Re: [Mesa-dev] [PATCH 4/6] nir: Introduce a nir_opt_move_comparisons() pass.

2016-08-10 Thread Jason Ekstrand 
On Wed, Aug 10, 2016 at 11:14 AM, Ian Romanick  wrote:

> On 08/09/2016 07:30 PM, Kenneth Graunke wrote:
> > This tries to move comparisons (a common source of boolean values)
> > closer to their first use.  For GPUs which use condition codes,
> > this can eliminate a lot of temporary booleans and comparisons
> > which reload the condition code register based on a boolean.
> >
> > Signed-off-by: Kenneth Graunke 
> > ---
> >  src/compiler/Makefile.sources   |   1 +
> >  src/compiler/nir/nir.h  |   2 +
> >  src/compiler/nir/nir_opt_move_comparisons.c | 173
> 
> >  3 files changed, 176 insertions(+)
> >  create mode 100644 src/compiler/nir/nir_opt_move_comparisons.c
> >
> > diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.
> sources
> > index 0ff9b23..008a101 100644
> > --- a/src/compiler/Makefile.sources
> > +++ b/src/compiler/Makefile.sources
> > @@ -226,6 +226,7 @@ NIR_FILES = \
> >   nir/nir_opt_gcm.c \
> >   nir/nir_opt_global_to_local.c \
> >   nir/nir_opt_peephole_select.c \
> > + nir/nir_opt_move_comparisons.c \
> >   nir/nir_opt_remove_phis.c \
> >   nir/nir_opt_undef.c \
> >   nir/nir_phi_builder.c \
> > diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> > index 9ce5be2..79511a7 100644
> > --- a/src/compiler/nir/nir.h
> > +++ b/src/compiler/nir/nir.h
> > @@ -2574,6 +2574,8 @@ bool nir_opt_dead_cf(nir_shader *shader);
> >
> >  void nir_opt_gcm(nir_shader *shader);
> >
> > +bool nir_opt_move_comparisons(nir_shader *shader);
> > +
> >  bool nir_opt_peephole_select(nir_shader *shader);
> >
> >  bool nir_opt_remove_phis(nir_shader *shader);
> > diff --git a/src/compiler/nir/nir_opt_move_comparisons.c
> b/src/compiler/nir/nir_opt_move_comparisons.c
> > new file mode 100644
> > index 000..74927c9
> > --- /dev/null
> > +++ b/src/compiler/nir/nir_opt_move_comparisons.c
> > @@ -0,0 +1,173 @@
> > +/*
> > + * Copyright © 2016 Intel Corporation
> > + *
> > + * Permission is hereby granted, free of charge, to any person
> obtaining a
> > + * copy of this software and associated documentation files (the
> "Software"),
> > + * to deal in the Software without restriction, including without
> limitation
> > + * the rights to use, copy, modify, merge, publish, distribute,
> sublicense,
> > + * and/or sell copies of the Software, and to permit persons to whom the
> > + * Software is furnished to do so, subject to the following conditions:
> > + *
> > + * The above copyright notice and this permission notice (including the
> next
> > + * paragraph) shall be included in all copies or substantial portions
> of the
> > + * Software.
> > + *
> > + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> EXPRESS OR
> > + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
> MERCHANTABILITY,
> > + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT
> SHALL
> > + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
> OTHER
> > + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> ARISING
> > + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
> DEALINGS
> > + * IN THE SOFTWARE.
> > + */
> > +
> > +#include "nir.h"
> > +
> > +/**
> > + * \file nir_opt_move_comparisons.c
> > + *
> > + * This pass moves ALU comparison operations just before their first
> use.
> > + *
> > + * It only moves instructions within a single basic block; cross-block
> > + * movement is left to global code motion.
> > + *
> > + * Many GPUs generate condition codes for comparisons, and use
> predication
> > + * for conditional selects and control flow.  In a sequence such as:
> > + *
> > + * vec1 32 ssa_1 = flt a b
> > + * 
> > + * vec1 32 ssa_2 = bcsel ssa_1 c d
> > + *
> > + * the backend would likely do the comparison, producing condition
> codes,
> > + * then save those to a boolean value.  The intervening operations might
> > + * trash the condition codes.  Then, in order to do the bcsel, it would
> > + * need to re-populate the condition code register based on the boolean.
> > + *
> > + * By moving the comparison just before the bcsel, the condition codes
> could
> > + * be used directly.  This eliminates the need to reload them from the
> boolean
> > + * (generally eliminating an instruction).  It may also eliminate the
> need to
> > + * create a boolean value altogether (unless it's used elsewhere),
> which could
> > + * lower register pressure.
> > + */
> > +
> > +static bool
> > +is_comparison(nir_op op)
> > +{
> > +   switch (op) {
> > +   case nir_op_flt:
> > +   case nir_op_fge:
> > +   case nir_op_feq:
> > +   case nir_op_fne:
> > +   case nir_op_ilt:
> > +   case nir_op_ult:
> > +   case nir_op_ige:
> > +   case nir_op_uge:
> > +   case nir_op_ieq:
> > +   case nir_op_ine:
> > +  return true;
> > +   default:
> > +  return false;
> > +   }
> > +}
> > +
> > +static bool
>

Re: [Mesa-dev] [PATCH 4/6] nir: Introduce a nir_opt_move_comparisons() pass.

2016-08-10 Thread Ian Romanick 
On 08/09/2016 07:30 PM, Kenneth Graunke wrote:
> This tries to move comparisons (a common source of boolean values)
> closer to their first use.  For GPUs which use condition codes,
> this can eliminate a lot of temporary booleans and comparisons
> which reload the condition code register based on a boolean.
> 
> Signed-off-by: Kenneth Graunke 
> ---
>  src/compiler/Makefile.sources   |   1 +
>  src/compiler/nir/nir.h  |   2 +
>  src/compiler/nir/nir_opt_move_comparisons.c | 173 
> 
>  3 files changed, 176 insertions(+)
>  create mode 100644 src/compiler/nir/nir_opt_move_comparisons.c
> 
> diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
> index 0ff9b23..008a101 100644
> --- a/src/compiler/Makefile.sources
> +++ b/src/compiler/Makefile.sources
> @@ -226,6 +226,7 @@ NIR_FILES = \
>   nir/nir_opt_gcm.c \
>   nir/nir_opt_global_to_local.c \
>   nir/nir_opt_peephole_select.c \
> + nir/nir_opt_move_comparisons.c \
>   nir/nir_opt_remove_phis.c \
>   nir/nir_opt_undef.c \
>   nir/nir_phi_builder.c \
> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> index 9ce5be2..79511a7 100644
> --- a/src/compiler/nir/nir.h
> +++ b/src/compiler/nir/nir.h
> @@ -2574,6 +2574,8 @@ bool nir_opt_dead_cf(nir_shader *shader);
>  
>  void nir_opt_gcm(nir_shader *shader);
>  
> +bool nir_opt_move_comparisons(nir_shader *shader);
> +
>  bool nir_opt_peephole_select(nir_shader *shader);
>  
>  bool nir_opt_remove_phis(nir_shader *shader);
> diff --git a/src/compiler/nir/nir_opt_move_comparisons.c 
> b/src/compiler/nir/nir_opt_move_comparisons.c
> new file mode 100644
> index 000..74927c9
> --- /dev/null
> +++ b/src/compiler/nir/nir_opt_move_comparisons.c
> @@ -0,0 +1,173 @@
> +/*
> + * Copyright © 2016 Intel Corporation
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice (including the next
> + * paragraph) shall be included in all copies or substantial portions of the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 
> DEALINGS
> + * IN THE SOFTWARE.
> + */
> +
> +#include "nir.h"
> +
> +/**
> + * \file nir_opt_move_comparisons.c
> + *
> + * This pass moves ALU comparison operations just before their first use.
> + *
> + * It only moves instructions within a single basic block; cross-block
> + * movement is left to global code motion.
> + *
> + * Many GPUs generate condition codes for comparisons, and use predication
> + * for conditional selects and control flow.  In a sequence such as:
> + *
> + * vec1 32 ssa_1 = flt a b
> + * 
> + * vec1 32 ssa_2 = bcsel ssa_1 c d
> + *
> + * the backend would likely do the comparison, producing condition codes,
> + * then save those to a boolean value.  The intervening operations might
> + * trash the condition codes.  Then, in order to do the bcsel, it would
> + * need to re-populate the condition code register based on the boolean.
> + *
> + * By moving the comparison just before the bcsel, the condition codes could
> + * be used directly.  This eliminates the need to reload them from the 
> boolean
> + * (generally eliminating an instruction).  It may also eliminate the need to
> + * create a boolean value altogether (unless it's used elsewhere), which 
> could
> + * lower register pressure.
> + */
> +
> +static bool
> +is_comparison(nir_op op)
> +{
> +   switch (op) {
> +   case nir_op_flt:
> +   case nir_op_fge:
> +   case nir_op_feq:
> +   case nir_op_fne:
> +   case nir_op_ilt:
> +   case nir_op_ult:
> +   case nir_op_ige:
> +   case nir_op_uge:
> +   case nir_op_ieq:
> +   case nir_op_ine:
> +  return true;
> +   default:
> +  return false;
> +   }
> +}
> +
> +static bool
> +move_comparison_source(nir_src *src, nir_block *block, struct exec_node 
> *before)
> +{
> +   if (src->is_ssa && src->ssa->parent_instr->block == block &&
> +   src->ssa->parent_instr->type == nir_instr_type_alu &&
> +   is_comparison(nir_instr_as_alu(src->ssa->parent_instr)->op)) {
> +
> +  struct exec_node *src_node =

[Mesa-dev] [PATCH 4/6] nir: Introduce a nir_opt_move_comparisons() pass.

2016-08-09 Thread Kenneth Graunke 
This tries to move comparisons (a common source of boolean values)
closer to their first use.  For GPUs which use condition codes,
this can eliminate a lot of temporary booleans and comparisons
which reload the condition code register based on a boolean.

Signed-off-by: Kenneth Graunke 
---
 src/compiler/Makefile.sources   |   1 +
 src/compiler/nir/nir.h  |   2 +
 src/compiler/nir/nir_opt_move_comparisons.c | 173 
 3 files changed, 176 insertions(+)
 create mode 100644 src/compiler/nir/nir_opt_move_comparisons.c

diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
index 0ff9b23..008a101 100644
--- a/src/compiler/Makefile.sources
+++ b/src/compiler/Makefile.sources
@@ -226,6 +226,7 @@ NIR_FILES = \
nir/nir_opt_gcm.c \
nir/nir_opt_global_to_local.c \
nir/nir_opt_peephole_select.c \
+   nir/nir_opt_move_comparisons.c \
nir/nir_opt_remove_phis.c \
nir/nir_opt_undef.c \
nir/nir_phi_builder.c \
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 9ce5be2..79511a7 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2574,6 +2574,8 @@ bool nir_opt_dead_cf(nir_shader *shader);
 
 void nir_opt_gcm(nir_shader *shader);
 
+bool nir_opt_move_comparisons(nir_shader *shader);
+
 bool nir_opt_peephole_select(nir_shader *shader);
 
 bool nir_opt_remove_phis(nir_shader *shader);
diff --git a/src/compiler/nir/nir_opt_move_comparisons.c 
b/src/compiler/nir/nir_opt_move_comparisons.c
new file mode 100644
index 000..74927c9
--- /dev/null
+++ b/src/compiler/nir/nir_opt_move_comparisons.c
@@ -0,0 +1,173 @@
+/*
+ * Copyright © 2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "nir.h"
+
+/**
+ * \file nir_opt_move_comparisons.c
+ *
+ * This pass moves ALU comparison operations just before their first use.
+ *
+ * It only moves instructions within a single basic block; cross-block
+ * movement is left to global code motion.
+ *
+ * Many GPUs generate condition codes for comparisons, and use predication
+ * for conditional selects and control flow.  In a sequence such as:
+ *
+ * vec1 32 ssa_1 = flt a b
+ * 
+ * vec1 32 ssa_2 = bcsel ssa_1 c d
+ *
+ * the backend would likely do the comparison, producing condition codes,
+ * then save those to a boolean value.  The intervening operations might
+ * trash the condition codes.  Then, in order to do the bcsel, it would
+ * need to re-populate the condition code register based on the boolean.
+ *
+ * By moving the comparison just before the bcsel, the condition codes could
+ * be used directly.  This eliminates the need to reload them from the boolean
+ * (generally eliminating an instruction).  It may also eliminate the need to
+ * create a boolean value altogether (unless it's used elsewhere), which could
+ * lower register pressure.
+ */
+
+static bool
+is_comparison(nir_op op)
+{
+   switch (op) {
+   case nir_op_flt:
+   case nir_op_fge:
+   case nir_op_feq:
+   case nir_op_fne:
+   case nir_op_ilt:
+   case nir_op_ult:
+   case nir_op_ige:
+   case nir_op_uge:
+   case nir_op_ieq:
+   case nir_op_ine:
+  return true;
+   default:
+  return false;
+   }
+}
+
+static bool
+move_comparison_source(nir_src *src, nir_block *block, struct exec_node 
*before)
+{
+   if (src->is_ssa && src->ssa->parent_instr->block == block &&
+   src->ssa->parent_instr->type == nir_instr_type_alu &&
+   is_comparison(nir_instr_as_alu(src->ssa->parent_instr)->op)) {
+
+  struct exec_node *src_node = >ssa->parent_instr->node;
+  exec_node_remove(src_node);
+
+  if (before)
+ exec_node_insert_node_before(before, src_node);
+  else
+ exec_list_push_tail(>instr_list, src_node);
+
+  return true;
+   }
+
+   return false;
+}
+
+/* nir_foreach_src callback