[Bug c++/35393] Deep templates get 40x slower when depth goes from 17 to 18.

[jacob at math dot jussieu dot fr]

------- Comment #3 from jacob at math dot jussieu dot fr  2008-02-27 15:49 -------
Thank you for the quick reply. I'm glad to hear that it's fixed and much
improved in 4.3.

I tried what you said about __attribute__((flatten)) (I copied and pasted your
code) and that didn't change the performance at all (with 4.2.3). I compiled
with

g++ -O3 -DDEPTH18 test.cpp  -S -o test.s

and the generated assembly file starts as follows:

        .file   "test.cpp"
        .section        .ctors,"aw",@progbits
        .align 4
        .long   _GLOBAL__I__Z8evaluatev
        .text
        .align 2
        .p2align 4,,15
.globl _Z8evaluatev
        .type   _Z8evaluatev, @function
_Z8evaluatev:
.LFB1436:
        pushl   %ebp
.LCFI0:
        movl    %esp, %ebp
.LCFI1:
        pushl   %edi
.LCFI2:
        pushl   %esi
.LCFI3:
        xorl    %esi, %esi
        pushl   %ebx
.LCFI4:
        xorl    %ebx, %ebx
        subl    $172, %esp
.LCFI5:
        leal    -84(%ebp), %edi
        .p2align 4,,7
.L2:
        leal    -160(%ebp), %eax
        movl    %ebx, -16(%ebp)
        movl    %ebx, -20(%ebp)
        movl    %ebx, -24(%ebp)
        movl    %ebx, -28(%ebp)
        movl    %ebx, -32(%ebp)
        movl    %ebx, -36(%ebp)
        movl    %ebx, -40(%ebp)
        movl    %ebx, -44(%ebp)
        movl    %ebx, -48(%ebp)
        movl    %ebx, -52(%ebp)
        movl    %ebx, -56(%ebp)
        movl    %ebx, -60(%ebp)
        movl    %ebx, -64(%ebp)
        movl    %ebx, -68(%ebp)
        movl    %ebx, -72(%ebp)
        movl    %ebx, -76(%ebp)
        movl    %ebx, -80(%ebp)
        movl    %ebx, -84(%ebp)
        movl    $72, 8(%esp)
        movl    %edi, 4(%esp)
        movl    %eax, (%esp)
        call    memcpy
        movl    -92(%ebp), %eax
        addl    -96(%ebp), %eax
        addl    -100(%ebp), %eax
        addl    -104(%ebp), %eax
        addl    -108(%ebp), %eax
        addl    -112(%ebp), %eax
        addl    -116(%ebp), %eax
        addl    -120(%ebp), %eax
        addl    -124(%ebp), %eax
        addl    -128(%ebp), %eax
        addl    -132(%ebp), %eax
        addl    -136(%ebp), %eax
        addl    -140(%ebp), %eax
        addl    -144(%ebp), %eax
        addl    -148(%ebp), %eax
        addl    -152(%ebp), %eax
        addl    -156(%ebp), %eax
        addl    -160(%ebp), %eax
        addl    %esi, %eax
        leal    (%eax,%ebx), %esi
        addl    $1, %ebx
        cmpl    $1000000000, %ebx
        jne     .L2
        addl    $172, %esp
        movl    %esi, %eax
        popl    %ebx
        popl    %esi
        popl    %edi
        popl    %ebp
        ret
.LFE1436:
        .size   _Z8evaluatev, .-_Z8evaluatev

As you can see, there is a call to memcpy() which isn't present in your
version. What can I do (if anything) to improve speed with g++ 4.2.3?

Also, isn't it a bit surprising to have a call in a function that has
__attribute__((flatten)) ? The compilation doesn't produce any warning.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35393