[Bug c/108255] New: Repeated address-of (lea) not optimized for size.

[Bug c/108255] New: Repeated address-of (lea) not optimized for size.

[witold.baryluk+gcc at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108255

            Bug ID: 108255
           Summary: Repeated address-of (lea) not optimized for size.
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: witold.baryluk+gcc at gmail dot com
  Target Milestone: ---

https://godbolt.org/z/q5sx9e49j


void f(int *);

int g(int of) {
    int x = 13;
    f(&x);
    f(&x);
    f(&x);
    f(&x);
    f(&x);
    f(&x);
    f(&x);
    f(&x);
    return 0;
}


Got:

g(int):
        sub     rsp, 24
        lea     rdi, [rsp+12]
        mov     DWORD PTR [rsp+12], 13
        call    f(int*)
        lea     rdi, [rsp+12]             # compute, 5 bytes
        call    f(int*)
        lea     rdi, [rsp+12]             # recompute, 5 bytes
        call    f(int*)
        lea     rdi, [rsp+12]             # recompute, 5 bytes
        call    f(int*)
        lea     rdi, [rsp+12]
        call    f(int*)
        lea     rdi, [rsp+12]
        call    f(int*)
        lea     rdi, [rsp+12]
        call    f(int*)
        lea     rdi, [rsp+12]
        call    f(int*)
        xor     eax, eax
        add     rsp, 24
        ret


But, note that lea is 5 bytes.

Expected (generated by clang 3.0 - 15.0):

g(int):                                  # @g(int)
        push    rbx                              # extra, but just 1 byte
        sub     rsp, 16
        mov     dword ptr [rsp + 12], 13         # CSE temp
        lea     rbx, [rsp + 12]
        mov     rdi, rbx                         # use
        call    f(int*)@PLT
        mov     rdi, rbx                         # reuse, 3 bytes
        call    f(int*)@PLT
        mov     rdi, rbx                         # reuse, 3 bytes
        call    f(int*)@PLT
        mov     rdi, rbx
        call    f(int*)@PLT
        mov     rdi, rbx
        call    f(int*)@PLT
        mov     rdi, rbx
        call    f(int*)@PLT
        mov     rdi, rbx
        call    f(int*)@PLT
        mov     rdi, rbx
        call    f(int*)@PLT
        xor     eax, eax
        add     rsp, 16
        pop     rbx                          # extra, but just 1 byte
        ret


Technically this is more instructions.

But

mov rdi, rbx is 3 bytes, which is shorter than 5 bytes of lea. This is at minor
expense of needing to save and restore rbx.

PS. Same happens when using temporary `int *const y = &x;`

Also same when optimizing for size (`-Os`).

It looks like gcc 4.8.5 produced expected code, but gcc 4.9.0 does not.

It is possible that the code produced by gcc 4.9.0 is faster, but it is also
likely it contributes quite a bit to binary size.

clang uses CSE even if there are even just two uses of `&x` in the above
example. It is likely a bit higher threshold is (3 or 4) is actually optimal
(can be calculated knowing encoding sizes).


Weirdly tho, gcc -m32 does this:

g():
        push    ebp
        mov     ebp, esp
        push    ebx
        lea     ebx, [ebp-12]
        sub     esp, 32
        mov     DWORD PTR [ebp-12], 13
        push    ebx
        call    f(int*)
        mov     DWORD PTR [esp], ebx
        call    f(int*)
        mov     DWORD PTR [esp], ebx
        call    f(int*)
        mov     ebx, DWORD PTR [ebp-4]
        xor     eax, eax
        leave
        ret

Where, it does compute address and stores it in temporary. But does it on a
stack, instead in a register (my guess is there are no free register to store
it and it is spilled)., but in fact lea here would be likely faster (mov    
DWORD PTR [esp], ebx, but requires memory/cache access, lea is 5 bytes, but
does not require memory access)

[Bug target/108255] Repeated address-of (lea) not optimized for size.

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108255

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I suspect r0-127773-g3e7291458b96 changed the behavior for GCC 4.9+
I have not figured out what changed the behavior for GCC 4.8 yet though.

I suspect it was just a mistake that GCC 4.8 cost model was incorrect really.

LLVM might be not tuning correctly anyways ...

Also note ICC (not ICX) does the same as GCC ...
So I think this is just a LLVM issue rather than a GCC issue.

Someone who knows more about the x86 processors behavior can explain more.