https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111576
Bug ID: 111576
Summary: gcc generates conditional branch for bitwise "&"
Product: gcc
Version: 13.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: eggert at cs dot ucla.edu
Target Milestone: ---
Created attachment 55983
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55983&action=edit
source code for branch-free test for "." or ".."
Although this is low prioriy for me, I thought I'd mention it in case it would
help GCC optimize better for others.
I looked into implementing a test for "." or ".." that was branch free. In
other words, implement "strcmp (p, ".") == 0 || strcmp (p, "..") == 0" without
using conditional branches. I came up with an expression that should do this,
but GCC translates a bitwise "&" into code that involves conditional branches.
Normally I would think conditional branches would be better avoided for bitwise
"&".
To see the situation, compile the attached program t.c with 'gcc -O2 -S t.c'
using gcc (GCC) 13.2.1 20230728 (Red Hat 13.2.1-1). It generates the attached
assembly language output t.s. The code generated for 'f' contains two
conditional branches, even though the source uses '&'. Furthermore, the
generated code evaluates the more complicated side of the '&' first, to see
whether it should evaluate the easy part, and this is not likely to be faster
than just evaluating the whole thing.
The code generated for the logically equivalent function 'g' is branch free,
but g's source code is trickier as it substitutes "~+!!" for plain "!". (The
"+" is present to work around GCC bug 111715.)
