1/*
2 * Copyright © 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24#include "nir.h"
25#include "nir_builder.h"
26
27/**
28 * \file nir_lower_gs_intrinsics.c
29 *
30 * Geometry Shaders can call EmitVertex()/EmitStreamVertex() to output an
31 * arbitrary number of vertices.  However, the shader must declare the maximum
32 * number of vertices that it will ever output - further attempts to emit
33 * vertices result in undefined behavior according to the GLSL specification.
34 *
35 * Drivers might use this maximum number of vertices to allocate enough space
36 * to hold the geometry shader's output.  Some drivers (such as i965) need to
37 * implement "safety checks" which ensure that the shader hasn't emitted too
38 * many vertices, to avoid overflowing that space and trashing other memory.
39 *
40 * The count of emitted vertices can also be useful in buffer offset
41 * calculations, so drivers know where to write the GS output.
42 *
43 * However, for simple geometry shaders that emit a statically determinable
44 * number of vertices, this extra bookkeeping is unnecessary and inefficient.
45 * By tracking the vertex count in NIR, we allow constant folding/propagation
46 * and dead control flow optimizations to eliminate most of it where possible.
47 *
48 * This pass introduces a new global variable which stores the current vertex
49 * count (initialized to 0), and converts emit_vertex/end_primitive intrinsics
50 * to their *_with_counter variants.  emit_vertex is also wrapped in a safety
51 * check to avoid buffer overflows.  Finally, it adds a set_vertex_count
52 * intrinsic at the end of the program, informing the driver of the final
53 * vertex count.
54 */
55
56struct state {
57   nir_builder *builder;
58   nir_variable *vertex_count_var;
59   bool progress;
60};
61
62/**
63 * Replace emit_vertex intrinsics with:
64 *
65 * if (vertex_count < max_vertices) {
66 *    emit_vertex_with_counter vertex_count ...
67 *    vertex_count += 1
68 * }
69 */
70static void
71rewrite_emit_vertex(nir_intrinsic_instr *intrin, struct state *state)
72{
73   nir_builder *b = state->builder;
74
75   /* Load the vertex count */
76   b->cursor = nir_before_instr(&intrin->instr);
77   nir_ssa_def *count = nir_load_var(b, state->vertex_count_var);
78
79   nir_ssa_def *max_vertices =
80      nir_imm_int(b, b->shader->info.gs.vertices_out);
81
82   /* Create: if (vertex_count < max_vertices) and insert it.
83    *
84    * The new if statement needs to be hooked up to the control flow graph
85    * before we start inserting instructions into it.
86    */
87   nir_push_if(b, nir_ilt(b, count, max_vertices));
88
89   nir_intrinsic_instr *lowered =
90      nir_intrinsic_instr_create(b->shader,
91                                 nir_intrinsic_emit_vertex_with_counter);
92   nir_intrinsic_set_stream_id(lowered, nir_intrinsic_stream_id(intrin));
93   lowered->src[0] = nir_src_for_ssa(count);
94   nir_builder_instr_insert(b, &lowered->instr);
95
96   /* Increment the vertex count by 1 */
97   nir_store_var(b, state->vertex_count_var,
98                 nir_iadd(b, count, nir_imm_int(b, 1)),
99                 0x1); /* .x */
100
101   nir_pop_if(b, NULL);
102
103   nir_instr_remove(&intrin->instr);
104
105   state->progress = true;
106}
107
108/**
109 * Replace end_primitive with end_primitive_with_counter.
110 */
111static void
112rewrite_end_primitive(nir_intrinsic_instr *intrin, struct state *state)
113{
114   nir_builder *b = state->builder;
115
116   b->cursor = nir_before_instr(&intrin->instr);
117   nir_ssa_def *count = nir_load_var(b, state->vertex_count_var);
118
119   nir_intrinsic_instr *lowered =
120      nir_intrinsic_instr_create(b->shader,
121                                 nir_intrinsic_end_primitive_with_counter);
122   nir_intrinsic_set_stream_id(lowered, nir_intrinsic_stream_id(intrin));
123   lowered->src[0] = nir_src_for_ssa(count);
124   nir_builder_instr_insert(b, &lowered->instr);
125
126   nir_instr_remove(&intrin->instr);
127
128   state->progress = true;
129}
130
131static bool
132rewrite_intrinsics(nir_block *block, struct state *state)
133{
134   nir_foreach_instr_safe(instr, block) {
135      if (instr->type != nir_instr_type_intrinsic)
136         continue;
137
138      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
139      switch (intrin->intrinsic) {
140      case nir_intrinsic_emit_vertex:
141         rewrite_emit_vertex(intrin, state);
142         break;
143      case nir_intrinsic_end_primitive:
144         rewrite_end_primitive(intrin, state);
145         break;
146      default:
147         /* not interesting; skip this */
148         break;
149      }
150   }
151
152   return true;
153}
154
155/**
156 * Add a set_vertex_count intrinsic at the end of the program
157 * (representing the final vertex count).
158 */
159static void
160append_set_vertex_count(nir_block *end_block, struct state *state)
161{
162   nir_builder *b = state->builder;
163   nir_shader *shader = state->builder->shader;
164
165   /* Insert the new intrinsic in all of the predecessors of the end block,
166    * but before any jump instructions (return).
167    */
168   set_foreach(end_block->predecessors, entry) {
169      nir_block *pred = (nir_block *) entry->key;
170      b->cursor = nir_after_block_before_jump(pred);
171
172      nir_ssa_def *count = nir_load_var(b, state->vertex_count_var);
173
174      nir_intrinsic_instr *set_vertex_count =
175         nir_intrinsic_instr_create(shader, nir_intrinsic_set_vertex_count);
176      set_vertex_count->src[0] = nir_src_for_ssa(count);
177
178      nir_builder_instr_insert(b, &set_vertex_count->instr);
179   }
180}
181
182bool
183nir_lower_gs_intrinsics(nir_shader *shader)
184{
185   struct state state;
186   state.progress = false;
187
188   nir_function_impl *impl = nir_shader_get_entrypoint(shader);
189   assert(impl);
190
191   nir_builder b;
192   nir_builder_init(&b, impl);
193   state.builder = &b;
194
195   /* Create the counter variable */
196   state.vertex_count_var =
197      nir_local_variable_create(impl, glsl_uint_type(), "vertex_count");
198   /* initialize to 0 */
199   b.cursor = nir_before_cf_list(&impl->body);
200   nir_store_var(&b, state.vertex_count_var, nir_imm_int(&b, 0), 0x1);
201
202   nir_foreach_block_safe(block, impl)
203      rewrite_intrinsics(block, &state);
204
205   /* This only works because we have a single main() function. */
206   append_set_vertex_count(impl->end_block, &state);
207
208   nir_metadata_preserve(impl, 0);
209
210   return state.progress;
211}
212