McSema x86 runtime doesn't marshal FPU stack registers between native/lifted

[pgoodman]

I think that this should be pretty straightforward. For now, we can probably just push/pop off the stack.

[artemdinaburg]

Just checked and its still an issue. Guess we kind of forgot about the FPU stack when making those runtimes.

[Aiethel]

After this is fixed test for complex numbers abi can be enhanced with long double
https://github.com/trailofbits/mcsema/blob/master/tests/test_suite_generator/src/linux/complex_numbers/complex_numbers.c#L80

[abarbala]

Probably related to this. I have an application with floating point that uses x87 FPU. (Running Linux on x86_64. Current aim to decompile->lift->recompile.) As far as my current understanding of remill/mcsema is, __mcsema_attach_call does the following seq of instructions push %dx -> fldcw (%rsp) -> pop REL_ADDR(%rdi) where %rdx is the third argument to __mcsema_attach_call, which following the comments in the source code is the return address into native caller. This creates me a problem: I get FPU exceptions! What is the prototype of __mcsema_attach_call? Where in the code are you generating the arguments to pass to __mcsema_attach_call? (In my asm those are not generated correctly) ... I think that it is a great idea to pass the $fctrl value as a parameter, but probably it is better to make it fix for now ... Please suggest, thanks!

[abarbala]

My current solution (which work) is to add a single line in remill/tools/mcsema/mcsema/Arch/X86/Runtime/print_ELF_64_linux.cpp

static void PrintLoadFlags(FILE * out) {
// FPU control.

• fprintf(out, " mov edx, 0xf7f\n");

However, this will not allow to pass this value from the caller, i.e. the lifted code. Additionally, we are not conserving the state of the native code in this way. Any suggestion? BTW Can you explain why this function has this name?