[PATCH][RTL-ifcvt] PR rtl-optimization/67465: Handle pairs of complex+simple blocks and empty blocks more gracefully

[Kyrill Tkachov]

On 11/09/15 09:51, Rainer Orth wrote:
> Kyrill Tkachov <kyrylo.tkachov@arm.com> writes:
>
>> On 10/09/15 12:43, Rainer Orth wrote:
>>> Hi Kyrill,
>>>
>>>> Rainer, could you please check that this patch still fixes the SPARC
>>>> regressions?
>>> unfortunately, it breaks sparc-sun-solaris2.10 bootstrap: compiling
>>> stage2 libiberty/regex.c FAILs:
>>>
>>>
>> Thanks for providing the preprocessed file.
>> I've reproduced and fixed the ICE in this version of the patch.
>> The problem was that I was taking the mode of x before the check
>> of whether a and b are MEMs, after which we would change x to an
>> address_mode reg,
>> thus confusing emit_move_insn.
>>
>> The fix is to take the mode of x and perform the can_conditionally_move_p check
>> after that transformation.
>>
>> Bootstrapped and tested on aarch64 and x86_64.
>> The preprocessed regex.i that Rainer provided now compiles successfully for me
>> on a sparc-sun-solaris2.10 stage-1 cross-compiler.
>>
>> Rainer, thanks for your help so far, could you please try out this patch?
> While bootstrap succeeds again, the testsuite regression in
> gcc.c-torture/execute/20071216-1.c reoccured.
Right, so I dug into the RTL dumps and I think this is a separate issue that's being exacerbated by my patch.
The code tries to if-convert a block which contains a compare instruction i.e. sets the CC register.
Now, bb_valid_for_noce_process_p should have caught this, and in particular insn_valid_noce_process_p
which should check that the instruction doesn't set the CC register. However, on SPARC the
cc_in_cond returns NULL! This is due to the canonicalize_comparison implementation that seems to
remove the CC register from the condition expression and returns something like:
(leu (reg/v:SI 109 [ b ])
     (const_int -4096 [0xfffffffffffff000])

Therefore the set_of (cc_in_cond (cond), insn) check doesn't get triggered because cc_in_cond returns NULL.
Regardless of how the branch condition got canonicalized, I think we still want to reject any insn in the block
that sets a condition code register, so this patch checks the destination of every set in the block for a MODE_CC
expression and cancels if-conversion if that's the case.

Oleg pointed me to the older PR 58517 affecting SH which seems similar and I think my previous ifcvt patch would expose
this problem more.

Anyway, with this patch the failing SPARC testcase gcc.c-torture/execute/20071216-1.c generates the same assembly
as before r227368 and bootstrap and test on aarch64 and x86_64 passes ok for me.

Rainer, could you try this patch on top of the previous patch? (https://gcc.gnu.org/ml/gcc-patches/2015-09/msg00689.html)
The two together should fix all of PR 67456, 67464, 67465 and 67481.

Thanks,
Kyrill

2015-09-11  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     PR rtl-optimization/67481
     * ifcvt.c (contains_ccmode_rtx_p): New function.
     (insn_valid_noce_process_p): Use it.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: ifcvt-bug-2.patch
Type: text/x-patch
Size: 972 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20150911/272117cd/attachment.bin>