[Bug target/45835] New: Consider push simm8;pop reg for -Os

[Bug target/45835] New: Consider push simm8;pop reg for -Os

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45835

           Summary: Consider push simm8;pop reg for -Os
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: jakub@gcc.gnu.org
                CC: uros@gcc.gnu.org, hjl@gcc.gnu.org
            Target: x86_64-linux


http://embed.cs.utah.edu/embarrassing/jan_10/ snippets suggest that for -Os
(not sure if just for -m32 or even -m64) icc generates shorter sequences for
loading signed 8 bit immediates into registers.
movl $1, %eax is 5 bytes long, while pushl $1; popl %eax is 3 byte long for
-m32
(and similarly pushq $1; popq %rax for -m64).  For r8..r15 push/pop is 4 bytes,
while movl is 6 bytes.
Not sure about the performance implications and whether it shouldn't be
something controllable by some -m* switch for users like Linux kernel which
want -Os primarily to improve performance and if the push/pop would be
significantly slower they might not appreciate it.

[Bug target/45835] Consider push simm8;pop reg for -Os

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45835

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-26 08:46:48 UTC ---
As measurements show, it is indeed significantly slower.  So if we want to do
it, we'd need a switch to disable that.