https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65658
Bug ID: 65658
Summary: Jump threading too pessimistic when optimizing for
size
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: law at redhat dot com
Created attachment 35209
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35209&action=edit
testcase, compile with -O2 -Wall note uninitialized warnings for SRA's
variables and trivially threadable block in setup()
The jump threader will cancel jump threads when optimizing for size and the
block will need to be duplicated. But that heuristic fails for something a
block like this:
# _36 = PHI <1(2), 1(4), 0(7)>
# problem$l3_101 = PHI <problem$l3_8(D)(2), problem$l3_8(D)(4),
problem$l3_35(7)>
# problem$l4_100 = PHI <problem$l4_124(D)(2), problem$l4_124(D)(4),
problem$l4_34(7)>
amt ={v} {CLOBBER};
if (_36 != 0)
goto <bb 25>;
else
goto <bb 9>;
All incoming edges are threadable and the block produces no code. It's really
just a redirection block that's not recognized as such -- probably due to the
clobber statement.
The net result is we cancel the jump thread, get worse code and issue false
positive warnings for uninitialized variables. All around a lose.
