Re: testing for signed char / short / int addition overflow

[Paul Eggert]

On 3/14/21 7:59 PM, Bruno Haible wrote:
> The functions my_signed1_overflow, my_signed2_overflow
> in the attached file produce better machine code than the corresponding
> functions signed1_overflow, signed2_overflow that use intprops.h primitives.

That wasn't true on my platform (Ubuntu 20.10 x86-64, GCC 
10.2.0-13ubuntu1 with -O2). But I expect that doesn't matter much since 
the signed1_overflow and signed2_overflow functions were probably not 
what you wanted. For example, signed1_overflow simply did this:

        endbr64
        xorl    %eax, %eax
        ret

That is, it always returned 'false', which is a correct implementation 
of the macro since adding two signed char values cannot overflow on my 
platform.

I expect you meant something more like the attached foo1.c. On my 
platform, its signed1_overflow seems more efficient than 
my_signed1_overflow as it is 6 rather than 9 insns (no branches in 
either case). In contrast, signed2_overflow (typically 8 insns with one 
typically-not-taken conditional branch) seems in the same ballparck as 
as my_signed2_overflow (9 insns, no branches). And signed4_overflow (4 
insns, no branches) is definitely faster than my_signed4_overflow (9 
insns, no branches).

If I switch to clang 11.0.0-2 on the same platform (which omits a check 
that GCC includes), the numbers are:

signed1_overflow: 3 insns
my_signed1_overflow: 6 insns
signed2_overflow: 3 insns
my_signed2_overflow: 6 insns
signed4_overflow: 3 insns
my_signed4_overflow: 8 insns

so here, intprops.h seems the clear winner overall.

I'd say that for these platforms, intprops.h is OK as-is. Perhaps on 
other compilers it could be improved significantly, but I'm not sure how 
much we should worry about optimizing for compilers other than GCC.

I installed the attached commentary patch to intprops.h to try to clear 
up some of the confusion you mentioned in this and your other email. If 
the GCC-vs-clang inefficiency is of concern perhaps someone could file a 
GCC bug report.

0001-intprops-improve-commentary.patch
Description: Text Data