[Mesa-dev] [PATCH] ir_to_mesa: Don't optimize clamp into saturate when target is GL_VERTEX_PROGRAM_ARB

[Kenneth Graunke]

On Thursday, November 27, 2014 10:45:00 AM Abdiel Janulgue wrote:
> Add the the same restriction as in the previous try_emit_sat when trying
> to optimize clamp. Fixes an infinite loop in swrast where the lowering
> pass unpacks saturate into clamp but the opt_algebraic pass tries to do
> the opposite.
> 
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=83463
> 
> Tested-by: Vinson Lee <vlee at freedesktop.org>
> Signed-off-by: Abdiel Janulgue <abdiel.janulgue at linux.intel.com>
> ---
>  src/glsl/opt_algebraic.cpp      | 3 ++-
>  src/mesa/main/mtypes.h          | 1 +
>  src/mesa/program/ir_to_mesa.cpp | 3 ++-
>  3 files changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/src/glsl/opt_algebraic.cpp b/src/glsl/opt_algebraic.cpp
> index 430f5cb..62b14dd 100644
> --- a/src/glsl/opt_algebraic.cpp
> +++ b/src/glsl/opt_algebraic.cpp
> @@ -679,7 +679,8 @@ ir_algebraic_visitor::handle_expression(ir_expression *ir)
>  
>     case ir_binop_min:
>     case ir_binop_max:
> -      if (ir->type->base_type != GLSL_TYPE_FLOAT)
> +      if (ir->type->base_type != GLSL_TYPE_FLOAT ||
> +         options->EmitNoSat)
>           break;
>  
>        /* Replace min(max) operations and its commutative combinations with
> diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h
> index 7389baa..199443a 100644
> --- a/src/mesa/main/mtypes.h
> +++ b/src/mesa/main/mtypes.h
> @@ -2990,6 +2990,7 @@ struct gl_shader_compiler_options
>     GLboolean EmitNoMainReturn;            /**< Emit CONT/RET opcodes? */
>     GLboolean EmitNoNoise;                 /**< Emit NOISE opcodes? */
>     GLboolean EmitNoPow;                   /**< Emit POW opcodes? */
> +   GLboolean EmitNoSat;
>     GLboolean LowerClipDistance; /**< Lower gl_ClipDistance from float[8] to vec4[2]? */
>  
>     /**
> diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp
> index 5cd9058..44fd315 100644
> --- a/src/mesa/program/ir_to_mesa.cpp
> +++ b/src/mesa/program/ir_to_mesa.cpp
> @@ -2935,7 +2935,7 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
>  
>        bool progress;
>        exec_list *ir = prog->_LinkedShaders[i]->ir;
> -      const struct gl_shader_compiler_options *options =
> +      struct gl_shader_compiler_options *options =
>              &ctx->Const.ShaderCompilerOptions[prog->_LinkedShaders[i]->Stage];
>  
>        do {
> @@ -2949,6 +2949,7 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
>  				 | ((options->EmitNoPow) ? POW_TO_EXP2 : 0)
>  				 | ((target == GL_VERTEX_PROGRAM_ARB) ? SAT_TO_CLAMP
>                                      : 0)));
> +         options->EmitNoSat = (target == GL_VERTEX_PROGRAM_ARB);
>  
>  	 progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress;

NAK.  You don't get to edit ctx->Const.  It should be set up at context
initialization time (either in core mesa or in the driver), and never
touched after that.

You should:
1. Create an gl_shader_compiler_options::EmitNoSat field (hunk 2 of this patch).
2. Set it appropriately in all relevant drivers:
   - swrast
   - r200
   - i915
   - i965
   - Gallium (st_extensions.c)
   You can skip radeon and nouveau_vieux as they don't support shaders.
3. Make opt_algebraic pay attention to it (hunk 1 of this patch).
4. Change all callers of lower_instructions with SAT_TO_CLAMP to
   use options->EmitNoSat, not target == GL_VERTEX_PROGRAM stuff.

It's unclear to me that we're making the correct decisions today - in other
words, setting EmitNoSat /properly/ for drivers may be a change in behavior.

In particular, in commit cfa8c1cb, your st_glsl_to_tgsi.cpp hunk appears to be
doing SAT_TO_CLAMP lowering only for vertex shaders on pipe drivers that
*support SM3*.  It seems like you meant to do lowering for ones that
*don't* support SM3.

Please try to get this fixed ASAP - we have a release coming up in under a
week; we can't afford to wait two months to attend to simple regressions.

--Ken
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20141129/2146bbbf/attachment.sig>