[avr-libc-dev] Re: Testresults for [PATCH]: Rewrite AVR backend's rtx_costs (take 3)

[Andy Hutchinson]

Regression seems to be related to bug 18251.

Block move (pattern *movmemqi_insn) is being used to put structure on 
stack - that needs 2 of 3 available pointer registers. Which is a 
challenge when frame p (r28/r29) is the other one mixed up in all of this!

I guessed that it assigned the other one as move count. So I tried a 
quick hack on reduced testcase by forcing counter to avoid pointer 
registers (constraint "l" instead of "r"). Not pretty but it worked.

Now for better fix.....


address@hidden wrote:

> Hi,
>
> I have been testing Roger Sayles patch 
> (http://gcc.gnu.org/ml/gcc-patches/2005-01/msg02224.html) together with the 
> MODES_TIEABLE 1 change suggested by Andy Hutchinson. 
> Tests were run with the simulavr simulator for the atmega128 platform.
>
> Now for the first time head 4.0.0 outperforms 3.4.3 considerably concerning 
> the code size (e.g. 3.4.3 without -fnew-ra about 0x3A12, 3.4.3 with -fnew-ra 
> 0x3712, head 4.0.0 with both changes and -morder1: now 0x361C).
>
> Looking at the testsuite results with the simulator, I have observed a single 
> regression:
>
> Comparison of head 4.0.0 with head 4.0.0 and both, Roger Sayles patch *and* 
> MODES_TIEABLE defined to yield 1 in avr.h:
>
> > Tests that now fail, but worked before:
> >
> > gcc.dg/struct-by-value-1.c (test for excess errors)
>
>
> Considering the code-size reduction, I think that Roger Sayles patch is 
> clearly a win for the avr target. Maybe one could figure out the origin of 
> the one regression. Despite that, I think that the considerable code 
> size-reduction will probably motivate quite some avr-developers to test the 
> new 4.0.0. .
>
> Best regards,
>
> Björn
>
> P.S.:
>
> This is the additional error message from the testsuite log file:
>
> Executing on host: /home/bmh/gnucvs/head/build/gcc/xgcc 
> -B/home/bmh/gnucvs/head/build/gcc/ /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c   
> -O2 -DSTACK_SIZE=400 -DNO_TRAMPOLINES  -DSIGNAL_SUPPRESS 
> -mmcu=atmega128  /home/bmh/dejagnuboards/exit.c  -lm   
> -o ./struct-by-value-1.exe    (timeout = 300)
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c: In 
> function 'testit19':
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: error: 
> unable to find a register to spill in class 'BASE_POINTER_REGS'
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: error: 
> this is the insn:
> (insn 77 53 57 0 (set (reg:QI 24 r24)
>         (mem:QI (plus:HI (reg/f:HI 28 r28)
>                 (const_int 1 [0x1])) [42 S1 A8])) 9 {*movqi} (nil)
>     (nil))
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: 
> internal compiler error: in spill_failure, at reload1.c:1872
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <URL:http://gcc.gnu.org/bugs.html> for instructions.
> compiler exited with status 1
> output is:
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c: In 
> function 'testit19':
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: error: 
> unable to find a register to spill in class 'BASE_POINTER_REGS'
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: error: 
> this is the insn:
> (insn 77 53 57 0 (set (reg:QI 24 r24)
>         (mem:QI (plus:HI (reg/f:HI 28 r28)
>                 (const_int 1 [0x1])) [42 S1 A8])) 9 {*movqi} (nil)
>     (nil))
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: 
> internal compiler error: in spill_failure, at reload1.c:1872
> Please submit a full bug report,
> with preprocessed source if appropriate.
> See <URL:http://gcc.gnu.org/bugs.html> for instructions.
>
> FAIL: gcc.dg/struct-by-value-1.c (test for excess errors)
> Excess errors:
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: error: 
> unable to find a register to spill in class 'BASE_POINTER_REGS'
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: error: 
> this is the insn:
> (insn 77 53 57 0 (set (reg:QI 24 r24)
>         (mem:QI (plus:HI (reg/f:HI 28 r28)
>                 (const_int 1 [0x1])) [42 S1 A8])) 9 {*movqi} (nil)
>     (nil))
> /home/bmh/gnucvs/head/gcc/gcc/testsuite/gcc.dg/struct-by-value-1.c:58: 
> internal compiler error: in spill_failure, at reload1.c:1872
>
> WARNING: gcc.dg/struct-by-value-1.c compilation failed to produce executable