mesa.git
14 years agodri/nv04: Don't expose ARB_texture_env_combine/dot3.
Andrew Randrianasulu [Tue, 14 Sep 2010 18:59:44 +0000 (20:59 +0200)]
dri/nv04: Don't expose ARB_texture_env_combine/dot3.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agollvmpipe: add DEBUG_FS to dump variant information
Keith Whitwell [Thu, 16 Sep 2010 14:16:27 +0000 (15:16 +0100)]
llvmpipe: add DEBUG_FS to dump variant information

14 years agollvmpipe: add LP_PERF flag to disable various aspects of rasterization
Keith Whitwell [Thu, 16 Sep 2010 09:45:52 +0000 (10:45 +0100)]
llvmpipe: add LP_PERF flag to disable various aspects of rasterization

Allows disabling various operations (mainly texture-related, but
will grow) to try & identify bottlenecks.

Unlike LP_DEBUG, this is active even in release builds - which is
necessary for performance investigation.

14 years agogallivm: make lp_build_sample_nop public
Keith Whitwell [Thu, 16 Sep 2010 11:31:42 +0000 (12:31 +0100)]
gallivm: make lp_build_sample_nop public

14 years agogallivm: move i32_vec_type inside the #ifdef
Brian Paul [Thu, 16 Sep 2010 15:00:54 +0000 (09:00 -0600)]
gallivm: move i32_vec_type inside the #ifdef

14 years agogallivm: fix incorrect vector shuffle datatype
Brian Paul [Thu, 16 Sep 2010 14:56:22 +0000 (08:56 -0600)]
gallivm: fix incorrect vector shuffle datatype

The permutation vector must always be a vector of int32 values.

14 years agonv50: get shader fixups/relocations into working state
Christoph Bumiller [Thu, 16 Sep 2010 12:47:44 +0000 (14:47 +0200)]
nv50: get shader fixups/relocations into working state

14 years agonv50: don't segfault on shaders with 0 instructions
Christoph Bumiller [Wed, 15 Sep 2010 22:51:32 +0000 (00:51 +0200)]
nv50: don't segfault on shaders with 0 instructions

14 years agoglsl: Don't print blank (function ...) headers for built-ins.
Kenneth Graunke [Thu, 16 Sep 2010 09:55:02 +0000 (02:55 -0700)]
glsl: Don't print blank (function ...) headers for built-ins.

Fixes a regression caused when I added my GLSL ES support.

14 years agoglsl: Change from has_builtin_signature to has_user_signature.
Kenneth Graunke [Thu, 16 Sep 2010 09:52:25 +0000 (02:52 -0700)]
glsl: Change from has_builtin_signature to has_user_signature.

The print visitor needs this, and the only existing user can work with
has_user_signature just as well.

14 years agor600g: Use clamped math for RCP and RSQ.
Tilman Sauerbeck [Fri, 10 Sep 2010 16:24:01 +0000 (18:24 +0200)]
r600g: Use clamped math for RCP and RSQ.

This is likely only correct for OpenGL and not other state trackers.

Signed-off-by: Tilman Sauerbeck <tilman@code-monkey.de>
14 years agor600g: Fixed a bo leak in r600_blit_state_ps_shader().
Tilman Sauerbeck [Fri, 10 Sep 2010 13:20:15 +0000 (15:20 +0200)]
r600g: Fixed a bo leak in r600_blit_state_ps_shader().

We would leak the newly created bo if it cannot be mapped.

Signed-off-by: Tilman Sauerbeck <tilman@code-monkey.de>
14 years agost/xlib: Notify the context when the front/back buffers are swapped.
Chia-I Wu [Tue, 7 Sep 2010 16:30:27 +0000 (00:30 +0800)]
st/xlib: Notify the context when the front/back buffers are swapped.

The current context should be notified when the the front/back buffers
of the current drawable are swapped.  The notification was skipped when
xmesa_strict_invalidate is false (the default).

This fixes fdo bug #29774.

14 years agomesa: Update ES APIspec.xml.
Chia-I Wu [Wed, 15 Sep 2010 14:37:40 +0000 (22:37 +0800)]
mesa: Update ES APIspec.xml.

Enable some extensions now that the needed tokens are defined in
GLES/glext.h and GLES2/glext.h.  Update the prototype of MultiDrawArrays
now that the prototype of _mesa_MultiDrawArraysEXT has been updated.

14 years agor600g: fix texture bos and avoid doing depth blit on evergreen
Dave Airlie [Thu, 16 Sep 2010 11:48:02 +0000 (21:48 +1000)]
r600g: fix texture bos and avoid doing depth blit on evergreen

since the depth blit code is hardcoded hex yay \o/

14 years agor600g: fixup texture state on evergreen.
Dave Airlie [Thu, 16 Sep 2010 11:29:08 +0000 (21:29 +1000)]
r600g: fixup texture state on evergreen.

This whole set of state just seems wrong, another cut-n-paste nightmare.

14 years agomesa/st: Silence uninitialized variable warning.
Vinson Lee [Thu, 16 Sep 2010 01:47:17 +0000 (18:47 -0700)]
mesa/st: Silence uninitialized variable warning.

14 years agonv50: Fix 'control reaches end of non-void function' warning.
Vinson Lee [Thu, 16 Sep 2010 01:26:06 +0000 (18:26 -0700)]
nv50: Fix 'control reaches end of non-void function' warning.

14 years agonv50: Silence uninitialized variable warnings.
Vinson Lee [Thu, 16 Sep 2010 01:24:28 +0000 (18:24 -0700)]
nv50: Silence uninitialized variable warnings.

14 years agodraw: Remove unnecessary header.
Vinson Lee [Thu, 16 Sep 2010 01:17:51 +0000 (18:17 -0700)]
draw: Remove unnecessary header.

14 years agogallivm: Remove unnecessary headers.
Vinson Lee [Thu, 16 Sep 2010 01:14:18 +0000 (18:14 -0700)]
gallivm: Remove unnecessary headers.

14 years agonv50: Silence uninitialized variable warning.
Vinson Lee [Thu, 16 Sep 2010 00:27:50 +0000 (17:27 -0700)]
nv50: Silence uninitialized variable warning.

14 years agonv50: Silence uninitialized variable warning.
Vinson Lee [Thu, 16 Sep 2010 00:24:50 +0000 (17:24 -0700)]
nv50: Silence uninitialized variable warning.

14 years agonv50: Silence uninitialized variable warning.
Vinson Lee [Thu, 16 Sep 2010 00:09:59 +0000 (17:09 -0700)]
nv50: Silence uninitialized variable warning.

14 years agonv50: Remove unnecessary headers.
Vinson Lee [Wed, 15 Sep 2010 23:51:39 +0000 (16:51 -0700)]
nv50: Remove unnecessary headers.

14 years agonv50: Update files in SConscript to match Makefile.
Vinson Lee [Wed, 15 Sep 2010 23:46:04 +0000 (16:46 -0700)]
nv50: Update files in SConscript to match Makefile.

14 years agor600g: add vgt dma src defines
Dave Airlie [Wed, 15 Sep 2010 23:41:43 +0000 (09:41 +1000)]
r600g: add vgt dma src defines

14 years agor600g: use index min/max + index buffer offset.
Dave Airlie [Wed, 15 Sep 2010 23:39:29 +0000 (09:39 +1000)]
r600g: use index min/max + index buffer offset.

more prep work for fixing up buffer handling

14 years agor600g: pull r600_draw struct out into header
Dave Airlie [Wed, 15 Sep 2010 23:33:34 +0000 (09:33 +1000)]
r600g: pull r600_draw struct out into header

we need this for future buffer rework, it also makes the vtbl easier

14 years agogallivm: expand AoS sampling to cover all filtering modes
Brian Paul [Wed, 15 Sep 2010 23:04:26 +0000 (17:04 -0600)]
gallivm: expand AoS sampling to cover all filtering modes

...and all texture targets (1D/2D/3D/CUBE).

14 years agotgsi: fix incorrect usage_mask for shadow tex instructions
Brian Paul [Wed, 15 Sep 2010 19:56:00 +0000 (13:56 -0600)]
tgsi: fix incorrect usage_mask for shadow tex instructions

The shadow versions of the texture targets use an extra component
(Z) to express distance from light source to the fragment.
Fixes the shadowtex demo with llvmpipe.

14 years agonv50: use unsigned int for bitfields to silence warnings
Brian Paul [Wed, 15 Sep 2010 18:51:09 +0000 (12:51 -0600)]
nv50: use unsigned int for bitfields to silence warnings

14 years agollvmpipe: s/boolean/unsigned/ in bitfield to silence warning
Brian Paul [Wed, 15 Sep 2010 18:49:11 +0000 (12:49 -0600)]
llvmpipe: s/boolean/unsigned/ in bitfield to silence warning

Using non-int types for bitfields is a gcc extension.
The size of the struct is not effected by this change.

14 years agollvmpipe: cast to silence warning
Brian Paul [Wed, 15 Sep 2010 18:48:29 +0000 (12:48 -0600)]
llvmpipe: cast to silence warning

14 years agoglsl2: fix signed/unsigned comparison warning
Brian Paul [Wed, 15 Sep 2010 18:47:32 +0000 (12:47 -0600)]
glsl2: fix signed/unsigned comparison warning

14 years agor600g: misc cleanup
John Doe [Wed, 15 Sep 2010 15:47:32 +0000 (11:47 -0400)]
r600g: misc cleanup

Avoid using r600_screen structure to get ptr to radeon
winsys structure.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
14 years agoMerge remote branch 'origin/nv50-compiler'
Christoph Bumiller [Wed, 15 Sep 2010 15:34:40 +0000 (17:34 +0200)]
Merge remote branch 'origin/nv50-compiler'

Conflicts:
src/gallium/drivers/nouveau/nouveau_class.h
src/gallium/drivers/nv50/nv50_screen.c

14 years agollvmpipe: return zero from floor_pot(zero)
Keith Whitwell [Wed, 15 Sep 2010 15:28:49 +0000 (16:28 +0100)]
llvmpipe: return zero from floor_pot(zero)

14 years agonv50: put low limit on REG_ALLOC_TEMP and FP_RESULT_COUNT
Christoph Bumiller [Wed, 15 Sep 2010 13:21:41 +0000 (15:21 +0200)]
nv50: put low limit on REG_ALLOC_TEMP and FP_RESULT_COUNT

14 years agonv50: improve and fix modifier folding optimization
Christoph Bumiller [Wed, 15 Sep 2010 11:59:09 +0000 (13:59 +0200)]
nv50: improve and fix modifier folding optimization

Execute before folding loads, because we don't check if it's legal
in lower_mods.
Ensure that a value's insn pointer is updated when transferring it
to a different instruction.

14 years agonv50: consider address register in reload elimination
Christoph Bumiller [Mon, 13 Sep 2010 19:13:36 +0000 (21:13 +0200)]
nv50: consider address register in reload elimination

14 years agollvmpipe: remove duplicate code
Keith Whitwell [Wed, 15 Sep 2010 10:51:41 +0000 (11:51 +0100)]
llvmpipe: remove duplicate code

Bad rebase presumably.

14 years agollvmpipe: brackets around macro arg
Keith Whitwell [Sun, 5 Sep 2010 15:11:25 +0000 (16:11 +0100)]
llvmpipe: brackets around macro arg

14 years agoglapi: Fix ES build errors again.
Chia-I Wu [Wed, 15 Sep 2010 13:12:38 +0000 (21:12 +0800)]
glapi: Fix ES build errors again.

This fixes an error in GLAPI ES.  My build is ok with or without this
patch, and the error affects others' setups.

[Patch from Francesco Marella]

14 years agor600g: Silence unused variable warning.
Vinson Lee [Wed, 15 Sep 2010 13:13:31 +0000 (06:13 -0700)]
r600g: Silence unused variable warning.

The code that uses dname is currently ifdef'ed out.

14 years agor600g: Silence uninitialized variable warning.
Vinson Lee [Wed, 15 Sep 2010 13:09:28 +0000 (06:09 -0700)]
r600g: Silence uninitialized variable warning.

14 years agor600g: Silence uninitialized variable warning.
Vinson Lee [Wed, 15 Sep 2010 12:52:16 +0000 (05:52 -0700)]
r600g: Silence uninitialized variable warning.

14 years agor600g: Silence uninitialized variable warning.
Vinson Lee [Wed, 15 Sep 2010 12:46:34 +0000 (05:46 -0700)]
r600g: Silence uninitialized variable warning.

14 years agonvfx: Silence uninitialized variable warnings.
Vinson Lee [Wed, 15 Sep 2010 12:43:50 +0000 (05:43 -0700)]
nvfx: Silence uninitialized variable warnings.

14 years agor600g: Silence uninitialized variable warning.
Vinson Lee [Wed, 15 Sep 2010 12:34:29 +0000 (05:34 -0700)]
r600g: Silence uninitialized variable warning.

14 years agor600g: Silence uninitialized variable warning.
Vinson Lee [Wed, 15 Sep 2010 12:31:31 +0000 (05:31 -0700)]
r600g: Silence uninitialized variable warning.

14 years agoglsl: Fix 'format not a string literal and no format arguments' warning.
Vinson Lee [Wed, 15 Sep 2010 12:17:57 +0000 (05:17 -0700)]
glsl: Fix 'format not a string literal and no format arguments' warning.

Fix the following GCC warning.
loop_controls.cpp: In function 'int calculate_iterations(ir_rvalue*, ir_rvalue*, ir_rvalue*, ir_expression_operation)':
loop_controls.cpp:88: warning: format not a string literal and no format arguments

14 years agor300g: fix buffer reuse issue caused by previous commit
Dave Airlie [Wed, 15 Sep 2010 10:58:14 +0000 (20:58 +1000)]
r300g: fix buffer reuse issue caused by previous commit

caused by 0b9eb5c9bb03e5134d9a41786178100109e80c5a

test run glxgears, resize.

14 years agoglapi: Fix build errors for ES.
Chia-I Wu [Wed, 15 Sep 2010 09:43:21 +0000 (17:43 +0800)]
glapi: Fix build errors for ES.

The latest glext.h defines GL_FIXED.  Test GL_OES_fixed_point instead to
decide whether to define GLfixed and GLclampx.

This fixes fdo bug #30205.

14 years agor600c: fix buffer height setting in dri2 case
Andre Maasikas [Wed, 15 Sep 2010 08:08:24 +0000 (11:08 +0300)]
r600c: fix buffer height setting in dri2 case

fbHeight is 0 in this case

uncovered by changes in b0bc026c and should fix kernel rejecting command
streams after that commit

14 years agor300g: prevent creating multiple winsys BOs for the same handle
Marek Olšák [Wed, 15 Sep 2010 01:46:10 +0000 (03:46 +0200)]
r300g: prevent creating multiple winsys BOs for the same handle

This fixes a DRM deadlock in the cubestorm xscreensaver, because somehow
there must not be 2 different BOs relocated in one CS if both BOs back
the same handle. I was told it is impossible to happen, but apparently
it is not, or there is something else wrong.

14 years agomesa: Include missing header in program.h.
Vinson Lee [Wed, 15 Sep 2010 00:54:46 +0000 (17:54 -0700)]
mesa: Include missing header in program.h.

Include compiler.h for ASSERT symbol.

14 years agor600g: Remove unnecessary headers.
Vinson Lee [Wed, 15 Sep 2010 00:42:47 +0000 (17:42 -0700)]
r600g: Remove unnecessary headers.

14 years agoauxiliary: fix unintended fallthrough
Luca Barbieri [Tue, 14 Sep 2010 19:45:01 +0000 (21:45 +0200)]
auxiliary: fix unintended fallthrough

14 years agollvmpipe: Remove unnecessary header.
Vinson Lee [Tue, 14 Sep 2010 21:23:17 +0000 (14:23 -0700)]
llvmpipe: Remove unnecessary header.

14 years agoglx: add const qualifiers to __indirect_glMultiDrawArraysEXT()
Brian Paul [Tue, 14 Sep 2010 17:01:03 +0000 (11:01 -0600)]
glx: add const qualifiers to __indirect_glMultiDrawArraysEXT()

14 years agor600c: fix setting negative values to bitfields
Andre Maasikas [Tue, 14 Sep 2010 15:56:39 +0000 (18:56 +0300)]
r600c: fix setting negative values to bitfields

when setting negative integers to bitfields we could overwrite
other parts of it. So mask the value to be written correctly.
This is used quite often in the driver - hope it doesnt affect
performace or uncover behaviour relied before...

fixes strange effects when setting negative lodbias on evergreen

14 years agomesa: upgrade wglext.h to version 22
Brian Paul [Tue, 14 Sep 2010 15:40:23 +0000 (09:40 -0600)]
mesa: upgrade wglext.h to version 22

14 years agomesa: upgrade glxext.h to version 32
Brian Paul [Tue, 14 Sep 2010 15:39:44 +0000 (09:39 -0600)]
mesa: upgrade glxext.h to version 32

14 years agomesa: update to version 64 of GL/glext.h
Brian Paul [Tue, 14 Sep 2010 15:37:35 +0000 (09:37 -0600)]
mesa: update to version 64 of GL/glext.h

A number of other files had to be updated as well because const
qualifiers were added to the glMultiDrawArrays() function.
Also, GL_FIXED is now defined in glext.h.

14 years agomesa: move, redefine MESA_GEOMETRY_PROGRAM
Brian Paul [Tue, 14 Sep 2010 15:16:40 +0000 (09:16 -0600)]
mesa: move, redefine MESA_GEOMETRY_PROGRAM

14 years agoglsl2: add case for ir_unop_noise
Brian Paul [Tue, 14 Sep 2010 15:15:19 +0000 (09:15 -0600)]
glsl2: add case for ir_unop_noise

Silences a compiler warning.  Still need to add some assertions
for this case.

14 years agoglsl2: fix comments
Brian Paul [Tue, 14 Sep 2010 15:05:46 +0000 (09:05 -0600)]
glsl2: fix comments

14 years agomesa: Remove unnecessary FEATURE tests.
Chia-I Wu [Mon, 13 Sep 2010 05:20:25 +0000 (13:20 +0800)]
mesa: Remove unnecessary FEATURE tests.

Remove all FEATURE tests in mesa/drivers/common/.  They are not needed
and the code looks better without them.

14 years agomesa: Less FEATURE_ARB_sync tests.
Chia-I Wu [Mon, 13 Sep 2010 05:35:12 +0000 (13:35 +0800)]
mesa: Less FEATURE_ARB_sync tests.

Add dummy static inline definitions to syncobj.h when FEATURE_ARB_sync
is 0, and remove most FEATURE_ARB_sync tests.

14 years agomesa: Remove unused _MESA_INIT_*_FUNCTIONS.
Chia-I Wu [Mon, 13 Sep 2010 05:51:52 +0000 (13:51 +0800)]
mesa: Remove unused _MESA_INIT_*_FUNCTIONS.

They were intended to be used to build OpenGL ES only DRI drivers, but
that never happened.

14 years agogallivm: Remove unnecessary header.
Vinson Lee [Tue, 14 Sep 2010 07:42:20 +0000 (00:42 -0700)]
gallivm: Remove unnecessary header.

14 years agollvmpipe: Initialize variable for potentially unhandled switch case.
Vinson Lee [Tue, 14 Sep 2010 07:17:13 +0000 (00:17 -0700)]
llvmpipe: Initialize variable for potentially unhandled switch case.

14 years agoi965g: Fix 'control reaches end of non-void function' warning.
Vinson Lee [Tue, 14 Sep 2010 07:08:45 +0000 (00:08 -0700)]
i965g: Fix 'control reaches end of non-void function' warning.

Fixes the following GCC warning.
brw_screen.c: In function 'brw_get_shader_param':
brw_screen.c:241: warning: control reaches end of non-void function

14 years agoi915g: Fix 'control reaches end of non-void function' warning.
Vinson Lee [Tue, 14 Sep 2010 06:58:42 +0000 (23:58 -0700)]
i915g: Fix 'control reaches end of non-void function' warning.

Fixes the following GCC warning.
i915_screen.c: In function 'i915_get_shader_param':
i915_screen.c:184: warning: control reaches end of non-void function

14 years agoi915: Fix "implicit declaration of function 'draw_get_shader_param'" warning.
Vinson Lee [Tue, 14 Sep 2010 06:42:35 +0000 (23:42 -0700)]
i915: Fix "implicit declaration of function 'draw_get_shader_param'" warning.

Fixes the following GCC warning.
i915_screen.c: In function 'i915_get_shader_param':
i915_screen.c:147: warning: implicit declaration of function 'draw_get_shader_param'

14 years agoidentity: Fix 'assignment from incompatible pointer type' warning.
Vinson Lee [Tue, 14 Sep 2010 06:23:34 +0000 (23:23 -0700)]
identity: Fix 'assignment from incompatible pointer type' warning.

This is a follow-up to commit a508d2dddcc67d0f92cc36b9ed6f36a9bbfc579d.

Fixes the following GCC warning.
id_screen.c: In function 'identity_screen_create':
id_screen.c:317: warning: assignment from incompatible pointer type

14 years agorbug: Fix 'assignment from incompatible pointer type' warning.
Vinson Lee [Tue, 14 Sep 2010 06:02:37 +0000 (23:02 -0700)]
rbug: Fix 'assignment from incompatible pointer type' warning.

This is a follow-up to commit a508d2dddcc67d0f92cc36b9ed6f36a9bbfc579d.

Fixes the following GCC warning.
rbug_screen.c: In function 'rbug_screen_create':
rbug_screen.c:331: warning: assignment from incompatible pointer type

14 years agomesa/st: ask GLSL to not emit noise since we have a dummy implementation
Luca Barbieri [Tue, 14 Sep 2010 03:10:59 +0000 (05:10 +0200)]
mesa/st: ask GLSL to not emit noise since we have a dummy implementation

Note, BTW, that the Gallium implementation returns 0.5, which seems
to violate the GLSL spec, where it should return 0.0 instead.

Not sure whether changing it to 0 is correct or not.

14 years agomesa/st: set compiler options based on Gallium shader caps
Luca Barbieri [Mon, 6 Sep 2010 00:31:20 +0000 (02:31 +0200)]
mesa/st: set compiler options based on Gallium shader caps

This turns on if conversion and unlimited loop unrolling if control
flow is not supported.

NOTE: this will change the behavior of r300g and any other driver
that doesn't advertise control flow

14 years agogallium: introduce get_shader_param (ALL DRIVERS CHANGED) (v3)
Luca Barbieri [Sun, 5 Sep 2010 18:50:50 +0000 (20:50 +0200)]
gallium: introduce get_shader_param (ALL DRIVERS CHANGED) (v3)

Changes in v3:
- Also change trace, which I forgot about

Changes in v2:
- No longer adds tessellation shaders

Currently each shader cap has FS and VS versions.

However, we want a version of them for geometry, tessellation control,
and tessellation evaluation shaders, and want to be able to easily
query a given cap type for a given shader stage.

Since having 5 duplicates of each shader cap is unmanageable, add
a new get_shader_param function that takes both a shader cap from a
new enum and a shader stage.

Drivers with non-unified shaders will first switch on the shader
and, within each case, switch on the cap.

Drivers with unified shaders instead first check whether the shader
is supported, and then switch on the cap.

MAX_CONST_BUFFERS is now per-stage.
The geometry shader cap is removed in favor of checking whether the
limit of geometry shader instructions is greater than 0, which is also
used for tessellation shaders.

WARNING: all drivers changed and compiled but only nvfx tested

14 years agoglsl2: Port equal() and notEqual() to ir_unop_all_equal and ir_unop_any_nequal
Ian Romanick [Tue, 14 Sep 2010 00:53:32 +0000 (17:53 -0700)]
glsl2: Port equal() and notEqual() to ir_unop_all_equal and ir_unop_any_nequal

14 years agoglsl: introduce ir_binop_all_equal and ir_binop_any_equal, allow vector cmps
Luca Barbieri [Tue, 7 Sep 2010 23:31:39 +0000 (01:31 +0200)]
glsl: introduce ir_binop_all_equal and ir_binop_any_equal, allow vector cmps

Currently GLSL IR forbids any vector comparisons, and defines "ir_binop_equal"
and "ir_binop_nequal" to compare all elements and give a single bool.

This is highly unintuitive and prevents generation of optimal Mesa IR.

Hence, first rename "ir_binop_equal" to "ir_binop_all_equal" and
"ir_binop_nequal" to "ir_binop_any_nequal".

Second, readd "ir_binop_equal" and "ir_binop_nequal" with the same semantics
as less, lequal, etc.

Third, allow all comparisons to acts on vectors.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
14 years agoloop_unroll: unroll loops with (lowered) breaks
Luca Barbieri [Tue, 7 Sep 2010 15:03:43 +0000 (17:03 +0200)]
loop_unroll: unroll loops with (lowered) breaks

If the loop ends with an if with one break or in a single break unroll
it.  Loops that end with a continue will have that continue removed by
the redundant jump optimizer.  Likewise loops that end with an
if-statement with a break at the end of both branches will have the
break pulled out after the if-statement.

Loops of the form

   for (...) {
      do_something1();
      if (cond) {
 do_something2();
 break;
      } else {
 do_something3();
      }
   }

will be unrolled as

   do_something1();
   if (cond) {
      do_something2();
   } else {
      do_something3();
      do_something1();
      if (cond) {
 do_something2();
      } else {
 do_something3();
 /* Repeat inserting iterations here.*/
      }
   }

ir_lower_jumps can guarantee that all loops are put in this form
and thus all loops are now potentially unrollable if an upper bound
on the number of iterations can be found.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
14 years agoglsl2: Add pass to remove redundant jumps
Ian Romanick [Mon, 13 Sep 2010 21:25:26 +0000 (14:25 -0700)]
glsl2: Add pass to remove redundant jumps

14 years agoglsl: Explain file naming convention
Ian Romanick [Mon, 13 Sep 2010 20:46:29 +0000 (13:46 -0700)]
glsl: Explain file naming convention

14 years agoloop_controls: fix analysis of already analyzed loops
Luca Barbieri [Tue, 7 Sep 2010 15:02:37 +0000 (17:02 +0200)]
loop_controls: fix analysis of already analyzed loops

The loop_controls pass didn't look at the counter values it put in ir_loop
on previous iterations, so while the first iteration worked, subsequent
ones couldn't determine max_iterations.

14 years agoi965: Request that returns be lowered in shader main
Ian Romanick [Mon, 13 Sep 2010 18:05:05 +0000 (11:05 -0700)]
i965: Request that returns be lowered in shader main

Fixes piglit tests glsl-vs-main-return and glsl-fs-main-return.

14 years agoglsl: call ir_lower_jumps according to compiler options
Luca Barbieri [Tue, 7 Sep 2010 00:15:26 +0000 (02:15 +0200)]
glsl: call ir_lower_jumps according to compiler options

14 years agoglsl: add continue/break/return unification/elimination pass (v2)
Luca Barbieri [Mon, 6 Sep 2010 22:24:08 +0000 (00:24 +0200)]
glsl: add continue/break/return unification/elimination pass (v2)

Changes in v2:
- Base class renamed to ir_control_flow_visitor
- Tried to comply with coding style

This is a new pass that supersedes ir_if_return and "lowers" jumps
to if/else structures.

Currently it causes no regressions on softpipe and nv40, but I'm not sure
whether the piglit glsl tests are thorough enough, so consider this
experimental.

It can be asked to:
1. Pull jumps out of ifs where possible
2. Remove all "continue"s, replacing them with an "execute flag"
3. Replace all "break" with a single conditional one at the end of the loop
4. Replace all "return"s with a single return at the end of the function,
   for the main function and/or other functions

This gives several great benefits:
1. All functions can be inlined after this pass
2. nv40 and other pre-DX10 chips without "continue" can be supported
3. nv30 and other pre-DX10 chips with no control flow at all are better supported

Note that for full effect we should also teach the unroller to unroll
loops with a fixed maximum number of iterations but with the canonical
conditional "break" that this pass will insert if asked to.

Continues are lowered by adding a per-loop "execute flag", initialized to
TRUE, that when cleared inhibits all execution until the end of the loop.

Breaks are lowered to continues, plus setting a "break flag" that is checked
at the end of the loop, and trigger the unique "break".

Returns are lowered to breaks/continues, plus adding a "return flag" that
causes loops to break again out of their enclosing loops until all the
loops are exited: then the "execute flag" logic will ignore everything
until the end of the function.

Note that "continue" and "return" can also be implemented by adding
a dummy loop and using break.
However, this is bad for hardware with limited nesting depth, and
prevents further optimization, and thus is not currently performed.

14 years agoglsl: add ir_control_flow_visitor
Luca Barbieri [Mon, 6 Sep 2010 22:22:34 +0000 (00:22 +0200)]
glsl: add ir_control_flow_visitor

This is just a subclass of ir_visitor with empty implementations of all
the visit methods for non-control flow nodes.

Used to avoid duplicating that in ir_visitor subclasses.

ir_hierarchical_visitor is another way to solve this, but is less natural
for some applications.

14 years agollvmpipe: Fix non SSE2 builds.
José Fonseca [Mon, 13 Sep 2010 19:43:36 +0000 (20:43 +0100)]
llvmpipe: Fix non SSE2 builds.

Should fix fdo 30168.

14 years agor300g/swtcl: unlock VBO after draw_flush
Marek Olšák [Mon, 13 Sep 2010 19:08:48 +0000 (21:08 +0200)]
r300g/swtcl: unlock VBO after draw_flush

https://bugs.freedesktop.org/show_bug.cgi?id=29901
https://bugs.freedesktop.org/show_bug.cgi?id=30132

14 years agollvmpipe: Change asm to __asm__.
Witold Baryluk [Mon, 13 Sep 2010 17:57:35 +0000 (18:57 +0100)]
llvmpipe: Change asm to __asm__.

According to gcc documentation both are equivalent,
second are prefered as first can make conflict with existing symbols.

Signed-off-by: José Fonseca <jfonseca@vmware.com>
14 years agoEGL DRI2: 0xa011 is Pineview not Ironlake
Jesse Barnes [Mon, 13 Sep 2010 17:55:16 +0000 (10:55 -0700)]
EGL DRI2: 0xa011 is Pineview not Ironlake

Point about needing a better way to do this validated.

14 years agor600c: const buffer sizes must be a multiple of 16 consts
Alex Deucher [Mon, 13 Sep 2010 17:36:19 +0000 (13:36 -0400)]
r600c: const buffer sizes must be a multiple of 16 consts

This applies to r6xx/r7xx/evergreen

14 years agoEGL DRI2: add PCI ID for Ironlake mobile
Jesse Barnes [Mon, 13 Sep 2010 17:35:45 +0000 (10:35 -0700)]
EGL DRI2: add PCI ID for Ironlake mobile

Allows KMS EGL driver to load.  We need a better way of doing this.

14 years agor600c/eg: remove obselete comment
Alex Deucher [Mon, 13 Sep 2010 16:16:00 +0000 (12:16 -0400)]
r600c/eg: remove obselete comment

14 years agor600c/eg: remove unused emit timestamp function
Alex Deucher [Mon, 13 Sep 2010 16:14:24 +0000 (12:14 -0400)]
r600c/eg: remove unused emit timestamp function

14 years agor600c/eg: emit CB_BLEND_ALPHA with the other blend values
Alex Deucher [Mon, 13 Sep 2010 16:11:29 +0000 (12:11 -0400)]
r600c/eg: emit CB_BLEND_ALPHA with the other blend values

saves a few dwords