Re: ARM: Enable scheduling for Thumb-2

[Matthias Klose <doko at ubuntu dot com>]

On 14.11.2009 13:24, Daniel Jacobowitz wrote:
> On Fri, Nov 13, 2009 at 05:24:39PM +0000, Ramana Radhakrishnan wrote:
> > On Tue, 2009-11-10 at 16:11 -0500, Daniel Jacobowitz wrote:
> > > I came across this ??? comment... in my testing, enabling scheduling
> > > for Thumb-2 is size-neutral and has performance benefits ranging from
> > > 1% to 20%.  A couple of performance tests regress, but they are
> > > reasonable problems with non-clairvoyant register allocation
> > > (something was spilled because scheduling increased lifetimes).
> >
> > This broke CSiBe builds for -Os and Thumb2 - I've filed
> > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42031
> >
> > for this. Could you take a look at this ?
>
> Thanks for fixing this, Richard!
>
> I'm surprised this did not show up in my testing...

when applying these two patches to the 4.4 branch (configured --with-mode=thumb, 
the build fails in stage2 with:

/home/doko/gcc/4.4/gcc-4.4-4.4.2/build/./prev-gcc/xgcc 
-B/home/doko/gcc/4.4/gcc-4.4-4.4.2/build/./prev-gcc/ 
-B/usr/arm-linux-gnueabi/bin/ -c  -g -O2 -fno-stack-protector -DIN_GCC   -W 
-Wall -Wwrite-strings -Wstrict-prototypes -Wmissing-prototypes -Wcast-qual 
-Wold-style-definition -Wc++-compat -Wmissing-format-attribute -pedantic 
-Wno-long-long -Wno-variadic-macros -Wno-overlength-strings   -DHAVE_CONFIG_H 
-I. -I. -I../../src/gcc -I../../src/gcc/. -I../../src/gcc/../include 
-I../../src/gcc/../libcpp/include  -I../../src/gcc/../libdecnumber 
-I../../src/gcc/../libdecnumber/dpd -I../libdecnumber -I/include  -I/include 
-DCLOOG_PPL_BACKEND   ../../src/gcc/calls.c -o calls.o
../../src/gcc/calls.c: In function 'emit_library_call_value_1':
../../src/gcc/calls.c:3933: error: unable to find a register to spill in class 
'LO_REGS'
../../src/gcc/calls.c:3933: error: this is the insn:
(insn 856 854 849 94 ../../src/gcc/calls.c:3531 (parallel [
            (set (reg:DI 277 [ D.47483 ])
                (plus:DI (reg:DI 277 [ D.47483 ])
                    (reg:DI 278 [ D.47482 ])))
            (clobber (reg:CC 24 cc))
        ]) 1 {*arm_adddi3} (expr_list:REG_UNUSED (reg:CC 24 cc)
        (nil)))
../../src/gcc/calls.c:3933: confused by earlier errors, bailing out

building with -O1 or with -marm works around the failure. Any hint if an 
additional backport from the trunk is required? I didn't check trunk yet.

  Matthias