nir: fix dangling ssadef->name ptrs

In many places, the convention is to pass an existing ssadef name ptr
when construction/initializing a new nir_ssa_def.  But that goes badly
(as noticed by garbage in nir_print output) when the original string
gets freed.

Just use ralloc_strdup() instead, and add ralloc_free() in the two
places that would care (not that the strings wouldn't eventually get
freed anyways).

Also fixup the nir_search code which was directly setting ssadef->name
to use the parent instruction as memctx.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

glsl: Add propagate_invariance to the other makefile

This fixes the scons build

nir/glsl: Propagate invariant into NIR alu ops

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

glsl/rebalance_tree: Don't handle invariant or precise trees

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

glsl/opt_algebraic: Don't handle invariant or precise trees

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

glsl: Add a pass to propagate the "invariant" and "precise" qualifiers

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/alu_to_scalar: Propagate the "exact" bit

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

i965/peephole_ffma: Don't fuse exact adds

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/cse: Properly handle nir_ssa_def.exact

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/algebraic: Flag inexact optimizations

Many of our optimizations, while great for cutting shaders down to size,
aren't really precision-safe.  This commit tries to flag all of the
inexact floating-point optimizations so they don't get run on values that
are flagged "exact".  It's a bit conservative and maybe flags some safe
optimizations as unsafe but that's better than missing one.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/algebraic: Fix fmin detection to match the spec

The previous transformation got the arguments to fmin backwards.  When NaNs
are involved, the GLSL min/max aren't commutative so it matters.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/algebraic: Get rid of an invlid fxor optimization

The fxor opcode is required to return 1.0f or 0.0f but the input variable
may not be 1.0f or 0.0f.

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/algebraic: Allow for flagging operations as being inexact

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/search: Propagate exactness into newly created expressions

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/builder: Add a flag for setting exact

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir: Add an "exact" bit to nir_alu_instr

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

nir/clone: Export nir_variable_clone

Reviewed-by: Rob Clark <robclark@gmail.com>

nir/clone: Expose nir_constant_clone

Reviewed-by: Rob Clark <robclark@gmail.com>

nir: Fix whitespace

Reviewed-by: Rob Clark <robclark@gmail.com>

docs: use latest libDRM version

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

compiler/glsl: allow sequence op as a const expr in gles 1.0

Allow the sequence operator to be a constant expression in GLSL ES
versions prior to GLSL ES 3.0

Fixes the following piglit test:
/all/spec/glsl-es-1.0/compiler/array-sized-by-sequence-in-parenthesis.vert

This is similar to the logic from process_initializer() which performs
the same check for constant variable initialization with sequence
operators.

v2: Fixed regression pointed out by Eduardo Lima Mitev

Signed-off-by: Lars Hamre <chemecse@gmail.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>

radeonsi: fix out-of-bounds indexing of shader images

Results are undefined but may not crash. Without this change, out-of-bounds
indexing can lead to VM faults and GPU hangs.

Constant buffers, samplers, and possibly others will eventually need similar
treatment to support GL_ARB_robust_buffer_access_behavior.

Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-and-Tested-by: Michel Dänzer <michel.daenzer@amd.com>

radeonsi: cache flush/invalidation for missing PIPE_BARRIER_*_BUFFER bits (v2)

This fixes arb_shader_image_load_store-host-mem-barrier.

v2: flush TC L2 for index buffers on <= CIK (Marek)

Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

st/mesa: add missing MemoryBarrier bits and some explanations

Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium: add PIPE_BARRIER_STREAMOUT_BUFFER

Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: fix 2D array MSAA failures since image support landed

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Reviewed-and-Tested-by: Michel Dänzer <michel.daenzer@amd.com>

i965/fs: Don't constant-fold RCP

No shader-db changes on Broadwell

Reviewed-by: Matt Turner <mattst88@gmail.com>

i965: Remove the RCP+RSQ algebraic optimizations

NIR already has this optimization and it can do much better than the little
peephole in the backend.

No shader-db change on Haswell or Broadwell.

Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: Don't abs slt and friends

No shader-db changes, but this is symmetric with the previous commit.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: Don't abs the result of b2f or b2i

In the results below, 2 SIMD16 shaders in Trine are lost.

G4X
total instructions in shared programs: 4012279 -> 4011108 (-0.03%)
instructions in affected programs: 116776 -> 115605 (-1.00%)
helped: 339
HURT: 0

total cycles in shared programs: 84315862 -> 84313584 (-0.00%)
cycles in affected programs: 1767232 -> 1764954 (-0.13%)
helped: 274
HURT: 81

Ironlake
total instructions in shared programs: 6399073 -> 6396998 (-0.03%)
instructions in affected programs: 218050 -> 215975 (-0.95%)
helped: 600
HURT: 0

total cycles in shared programs: 128892088 -> 128888810 (-0.00%)
cycles in affected programs: 2867452 -> 2864174 (-0.11%)
helped: 422
HURT: 137

Sandy Bridge
total instructions in shared programs: 8462174 -> 8460759 (-0.02%)
instructions in affected programs: 178529 -> 177114 (-0.79%)
helped: 596
HURT: 0

total cycles in shared programs: 117542276 -> 117534098 (-0.01%)
cycles in affected programs: 1239166 -> 1230988 (-0.66%)
helped: 369
HURT: 150

Ivy Bridge
total instructions in shared programs: 7775131 -> 7773410 (-0.02%)
instructions in affected programs: 162903 -> 161182 (-1.06%)
helped: 590
HURT: 0

total cycles in shared programs: 65759882 -> 65747268 (-0.02%)
cycles in affected programs: 1004354 -> 991740 (-1.26%)
helped: 467
HURT: 141

Haswell
total instructions in shared programs: 7107786 -> 7106327 (-0.02%)
instructions in affected programs: 140954 -> 139495 (-1.04%)
helped: 590
HURT: 0

total cycles in shared programs: 64668028 -> 64655322 (-0.02%)
cycles in affected programs: 967080 -> 954374 (-1.31%)
helped: 452
HURT: 149

LOST:   2
GAINED: 0

Broadwell
total instructions in shared programs: 8980029 -> 8978287 (-0.02%)
instructions in affected programs: 197232 -> 195490 (-0.88%)
helped: 715
HURT: 0

total cycles in shared programs: 70070448 -> 70055970 (-0.02%)
cycles in affected programs: 975724 -> 961246 (-1.48%)
helped: 471
HURT: 111

LOST:   2
GAINED: 0

Skylake
total instructions in shared programs: 9115178 -> 9113436 (-0.02%)
instructions in affected programs: 203012 -> 201270 (-0.86%)
helped: 715
HURT: 0

total cycles in shared programs: 68848660 -> 68834004 (-0.02%)
cycles in affected programs: 993888 -> 979232 (-1.47%)
helped: 473
HURT: 116

LOST:   2
GAINED: 0

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: Simplify 0 < fabs(a)

Sandy Bridge / Ivy Bridge / Haswell
total instructions in shared programs: 8462180 -> 8462174 (-0.00%)
instructions in affected programs: 564 -> 558 (-1.06%)
helped: 6
HURT: 0

total cycles in shared programs: 117542462 -> 117542276 (-0.00%)
cycles in affected programs: 9768 -> 9582 (-1.90%)
helped: 12
HURT: 0

Broadwell / Skylake
total instructions in shared programs: 8980833 -> 8980826 (-0.00%)
instructions in affected programs: 626 -> 619 (-1.12%)
helped: 7
HURT: 0

total cycles in shared programs: 70077900 -> 70077714 (-0.00%)
cycles in affected programs: 9378 -> 9192 (-1.98%)
helped: 12
HURT: 0

G45 and Ironlake showed no change.

v2: Modify the comments to look more like a proof.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: Simplify 0 >= b2f(a)

This also prevented some regressions with other patches in my local
tree.

Broadwell / Skylake
total instructions in shared programs: 8980835 -> 8980833 (-0.00%)
instructions in affected programs: 45 -> 43 (-4.44%)
helped: 1
HURT: 0

total cycles in shared programs: 70077904 -> 70077900 (-0.00%)
cycles in affected programs: 122 -> 118 (-3.28%)
helped: 1
HURT: 0

No changes on earlier platforms.

v2: Modify the comments to look more like a proof.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: Simplify i2b with negated or abs operand

This enables removing ssa_201 and ssa_202 in sequences like:

                 vec1 ssa_200 = flt ssa_199, ssa_194
                 vec1 ssa_201 = b2i ssa_200
                 vec1 ssa_202 = i2b -ssa_201

shader-db results:

Sandy Bridge
total instructions in shared programs: 8462257 -> 8462180 (-0.00%)
instructions in affected programs: 3846 -> 3769 (-2.00%)
helped: 35
HURT: 0

total cycles in shared programs: 117542934 -> 117542462 (-0.00%)
cycles in affected programs: 20072 -> 19600 (-2.35%)
helped: 20
HURT: 1

Ivy Bridge
total instructions in shared programs: 7775252 -> 7775137 (-0.00%)
instructions in affected programs: 3645 -> 3530 (-3.16%)
helped: 35
HURT: 0

total cycles in shared programs: 65760522 -> 65760068 (-0.00%)
cycles in affected programs: 21082 -> 20628 (-2.15%)
helped: 25
HURT: 2

Haswell
total instructions in shared programs: 7108666 -> 7108589 (-0.00%)
instructions in affected programs: 3253 -> 3176 (-2.37%)
helped: 35
HURT: 0

total cycles in shared programs: 64675726 -> 64675272 (-0.00%)
cycles in affected programs: 21034 -> 20580 (-2.16%)
helped: 26
HURT: 1

Broadwell / Skylake
total instructions in shared programs: 8980912 -> 8980835 (-0.00%)
instructions in affected programs: 3223 -> 3146 (-2.39%)
helped: 35
HURT: 0

total cycles in shared programs: 70077926 -> 70077904 (-0.00%)
cycles in affected programs: 21886 -> 21864 (-0.10%)
helped: 21
HURT: 6

G45 and Ironlake showed no change.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Suggested-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: Lower flrp with Boolean interpolator to bcsel

On Intel platforms that don't set lower_flrp, using bcsel instead of
flrp seems to be a small amount worse.  On those platforms, the use of
flrp, bcsel, and multiply of b2f is still an active area of research.
In review, Matt suggested this is because bcsel turns into CMP+SEL, and
because of the flag register we can't schedule instructions well.

shader-db results:

G4X / Ironlake
total instructions in shared programs: 4016538 -> 4012279 (-0.11%)
instructions in affected programs: 161556 -> 157297 (-2.64%)
helped: 1077
HURT: 1

total cycles in shared programs: 84328296 -> 84315862 (-0.01%)
cycles in affected programs: 4174570 -> 4162136 (-0.30%)
helped: 926
HURT: 53

Unsurprisingly, no changes on later platforms.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

i965: Have NIR lower flrp on pre-GEN6 vec4 backend

Previously we were doing the lowering by hand in vec4_visitor::emit_lrp.
By doing it in NIR, we have the opportunity for NIR to do additional
optimization of the expanded code.

This also enables optimizations added by the next commit.

shader-db results:

G4X / Ironlake
total instructions in shared programs: 4024401 -> 4016538 (-0.20%)
instructions in affected programs: 447686 -> 439823 (-1.76%)
helped: 2623
HURT: 0

total cycles in shared programs: 84375846 -> 84328296 (-0.06%)
cycles in affected programs: 16964960 -> 16917410 (-0.28%)
helped: 2556
HURT: 41

Unsurprisingly, no changes on later platforms.

v2: Formatting and comment changes suggested by Matt.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

swrast: fix discarded const warning in s_texture.c

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

i965: fix invalid memory write

I noticed some heap corruption running virgl tests, and valgrind
helped me to track it down to the following error:

==29272== Invalid write of size 4
==29272==    at 0x90283D4: push_loop_stack (brw_eu_emit.c:1307)
==29272==    by 0x9029A7D: brw_DO (brw_eu_emit.c:1750)
==29272==    by 0x90554B0: fs_generator::generate_code(cfg_t const*, int) (brw_fs_generator.cpp:1999)
==29272==    by 0x904491F: brw_compile_fs (brw_fs.cpp:5685)
==29272==    by 0x8FC5DC5: brw_codegen_wm_prog (brw_wm.c:137)
==29272==    by 0x8FC7663: brw_fs_precompile (brw_wm.c:638)
==29272==    by 0x8FA4040: brw_shader_precompile(gl_context*, gl_shader_program*) (brw_link.cpp:51)
==29272==    by 0x8FA4A9A: brw_link_shader (brw_link.cpp:260)
==29272==    by 0x8DEF751: _mesa_glsl_link_shader (ir_to_mesa.cpp:3006)
==29272==    by 0x8C84325: _mesa_link_program (shaderapi.c:1042)
==29272==    by 0x8C851D7: _mesa_LinkProgram (shaderapi.c:1515)
==29272==    by 0x4E4B8E8: add_shader_program (vrend_renderer.c:880)
==29272==  Address 0xf2f3cb0 is 0 bytes after a block of size 112 alloc'd
==29272==    at 0x4C2AA98: calloc (vg_replace_malloc.c:711)
==29272==    by 0x8ED11F7: ralloc_size (ralloc.c:113)
==29272==    by 0x8ED1282: rzalloc_size (ralloc.c:134)
==29272==    by 0x8ED14C0: rzalloc_array_size (ralloc.c:196)
==29272==    by 0x9019C7B: brw_init_codegen (brw_eu.c:291)
==29272==    by 0x904F565: fs_generator::fs_generator(brw_compiler const*, void*, void*, void const*, brw_stage_prog_data*, unsigned int, bool, gl_shader_stage) (brw_fs_generator.cpp:124)
==29272==    by 0x9044883: brw_compile_fs (brw_fs.cpp:5675)
==29272==    by 0x8FC5DC5: brw_codegen_wm_prog (brw_wm.c:137)
==29272==    by 0x8FC7663: brw_fs_precompile (brw_wm.c:638)
==29272==    by 0x8FA4040: brw_shader_precompile(gl_context*, gl_shader_program*) (brw_link.cpp:51)
==29272==    by 0x8FA4A9A: brw_link_shader (brw_link.cpp:260)
==29272==    by 0x8DEF751: _mesa_glsl_link_shader (ir_to_mesa.cpp:3006)

if_depth_in_loop is an array of size p->loop_stack_array_size, and
push_loop_stack() will access if_depth_in_loop[p->loop_stack_depth+1],
thus the condition to grow the array should be
p->loop_stack_array_size <= (p->loop_stack_depth + 1) (it's currently
off by 2...)

This can be reproduced by running the following test with virgl test
server:
LIBGL_ALWAYS_SOFTWARE=y GALLIUM_DRIVER=virpipe bin/shader_runner
./tests/shaders/glsl-fs-unroll-explosion.shader_test -auto

Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

tgsi: drop unused set_exec/kill_mask interfaces.

These don't get used and haven't been in git history from what I can
see, so drop them.

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

docs/relnotes: update ARB_internalformat_query2 status.

Signed-off-by: Dave Airlie <Airlied@redhat.com>

st/mesa: add support for internalformat query2.

Add code to handle GL_INTERNALFORMAT_PREFERRED.
Add code to deal with GL_RENDERBUFFER being passes into ChooseTextureFormat.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

i965: Fix assert conditions for src/dst x/y offsets

Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

swrast: Move assert for 'slice' in to check_map_teximage

Signed-off-by: Anuj Phogat <anuj.phogat@gmail.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>

r600/sb: Do not distribute neg in expr_handler::fold_assoc() when folding multiplications.

Previously it was doing this transformation for a Trine 3 shader:
     MUL     R6.x.12,    R13.x.23, 0.5|3f000000
-    MULADD     R4.x.12,    -R6.x.12, 2|40000000, 1|3f800000
+    MULADD     R4.x.12,    -R13.x.23, -1|bf800000, 1|3f800000

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94412
Signed-off-by: Xavier Bouchoux <xavierb@gmail.com>
Cc: "11.0 11.1 11.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

nvc0: make sure to delete samplers used by compute shaders

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.1 11.2" <mesa-stable@lists.freedesktop.org>

i965/blorp: Make BlitFramebuffer() do sRGB encoding in ES 3.x.

According to the ES 3.0 and GL 4.4 specifications, glBlitFramebuffer
is supposed to perform sRGB decoding and encoding whenever sRGB formats
are in use.  The ES 3.0 specification is completely clear, and has
always stated this.

However, the GL specification has changed behavior in 4.1, 4.2, and
4.4.  The original behavior stated that no sRGB encoding should occur.
The 4.4 behavior matches ES 3.0's wording.  However, implementing the
new behavior appears to break applications such as Left 4 Dead 2.

This patch changes Meta to apply the ES 3.x rules in ES 3.x, but
leaves OpenGL alone for now, to avoid breaking applications.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>

i965/blorp: Refactor sRGB encoding/decoding.

Because the rules for sRGB are so insane, we change brw_blorp_miptrees
to take decode_srgb and encode_srgb flags, which control linearization
of the source and destination separately.

This should make it easy to implement whatever crazy combination of
rules people throw at us.  For now, it should be equivalent.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>

meta: Make BlitFramebuffer() do sRGB encoding in ES 3.x.

According to the ES 3.0 and GL 4.4 specifications, glBlitFramebuffer
is supposed to perform sRGB decoding and encoding whenever sRGB formats
are in use.  The ES 3.0 specification is completely clear, and has
always stated this.

However, the GL specification has changed behavior in 4.1, 4.2, and
4.4.  The original behavior stated that no sRGB encoding should occur.
The 4.4 behavior matches ES 3.0's wording.  However, implementing the
new behavior appears to break applications such as Left 4 Dead 2.

This patch changes Meta to apply the ES 3.x rules in ES 3.x, but
leaves OpenGL alone for now, to avoid breaking applications.

Meta implements several other functions in terms of BlitFramebuffer,
and many of those explicitly do not perform sRGB encoding.  So, this
patch explicitly disables sRGB encoding in those other functions,
preserving the existing (correct) behavior.

If you're from the future and are reading this, hi!  Welcome to
the "fun" of debugging sRGB problems!  Best of luck!

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

docs: mark GL_ARB_shader_image_load_store/_size as done for radeonsi

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

radeonsi: Set PIPE_SHADER_CAP_MAX_SHADER_IMAGES

This enables ARB_shader_image_load_store and ARB_shader_image_size.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
[allow the same number of images for all shader stages and require LLVM 3.9]

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

radeonsi: disable early Z if the fragment shader writes to memory

Empirically, both the EXEC_ON_* flags and LATE_Z are necessary.

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

tgsi/scan: add writes_memory to flag presence of stores or atomics

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

radeonsi: force the DCC enable bit off in image descriptors for writing (v2)

This avoids a lockup at least on Tonga.

v2: only force DCC off on VI+ (Marek)

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

radeonsi: implement MemoryBarrier (v2)

v2: invalidate both constant and VMEM/TC L1 for constant buffers (Marek)

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

radeonsi: implement volatile memory access

Prevent loads from being re-ordered or coalesced.

Atomics don't need special handling by definition, and stores don't need
special handling because LLVM is unable to detect dead image or buffer
stores.

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

radeonsi: implement coherent memory access (v2)

v2: set glc=1 for volatile also on buffers

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

radeonsi: Lower TGSI_OPCODE_MEMBAR down to LLVM op

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

radeonsi: Lower TGSI_OPCODE_ATOM* down to LLVM op

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

radeonsi: Lower TGSI_OPCODE_STORE down to LLVM op

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

radeonsi: Lower TGSI_OPCODE_LOAD down to LLVM op (v3)

v2: new signature style for buffer intrinsics (offsets)
v3: new signature style for llvm.amdgcn.buffer.load.format (overloaded return)

Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v2)

radeonsi: extract the LLVM type name construction into its own function

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

radeonsi: Lower TGSI_OPCODE_RESQ down to LLVM op

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

radeonsi: extract TXQ buffer size computation into its own function

This will allow it to be reused for RESQ.

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

radeonsi: decompress shader images

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

radeonsi: update shader image descriptor for invalidated buffer

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

radeonsi: implement set_shader_images (v2)

Whether DCC is disabled depends on the access flags with which the image
is bound: image_load supports DCC, but store and atomic don't.

v2: remove an unnecessary masking of images->desc.enabled_mask

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

gallium/radeon: make r600_texture_disable_dcc externally accessible

We will need it in radeonsi for shader images.

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

tgsi/scan: track which shader images are really buffers

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

tgsi/scan: add images_writemask

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

st/mesa: translate additional flags in MemoryBarrier

Re-order flags in the order in which they appear in the OpenGL spec in the
description of MemoryBarrier().

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium: add additional PIPE_BARRIER_* bits

Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

svga: add svga_winsys_context::pipe_debug_callback pointer

The svga winsys modules can use this to send debug messages to the
state tracker and Mesa.

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

svga: Fix the index buffer rebind regression

The index buffer handle saved in the hw_state structure could
be invalid after the index buffer is destroyed. Instead of
rebinding the index buffer using the saved index buffer handle,
we will reset the index buffer handle in the hw_state structure
to force resending of the index buffer.

Fixes bug 1593320

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

svga: rebind stream output targets

To ensure stream output target surfaces are available for the draw commands,
we need to rebind the current stream output targets at the first draw in the
command buffer.

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

svga: rebind index buffer

Similar to other resources, current index buffer needs to be
rebound at the first draw of the current command buffer to make
sure the buffer is available for the draw command.

Fixes bug 1587263.

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

svga: minor formatting fix, comment addition

To sync with our internal tree.

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

svga: optimize constant buffer uploads

When a constant buffer slot is allocated in the upload buffer,
the allocated slot size is always in multiple of 256. But the actual buffer
size might not be in multiple of 256. This causes a gap between
the ending offset of a slot and the starting offset of the next slot.
The gap will prevent the two slots to be updated in a single update command.
In order to maximize the chance of merging the contiguous dirty ranges,
when a slot is to be allocated in the constant upload buffer,
specify a buffer size in multiple of 256.

There is about 10% performance improvement with Lightsmark2008 and
30% with Cinebench R11.

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

svga: add a few more resource updates HUD query

This patch adds the following HUD queries:
.num-resource-updates  -- number of resource update. Commands include
                          UPDATE_SUBRESOURCE, UPDATE_GB_IMAGE.
.num-buffer-uploads    -- number of buffer uploads.
.num-const-buf-updates -- number of set constant buffer.
.num-const-updates     -- number of set shader constant.

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

svga: add new num-readbacks HUD query

To find out how many image readback command is issued.

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

svga: use shader sampler view declarations

Previously, we looked at the bound textures (via the pipe_sampler_views)
to determine texture dimensions (1D/2D/3D/etc) and datatype (float vs.
int).  But this could fail in out of memory conditions.  If we failed to
allocate a texture and didn't create a pipe_sampler_view, we'd default
to using 0 (PIPE_BUFFER) as the texture type.  This led to device errors
because of inconsistent shader code.

This change relies on all TGSI shaders having an SVIEW declaration for
each SAMP declaration.  The previous patch series does that.

Reviewed-by: Charmaine Lee <charmainel@vmware.com>

gallium/tests: declare sampler views in shaders

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

gallium/util: declare sampler view in util_make_fs_blit_msaa_depthstencil()

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

postprocess: declare sampler views in shaders

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

hud: add sampler view declaration in text fragment shader

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

st/mesa: emit sampler view decls in drawpixels code

v2: support both TGSI_TEXTURE_2D and _RECT

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

st/mesa: emit sampler view declaration in bitmap shader

In June 2015, Rob Clark started updating the tgsi utility code to emit
SVIEW declarations in various shaders (for polygon stipple, blitting,
etc).  These patches do the same for the Mesa state tracker.

The VMware driver will use this.

v2: support both TGSI_TEXTURE_2D and _RECT

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

st/mesa: emit sampler view declarations for ARB vert/frag programs

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

st/mesa: use correct TGSI texture target in drawpix fragment shader

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

st/mesa: use correct TGSI texture target in bitmap fragment shader

Depending on the driver's support for NPOT textures, we might use
a RECT texture instead of 2D texture.  We should propogate that info
to the fragment shader's TEX instruction.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

gallium/tgsi: pass TGSI tex target to tgsi_transform_tex_inst()

Instead of hard-coded 2D tex target in tgsi_transform_tex_2d_inst()

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Charmaine Lee <charmainel@vmware.com>

st/mesa: use the texture view's format for render-to-texture

Aside from the bug below, it fixes a simplistic test I've written locally,
and I see no regression in Piglit for radeonsi.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94595
Cc: "11.0 11.1 11.2" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

gallium: Remove unused TGSI_RESOURCE_ defines

These magic file-index defines where only ever used in the nouveau code
and that no longer uses them.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v2)

nouveau: codegen: Do not silently fail in handeLOAD / handleSTORE / handleATOM

handeLOAD / handleSTORE / handleATOM can only handle TGSI_FILE_BUFFER
and TGSI_FILE_MEMORY. Make things fail explictly when another
register-file is used in these functions.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)

nouveau: codegen: Disable more old resource handling code

Commit c3083c7082 ("nv50/ir: add support for BUFFER accesses") disabled /
commented out some of the old resource handling code, but not all of it.

Effectively all of it is dead already, if we ever enter the old code
paths in handeLOAD / handleSTORE / handleATOM we will get an exception
due to trying to access the now always zero-sized resources vector.

Disable all the dead code.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)

nouveau: codegen: gk110: Make emitSTORE offset handling identical to emitLOAD

Make the store offset handling in CodeEmitterGK110::emitSTORE identical
to the one in CodeEmitterGK110::emitLOAD handling.

This is just a cleanup, it does not cause any functional changes.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

nouveau: codegen: Slightly refactor Source::scanInstruction() dst handling

Use the dst temp variable which was used in the TGSI_FILE_OUTPUT
case everywhere. This makes the code somewhat easier to reads
and helps avoiding going over 80 chars with upcoming changes.

This also brings the dst handling more in line with the src
handling.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

nouveau: codegen: Add support for clover / OpenCL kernel input parameters

Add support for clover / OpenCL kernel input parameters.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu> (v1)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)

tgsi: Add support for global / private / input MEMORY

Extend the MEMORY file support to differentiate between global, private
and shared memory, as well as "input" memory.

"MEMORY[x], INPUT" is intended to access OpenCL kernel parameters, a
special memory type is added for this, since the actual storage of these
(e.g. UBO-s) may differ per implementation. The uploading of kernel
parameters is handled by launch_grid, "MEMORY[x], INPUT" allows drivers
to use an access mechanism for parameter reads which matches with the
upload method.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu> (v1)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)

tgsi: Fix decl.Atomic and .Shared not propagating when parsing tgsi text

When support for decl.Atomic and .Shared was added, tgsi_build_declaration
was not updated to propagate these properly.

Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu> (v1)
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> (v2)

doc: document spilling options accepted by INTEL_DEBUG

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

tgsi: Fix return of uninitialized memory in tgsi_*_instruction_memory

tgsi_default_instruction_memory / tgsi_build_instruction_memory were
returning uninitialized memory for tgsi_instruction_memory.Texture and
tgsi_instruction_memory.Format. Note 0 means not set, and thus is a
correct default initializer for these.

Fixes: 3243b6fc97 ("tgsi: add Texture and Format to tgsi_instruction_memory")
Cc: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>