On 21 May 2014 00:54, Jeff Law <l...@redhat.com> wrote:
> On 05/20/14 01:11, Zhenqiang Chen wrote:
>>
>> Hi,
>>
>> The patch fix ICE issue triggered by shrink-wrapping enhancement.
>>
>> Bootstrap and no make check regression on X86-64.
>>
>> OK for trunk?
>>
>> Thanks!
>> -Zhenqiang
>>
>>
>> 2014-05-20 Zhenqiang Chen <zhenqiang.c...@linaro.org>
>>
>> PR rtl-optimization/61220
>> Part of PR rtl-optimization/61225
>> * shrink-wrap.c (move_insn_for_shrink_wrap): Skip SP and FP
>> adjustment
>> insn; skip split_edge for a block without only on successor.
>>
>> testsuite/ChangeLog:
>> 2014-05-20 Zhenqiang Chen <zhenqiang.c...@linaro.org>
>>
>> * gcc.dg/pr61220.c: New test.
>> * gcc.dg/shrink-wrap-loop.c: Disable for x86_64 -m32 mode.
>>
> [ ... ]
>
>
>
>
>> /* Make sure that the source register isn't defined later in BB. */
>> @@ -204,6 +209,10 @@ move_insn_for_shrink_wrap (basic_block bb, rtx insn,
>> /* Create a new basic block on the edge. */
>> if (EDGE_COUNT (next_block->preds) == 2)
>> {
>> + /* split_edge for a block with only one successor is meaningless.
>> */
>> + if (EDGE_COUNT (bb->succs) == 1)
>> + return false;
>> +
>> next_block = split_edge (live_edge);
>
> So don't you just want to use:
>
> if (EDGE_CRITICAL_P (live_edge))
>
> to replace the two explicit checks of EDGE_COUNT on the preds & succs?
Thanks for the comment.
In the code, there are 4 combinations of EDGE_COUNT: <1, 1>, <1, 2>,
<2, 1> and <2, 2>. <2, 1> is "illegal". <2, 2> is legal, but need
split_edge. <1, *> can bypass the second check. EDGE_CRITICAL_P can
only distinguish <2, 2> from others.
So I think two explicit checks is more efficient than EDGE_CRITICAL_P.
Thanks!
-Zhenqiang
