nir: Add a lower_vec_to_movs pass
authorJason Ekstrand <jason.ekstrand@intel.com>
Wed, 22 Oct 2014 19:57:28 +0000 (12:57 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Thu, 15 Jan 2015 15:19:00 +0000 (07:19 -0800)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
src/glsl/Makefile.sources
src/glsl/nir/nir.h
src/glsl/nir/nir_lower_vec_to_movs.c [new file with mode: 0644]

index 20bf272569c1ba34661ad96b85380fec460552a8..d3b17bd35f6aa68346ba4d7e696f3977d6af214b 100644 (file)
@@ -24,6 +24,7 @@ NIR_FILES = \
        $(GLSL_SRCDIR)/nir/nir_lower_samplers.cpp \
        $(GLSL_SRCDIR)/nir/nir_lower_system_values.c \
        $(GLSL_SRCDIR)/nir/nir_lower_variables_scalar.c \
+       $(GLSL_SRCDIR)/nir/nir_lower_vec_to_movs.c \
        $(GLSL_SRCDIR)/nir/nir_opcodes.c \
        $(GLSL_SRCDIR)/nir/nir_opcodes.h \
        $(GLSL_SRCDIR)/nir/nir_opt_copy_propagate.c \
index 0ddd44a9905fd0912392d6a63a724ae5d9b52172..5c303edb4c0d870c7718acd54e8129a07575b0c4 100644 (file)
@@ -1286,6 +1286,8 @@ void nir_lower_variables_scalar(nir_shader *shader, bool lower_globals,
 
 void nir_remove_dead_variables(nir_shader *shader);
 
+void nir_lower_vec_to_movs(nir_shader *shader);
+
 void nir_lower_samplers(nir_shader *shader,
                         struct gl_shader_program *shader_program,
                         struct gl_program *prog);
diff --git a/src/glsl/nir/nir_lower_vec_to_movs.c b/src/glsl/nir/nir_lower_vec_to_movs.c
new file mode 100644 (file)
index 0000000..a3120b6
--- /dev/null
@@ -0,0 +1,96 @@
+/*
+ * Copyright © 2014 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.
+ *
+ * Authors:
+ *    Jason Ekstrand (jason@jlekstrand.net)
+ *
+ */
+
+#include "nir.h"
+
+/*
+ * Implements a simple pass that lowers vecN instructions to a series of
+ * moves with partial writes.
+ */
+
+static bool
+lower_vec_to_movs_block(nir_block *block, void *mem_ctx)
+{
+   nir_foreach_instr_safe(block, instr) {
+      if (instr->type != nir_instr_type_alu)
+         continue;
+
+      nir_alu_instr *vec = (nir_alu_instr *)instr;
+
+      switch (vec->op) {
+      case nir_op_vec2:
+      case nir_op_vec3:
+      case nir_op_vec4:
+         break;
+      default:
+         continue; /* The loop */
+      }
+
+      for (unsigned i = 0, src_idx = 0; i < 4; i++) {
+         if (!(vec->dest.write_mask & (1 << i)))
+            continue;
+
+         assert(src_idx < nir_op_infos[vec->op].num_inputs);
+
+         nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov);
+         mov->src[0].src = nir_src_copy(vec->src[src_idx].src, mem_ctx);
+         mov->src[0].negate = vec->src[src_idx].negate;
+         mov->src[0].abs = vec->src[src_idx].abs;
+
+         /* We only care about the one swizzle */
+         mov->src[0].swizzle[i] = vec->src[src_idx].swizzle[0];
+
+         mov->dest.dest = nir_dest_copy(vec->dest.dest, mem_ctx);
+         mov->dest.saturate = vec->dest.saturate;
+         mov->dest.write_mask = (1u << i);
+
+         nir_instr_insert_before(&vec->instr, &mov->instr);
+
+         src_idx++;
+      }
+
+      nir_instr_remove(&vec->instr);
+      ralloc_free(vec);
+   }
+
+   return true;
+}
+
+static void
+nir_lower_vec_to_movs_impl(nir_function_impl *impl)
+{
+   nir_foreach_block(impl, lower_vec_to_movs_block, ralloc_parent(impl));
+}
+
+void
+nir_lower_vec_to_movs(nir_shader *shader)
+{
+   nir_foreach_overload(shader, overload) {
+      if (overload->impl)
+         nir_lower_vec_to_movs_impl(overload->impl);
+   }
+}