glsl/ast: Stop processing a switch-statement after an error in the init-expression

This happens to work now because ir_binop_all_equal is used.  This
causes vector typed init-expressions to produce scalar Boolean values
after comparison.

The next commit changes ir_binop_all_equal to ir_binop_equal.  Vector
typed init-expressions will then produce vector Boolean values, and, in
debug builds, the ir_assignment constructor will fail an assertion.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>

glsl: Don't pass NULL to ir_assignment constructor when not necessary

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>

glsl: Convert lower_variable_index_to_cond_assign to ir_builder

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>

glsl: Fix coding standards issues in lower_variable_index_to_cond_assign

Mostly tabs-before-spaces, but there was some other trivium too.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>

glsl: Convert lower_vec_index_to_cond_assign to using ir_builder

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>

glsl: Return ir_variable from compare_index_block

This is basically a wash now, but it simplifies later patches that
convert to using ir_builder.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>

glsl: Fix coding standards issues in lower_vec_index_to_cond_assign

Mostly tabs-before-spaces, but there was some other trivium too.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>

glsl: Fix coding standards issues in lower_if_to_cond_assign

Mostly tabs-before-spaces issues.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>

nir/spirv: Allow loop breaks in a switch body.

Per the SPIR-V spec 2.11 Structured Control Flow:

"The only blocks in a construct that can branch outside the construct are

...
- a break block for the innermost loop it is inside of.
..."

With

"Break block: A block containing a branch to the Merge Block of a loop header's merge instruction."

Note that it puts no restriction on not being in an if or switch within the innermost loop.

This passes the loop_break block to the switch body so it can properly detect loop breaks.

CC: <mesa-stable@lists.freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

freedreno/a5xx: fix missing restore state

RB_CLEAR_CNTL seems to be in a funny state after boot (at least on
8x96/a530).

Cc: 17.2 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>

radv: make radv_dynamic_state_copy() static

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

meson: change vulkan icd config to - instead of _

Just to be consistent.

v2: - update meson.build too
v3: - remove unrelated whitespace change

Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>

meson: check for python2 mako

Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>

docs: update calendar, add news item and link release notes for 17.2.2

Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>

docs: add sha256 checksums for 17.2.2

Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
(cherry picked from commit 5a71ed6fa5b78f04b29e972e0759fa15cf0247b2)

docs: add release notes for 17.2.2

Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
(cherry picked from commit bc12538a8e79113b733381ffdc0f6c89d59d0a50)

wayland-egl: rework and simplify wl_egl_window initialization

Use calloc instead of malloc + explicitly zeroing the different fields.
We need special handling for the version field which is of type
const intptr_t.

As we're here document why keeping the constness is a good idea.

The wl_egl_window_resize() call is replaced with an explicit set of the
width/height.

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

wayland-egl: move WL_EGL_EXPORT declaration to where it's used

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

wayland-egl: use C99 comments

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

wayland-egl: remove no longer needed wayland-client dependency

Was required for wl_surface, which is opaque and forward declared with
earlier patch.

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

wayland-egl: add stdint.h include for intptr_t

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

wayland-egl: forward declare struct wl_surface

It makes the header self-contained and with later commit we'll remove
the unnecessary wayland-client.h include.

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

wayland-egl: rename wayland-egl-{priv,backend}.h

In preparation to lifting the whole thing out as a separate library.

Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Reviewed-by: Miguel A. Vico <mvicomoya@nvidia.com>

egl/dri: link directly to libglapi.so

Shared glapi (libglapi.so) has been a requirement for years, in order
to build EGL.

Remove the no longer necessary dlopen/dlsym dance and link to the
library directly.

This allows us to remove a handful of platform specific workarounds, due
to the different name of the library.

v2:
 - Android: export the include dir (RobH)
 - Drop unused local variable (Eric)

Cc: Jonathan Gray <jsg@jsg.id.au>
Cc: Jon Turney <jon.turney@dronecode.org.uk>
Cc: Julien Isorce <julien.isorce@gmail.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com> (v1)
Tested-by: Tomasz Figa <tfiga@chromium.org> (v1)
Tested-by: Rob Herring <robh@kernel.org>

swr/rast: do not crash on NULL strings returned by getenv

The current convenience function GetEnv feeds the results of getenv
directly into std::string(). That is a bad idea, since the variable
may be unset, thus we feed NULL into the C++ construct.

The latter of which is not allowed and leads to a crash.

v2: Better variable name, implicit char* -> std::string conversion (Eric)

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101832
Fixes: a25093de718 ("swr/rast: Implement JIT shader caching to disk")
Cc: Tim Rowley <timothy.o.rowley@intel.com>
Cc: Laurent Carlier <lordheavym@gmail.com>
Cc: Bernhard Rosenkraenzer <bero@lindev.ch>
[Emil Velikov: make an actual commit from the misc diff]
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com> (v1)
Reviewed-by: Laurent Carlier <lordheavym@gmail.com> (v1)

freedreno/a5xx: align height to GMEM

Similar to the way width/pitch alignment works, it seems like we need to
do similar for height.  Otherwise the BLIT from system memory to GMEM
can over-fetch beyond the end of the buffer, triggering a fault.

I'm not sure if there is a better solution yet.  Possibly we could fall
back to pre-a5xx style DRAW packets for cases where BLIT might over-
fetch.  (We in theory have that problem already with rendering to higher
mipmap levels, although fortunately those tend to use GMEM bypass.)

This fixes issues reported with glamor.

Reported-by: don.harbin@linaro.org
Cc: 17.2 <mesa-stable@lists.freedesktop.org>
Signed-off-by: Rob Clark <robdclark@gmail.com>

radeonsi: adjust clip discard based on line width / point size

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: remove si_context::{scissor_enabled,clip_halfz}

They are just copies of the rasterizer state.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: simplify the signature of si_update_vs_writes_viewport_index

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: move current_rast_prim into si_context

v2: rebase fixes

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: move and rename scissor and viewport state and functions

v2: change GET_MAX_SCISSOR to SI_MAX_SCISSOR

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: remove si_apply_scissor_bug_workaround

It only affects pre-SI chips.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: move r600_viewport.c to si_viewport.c

This is purely a file-move + #include fixup + build system changes.
Other cleanups will follow in subsequent commits.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: fix maximum advertised point size / line width

The hardware registers store the half-size/width in 12.4 fixed point
format, so 8192 is the maximum.

Fixes dEQP-GLES3.functional.rasterization.*

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: deduce rast_prim correctly for tessellation point mode

Together with the previous patches, this fixes
dEQP-GLES31.functional.primitive_bounding_box.wide_points.*

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: don't discard points and lines

This is a bit conservative, but a more precise solution requires access
to the rasterizer state. This is something to tackle after the fork between
r600 and radeonsi.

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: move current_rast_prim to r600_common_context

We'll use it in the scissors / clip / guardband state.

v2: avoid a performance regression on r600 when applied to
    (pre-fork) stable branches

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

st/mesa: use R10G10B10X2 format where applicable

This is the last step of fixing
dEQP-GLES3.functional.fbo.completeness.renderable.texture.color0.rgb_unsigned_int_2_10_10_10_rev
for radeonsi.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium: add PIPE_FORMAT_R10G10B10X2_UNORM

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

mesa/main: R10G10B10_(A2) formats are not color renderable in ES

The EXT_texture_type_2_10_10_10_REV (ES only) states the following issue:

   "1. Should textures specified with this type be renderable?

    UNRESOLVED: No.  A separate extension could provide this functionality."

This partially fixes
dEQP-GLES3.functional.fbo.completeness.renderable.texture.color0.{rgb,rgba}_unsigned_int_2_10_10_10_rev

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

mesa/main: select the R10G10B10X2_UNORM internal format based on data type

ES requires it. This is a partial fix for
dEQP-GLES3.functional.fbo.completeness.renderable.texture.color0.rgb_unsigned_int_2_10_10_10_rev

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

glsl: do not set the 'smooth' qualifier by default on ES shaders

It leads to surprising states with integer inputs and outputs on
vertex processing stages (e.g. geometry stages). Instead, rely on the
driver to choose smooth interpolation by default.

We still allow varyings to match when one stage declares it as smooth
and the other declares it without interpolation qualifiers.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

freedreno: fix PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE

Fixes an assert in fd_acc_query_register_provider() about query provider
not already registered.

Fixes: 3f6b3d9d ("gallium: add PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE")
Signed-off-by: Rob Clark <robdclark@gmail.com>

egl/wayland: simplify LIBGL_ALWAYS_SOFTWARE logic

Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>

radeonsi: fix a regression in integer cube map handling

A recent commit fixed the case of 8888 integer cube maps, which need the
workaround of replacing the data format with USCALED/SSCALED. However,
this broke the case of non-8888 integer cube maps; those still need the
fix of shifting the texture coordinates.

Fixes KHR-GL45.texture_gather.plain-gather-int-cube-array and similar.

Fixes: 6fb0c1013b35 ("radeonsi: workaround for gather4 on integer cube maps")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

amd/common: move ac_build_phi from radeonsi

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radv: remove unused radv_meta_state::btoi::render_pass handle

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: do not check the number of levels when doing fast htile

We shouldn't reach this point because HTILE is only enabled
when the number of levels is 1.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: cleanup radv_device_finish_meta_XXX() helpers

Unnecessary to double check that handles are not NULL.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: select the pipeline outside of emit_fast_clear_flush()

It can't change during the decompression pass.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: drop useless param in emit_depth_decomp()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: drop useless check in depth_view_can_fast_clear()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: add radv_subpass_clear_attachment() helper

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: add radv_attachment_needs_clear() helper

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: remove unused param in radv_handle_{cmask,dcc}_image_transition()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: add radv_vi_dcc_enabled() helper

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: do not need to double zero-init the meta state structures

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: inline destroy_render_pass()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: use pipeline handles instead of objects for meta clear operations

To be consistent with other meta operations.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: inline blit2d_unbind_dst()

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

radv: rework DCC/CMASK/FMASK/HTILE allocations

Add helpers and some comments to make the thing more readable.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

meson: fix version typo + grammar

Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>

i965: skip reading unused slots at the begining of the URB for the FS

We can start reading the URB at the first offset that contains varyings
that are actually read in the URB. We still need to make sure that we
read at least one varying to honor hardware requirements.

This helps alleviate a problem introduced with 99df02ca26f61 for
separate shader objects: without separate shader objects we assign
locations sequentially, however, since that commit we have changed the
method for SSO so that the VUE slot assigned depends on the number of
builtin slots plus the location assigned to the varying. This fixed
layout is intended to help SSO programs by avoiding on-the-fly recompiles
when swapping out shaders, however, it also means that if a varying uses
a large location number close to the maximum allowed by the SF/FS units
(31), then the offset introduced by the number of builtin slots can push
the location outside the range and trigger an assertion.

This problem is affecting at least the following CTS tests for
enhanced layouts:

KHR-GL45.enhanced_layouts.varying_array_components
KHR-GL45.enhanced_layouts.varying_array_locations
KHR-GL45.enhanced_layouts.varying_components
KHR-GL45.enhanced_layouts.varying_locations

which use SSO and the the location layout qualifier to select such
location numbers explicitly.

This change helps these tests because for SSO we always have to include
things such as VARYING_SLOT_CLIP_DIST{0,1} even if the fragment shader is
very unlikely to read them, so by doing this we free builtin slots from
the fixed VUE layout and we avoid the tests to crash in this scenario.

Of course, this is not a proper fix, we'd still run into problems if someone
tries to use an explicit max location and read gl_ViewportIndex, gl_LayerID or
gl_CullDistancein in the FS, but that would be a much less common bug and we
can probably wait to see if anyone actually runs into that situation in a real
world scenario before making the decision that more aggresive changes are
required to support this without reverting 99df02ca26f61.

v2:
- Add a debug message when we skip clip distances (Ilia)
- we also need to account for this when we compute the urb setup
  for the fragment shader stage, so add a compiler util to compute
  the first slot that we need to read from the URB instead of
  replicating the logic in both places.

v3:
- Make the util more generic so it can account for all unused slots
  at the beginning of the URB, that will make it more useful (Ken).
- Drop the debug message, it was not what Ilia was asking for.

Suggested-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Normalize types for FBL, FBH, etc

Allows the instructions to be compacted. The documentation claims that
some of these only accept UD types, even though the type doesn't change
the operation performed. Just normalize the types to ensure we get
instruction compaction.

The only functional changes are for FBL and CBIT (always use UD types)
and FBH (always use the same types).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

radeonsi: don't use the template keyword

for C++ editors

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

glx: don't use the template keyword

for C++ editors

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

gallium/vl: don't use the template keyword

for C++ editors

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

egl/dri2: don't use the template keyword

for C++ editors

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

radeonsi/uvd: clean up si_video_buffer_create

V2: remove code duplication and one unnessecary variable, minor whitespace fix

Signed-off-by: Marek Olšák <marek.olsak@amd.com>

radeonsi/uvd: fix planar formats broken since f70f6baaa3bb0f8b280ac2eaea69bb

Tested-by: Benedikt Schemmer <ben@besd.de>
Reviewed-by: Christian König <christian.koenig@amd.com>

gallium: add new LOD opcode

The operation performed is all the same as LODQ, but with the usual
differences between dx10 and GL texture opcodes, that is separate resource
and sampler indices (plus result swizzling, and setting z/w channels
to zero).

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Acked-by: Nicolai Hähnle <nicolai.haehnle@amd.com>

drirc: whitelist glthread for Outlast

FPS increase 10-20% in starting locations on Core i5-4570 +
Radeon R9 270.

travis: Add clover build using llvm-5.0

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>

travis: Add clover build using llvm-4.0

llvm-4 needs gcc 4.8:
http://releases.llvm.org/4.0.1/docs/ReleaseNotes.html#non-comprehensive-list-of-changes-in-this-release

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>

travis: Add clover build using llvm-3.9

Use r600,radeonsi instead of i915
Update binutils, new linker is required for llvm-3.9:
https://www.ubuntuupdates.org/package/core/trusty/universe/updates/binutils-2.26

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>

st/va: add dst rect to avoid scale on deint

For 1080p video transcode, the height will be scaled to 1088 when deint
to progressive buffer. Set dst rect to make sure no scale.

Fixes: 3ad8687 "st/va: use new vl_compositor_yuv_deint_full() to deint"
Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Acked-by: Andy Furniss <adf.lists@gmail.com>

radeonsi: emit DLDEXP and DFRACEXP TGSI opcodes

Note: this causes spurious regressions in some current piglit tests,
because the tests incorrectly assume that there is no denorm support for
doubles. I'm going to send out a fix for those tests as well.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

radeonsi: emit LDEXP opcode

The LLVM intrinsic has existed for a long time. The current name was
established in LLVM 3.9.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

st/glsl_to_tgsi: use LDEXP when available

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

gallium: add LDEXP TGSI instruction and corresponding cap

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

tgsi: infer that dst[1] of DFRACEXP is an integer

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

gallivm: add support for TGSI instructions with two outputs

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

gallivm: add dst register index to lp_build_tgsi_context::emit_store

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

tgsi: clarify the semantics of DFRACEXP

The status quo is quite the mess:

1. tgsi_exec will do a per-channel computation, and store the dst[0]
   result (significand) correctly for each channel. The dst[1] result
   (exponent) will be written to the first bit set in the writemask.
   So per-component calculation only works partially.

2. r600 will only do a single computation. It will replicate the
   exponent but not the significand.

3. The docs pretend that there's per-component calculation, but even
   get dst[0] and dst[1] confused.

4. Luckily, st_glsl_to_tgsi only ever emits single-component instructions,
   and kind-of assumes that everything is replicated, generating this for
   the dvec4 case:

     DFRACEXP TEMP[0].xy, TEMP[1].x, CONST[0][0].xyxy
     DFRACEXP TEMP[0].zw, TEMP[1].y, CONST[0][0].zwzw
     DFRACEXP TEMP[2].xy, TEMP[1].z, CONST[0][1].xyxy
     DFRACEXP TEMP[2].zw, TEMP[1].w, CONST[0][1].zwzw

Settle on the simplest behavior, which is single-component calculation
with replication, document it, and adjust tgsi_exec and r600.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

tgsi: fix the documentation of DLDEXP

Sourcing the exponent for the zw destination pair from Z is consistent
with both tgsi_exec and gallivm. In practice, st_glsl_to_tgsi always
generates per-channel instructions anyway.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

tgsi: infer that DLDEXP's second source has an integer type

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

glsl/lower_instruction: handle denorms and overflow in ldexp correctly

GLSL ES requires both, and while GLSL explicitly doesn't require correct
overflow handling, it does appear to require handling input inf/denorms
correctly.

Fixes dEQP-GLES31.functional.shaders.builtin_functions.precision.ldexp.*

Cc: mesa-stable@lists.freedesktop.org
Acked-by: Matt Turner <mattst88@gmail.com>
Acked-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

util/queue: fix a race condition in the fence code

A tempting alternative fix would be adding a lock/unlock pair in
util_queue_fence_is_signalled. However, that wouldn't actually
improve anything in the semantics of util_queue_fence_is_signalled,
while making that test much more heavy-weight. So this lock/unlock
pair in util_queue_fence_destroy for "flushing out" other threads
that may still be in util_queue_fence_signal looks like the better
fix.

v2: rephrase the comment

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Gustaw Smolarczyk <wielkiegie@gmail.com>

r600: cleanup set_occlusion_query_state

This fixes a warning caused by the fork (note the change in the function
signature):

../../../../../mesa-src/src/gallium/drivers/r600/r600_state_common.c: In function ‘r600_init_common_state_functions’:
../../../../../mesa-src/src/gallium/drivers/r600/r600_state_common.c:2974:36: warning: assignment from incompatible pointer type [-Wincompatible-pointer-types]
  rctx->b.set_occlusion_query_state = r600_set_occlusion_query_state;

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

r300: add missing case PIPE_SHADER_CAP_INT64_ATOMICS

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: fix border color translation for integer textures

This fixes the extremely unlikely case that an application uses
0x80000000 or 0x3f800000 as border color for an integer texture and
helps in the also, but perhaps slightly less, unlikely case that 1 is
used as a border color.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

radeonsi: clamp border colors for upgraded depth textures

The hardware does this automatically for unorm formats, but we need to
do it manually for unorm depth formats that have been upgraded to
Z32_FLOAT.

Fixes dEQP-GLES31.functional.texture.border_clamp.range_clamp.nearest_unorm_depth
and others.

Fixes: d4d9ec55c589 ("radeonsi: implement TC-compatible HTILE")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

radeonsi: clamp depth comparison value only for fixed point formats

The hardware usually does this automatically. However, we upgrade
depth to Z32_FLOAT to enable TC-compatible HTILE, which means the
hardware no longer clamps the comparison value for us.

The only way to tell in the shader whether a clamp is required
seems to be to communicate an additional bit in the descriptor
table. While VI has some unused bits in the resource descriptor,
those bits have unfortunately all been used in gfx9. So we use
an unused bit in the sampler state instead.

Fixes dEQP-GLES3.functional.texture.shadow.2d.linear.equal_depth_component32f
and many other tests in dEQP-GLES3.functional.texture.shadow.*

Fixes: d4d9ec55c589 ("radeonsi: implement TC-compatible HTILE")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

radeonsi/gfx9: fix geometry shaders without output vertices

Not that those are super common or useful, but hey! Fun corner cases
of the API...

Fixes dEQP-GLES31.functional.geometry_shading.emit.*

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

amd/common: save an instruction in the build_cube_select sequence

Avoid a v_cndmask: the absolute value is free due to input modifiers.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

amd/common: fix build_cube_select

Fix the custom cube coord selection sequence to be identical to
the hardware v_cubesc/tc and OpenGL spec. Affects texture sampling
with user-provided derivatives.

Fixes dEQP-GLES3.functional.shaders.texture_functions.texturegrad.*

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

st/glsl_to_tgsi: fix conditional assignments to packed shader outputs

Overriding the default (no-op) swizzle is clearly counter-productive,
since the whole point is putting the destination register as one of
the source operands so that it remains unmodified when the assignment
condition is false.

Fragment depth and stencil outputs are a special case due to how their
source swizzles are manipulated in translate_src when compiling to
TGSI.

Fixes dEQP-GLES2.functional.shaders.conditionals.if.*_vertex
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

st/glsl_to_tgsi: fix a use-after-free in merge_two_dsts

Found by address sanitizer.

The loop here tries to be safe, but in doing so, it ends up doing
exactly the wrong thing: the safe foreach is for when the loop
variable (inst) could be deleted and nothing else. However, this
particular can delete inst's successor, but not inst itself.

Fixes: 8c6a0ebaad72 ("st/mesa: add st fp64 support (v7.1)")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>

radeonsi: move descriptor logs to after corresponding draw/compute packet

It has to happen after descriptor uploads since otherwise we'll print out
the wrong GPU list / incorrectly claim descriptor corruption.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

amd/common: remove ac_shader_abi::chip_class

Redundant with the recently added ac_llvm_context::chip_class.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>