On 07/20/2011 01:23 PM, Jakub Jelinek wrote:
On Wed, Jul 20, 2011 at 01:07:40PM -0700, Michael Eager wrote:
I've run into a problem with this change when building microblaze-xilinx-elf.
When compiling _divdi3.o, cselib_expand_value_rtx_1 returns a huge rtx
tree for variable _r1 when max_depth is greater than 17. If -g is
specified, this later results in attempting to generate a DWARF location
list much larger than the 0xffff size limit, resulting in an assert failure.
I think Alex is working on a patch which will hopefully improve it.
In 4.6-RH in the mean time we are defaulting max-vartrack-expr-depth to 12
instead, perhaps that would be better than 10.
I'm OK with a value of 12.
As for the 0xffff size limit, I'd say we should handle it instead of
asserting, e.g. if the size is 64K or more, we could emit there a single
DW_OP_call4 + DW_TAG_dwarf_procedure that would contain it.
Or just drop the range on the floor, still better than ICE.
I don't think that a huge location list is a good result, even if you
package it into a DWARF procedure.
--
Michael Eager ea...@eagercon.com
1960 Park Blvd., Palo Alto, CA 94306 650-325-8077
