[PATCH 3/6] Allow jumps in epilogues

[Richard Henderson]

On 03/31/2011 12:59 PM, Bernd Schmidt wrote:
>> So long as late late compilation passes continue to not move frame-related
>> insns across basic block boundaries, we should be fine.
> 
> I'm nervous about this as the reorg pass can do arbitrary
> transformations. On Blackfin for example, we can reorder basic blocks
> for the sake of loop optimizations; sched-ebb can create new blocks,
> etc. I think it would be best if we can somehow make it work during
> final, without a CFG.

I guess that's the best thing for now.  I'm sure we all agree that long
term all transformations should preserve the CFG all the way through final.
At which point this could be implemented as a pass on the function after
all transformations are complete.

> Rather than use a CFG, I've tried to do something similar to
> compute_barrier_args_size, using JUMP_LABELs etc.

A reasonable solution for now, I suppose.

> 
> Summary of the patches:
> 001 - just create a dwarf2out_frame_debug_init function.

Ok.

> 002 - Make it walk the function in a first pass and record CFIs to
>       be output later

Do I correctly understand that NOTE_INSN_CFI isn't actually being
used in this patch?

> 003 - Store dw_cfi_refs in VECs rather than linked lists. Looks
>       larger than it is due to reindentation

Like 001, this one looks like it's totally independent of and of
the other changes, and a good cleanup.  Please go ahead and test
and commit this one independently.

> 004 - Change the function walk introduced in 002 so that it records
>       and restores state when reaching jumps/barriers

I'm not too fond of vec_is_prefix_of.  The Problem is that you're
not applying any understanding of the actual data, just doing a
string comparison (effectively).

Imagine two code paths A and B that both save R2 and R3 into their
respective stack slots.  Imagine that -- for whatever reason -- the
stores have been scheduled differently such that on path A R2 is 
saved before R3, and the reverse on path B.

Your prefix test will conclude that paths A and B end with different
unwind info, even though they are in fact compatible.

Using some mechanism by which we can compare aggregate CFI information
on a per-register basis ought to also vastly improve the efficiency in
adjusting the cfi info between code points.  It should also enable
proper information in the -freorder-blocks-and-partition case.

> * i386.c uses dwarf2out_frame_debug directly in some cases and is
>   unconverted

Hum.  I wonder what the best way to attack this.  It's a local change,
adjusting and then restoring the unwind state between two insns that
should not be scheduled separately.

We could turn them into two unspec_volatiles, and lose scheduling 
across this pattern.  But ideally this is a value that ought to be
shrink-wrapped.  It's expensive to compute, and there are many
early-return situations in which we don't need it.

I suppose we could split this pattern manually in i386 reorg; 
forcing this to be split before final even at -O0.  At that point
all shrink-wrapping would be done and an unspecv replacement 
would be ok.

> * I haven't tested whether my attempt to use
>   get_eh_landing_pad_from_rtx in the absence of a CFG actually works

It will.  This information is stored in cfun->eh.  By design this
information must survive until final, so that we can emit the 
actual eh info into the appropriate tables.

> * Computed jumps and nonlocal gotos aren't handled. I think this
>   could be done by recording the state at NOTE_INSN_PROLOGUE_END
>   and using that for all labels we can't otherwise reach.

That should be reasonable.  You could assert that all of these 
labels are in forced_labels.  All computed branch targets should
be listed therein.


r~