Bug 74298

Summary:	the signal handler's ucontext doesn't get any FPU state
Product:	[Developer tools] valgrind	Reporter:	Jeremy Fitzhardinge <jeremy>
Component:	general	Assignee:	Jeremy Fitzhardinge <jeremy>
Status:	RESOLVED FIXED
Severity:	normal	CC:	johan.walles, mcguire, tjt
Priority:	NOR
Version First Reported In:	2.1 CVS
Target Milestone:	---
Platform:	Unlisted Binaries
OS:	Linux
Latest Commit:		Version Fixed/Implemented In:
Sentry Crash Report:

Description Jeremy Fitzhardinge 2004-02-05 23:51:28 UTC

When a signal handler for CPU excepton is called, it gets passed a ucontext
structure, which contains the CPU state at the time of the exception.  Valgrind
doesn't pass the handler any FPU/SSE related information; the fpstate pointer is
left NULL.

This makes Mesa-6.0 crash, because it looks at the fpstate when it's checking
that SSE is properly supported.

Comment 1 Dirk Mueller 2004-02-14 20:27:04 UTC

confirmed.

Comment 2 Jeremy Fitzhardinge 2004-03-12 04:54:53 UTC

Well, this turns out to be tricky to fix, at least for Mesa.  Mesa wants to do the hardest of all things - it wants to take an FPU exception, inspect the up-to-date FP state in the handler, then modify it, returning to the original instruction.

Valgrind really isn't set up for this.  Firstly, because of bug 69530, we can't give up-to-date FPU state to the handler.  Secondly, the signal frame is constructed so that there are two copies of the VCPU state - one for the handler to see, and one which is used to restore the state when the handler returns, so if the handler does change the state, it isn't the one which takes effect.

The second problem can be fixed reasonably easily, but the first is much harder.  The workaround we can apply now is to save the FP state before each potentially faulting instruction, but this is hard work and potentially a huge performance hit.

Not really sure what to do here.

Comment 3 Jeremy Fitzhardinge 2004-03-13 03:30:57 UTC

Well, it can be made to work, but only if you use --single-step=yes.

Comment 4 Tom Hughes 2004-08-05 22:04:15 UTC

*** Bug 86641 has been marked as a duplicate of this bug. ***

Comment 5 Tom Hughes 2005-01-04 16:33:28 UTC

*** Bug 95667 has been marked as a duplicate of this bug. ***

Comment 6 Jeremy Fitzhardinge 2005-01-18 23:45:20 UTC

OK, signal handlers now get state.  You still need to use --single-step=yes if you want accurately inspect/modify it from the handler (as Mesa wants to do).