[Bug tree-optimization/104484] New: -freorder-block-and-partition not splitting into sections with __builin_expect()

[Bug tree-optimization/104484] New: -freorder-block-and-partition not splitting into sections with __builin_expect()

[avi at scylladb dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104484

            Bug ID: 104484
           Summary: -freorder-block-and-partition not splitting into
                    sections with __builin_expect()
           Product: gcc
           Version: 11.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: avi at scylladb dot com
  Target Milestone: ---

I expected code guarded by __builtin_expect() to be pushed into .text.cold, but
that's not happening:

int f1, f2;

inline int is() {
    return __builtin_expect(f1, 0) && f2;
}

void heavy() {
    extern void very_heavy();
    if (is()) {
        very_heavy();
    }
}


void fun();

void light() {
    fun();
    heavy();
    fun();
}


with -O3 -freorder-blocks-and-partition:


light:
.LFB2:
        .cfi_startproc
        subq    $8, %rsp
        .cfi_def_cfa_offset 16
        xorl    %eax, %eax
        call    fun
        movl    f1(%rip), %edx
        testl   %edx, %edx
        jne     .L8        .p2align 4,,10
        .p2align 3
.L8:
        .cfi_restore_state
        movl    f2(%rip), %eax
        testl   %eax, %eax
        je      .L6
        xorl    %eax, %eax
        call    very_heavy
        jmp     .L6
        .cfi_endproc

.L6:
        xorl    %eax, %eax
        addq    $8, %rsp
        .cfi_remember_state
        .cfi_def_cfa_offset 8
        jmp     fun


I expected .L8 to be in another section, but it isn't.

[Bug rtl-optimization/104484] -freorder-block-and-partition not splitting into sections with __builin_expect()

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104484

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
There are heuristics used here. It might be also the data got lost when moving
to rtl from gimple.
But the assembly output does not make sense to the code you gave either.

[Bug rtl-optimization/104484] -freorder-block-and-partition not splitting into sections with __builin_expect()

[avi at scylladb dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104484

--- Comment #2 from Avi Kivity <avi at scylladb dot com> ---
(In reply to Andrew Pinski from comment #1)
> But the assembly output does not make sense to the code you gave either.


Apart from the missing .section directives, I didn't notice anything odd. Maybe
movl/testl could be optimized into cmpl.

Ah, I moved this code from C++ so the function signatures are bad, should be
fun(void). With the signatures adjusted, I get

light:
.LFB2:
        .cfi_startproc
        subq    $8, %rsp
        .cfi_def_cfa_offset 16
        call    fun
        movl    f1(%rip), %edx
        testl   %edx, %edx
        jne     .L8
.L6:
        addq    $8, %rsp
        .cfi_remember_state
        .cfi_def_cfa_offset 8
        jmp     fun
        .p2align 4,,10
        .p2align 3
.L8:
        .cfi_restore_state
        movl    f2(%rip), %eax
        testl   %eax, %eax
        je      .L6
        call    very_heavy
        jmp     .L6
        .cfi_endproc


Which still misses the .section directives and the peephole optimization to
merge movl/testl.

[Bug rtl-optimization/104484] -freorder-block-and-partition not splitting into sections with __builin_expect()

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104484

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
There is some heuristics going on here.  If we mark the function very_heavy as
cold, then GCC does almost the right thing.

[Bug rtl-optimization/104484] -freorder-block-and-partition not splitting into sections with __builin_expect()

[avi at scylladb dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104484

--- Comment #4 from Avi Kivity <avi at scylladb dot com> ---
gcc could infer that the check for f2 and the call to very_heavy are cold, and
move that little block to a .cold section, whether or not very_heavy is cold.