https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102906

--- Comment #6 from Christophe Lyon <clyon at gcc dot gnu.org> ---
(In reply to Aldy Hernandez from comment #5)
> (In reply to Christophe Lyon from comment #4)
> > If I'm not mistaken, if you click on "REGRESSED" for that specific line,
> > you'll see that only ssa-dom-thread-7.c actually regressed with the
> > corresponding flags.
> > 
> > For ivopts-4.c, if seems you need -mthumb and one of these values for 
> > -march:
> > armv7-a+mp+sec+neon-fp16
> > armv7ve+simd
> > armv7-a+mp+sec+vfpv3-d16-fp16
> > armv8-a+crc+simd+crypto
> > (possibly with -mfpu=auto)
> > or -march=armv8-a -mfpu=crypto-neon-fp-armv8
> 
> Works for:
> -Os -fdump-tree-ivopts -save-temps -S  -march=armv8-a
> -mfpu=crypto-neon-fp-armv8
> -Os -fdump-tree-ivopts -save-temps -S -mthumb -march=armv8-a
> -mfpu=crypto-neon-fp-armv8
> -Os -fdump-tree-ivopts -save-temps -S -mthumb -march=armv7ve+simd -mfpu=auto
> 
> Similarly for ssa-dom-thread-7.c.
> 
> Perhaps someone else can reproduce.

It fails for me with:
cc1 -fpreprocessed ivopts-4.i -quiet -dumpbase ivopts-4.c -dumpbase-ext .c
-mthumb -mfloat-abi=hard -mfpu=neon -mtls-dialect=gnu
-mlibarch=armv7-a+mp+sec+neon-fp16 -march=armv7-a+mp+sec+neon-fp16 -Os -version
-fdiagnostics-color=never -fno-diagnostics-show-caret
-fno-diagnostics-show-line-numbers -fdiagnostics-urls=never
-fdiagnostics-path-format=separate-events -fdump-tree-ivopts -o ivopts-4.s


the main compilation line is:
arm-none-linux-gnueabihf-gcc -mthumb -march=armv7-a+mp+sec+neon-fp16
-fdiagnostics-plain-output -Os -fdump-tree-ivopts -save-temps -c -o ivopts-4.o
/home/christophe.lyon/src/GCC/sources/gcc-fsf-git/fixes/gcc/testsuite/gcc.target/arm/ivopts-4.c
then size ivopts-4.o:
   text    data     bss     dec     hex filename
     38       0       0      38      26 ivopts-4.o
where the testcase expects text <= 36

My GCC is configured with --with-float=hard --with-cpu=cortex-a9
--with-fpu=neon --with-mode=arm

>
> And just to be clear, it would be easier if someone could reproduce with a
> cc1 / gcc command line, and not by pointing us to some table in a web page.

Sorry, these automated validations builds are deleted after the reports are
generated,
so it's not really easier for me to manually reproduce than for anyone else.

Reply via email to