https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66179

            Bug ID: 66179
           Summary: Sub-optimal code generation with __attribute__((leaf))
           Product: gcc
           Version: 5.1.0
            Status: UNCONFIRMED
          Severity: minor
          Priority: P3
         Component: other
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mednafen at sent dot com
  Target Milestone: ---

Created attachment 35556
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35556&action=edit
C code which demonstrates the issue.

The attached code, when compiled like(for example):

gcc -fno-asynchronous-unwind-tables -O2 -S -o leaf.s -c leaf.c

(with 4.9.2 or 5.1.0, Linux x86_64)

produces bloated, sub-optimal assembly like:

test_call_leaf:
        pushq   %r12
        pushq   %rbp
        pushq   %rbx
        movl    a(%rip), %r12d
        movl    c(%rip), %ebx
        call    function_leaf
        movl    b(%rip), %ebp
        addl    $2, %r12d
        addl    $2, %ebx
        movl    %r12d, a(%rip)
        call    function_leaf
        addl    $2, %ebp
        movl    %ebx, c(%rip)
        movl    %ebp, b(%rip)
        popq    %rbx
        popq    %rbp
        popq    %r12
        ret

compared to a more optimal possibility:

test_call_normal:
        subq    $8, %rsp
        addl    $2, a(%rip)
        call    function_normal
        addl    $2, b(%rip)
        call    function_normal
        addl    $2, c(%rip)
        addq    $8, %rsp
        ret

Reply via email to