Now that draw depends on llvm it is very difficult to correctly handle
broken llvm installations. Either the user requests LLVM and it needs to
supply a working installation. Or it doesn't, and it gets no LLVM
accelerate pipe drivers.
scons: Control caching via new SCONS_CACHE_DIR environment variable.
This serves several purposes:
- disable caching in situations were is it useless or undesired
- share caches among all trees
- simplify purging the cache (when it's a single location)
- move the cache out of the tree, since that slows downs IDEs considerably
To retain previous behavior just define do
export SCONS_CACHE_DIR=$PWD/build/cache
before invoking scons.
gallium: Provide alternative stdint.h and stdbool.h C99 headers, instead of ad-hoc definitions.
Everybody is using the C99's integer types. Everybody except Microsoft,
which in turns means everybody is including their own definitions of C99
integer types for MSVC, causing duplicate definitions when linking two
projects. This is the case of building Gallium and LLVM with MSVC.
Shipping alternative stdint.h and stdbool.h headers for MSVC allows us
to share a single definition. It also removes clutter from the Gallium
headers.
Some of the demo progams legitimately need the functionality
that's disabled by WIN32_LEAN_AND_MEAN.
Instead the solution should be to define WIN32_LEAN_AND_MEAN just before
including windows.h on a case by case basis.
Strict aliasing tule violations were fixed on master, but
they're still causing problem in this branch, so disable this assumptions.
Do not apply this fix to master (revert when you merge).
-mstackrealign causes stack corruption on MinGW. And without it the ability
to use SSE instrinsics goes down the drain. Even if we use
__attribute__((force_align_arg_pointer)) for the functions we explicitly
use SSE instrinsics, the SSE code automatically generated by gcc will
cause assertion failures. What a nightmare.
Thankfully LLVM gets this right, so all runtime generated SSE code just
works. rtasm code doesn't assume 16byte alignment. Therefore the bulk of
our performance sensitive code is not affected by this.
Still, intrinsics can be convenient, and it would be nice
to get this working again some day, sp will try to get a reduced test
case.
scons: Drop gprof support for profile builds; tweak optimization flags instead.
gprof is useful for shared libraries, hence our drivers. Nevertheless
profilers like oprofile can benefit from disabling some relatively
minor optimizations for more accurate / complete results.
José Fonseca