[PATCH] Fix SRA problem with variable-sized field

[Eric Botcazou]

Hi,

This is a case where a record with a variable-sized field slips through the 
cracks of the guards put in place in SRA against them, leading to an ICE 
because iterative_hash_expr is not supposed to see PLACEHOLDER_EXPRs.

SRA is trying to decide whether to scalarize a record m (unfortunately I have 
no testcase for the mainline compiler).

First, although m contains a record with variable size F, it is of fixed size 
because of the padding; so the check in sra_type_can_be_decomposed_p doesn't 
catch it because the predicate is not recursive (by design).

Second, although sra_walk_expr has a check against fields at variable offsets,
it doesn't catch F because it is at offset 0 despite being of variable size.

When maybe_lookup_element_for_expr is invoked, it will try to compute a hash 
value for the field F and runs into the PLACEHOLDER_EXPR.

Fixed by dealing with variable-sized records exactly like with unions in SRA, 
i.e. by explicitly balking at them.

Tested on i586-suse-linux, OK for mainline?


2007-09-16  Eric Botcazou  <ebotcazou@adacore.com>

	* tree-sra.c (maybe_lookup_element_for_expr) <COMPONENT_REF>: Return
	NULL for variable-sized records too.
	(sra_walk_expr) <COMPONENT_REF>: Stop at variable-sized records too.


-- 
Eric Botcazou
-------------- next part --------------
A non-text attachment was scrubbed...
Name: p.diff
Type: text/x-diff
Size: 1622 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20070916/f383d7c6/attachment.bin>