[cfe-dev] Even abs() comes with a performance penalty

[Stefan Kanthak via cfe-dev]

"Craig Topper" <craig.topper at gmail.com> wrote:

> cmov has been 1 cycle since Sandy Bridge.

That doesn't matter here. It's the data dependency it introduces.

> Moves execute in the register renamer since Ivy Bridge.

That's why my code shows 2 of them.

> So mov+neg+cmov should be faster than cdq+add+xor on modern CPUs.

You but forgot sbb+ test, and the data dependency: how well does the CPU
speculate about cmovs?

> Furthermore, cdq really ties the hands of the register allocator so probably
> doesn't make sense in a larger function with abs mixed with other code.

The optimiser is free to use mov+sar then, at the expense of +4 or +6 bytes.

Ever heard of trade-off?

Stefan

On Sun, Sep 6, 2020 at 12:30 PM Stefan Kanthak via cfe-dev <
cfe-dev at lists.llvm.org> wrote:

> --- bugs-bunny.c ---
> // Copyleft © 2014-2020, Stefan Kanthak <stefan.kanthak at nexgo.de>
>
> #ifdef __amd64__
> __int128_t __absti2(__int128_t argument) {
>     return argument < 0 ? -argument : argument;
> }
> #else
> long long __absdi2(long long argument) {
> #ifdef BUNNY
>     return __builtin_llabs(argument);
> #else
>     return argument < 0 ? -argument : argument;
> #endif // BUNNY
> }
>
> long __abssi2(long argument) {
> #ifdef BUNNY
>     return __builtin_labs(argument);
> #else
>     return argument < 0 ? -argument : argument;
> #endif // BUNNY
> }
> #endif // __amd64__
> --- EOF ---
>
> Run clang -c -o- -O3 -S -target amd64-pc-linux bugs-bunny.c
>
> Left: inperformant original code # right: proper code,
>                                  #        faster and 3 bytes shorter
>
> __absti2:      # @__absti2
> # %bb.0:                         # .intel_syntax noprefix
>       xorl     %edx, %edx        #        mov    rax, rsi
>       movq     %rdi, %rax        #        cqo
>       negq     %rax              #        mov    rax, rdx
>       sbbq     %rsi, %rdx        #        add    rdi, rdx
>       testq    %rsi, %rsi        #        adc    rsi, rdx
>       cmovnsq  %rdi, %rax        #        xor    rax, rdi
>       cmovnsq  %rsi, %rdx        #        xor    rdx, rsi
>       retq                       #        ret
>
> CMOVcc introduces a data dependency here, WITHOUT necessity!
>
>
> Run clang -c -o- -O3 -S -target i386-pc-linux bugs-bunny.c
>
> Left: inperformant original code # right: proper code, runs even on real
>                                  #        i386, not just PentiumPro+
>
> ___abssi2:    # @__abssi2
> # %bb.0:                         # .intel_syntax noprefix
>       movl    4(%esp), %ecx      #        mov    eax, [esp+4]
>       movl    %ecx, %eax         #        cdq
>       negl    %eax               #        add    eax, edx
>       cmovll  %ecx, %eax         #        xor    eax, edx
>       retl                       #        ret
>
>
> Writing shorter code for __absdi2() for i386 is left as an
> exercise to the reader.
>
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>