[Bug middle-end/49035] New: Avoid setting up stack frame for short, hot code paths

["scovich at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49035

           Summary: Avoid setting up stack frame for short, hot code paths
           Product: gcc
           Version: 4.5.2
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: scovich@gmail.com


I often find myself writing functions of the following form:

void foo () {
    if (common_case) {
        /* do a little work and return */
    }
    /* uncommon case: do a lot of work, call functions, etc. */
}

The resulting assembly code always sets up a stack frame in the function
prologue, even though the function usually executes as a leaf using few (or
zero) of the callee-save registers and stack slots it saves. 

Here's an example which is only slightly contrived:

=== rfe.cpp ============
struct link {
    link* prev;
    long go_slow;
    void frob(link* parent, link* grandparent);
};

link* foo(link* list) {
    link* prev = list->prev;
    while (__builtin_expect(prev->go_slow, 0)) {
        link* pprev = __sync_lock_test_and_set(&prev->prev, 0);
        pprev->frob(prev, list);
        prev = pprev;
    }
    return prev;
}
=== rfe.cpp ============

Compiling the above with `x86_64-unknown-linux-gnu-g++-4.5.2 -O3 -S' yields

_Z3fooP4link:
.LFB0:
        movq    %rbx, -24(%rsp)
        movq    %rbp, -16(%rsp)
        movq    %rdi, %rbx
        movq    %r12, -8(%rsp)
        subq    $24, %rsp
        movq    (%rdi), %rax
        cmpq    $0, 8(%rax)
        jne     .L8
.L2:
        movq    (%rsp), %rbx
        movq    8(%rsp), %rbp
        movq    16(%rsp), %r12
        addq    $24, %rsp
        ret
.L8:
        xorl    %r12d, %r12d
.L6:
        movq    %r12, %rbp
        xchgq   (%rax), %rbp
        movq    %rbx, %rdx
        movq    %rax, %rsi
        movq    %rbp, %rdi
        call    _ZN4link4frobEPS_S0_
        cmpq    $0, 8(%rbp)
        jne     .L4
        movq    %rbp, %rax
        jmp     .L2
.L4:
        movq    %rbp, %rax
        jmp     .L6


Ideally, it would look like this instead:

_Z3fooP4link:
.LFB0:
        ;; *** hot path executes as leaf ***
        movq    (%rdi), %rax
        cmpq    $0, 8(%rax)
        jne     .L8
        ret
.L8:
        ;; *** set up stack frame ***
        movq    %rbx, -24(%rsp)
        movq    %rbp, -16(%rsp)
        movq    %rdi, %rbx
        movq    %r12, -8(%rsp)
        subq    $24, %rsp
        ;; ***
        xorl    %r12d, %r12d
.L6:
        movq    %r12, %rbp
        xchgq   (%rax), %rbp
        movq    %rbx, %rdx
        movq    %rax, %rsi
        movq    %rbp, %rdi
        call    _ZN4link4frobEPS_S0_
        cmpq    $0, 8(%rbp)
        jne     .L4
        ;; *** tear down stack frame ***
        movq    %rbp, %rax
        movq    (%rsp), %rbx
        movq    8(%rsp), %rbp
        movq    16(%rsp), %r12
        addq    $24, %rsp
        ;; *** 
        ret
.L4:
        movq    %rbp, %rax
        jmp     .L6


The effect can sometimes be simulated using an inlined foo which includes the
fast path and a call to the (non-inlined) slow path, but the whims of function
inlining often conspire against it even when callers are able to inline foo
(e.g. foo is not a library function).

There's probably some overlap with partial inlining here: the ideal case
essentially splits the slow path off into its own function (called using tail
recursion); presumably partial inlining would inline the fast path while having
all callers jump to the same copy of the slow path function. However, the
optimization is arguably useful even if foo is never inlined at all.

Thoughts?
Ryan