[PATCH] x86 interrupt attribute

[H.J. Lu]

On Tue, Sep 29, 2015 at 5:23 PM, Mike Stump <mikestump@comcast.net> wrote:
> On Sep 29, 2015, at 3:40 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> How about adding a "no_caller_saved_registers" attribute?
>
> You can save all call clobbered registers with 3 instructions?  Really?  I’m skeptical.  Anyway, if you do this by turning off great swaths of registers, then, I guess that doesn’t surprise me.  I try and not turn any off in mine.

We need one instruction to save one register.

> Now, if you turn off great swaths of registers, then you can’t actually make any calls to an abi that supports those registers as call clobbered:

We don't need to save and restore any registers when calling a
function marked with
"no_caller_saved_registers" attribute except for registers used to
pass parameters.
Maybe "no_caller_saved_registers" isn't a good name.

>> +  if (cfun->machine->is_interrupt)
>> +    return ((df_regs_ever_live_p (regno)
>> +          || (call_used_regs[regno] && cfun->machine->make_calls))
>> +         && !fixed_regs[regno]
>> +         && !STACK_REGNO_P (regno)
>> +         && !MMX_REGNO_P (regno)
>> +         && regno != BP_REG
>> +         && regno != SP_REG
>> +         && (regno <= ST7_REG || regno >= XMM0_REG));
>
> So, any calls to a function that uses any excluded register won’t work, if that register is not fixed and is call clobbered.  If you call such a function, you _must_ save and restore all those registers.  Code like:

We save all changed registers as well as all caller-saved registers if we make
any function call.  Shouldn't it be sufficient?

>> +  if (cfun->machine->is_interrupt && VALID_MMX_REG_MODE (mode))
>> +    {
>> +      error ("MMX/3Dnow instructions aren't allowed in %s service routine",
>> +          (cfun->machine->is_exception ? "exception" : "interrupt"));
>> +      return;
>> +    }
>
> Does not save your from the obligation of saving such registers, if you support function calls inside an interrupt routine.  Once you add that support to make function calls work, then, you might as well lift this restriction, cause it would already just work?

Due to the way how MMX/x87 registers work, we don't support using
MMX/x87 registers in
interrupt handler.  People have to write assembly codes if they have
to use them.

> If you really can save everything in 3 instructions, then there is no point to trying to enhance it more on your port.  3 instructions execute so fast as to not matter.
>
> Now, if you ask me how I know all this, I had to debug a failure to save 1 register class in the prologue from large multi core instruction trace, and that class was the second most important class for general code gen right after the gprs.  Turns out that I killed a live variable in that class from a packet handler interrupt routine cause it failed to save/restore.  After that, I tested every register class and fixed all the issues.

We include one testcase for each register class in x86, including MMX
and x87, which
gives an error.


-- 
H.J.