Patchwork ARM: exclude fixed_regs for stack-alignment save/restore

login
register
mail settings
Submitter Roland McGrath
Date June 14, 2012, 5:16 p.m.
Message ID <x57jboklg6f9.fsf@frobland.mtv.corp.google.com>
Download mbox | patch
Permalink /patch/164974/
State New
Headers show

Comments

Roland McGrath - June 14, 2012, 5:16 p.m.
When the ARM compiler needs to ensure the stack pointer stays aligned
and it's already doing a multi-register push/pop in the prologue and
epilogue, it chooses some arbitrary register to add to the register set
in that push and pop just to increase the size of the stack used by 4
bytes.  This is presumed to be harmless, since some register that is
either call-clobbered or not touched by the function at all is just
getting pushed and then the same value popped into it.  

But if e.g. I use -ffixed-r9 then I think it's a reasonable expectation
that no code is generated that touches r9 in any way, shape, or form.
(My actual concern is a variant target port still in progress, where
the ABI specifies that r9 is reserved, and the system enforces that no
instruction may modify r9.)

I haven't managed to come up with an isolated test case to demonstrate
the bug.  Apparently I just don't understand the stack and register
pressure requirements that make the compiler get into the situation
where it wants to add a random register for alignment padding purposes.

I don't have a setup where I can do a proper regression test for ARM.
(My system has a /usr/arm-linux-gnueabi/include/ but configuring with
--target=arm-linux-gnueabi --with-headers=/usr/arm-linux-gnueabi/include
did not succeed in building libgcc.)

But the change seems pretty obviously correct IMHO.


Thanks,
Roland


gcc/
2012-06-14  Roland McGrath  <mcgrathr@google.com>

	* config/arm/arm.c (arm_get_frame_offsets): Never use a fixed register
	as the extra register to save/restore for stack-alignment padding.
Mike Stump - June 14, 2012, 8:13 p.m.
On Jun 14, 2012, at 10:16 AM, Roland McGrath wrote:
> But if e.g. I use -ffixed-r9 then I think it's a reasonable expectation
> that no code is generated that touches r9 in any way, shape, or form.

Also,  I can't help but wonder if global_regs is respected.  In theory, people are allowed to declare global registers, and nothing should be stopping them, though, this is abi breaking, and one does need to recompile the world as I recall to use it, so, most people don't use it and can't use it, but the bare metal people can.

Your change looks good to me.

I'll note in passing that cse.c does:

/* Determine whether register number N is considered a fixed register for the                             
   purpose of approximating register costs.                                                               
   It is desirable to replace other regs with fixed regs, to reduce need for                              
   non-fixed hard regs.                                                                                   
   A reg wins if it is either the frame pointer or designated as fixed.  */
#define FIXED_REGNO_P(N)  \
  ((N) == FRAME_POINTER_REGNUM || (N) == HARD_FRAME_POINTER_REGNUM \
   || fixed_regs[N] || global_regs[N])
Roland McGrath - June 14, 2012, 8:21 p.m.
On Thu, Jun 14, 2012 at 1:13 PM, Mike Stump <mikestump@comcast.net> wrote:
> On Jun 14, 2012, at 10:16 AM, Roland McGrath wrote:
>> But if e.g. I use -ffixed-r9 then I think it's a reasonable expectation
>> that no code is generated that touches r9 in any way, shape, or form.
>
> Also, I can't help but wonder if global_regs is respected.

It's clearly not.  Making the added condition !fixed_regs[i] && !global_regs[i]
seems sensible to me.


Thanks,
Roland
Richard Sandiford - June 16, 2012, 12:42 p.m.
Roland McGrath <mcgrathr@google.com> writes:
> On Thu, Jun 14, 2012 at 1:13 PM, Mike Stump <mikestump@comcast.net> wrote:
>> On Jun 14, 2012, at 10:16 AM, Roland McGrath wrote:
>>> But if e.g. I use -ffixed-r9 then I think it's a reasonable expectation
>>> that no code is generated that touches r9 in any way, shape, or form.
>>
>> Also, I can't help but wonder if global_regs is respected.
>
> It's clearly not.  Making the added condition !fixed_regs[i] &&
> !global_regs[i] seems sensible to me.

All global registers have to be fixed though.  The original seemed
fine to me FWIW.

Richard
Richard Earnshaw - June 18, 2012, 3 p.m.
On 16/06/12 13:42, Richard Sandiford wrote:
> Roland McGrath <mcgrathr@google.com> writes:
>> On Thu, Jun 14, 2012 at 1:13 PM, Mike Stump <mikestump@comcast.net> wrote:
>>> On Jun 14, 2012, at 10:16 AM, Roland McGrath wrote:
>>>> But if e.g. I use -ffixed-r9 then I think it's a reasonable expectation
>>>> that no code is generated that touches r9 in any way, shape, or form.
>>>
>>> Also, I can't help but wonder if global_regs is respected.
>>
>> It's clearly not.  Making the added condition !fixed_regs[i] &&
>> !global_regs[i] seems sensible to me.
> 
> All global registers have to be fixed though.  The original seemed
> fine to me FWIW.
> 
> Richard
> 
Indeed, see globalize_reg() in reginfo.c.

R.
Roland McGrath - June 18, 2012, 4:34 p.m.
OK then.  If you like the original patch, would you like to commit it for me?

Thanks,
Roland
Roland McGrath - June 20, 2012, 5:17 p.m.
On Mon, Jun 18, 2012 at 9:34 AM, Roland McGrath <mcgrathr@google.com> wrote:
> OK then.  If you like the original patch, would you like to commit it for me?

ping?

Patch

diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 092e202..bcfef3e 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -16752,7 +16752,12 @@  arm_get_frame_offsets (void)
 	  else
 	    for (i = 4; i <= (TARGET_THUMB1 ? LAST_LO_REGNUM : 11); i++)
 	      {
-		if ((offsets->saved_regs_mask & (1 << i)) == 0)
+		/* While the gratuitous register save/restore is ordinarily
+		   harmless, if a register is marked as fixed it may be
+		   entirely forbidden by the system ABI to touch it, so we
+		   should avoid those registers.  */
+		if (!fixed_regs[i]
+		    && (offsets->saved_regs_mask & (1 << i)) == 0)
 		  {
 		    reg = i;
 		    break;