Dave,
I actually do not know...
You can generate an assembly file -S and -dP options.
You'll see, that every register move being performed to save a value and
clobber this register in the future.
So, in the example below, r8 being used at line 120.
r7 - 169. Instead of using r7, the value nextMux can be saved on stack... but
this is cheaper to do in a register.
r14 - in bittests.
Yes, I agree there are some stupid register moves, but currently I can do
nothing with it.
~d
On Wednesday 18 December 2002 12:00, David Brown wrote:
> Yes - I skipped a fair amount of the code that was in the middle of the
> function, since it would make the code snippet far bigger. But I've now
> made a test example with pretty much the same effects. The exact registers
> used depend on what other variables are declared, and when they are
> declared. This time the bit I'm looking at uses r14 rather than r15, but
> it is still doing so unnecessarily. The code does use r14 in the following
> few lines of assembly, but it could just as easily have used r7.
>
> I get the same code as you for stand-alone lines. This is only an issue
> when the results of the calculation are to be held in a variable and only
> written out later on in the function.
>
> Extract from the attached list file:
>
> 13:t.c **** word nextMux = (mux + 1) & 0x07;
>
> 82 000a 1842 0000 mov &mux, r8 ; mux, nextMux
> 83 000e 0E48 mov r8, r14 ; nextMux
> 84 0010 1E53 add #llo(1), r14
> 85 0012 074E mov r14, r7 ; nextMux
> 86 0014 37F0 0700 and #llo(7), r7 ; nextMux
>
> mvh.
>
> David
>
> > Something is _really_ missing here.
> > gcc will never use registers below r12 for such a code.
> >
> >
> > The code will be:
> >
> > mov &muxSelect, r15
> > add #llo(1), r15
> > and #llo(7), r15
> > mov r15, &muxSelect
> > ret
> >
> > However, slightly better result can be obtained if you'll write:
> >
> > muxSelect++;
> > muxSelect &= 0x07;
> >
> > ~d
> >
> > On Tuesday 17 December 2002 17:35, David Brown wrote:
> > > Hi,
> > >
> > > I am looking at the code generated for a function, and I feel there is
> > > often unnecessary movement of data between registers, especially r15.
>
> Is
>
> > > there some reason while small calcuations always seem to use r15? I'm
> > > compiling with -O2.
> > >
> > > In this particular code snippet, I have a variable "muxSelect" defined
>
> as a
>
> > > "word" (typedef for unsigned int). The code includes:
> > >
> > > void foo(void) {
> > > word i, j;
> > > word nextMux = (muxSelect + 1) & 0x07;
> > >
> > > // More code, using nextMux and muxSelect
> > >
> > > muxSelect = nextMux;
> > > }
> > >
> > > The "nextMux = (muxSelect + 1) & 0x07" line is compiled to:
> > > mov &muxSelect, r9 // Puts muxSelect in a register,
>
> as
>
> > > it will be used later
> > > mov r9, r15
> > > add #llo(1), r15
> > > mov r15, r7 // Why go through r15, and
>
> not
>
> > > just use r7 directly?
> > > and #llo(7), r7
> > >
> > > I've seen this sort of thing on several occasions - r15 is used as a
> > > scratch register, when the original calculation could have been done
> > > directly in the destination register. Am I missing something here, or
>
> is
>
> > > there the possibility of optomising this more?
> > >
> > > Just for fun, I added the "-frename-registers" flag to the compile.
> > > The result was that r8 was used instead of r15, but basically the same
> > > code
>
> was
>
> > > produced.
> > >
> > > mvh.
> > >
> > > David
> > >
> > >
> > >
> > >
> > > -------------------------------------------------------
> > > This sf.net email is sponsored by:
> > > With Great Power, Comes Great Responsibility
> > > Learn to use your power at OSDN's High Performance Computing Channel
> > > http://hpc.devchannel.org/
> > > _______________________________________________
> > > Mspgcc-users mailing list
> > > Mspgcc-users@lists.sourceforge.net
> > > https://lists.sourceforge.net/lists/listinfo/mspgcc-users
> >
> > -------------------------------------------------------
> > This sf.net email is sponsored by:
> > With Great Power, Comes Great Responsibility
> > Learn to use your power at OSDN's High Performance Computing Channel
> > http://hpc.devchannel.org/
> > _______________________________________________
> > Mspgcc-users mailing list
> > Mspgcc-users@lists.sourceforge.net
> > https://lists.sourceforge.net/lists/listinfo/mspgcc-users
--
/********************************************************************
("`-''-/").___..--''"`-._ (\ Dimmy the Wild UA1ACZ
`6_ 6 ) `-. ( ).`-.__.`) Enterprise Information Sys
(_Y_.)' ._ ) `._ `. ``-..-' Nevsky prospekt, 20 / 44
_..`--'_..-_/ /--'_.' ,' Saint Petersburg, Russia
(il),-'' (li),' ((!.-' +7 (812) 3468202, 5585314
********************************************************************/
