[gfortran,patch] Make MIN and MAX handle NaNs correctly (PR32048)

[Tim Prince]

pinskia@gmail.com wrote:
> On 7/28/07, Tim Prince <tprince@myrealbox.com> wrote:
>> FX Coudert wrote:
>> Does this affect minval() and maxval() intrinsics?  Up to now, they
>> performed as well as equivalent C code with icpc, better than g++, but
>> not as well as ifort.  I suppose it removes the prospect of vectorizing
>> code with min() and max(), unless that will be enabled with -ffast-math.
> 
> Actually the G++ code should be better on the trunk than before.  If
> it is not, can you file a new bug?
> 

I did find several cases where g++ x86-64 performance improved greatly 
in the last month, but this is not one of them.
I didn't find anything clear to complain about.  g++ is generating a lot 
of useless (on Core 2) p2align, but that does not correlate with 
performance.  g++ uses minps memory to register operations, while 
gfortran loads each operand first.  This looks like a way for g++ to 
favor reduced code size over performance (somewhat contradictory, after 
consuming space with p2align).  It might be better to unroll less and 
pre-load the operands.  I didn't think this is worth a PR, when I have 
plenty of long-standing PRs.