https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119456
Robert Dubner <rdubner at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|ASSIGNED |RESOLVED
Resolution|--- |FIXED
--- Comment #3 from Robert Dubner <rdubner at gcc dot gnu.org> ---
This is by no means the end of the effort to produce more efficient code. But
in this particular case, the program
~~~~~~~
PROGRAM-ID. STRMOV.
DATA DIVISION.
WORKING-STORAGE SECTION.
77 foo pic x(10).
PROCEDURE DIVISION.
move 'A' to foo (1:1)
move 'B' to foo (2:1)
move 'C' to foo (3:1)
STOP RUN.
~~~~~~~~
compiled with -O0, now produces this fragment of x86_64 machine language:
.loc 1 6 10 is_stmt 1
movq foo.73.0(%rip), %rax
movl $0, %edx
addq %rdx, %rax
movb $65, (%rax)
.loc 1 7 10
movq foo.73.0(%rip), %rax
movl $1, %edx
addq %rdx, %rax
movb $66, (%rax)
.loc 1 8 10
movq foo.73.0(%rip), %rax
movl $2, %edx
addq %rdx, %rax
movb $67, (%rax)
.loc 1 9 10
movzwl __gg__data_return_code(%rip), %eax
That's four instructions per MOVE. Compiled with -O2, it gets down to two
instructions per MOVE:
.loc 1 6 10 view .LVU38
movq foo.73.0(%rip), %rax
movb $65, (%rax)
.loc 1 7 10 view .LVU39
movq foo.73.0(%rip), %rax
movb $66, 1(%rax)
.loc 1 8 10 view .LVU40
movq foo.73.0(%rip), %rax
movb $67, 2(%rax)
.loc 1 9 10 view .LVU41
movswl __gg__data_return_code(%rip), %edi
I feel justified in closing this PR.
