL as-is and not change it at all, and then directly switch Emscripten headers to start using bool in function signatures where there is no risk of compatibility issues, and in new APIs?

Although given that this has already happened, I don't wish to cause any more disruptions.

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 4, 2024

Unless I have missed a development(?), Emscripten end users cannot use these {{{ C_STRUCTS.emscripten_fetch_t.numBytes }}} types of directives. These are an Emscripten authors only thing. So this "somebody" would be all Emscripten users?

Emscripten users can use the {{{ C_STRUCTS }}} stuff. The limitation is that they cannot add new structs of their own that don't exist in emscripten. So I would phrase that as "emscripten users who uses EM_BOOL in a struct that they define and then use a hardcoded offset in JS to access members of this stuct". This is a very long way from all emscripten users. In fact, I've yet to hear of a single user effected in this way.

On that matter there has been some progress towards allowing users to extend C_STRUCTS system. The first part of this already landed in: #22251.

@juj
Copy link
Collaborator

juj commented Sep 4, 2024

In fact, I've yet to hear of a single user effected in this way.

I was a single user who was affected this way! That change totally broke my wasm_webgpu project for my downstream users, and I didn't know about it until someone reported it as a bug to me.

@juj
Copy link
Collaborator

juj commented Sep 4, 2024

On that matter there has been some progress towards allowing users to extend C_STRUCTS system. The first part of this already landed in: #22251.

That's cool - it would definitely be nice to have "equal" footing for Emscripten end user JS libraries and Emscripten system JS libraries. I understand that user JS libraries can refer to system struct fields, though the usefulness of that is quite limited. Rather, end users most often would like to use the offset mechanism for filling structs in their own projects.

The reason I was using EM_BOOL was that it was already a handy way for the "system" to provide a standard boolean type.

@sbc100
Copy link
Collaborator Author

sbc100 commented Sep 4, 2024

In fact, I've yet to hear of a single user effected in this way.

I was a single user who was affected this way! That change totally broke my wasm_webgpu project for my downstream users, and I didn't know about it until someone reported it as a bug to me.

I see, sorry I didn't process that this was the type of failure you had. Was it this exact issue? i.e. hardcoded offsets in your JS code?

@juj
Copy link
Collaborator

juj commented Sep 4, 2024

To marshall WebGPU API calls, structs on C side need to be converted to JS side descriptor objects, e.g.

C header code:

image

and JS code:

image

and I had assumed that the memory layout would stay fixed.

Fortunately I had added this at some point though: juj/wasm_webgpu@3345174

@sbc100 sbc100 force-pushed the em_bool branch 3 times, most recently from d3a27c2 to c59fd01 Compare October 1, 2024 22:03
These macros were already updated to expand to the C/C++ `bool` type
back in emscripten-core#22157.  This is the natural NFC followup to that change.
@sbc100 sbc100 merged commit a5c8a0c into emscripten-core:main Oct 2, 2024
28 checks passed
@sbc100 sbc100 deleted the em_bool branch October 2, 2024 00:20
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants