Saving the FPU state IA32

[giovanig]

Hi,

I would like to save the FPU state onto a context pointer in order to switch the context between two threads. I have the following context:

Code: Select all

      struct _fpstate _fpu_ctx; //has 108 bytes
    
	Reg32 _edi; //GP registers + eflags + eip = 40 bytes
	Reg32 _esi;
	Reg32 _ebp;
	Reg32 _esp;
	Reg32 _ebx;
	Reg32 _edx;
	Reg32 _ecx;
	Reg32 _eax;
	Reg32 _eflags;
	Reg32 _eip;

The total context size is 148 bytes.

My switch context method is below. I want to save the context onto "*o" and restore the previous saved context from "n".

Code: Select all

void IA32::switch_context(Context * volatile * o, Context * volatile n)
{
    ASM("	pushfl				\n"
	"	pushal				\n" //the first two instructions pushes 40 bytes
	"	sub %esp, 108                  \n"
	"	fnsave      -108(%esp)       \n"
	"	movl    148(%esp), %eax	# old	\n" 
	"	movl    %esp, (%eax)		\n"
	"	movl    152(%esp), %esp	# new	\n"
	"	frstor      (%esp)               \n"
	"	add     %esp, 108             \n"
	"	popal				\n"
	"	popfl				        \n");
}

I am getting a GP fault on this code. Any clues of what I am doing wrong? I am considering that fnsave pushes 108 bytes, as states the manual.

Thanks in advance,
Giovani

[Combuster]

I don't trust any CPU task switcher at all when the save and restore parts aren't properly matched. Regardless of what it should be, this is what it's doing right now:

Code: Select all

old_context->first_field = esp
esp = &(new_context->first_field)

After that, there's no telling what really happens, evenso when you

Code: Select all

fpu_state[-1] = current_state
(...)
current_state = fpu_state[0]

Which evenso tosses things in places where they don't belong.

[Owen]

Also you're not even accessing the right stack addresses (GCC built a function call frame around your ASM block!)

[giovanig]

I'm not using the ebp. g++ generated the exact asm code I posted. The past code was:

Code: Select all

ASM("       pushfl                          \n"
        "       pushal                          \n"
        "       movl    40(%esp), %eax  # old   \n" 
        "       movl    %esp, (%eax)            \n"
        "       movl    44(%esp), %esp  # new   \n"
        "       popal                           \n"
        "       popfl                           \n")

It worked perfectly, saving and restoring only GP registers, eip, and eflags. Then, I added the space for the fpu state to the Context and added the fnsave and frstor instructions to swtich_context method.

Does FPU really save 108 bytes?

[Owen]

Doesn't matter. GCC will build its own function call frame for all functions - having just an asm block in there doesn't stop it...

either use the parameter addresses GCC passes (not possible in this case) or don't use inline assembly.

Modern AVX state is over 1kb and has tight alignment requirements

[giovanig]

Even if the file is compiled with -fomit-frame-pointer?

[Owen]

Yes.