radeon/llvm: ExpandSpecialInstrs - Add support for vector instructions

radeon/llvm: Add R600ExpandSpecialInstrs pass

This pass expends reduction instructions into a MachineInstrBundle that
contains 4 instruction, one for each instruction slot.

radeon/llvm: Add helper function for getting sub reg indices

radeonsi: Handle NULL sampler views getting passed in by the state tracker.

Don't dereference NULL pointers, and if all views are NULL, don't generate an
invalid PM4 packet which locks up the GPU.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

APIspec: Remove cruft about AMD_compressed_???_texture

Mesa doesn't support these extensions, and it seems unlikely that it
ever will

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glFramebufferTexture3D textarget validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glGetShaderiv pname validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glCompressedTexImage border validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glPointSizePointer type validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glGetBufferPointer pname validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glGetVertexAttribPointer pname validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant element type validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant glGetShaderPrecisionFormat shader type validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant depth func validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant stencil op fail/zfail/zpass validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant shade model mode validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant light pname and light validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant hint mode validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant separate stencil face validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant stencil function validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant logic op operand validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant alpha function validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant separate stencil mask face validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant front-face mode validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant face culling mode validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant blend equation mode validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant texture target validation

Mesa doesn't check the parameter passed to glMultiTexCoord*.  It does,
however, mask the texture value to prevent out-of-bounds writes.  This
patch will promote this non-conformant behavior to OpenGL ES 1.  I don't
think anyone will care, and the gets some silly code out of a hot path.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Rearrange placement of GL_TEXTURE_MAX_ANISOTROPY_EXT in APIspec

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

mesa/es: Remove redundant min/mag filter validation

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Brian Paul <brianp@vmware.com>

radeon-llvm: Start multithreaded before using llvm.

This is required to make some of llvm's api calls
thread save. In particular the PassRegistry, which is
implicitly accessed while compiling shader programs.
The PassRegistry uses a mutex that is only active if
the llvm_is_multithreaded() returns true.
Calling llvm_start_multithreading() makes this happen
and by calling this function we try to make sure that
we can savely compile shaders in paralell.
Since there is also a call llvm_stop_multithreading()
in the llvm api, we cannot guarantee that this does
not get switched off while we are relying on this being
set, but for the easier use cases this fixes a race with
the radeon llvm compiler we have as of today.

Signed-off-by: Mathias Froehlich <Mathias.Froehlich@web.de>
Signed-off-by: Tom Stellard <thomas.stellard@amd.com>

r600g: Move common compute/3D register init to its own function

Signed-off-by: Tom Stellard <thomas.stellard@amd.com>

nv50/ir/tgsi: handle DP2 in tgsi Instruction srcMask

Solved by Tiziano Bacocco on IRC.

nv50/ir/emit: don't forget saturation bit on f32 add immediate

Solved by Maxim Levitsky on IRC.

mesa: use #if over #ifdef in the FEATURE_ES1 check to fix a build failure.

mfeatures.h will define FEATURE_ES1 to 0 if it's not defined yet.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=53664

Signed-off-by: Brian Paul <brianp@vmware.com>

st/mesa: fix sampler view counting

In the past, when we called pipe::set_sampler_views(n) the drivers set
samplers [n..MAX] to NULL.  We no longer do that.  The state tracker
code was already trying to set unused sampler views to NULL to cover
that case, but the logic was broken and unnoticed until now.  This patch
fixes it.

Strictly speaking, this patch shouldn't be necessary.  Drivers should simply
ignore unused samplers and sampler views.  But some drivers like llvmpipe (and
others?) count those things and they figure into state validation.  That could
be fixed in the future.

Fixes http://bugs.freedesktop.org/show_bug.cgi?id=53617

Reviewed-by: Marek Olšák <maraeo@gmail.com>

util: update and fix u_upload_mgr.h comments

st/mesa: use Elements() instead of hard-coded number

And add a comment about the velems_util_draw[] array.

mesa: remove unused params, add const qualifiers

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: querying GL_TEXTURE_COMPRESSED_IMAGE_SIZE for a buffer obj is illegal

GL_INVALID_OPERATION is to be raised when querying a non-compressed
image/buffer.  Since a buffer object can't have a compressed format this
query always generates an error.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa/es: Don't generate ES1 type conversion wrappers

These are gradually going to get whittled away and eventually folded into the
source files with the native type functions.

v2: Add (speculative) SConscript changes.  These may be broken.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

i965: Fix bug in the old FS backend's projtex() calculation.

In the old backend, we looked at any FS attribute's proj_attrib_mask bits, not
just texcoords.  Now that we have _mesa_vert_result_to_frag_attrib(), we can
fill in the other FS inputs with correct proj_attrib_mask info.

NOTE: This is a candidate for stable branches.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=46644
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: Support GL_TEXTURE_BUFFER in GetTexLevelParameter[if]v in GL 3.1+.

The OpenGL 3.1 specification explicitly allows this.  Oddly, the
ARB_texture_buffer_object spec's issues section claims this isn't
allowed, but proceeds to explain that the extension simply doesn't edit
the underlying spec to allow it, and thus it didn't appear in the list
of legal texture targets.

Thus, this patch legalizes it only in 3.1+ contexts, but still returns
INVALID_ENUM in earlier contexts that expose ARB_texture_buffer_object.

Unfortunately, the behavior of the call is horrendously undefined.

Fixes oglconform's tbo/negative.textureParams test.

v2: Require desktop OpenGL.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Split out part of glGetTexLevelParameter into a helper function.

Move the _mesa_GetTexLevelParameter[iv] functions below the helper
function so the prototype is available.

This will be useful in the next commit.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Add GL_TEXTURE_CUBE_MAP to _mesa_max_texture_levels(). [v2]

For cube maps, _mesa_generate_mipmap() calls this with
GL_TEXTURE_CUBE_MAP (the gl_texture_object's Target) rather than one
of the faces.  This caused _mesa_max_texture_levels() to return 0, which
resulted in maxLevels == -1 and the next line's assertion to fail.

This function is called from seven places:
- fbobject.c: framebuffer_texture()
- mipmap.c: _mesa_generate_mipmap()
- texgetimage.c:
  - getteximage_error_check()
  - getcompressedteximage_error_check()
- texparam.c: _mesa_GetTexLevelParameteriv()
- texstorage.c: tex_storage_error_check()

All of these (or their callers) now explicitly check for invalid targets
already, so this shouldn't cause invalid targets to slip through.
(Technically _mesa_generate_mipmap() doesn't check for invalid targets,
but the API-facing _mesa_GenerateMipmapEXT() function does.)

+2 oglconforms (float-texture/mipmap.automatic and mipmap.manual)

In addition to fixing the mipmap bug, it should also cause glTexStorage
to accept GL_TEXTURE_CUBE_MAP, which is explicitly allowed by the spec.

v2: Drop alterations to callers; this is now in a patch series that adds
    explicit checking to API functions.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Add explicit target checking to GetTexLevelParameter[if]v().

Previously, it relied on _mesa_max_texture_levels() for texture target
error checking.  This was somewhat dodgy, as _mesa_max_texture_levels()
is called in seven diferent places, not all of which necessarily accept
the same list of targets.

I copied the list of legal targets from _mesa_max_texture_levels(), so
this patch should not introduce any change in behavior.  Future patches
will cause the two to diverge.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Add explicit target checking to Get[Compressed]TexImage().

Previously, they relied on _mesa_max_texture_levels() for texture target
error checking.  This was somewhat dodgy, as _mesa_max_texture_levels()
is called in seven diferent places, not all of which necessarily accept
the same list of targets.

I copied the list of legal targets from _mesa_max_texture_levels() but
removed the proxy targets, as both functions explicitly rejected those
targets.  This changes the order in which we check errors, which could
change whether we return INVALID_VALUE or INVALID_ENUM.  However, it
shouldn't change the list of accepted targets.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

llvmpipe: remove polygon stipple assertion

It's possible for us to have an unused sampler bound when the fragment
shader itself doesn't use any samplers.  So the assertion isn't valid.

Fixes http://bugs.freedesktop.org/show_bug.cgi?id=53616

svga: minor code reformatting

To be consistent with other functions.

build: Remove -shared from OSMesa's LDFLAGS

Would break the static build.

build: Remove EXTRA_LIB_PATH

You can add extra library paths to LDFLAGS directly.

build: Require X11 pkg-config files

r600g: disable tiling for 422 formats again

r600g: fix blits of subsampled formats

r600g: fix copying between NPOT mipmapped compressed textures

We aligned the dimensions to the blocksize, then divided by it
(in r600_blit.c), then minified, which was wrong.

The minification must be done first, not last.
This fixes piglit/fbo-generatemipmap-formats with S3TC and maybe
a bunch of other tests too. Tested on RV730.

r600g: make F2U trans-only on r600-r700

This fixes a failing assertion in r600_asm.c.

r600g: set CB_COLOR_INFO to INVALID for disabled colorbuffers on r600-r700

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

r600g: rename r600_resource_texture to r600_texture

r600g: always put tiled textures in VRAM

r600g: cleanup r600_resource_texture in favor of radeon_surface

r600g: remove unused parameter in r600_texture_create_object

r600g: fixup the usage flag for the flushed depth texture

wayland-drm: close fd after the display is uninitialized

This fixes a "kernel rejected pushbuf: Bad file descriptor" error on
wl_drm display destruction.

Reviewed-by: Kristian Høgsberg <krh@bitplanet.net>

scons: Fix MinGW cross compilation.

Compensate for the recent changes and assumptions added to
Makefiles.sources

radeon/llvm: Lower implicit parameters before ISel

gallium/draw: move misplaced brace

mesa: raise GL_INVALID_OPERATION in glGenerateMipmap for missing base image

This seems to be expected by the WebGL texture-mips test.  The error makes
sense, but I haven't found (yet) any OpenGL documentation specifying this
error condition.

See http://bugs.freedesktop.org/show_bug.cgi?id=44912

Note: This is a candidate for the 8.0 branch.

Reviewed-by: José Fonseca <jfonseca@vmware.com>

r600: update sampler, sampler_view code for the future

For when we have pipe->set_sampler_states(pipe, shader, start, num, samplers),
etc.

Reviewed-by: Marek Olšák <maraeo@gmail.com>

rbug: update data structures, functions for future changes

To support geom/compute/etc shaders, samplers, sampler views, etc.
To support pipe->bind_sampler_states() w/ start_slot.

gallium/trace: add 'start' parameter to bind_sampler_states/views()

gallium/identity: add 'start' parameter to bind_sampler_states/views()

galahad: add 'start' parameter to bind_sampler_states/views()

svga: add 'start' parameter to bind_sampler_states/views()

llvmpipe: add 'start' parameter to bind_sampler_states/views()

softpipe: add 'start' parameter to bind_sampler_states/views()

To support updating a sub-range of sampler states/views in the future.
Note that we always pass start=0 at this time.

gallium/trace: consolidate sampler, sampler_view code

gallium/identity: consolidate sampler, sampler_view code

This will simplify things when the pipe_context functions are consolidated.

st/mesa: add support for GS textures and samplers

st/mesa: combine vertex/fragment sampler state in arrays

As with other recent changes, put the vertex and fragment sampler state
into arrays indexed by the shader type.  This will let us easily add
support for other types of shaders in the future.

gallium: remove PIPE_MAX_VERTEX/GEOMETRY_SAMPLERS #define

PIPE_MAX_SAMPLERS, PIPE_MAX_VERTEX_SAMPLERS and PIPE_MAX_GEOMETRY_SAMPLERS
were all defined to the same value (16).

In various places we're creating arrays such as
sampler_views[PIPE_SHADER_TYPES][PIPE_MAX_SAMPLERS] so we were assuming
the same number of max samplers for all shader stages anyway.

Of course, drivers are still free to advertise different numbers of max
samplers for different shaders.

draw: index samplers and sampler_view state by shader type

So that we can handle GS state and other types of shaders in the future.

draw: move tgsi-related state into a tgsi sub-struct

To better organize things a bit.

gallium: add a shader stage/type param to some draw functions

To prepare for geometry shader texture support in the draw module.
Note: we still only handle the vertex shader case.

st/mesa: silence signed/unsigned comparison warning

svga: move result->key expression after result != NULL check

svga: fix result==NULL logic in emit_fs_consts()

The previous test for result != NULL was kind of bogus since we dereferenced
the pointer earlier in the code.  Now, check for result != NULL first, then
get the result->key info.

Also, remove the useless "offset +=" code at the end.

svga: update comment (s/SVGA_NEW_VS_RESULT/SVGA_NEW_VS_PRESCALE/)

svga: rename svga_hw_vs_parameters -> svga_hw_vs_constants

and similarly for svga_hw_fs_parameters

st/mesa: index can be negative in the PROGRAM_CONSTANT case

NOTE: This is a candidate for the 8.0 branch.

Reviewed-by: Brian Paul <brianp@vmware.com>

mesa: add cast to silence warning in _mesa_pack_rgba_span_from_ints()

meta: remove unused variable

radeonsi: Fix symbol conflicts with r600g.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=50389

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

radeonsi: Fix memory leaks if returning early from some state functions.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

radeonsi: Fix LLVM context leak.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Don't assign virtual address space for BO that already has one.

We'd end up re-using the old one and throwing away the new one anyway, but only
after a roundtrip to the kernel.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Create hole for waste when allocating from va_offset.

Otherwise, the wasted area could never be used for an allocation again.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Fix potential address space loss in radeon_bomgr_force_va().

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Delete uppermost virtual address space hole if it's at the top.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Fix losing holes when allocating virtual address space.

If a hole exactly matches the allocated size plus alignment, we would fail to
preserve the alignment as a hole. This would result in never being able to use
the alignment area for an allocation again.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Merge holes when freeing virtual address space.

Otherwise we'll likely end up with an ever increasing amount of ever smaller
holes.

Requires keeping the list ordered wrt offsets.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: Make va_offset 64 bits wide.

Otherwise we'd wrap around after 32 bits. The kernel currently limits GPU
virtual address space to 4GB anyway, but that will probably change sooner or
later, and this would result in confusing error messages when running out of
virtual address space even now.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>