Re: PR 15492: floating-point arguments are loaded too early to x87stack

[Uros Bizjak <uros at kss-loka dot si>]

Florian Weimer wrote:

> On modern x86 CPUs, fxch is executed at instruction decoding time by
> renaming floating point registers.  It only costs execution time if
> the instruction decoder cannot keep up with the remaining pipeline (or
> if the working set exceeds the size of the processor's trace cache, if
> there is one).
>  
fxch is only part of problem. Another problem is, that more fp-stack 
registers are used in -O2 case. Consider another example from PR 15492:

double test1 (double a, int x, double b, int y, double c)
{
       return sin (c) + tan (b) / sqrt (a) + x * fabs (b) + y;
}

test1 ( with -O2 ):
       fldl    4(%esp)        #st(0)
       fsqrt
       fldl    16(%esp)    #st(0) st(1)
       fld     %st(0)        #st(0) st(1) st(2)
       fldl    28(%esp)    #st(0) st(1) st(2) st(3)
       fxch    %st(2)
       fabs
       fxch    %st(3)
       fdivrl  .LC1        #st(0) st(1) st(2) st(3)
       fxch    %st(1)
       fptan            #st(0) st(1) st(2) st(3) st(4)
       fstp    %st(0)        #st(0) st(1) st(2) st(3)
       fxch    %st(2)
       fsin            #st(0) st(1) st(2) st(3)
       fxch    %st(2)
       fmulp   %st, %st(1)    #st(0) st(1) st(2)
       faddp   %st, %st(1)    #st(0) st(1)
       fildl   12(%esp)    #st(0) st(1) st(2)
       fmulp   %st, %st(2)    #st(0) st(1)
       faddp   %st, %st(1)    #st(0)
       fildl   24(%esp)    #st(0) st(1)
       faddp   %st, %st(1)    #st(0)
       ret

test1 ( without -O2 ):
       fldl    28(%esp)    #st(0)
       fsin
       fldl    16(%esp)    #st(0) st(1)
       fptan            #st(0) st(1) st(2)
       fstp    %st(0)        #st(0) st(1)
       fldl    4(%esp)        #st(0) st(1) st(2)
       fsqrt
       fdivrp  %st, %st(1)    #st(0) st(1)
       faddp   %st, %st(1)    #st(0)
       fildl   12(%esp)    #st(0) st(1)
       fldl    16(%esp)    #st(0) st(1) st(2)
       fabs
       fmulp   %st, %st(1)    #st(0) st(1)
       faddp   %st, %st(1)    #st(0)
       fildl   24(%esp)    #st(0) st(1)
       faddp   %st, %st(1)    #st(0)
       ret

If functions are more complex, more stack space is wasted. There was no 
problem with fadd, fsub, fmul and fdiv insn, because they can operate 
with one of their arguments off stack. OTOH, arguments to fsin, fcos, 
etc are always loaded at the beginning of the function. In optimized 
(-O2) case, argument to fsin insn is filling up fp-stack slot for half 
of the function calculation, but it could be loaded just in front of 
fsin insn.

Perhaps this problem could be solved by introducing an expander for fsin 
& co. in i386.md to expand "fsin (memory)" pattern into "arg load; sin" 
sequence? Then combiner could combine "arg load; sin" sequence into 
"*fsin_m" pattern when appropriate and this new expander would expand 
this pseudo-insn back to "fldl;fsin". This would IMHO move a load in 
front of fsin when appropriate.

Uros.