On 11/18/14 09:57, Tom de Vries wrote:
Richard,
this (trunk) patch fixes PR62167.
The patch fixes a problem that triggers with the test-case on the 4.8
branch, when tail-merge makes a dead type-unsafe load alive.
I'm not able to reproduce this bug on 4.9 and trunk with the same
test-case. On those branches, the tail-merge already does not happen.
The reason for the difference is as follows: With 4.8 the two phi
arguments of the phi in the tail block are value-numbered identically:
...
SCC consists of: p_14
Value numbering p_14 stmt = p_14 = MEM[(struct head *)_13].first;
Setting value number of p_14 to p_14 (changed)
SCC consists of: p_15
Value numbering p_15 stmt = p_15 = _13->next;
Setting value number of p_15 to p_14 (changed)
...
With 4.9 (and trunk), that's not the case:
...
SCC consists of: p_14
Value numbering p_14 stmt = p_14 = MEM[(struct head *)&heads][k.1_9].first;
Setting value number of p_14 to p_14 (changed)
SCC consists of: p_15
Value numbering p_15 stmt = p_15 = _13->next;
Setting value number of p_15 to p_15 (changed)
...
I'm not sure the bug triggers on trunk and 4.9, but I see no reason why
it could not trigger, so I'd prefer to apply the patch to 4.9 and trunk
as well.
The patch introduces an xfail for pr51879-12.c. I can follow up with a
patch to improve upon that, but I think that's better limited to trunk
only.
Bootstrapped and reg-tested on x86_64/trunk.
So instead of simply disabling for anything with virtual operands,
shouldn't instead we be comparing the virtual operands and alias
information? Seems to me that if we did that properly, this would "just
work". Right?
jeff
