https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69702
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P2
Status|UNCONFIRMED |NEW
Keywords| |missed-optimization
Last reconfirmed| |2016-02-08
Component|c |tree-optimization
CC| |rguenth at gcc dot gnu.org
Ever confirmed|0 |1
Summary|excessive stack usage with |[4.9/5/6 Regression]
|-fprofile-arcs |excessive stack usage with
| |-fprofile-arcs, LIM store
| |motion lacks a register
| |pressure aware cost model
Target Milestone|--- |4.9.4
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
We're more aggressively moving coverage counter updates out of loops by
promoting
them to registers. That might cause extra spilling due to increased register
pressure.
Confirmed. -fno-tree-loop-im "fixes" the regression.
Caused by (I'm quite sure, but not actually verified):
2014-03-21 Richard Biener <[email protected]>
PR tree-optimization/60577
* tree-core.h (struct tree_base): Document nothrow_flag use
in DECL_NONALIASED.
* tree.h (DECL_NONALIASED): New.
(may_be_aliased): Adjust.
* coverage.c (build_var): Set DECL_NONALIASED.
which fixed a regression from earlier releases that did not properly do
alias analysis for those (and thus applied store-motion to them anyway).
In general store motion lacks a cost model taking into account register
pressure increase of its transform. That's a know deficiency. Ideally we'd
have a later
pass undo the transform rather than limiting the transform here.
