add integer pointer memory strategy #8423
Conversation
|
@headius I've added a naive integer pointer parameter strategy, but it's causing a segfault. I thought simply wrapping the integer using NativeMemoryIO would work, but apparently not. Any ideas? On a promising note, it seems to allow some of mittsu's examples to run, although most of them do segfault. |
|
Hmm I would have expected that to work ok. Can you point me toward a segfault report or show me how to generate the same error? |
danini-the-panini
force-pushed
the
ds-ffi-fixes
branch
from
f8a360e to
9f6e0c5
Compare
|
@headius You can trigger a segfault by just running the ffi specs. I don't even know which spec is actually failing since it crashes the whole VM. Also I haven't figured out how to debug JNI yet, any tips? In the meantime, I've managed a workaround in Fiddle's FFI backend: ruby/fiddle#162 |
|
@danini-the-panini Looking into this today. If you can stop by our Matrix channel I'll be in the office. |
|
Deleted previous repro and stack because it was reflecting known failures on macOS with function callbacks (see #6995). I dug out the first failure in the FFI suite, and I think this is a failure due to integers now looking like they could be used as a valid pointer. New repro and stack: require 'ffi'
require 'delegate'
module PointerTestLib
extend FFI::Library
ffi_lib "spec/ffi/fixtures/libtest.so"
attach_function :ptr_ret_int32_t, [ :pointer, :int ], :int
end
PointerTestLib.ptr_ret_int32_t(0, 0)This is from the "Fixnum cannot be used as a Pointer argument" spec in spec/ffi/pointer_spec.rb. It properly raised the error before because Integers (Fixnum is the old name for 64-bit range Integers) could not be converted to a pointer. Now that you added such a strategy, it goes ahead and tries to pointerify the zero value in the first argument to ptr_ret_int32, and that naturally blows up when accessed. The JVM trace is below: @danini-the-panini Everything seems like it's actually working correctly! The problem is that now it accepts integer values for pointer arguments, but most places you pass a pointer it should apparently NOT accept an integer (according to FFI specs). |
|
@danini-the-panini Can you show me what this change fixed in your codebase? I'm wondering if it's perhaps a feature that Fiddle supports but FFI is supposed to error, so this spec is right for FFI but prevents some functionality Fiddle needs. |
|
@headius ok that makes sense, I had a sneaky suspicion this wasn't going to be the correct fix. I had tried something similar by just adding Fiddle does seem to allow integers as pointers in CRuby, and that is used extensively in mittsu-opengl since OpenGL has a few methods that have VOIDP arguments that are actually just used as byte offsets e.g. glVertexAttribPointer Here is an example use in Mittsu: https://github.com/danini-the-panini/mittsu-opengl/blob/a69434802b183577a6f0f770ca17de9d9a48d37f/lib/mittsu/opengl/geometry_like.rb#L43 In light of FFI's spec, I'll close this and finish up the fix in Fiddle's ffi_backend |
|
Interestingly, I managed yesterday to run all of Mittsu's examples using this patch instead of the ffi_backend fix, and I didn't get any segfaults. This is on Apple Silicon (M2 Pro) with callbacks 🤔 |
|
@danini-the-panini I don't think anything is wrong with the patch other than suddenly allowing integers to be passed for pointer arguments. That may be a valid feature to add, or it may lead to confusion if someone passes an integer they don't expect to become a pointer. Perhaps open an issue with ffi/ffi and we can discuss there with other stakeholders? |
|
@headius Fair enough. It appears this was a feature that was removed for "JRuby compat" ffi/ffi@59b48a9 |
danini-the-panini
force-pushed
the
ds-ffi-fixes
branch
2 times, most recently
from
cc383e4 to
628a83e
Compare
danini-the-panini
force-pushed
the
ds-ffi-fixes
branch
from
628a83e to
078e4e2
Compare
|
Closing this since ffi/ffi#1130 was closed |
This allows for passing integers as pointer arguments to functions when using the FFI backend. This is a workaround until we can get JRuby's FFI implementation to allow for it directly (see also jruby/jruby#8423) --------- Co-authored-by: Benoit Daloze <[email protected]>
(ruby/fiddle#162) This allows for passing integers as pointer arguments to functions when using the FFI backend. This is a workaround until we can get JRuby's FFI implementation to allow for it directly (see also jruby/jruby#8423) --------- ruby/fiddle@e2f0952e9b Co-authored-by: Benoit Daloze <[email protected]>
Allows implicitly passing integers into FFI functions that take pointers, using the passed in integer as the address.
This mimics the behavior of CRuby.