[Mesa-dev] [PATCH v4 17/18] configure.ac: Only add default LLVM components if needed

[Tobias Droste]

Am Donnerstag, 20. Oktober 2016, 18:30:42 CEST schrieb Emil Velikov:
> On 19 October 2016 at 18:55, Tobias Droste <tdroste at gmx.de> wrote:
> > Am Mittwoch, 19. Oktober 2016, 14:17:53 CEST schrieb Emil Velikov:
> >> On 16 October 2016 at 01:20, Tobias Droste <tdroste at gmx.de> wrote:
> >> > LLVM components are only added when LLVM is needed.
> >> > This means gallium adds this as soon as "--enable-gallium-llvm"
> >> > is "yes" and radv + opencl add it explicitly.
> >> > 
> >> > This slightly changes the "auto" value of "--enable-gallium-llvm":
> >> > If no LLVM is installed, it is set to "no". If later drivers need
> >> > this they get the error message that "--enable-gallium-llvm" is
> >> > needed.
> >> > This way if someone is only building softpipe he doesn't need to
> >> > specify "--disable-gallium-llvm" and doesn't have to have LLVM
> >> > installed.
> >> > 
> >> > Signed-off-by: Tobias Droste <tdroste at gmx.de>
> >> > ---
> >> > 
> >> >  configure.ac | 21 +++++++++++++++++----
> >> >  1 file changed, 17 insertions(+), 4 deletions(-)
> >> > 
> >> > diff --git a/configure.ac b/configure.ac
> >> > index bbc1cb5..9e7130ad 100644
> >> > --- a/configure.ac
> >> > +++ b/configure.ac
> >> > @@ -897,7 +897,11 @@ llvm_add_component() {
> >> > 
> >> >      new_llvm_component=$1
> >> >      driver_name=$2
> >> > 
> >> > -    LLVM_COMPONENTS="$LLVM_COMPONENTS $new_llvm_component"
> >> > +    if $LLVM_CONFIG --components | grep -iqw $new_llvm_component ;
> >> > then
> >> > +        LLVM_COMPONENTS="${LLVM_COMPONENTS} ${new_llvm_component}"
> >> > +    else
> >> > +        AC_MSG_ERROR([LLVM component '$new_llvm_component' not enabled
> >> > in
> >> > your LLVM build. Required by $driver_name.]) +    fi
> >> > 
> >> >  }
> >> >  
> >> >  llvm_add_default_components() {
> >> > 
> >> > @@ -980,8 +984,6 @@ llvm_set_environment_variables() {
> >> > 
> >> >              LLVM_VERSION_INT=`echo $LLVM_VERSION | sed -e
> >> >              's/\([[0-9]]\)\.\([[0-9]]\)/\10\2/g'`>
> >> >          
> >> >          fi
> >> > 
> >> > -        llvm_add_default_components "gallium"
> >> > -
> >> > 
> >> >          DEFINES="${DEFINES} -DHAVE_LLVM=0x0$LLVM_VERSION_INT
> >> >          -DMESA_LLVM_VERSION_PATCH=$LLVM_VERSION_PATCH" MESA_LLVM=1
> >> >      
> >> >      else
> >> > 
> >> > @@ -1007,6 +1009,7 @@ radeon_llvm_check() {
> >> > 
> >> >      llvm_add_target $amdgpu_llvm_target_name $4
> >> > 
> >> > +    llvm_add_default_components $4
> >> > 
> >> >      llvm_add_component "bitreader" $4
> >> >      llvm_add_component "ipo" $4
> >> > 
> >> > @@ -2175,6 +2178,7 @@ if test "x$enable_opencl" = xyes; then
> >> > 
> >> >      llvm_check_version_for "3" "6" "0" "opencl"
> >> > 
> >> > +    llvm_add_default_components "opencl"
> >> > 
> >> >      llvm_add_component "all-targets" "opencl"
> >> >      llvm_add_component "linker" "opencl"
> >> >      llvm_add_component "instrumentation" "opencl"
> >> > 
> >> > @@ -2322,9 +2326,18 @@ if test -z "$with_gallium_drivers"; then
> >> > 
> >> >  fi
> >> >  if test "x$enable_gallium_llvm" = xauto; then
> >> >  
> >> >      case "$host_cpu" in
> >> > 
> >> > -    i*86|x86_64|amd64) enable_gallium_llvm=yes;;
> >> > +    i*86|x86_64|amd64)
> >> > +        if test "x$MESA_LLVM" = "x1"; then
> >> > +            enable_gallium_llvm=yes
> >> > +        else
> >> > +            enable_gallium_llvm=no
> >> > +        fi
> >> 
> >> This won't work as expected I'm afraid.
> >> 
> >> For example: if we've failed during the autodetection of llvm,
> >> MESA_LLVM will be 0, thus instead of enabling enable_gallium_llvm [on
> >> supported platforms] and error-ing out you'll get a silent disable.
> >> The latter of which is not what the llvmpipe/r300 folk would
> >> appreciate.
> > 
> > If you only build gallium swrast you would get softpipe instead of
> > llvmpipe, that's true and intended.
> > I think it's better to build softpipe instead of erroring out. If the user
> > specifies --enable-gallium-llvm to make sure he gets llvmpipe this hunk
> > doesn't do anything.
> 
> Atm, enable_gallium_llvm will default to yes on x86 even if you don't
> have LLVM available. Admittedly one can call this heuristics
> "sub-optimal" or worse, but regardless this is not something we should
> address here. This patch will change the behaviour and considering the
> serious objections* on the topic (esp. on the r300 side) things will
> end up being nasty.
> 
> *Devs are/were concerned about user miss-configuration leading to "OMG
> this sh*t is dead slow" reports.

I'm with you on not changing the behaviour for r300 and I assure you it is 
not changed. You cannot build r300 without LLVM (before and after this patch).

But ok, I remove this hunk and will probably need to change something else.
There was a reason for doing this :-D But I can't remember right now.

> 
> Thanks
> Emil