|Anonymous | Login | Signup for a new account||2016-10-24 22:09 CEST|
|Main | My View | View Issues | Change Log | Roadmap|
|View Issue Details|
|ID||Project||Category||View Status||Date Submitted||Last Update|
|0005008||OCaml||OCaml general||public||2010-03-24 16:52||2011-05-29 12:19|
|Target Version||Fixed in Version||3.12.0+dev|
|Summary||0005008: Stack smash(?) on Win64 causes C "float" function to return random results|
On ocaml-3.11.2 (also verified on ocaml-3.11.1) compiled with Visual Studio 9.0 (i.e., 2008) and Windows SDK 6.0A on Windows 7 in 64-bit mode.
We define a C function with "float" calling convention:
double ml_single_precision_add_float(double a, double b)
return (float) (a + b);
external add : t->t->t = "ml_single_precision_add" "ml_single_precision_add_float" "float"
let x = add 1. 2. in ...
within a large OCaml program whose source I cannot disclose, unfortunately.
Then, non-deterministically (approx. 2 out of 10 runs), x will get a value which is not 3. (checked with Printf.printf), although the C function gets the right arguments and computes the right result (checked with printf).
I was not able to construct a small program exhibiting the same behavior, unfortunately.
2) Tentative explanation
I am not sure I understand fully what happens, but here is a guess.
After add returns successfully its result in XMM0, caml_call_gc is called.
It saves XMM0 in [rsp] and calls caml_garbage_collection, a C function without any argument.
On Win64, I was led to believe that a function must reserve 32 bytes on its stack, even a void function, as the space might be used by the callee as a scratchpad.
I guess some C code in caml_garbage_collection may smash the saved XMM0 in [rsp].
3) Tentative fix
In amd64nt.asm, caml_call_gc, change:
sub rsp, 32
add rsp, 32
On my program, this seems to fix the issue.
Again, I do not claim to fully understand this.
I've also heard that the stack should be kept aligned on 16-bytes.
Here, rsp is decreased by an odd number of qwords.
I don't know whether this might cause a problem.
|Tags||No tags attached.|
Thanks for the detective work. You are correct that the bottom 32 bytes of stack are scratch space in the Win64 calling conventions. ocamlopt-generated code as well as other functions in amd64nt.asm carefully reserve this space, but caml_call_gc did not. I've applied your fix; it will go in release 3.12.0.
|2010-03-24 16:52||Antoine Mine||New Issue|
|2010-03-28 10:09||xleroy||Note Added: 0005273|
|2010-03-28 10:09||xleroy||Assigned To||=> xleroy|
|2010-03-28 10:09||xleroy||Priority||normal => none|
|2010-03-28 10:09||xleroy||Status||new => resolved|
|2010-03-28 10:09||xleroy||Resolution||open => fixed|
|2010-03-28 10:09||xleroy||Fixed in Version||=> 3.12.0+dev|
|2011-05-29 12:19||xleroy||Status||resolved => closed|
|Copyright © 2000 - 2011 MantisBT Group|