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#include "nir_control_flow.h"
27#include "nir_vla.h"
28
29static bool function_ends_in_jump(nir_function_impl *impl)
30{
31   nir_block *last_block = nir_impl_last_block(impl);
32   return nir_block_ends_in_jump(last_block);
33}
34
35void nir_inline_function_impl(struct nir_builder *b,
36                              const nir_function_impl *impl,
37                              nir_ssa_def **params,
38                              struct hash_table *shader_var_remap)
39{
40   nir_function_impl *copy = nir_function_impl_clone(b->shader, impl);
41
42   exec_list_append(&b->impl->locals, &copy->locals);
43   exec_list_append(&b->impl->registers, &copy->registers);
44
45   nir_foreach_block(block, copy) {
46      nir_foreach_instr_safe(instr, block) {
47         switch (instr->type) {
48         case nir_instr_type_deref: {
49            nir_deref_instr *deref = nir_instr_as_deref(instr);
50            if (deref->deref_type != nir_deref_type_var)
51               break;
52
53            /* We don't need to remap function variables.  We already cloned
54             * them as part of nir_function_impl_clone and appended them to
55             * b->impl->locals.
56             */
57            if (deref->var->data.mode == nir_var_function_temp)
58               break;
59
60            /* If no map is provided, we assume that there are either no
61             * shader variables or they already live b->shader (this is the
62             * case for function inlining within a single shader.
63             */
64            if (shader_var_remap == NULL)
65               break;
66
67            struct hash_entry *entry =
68               _mesa_hash_table_search(shader_var_remap, deref->var);
69            if (entry == NULL) {
70               nir_variable *nvar = nir_variable_clone(deref->var, b->shader);
71               nir_shader_add_variable(b->shader, nvar);
72               entry = _mesa_hash_table_insert(shader_var_remap,
73                                               deref->var, nvar);
74            }
75            deref->var = entry->data;
76            break;
77         }
78
79         case nir_instr_type_intrinsic: {
80            nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
81            if (load->intrinsic != nir_intrinsic_load_param)
82               break;
83
84            unsigned param_idx = nir_intrinsic_param_idx(load);
85            assert(param_idx < impl->function->num_params);
86            assert(load->dest.is_ssa);
87            nir_ssa_def_rewrite_uses(&load->dest.ssa,
88                                     params[param_idx]);
89
90            /* Remove any left-over load_param intrinsics because they're soon
91             * to be in another function and therefore no longer valid.
92             */
93            nir_instr_remove(&load->instr);
94            break;
95         }
96
97         case nir_instr_type_jump:
98            /* Returns have to be lowered for this to work */
99            assert(nir_instr_as_jump(instr)->type != nir_jump_return);
100            break;
101
102         default:
103            break;
104         }
105      }
106   }
107
108   bool nest_if = function_ends_in_jump(copy);
109
110   /* Pluck the body out of the function and place it here */
111   nir_cf_list body;
112   nir_cf_list_extract(&body, &copy->body);
113
114   if (nest_if) {
115      nir_if *cf = nir_push_if(b, nir_imm_bool(b, true));
116      nir_cf_reinsert(&body, nir_after_cf_list(&cf->then_list));
117      nir_pop_if(b, cf);
118   } else {
119      /* Insert a nop at the cursor so we can keep track of where things are as
120       * we add/remove stuff from the CFG.
121       */
122      nir_intrinsic_instr *nop = nir_nop(b);
123      nir_cf_reinsert(&body, nir_before_instr(&nop->instr));
124      b->cursor = nir_instr_remove(&nop->instr);
125   }
126}
127
128static bool inline_function_impl(nir_function_impl *impl, struct set *inlined);
129
130static bool
131inline_functions_block(nir_block *block, nir_builder *b,
132                       struct set *inlined)
133{
134   bool progress = false;
135   /* This is tricky.  We're iterating over instructions in a block but, as
136    * we go, the block and its instruction list are being split into
137    * pieces.  However, this *should* be safe since foreach_safe always
138    * stashes the next thing in the iteration.  That next thing will
139    * properly get moved to the next block when it gets split, and we
140    * continue iterating there.
141    */
142   nir_foreach_instr_safe(instr, block) {
143      if (instr->type != nir_instr_type_call)
144         continue;
145
146      progress = true;
147
148      nir_call_instr *call = nir_instr_as_call(instr);
149      assert(call->callee->impl);
150
151      /* Make sure that the function we're calling is already inlined */
152      inline_function_impl(call->callee->impl, inlined);
153
154      b->cursor = nir_instr_remove(&call->instr);
155
156      /* Rewrite all of the uses of the callee's parameters to use the call
157       * instructions sources.  In order to ensure that the "load" happens
158       * here and not later (for register sources), we make sure to convert it
159       * to an SSA value first.
160       */
161      const unsigned num_params = call->num_params;
162      NIR_VLA(nir_ssa_def *, params, num_params);
163      for (unsigned i = 0; i < num_params; i++) {
164         params[i] = nir_ssa_for_src(b, call->params[i],
165                                     call->callee->params[i].num_components);
166      }
167
168      nir_inline_function_impl(b, call->callee->impl, params, NULL);
169   }
170
171   return progress;
172}
173
174static bool
175inline_function_impl(nir_function_impl *impl, struct set *inlined)
176{
177   if (_mesa_set_search(inlined, impl))
178      return false; /* Already inlined */
179
180   nir_builder b;
181   nir_builder_init(&b, impl);
182
183   bool progress = false;
184   nir_foreach_block_safe(block, impl) {
185      progress |= inline_functions_block(block, &b, inlined);
186   }
187
188   if (progress) {
189      /* SSA and register indices are completely messed up now */
190      nir_index_ssa_defs(impl);
191      nir_index_local_regs(impl);
192
193      nir_metadata_preserve(impl, nir_metadata_none);
194   } else {
195      nir_metadata_preserve(impl, nir_metadata_all);
196   }
197
198   _mesa_set_add(inlined, impl);
199
200   return progress;
201}
202
203/** A pass to inline all functions in a shader into their callers
204 *
205 * For most use-cases, function inlining is a multi-step process.  The general
206 * pattern employed by SPIR-V consumers and others is as follows:
207 *
208 *  1. nir_lower_variable_initializers(shader, nir_var_function_temp)
209 *
210 *     This is needed because local variables from the callee are simply added
211 *     to the locals list for the caller and the information about where the
212 *     constant initializer logically happens is lost.  If the callee is
213 *     called in a loop, this can cause the variable to go from being
214 *     initialized once per loop iteration to being initialized once at the
215 *     top of the caller and values to persist from one invocation of the
216 *     callee to the next.  The simple solution to this problem is to get rid
217 *     of constant initializers before function inlining.
218 *
219 *  2. nir_lower_returns(shader)
220 *
221 *     nir_inline_functions assumes that all functions end "naturally" by
222 *     execution reaching the end of the function without any return
223 *     instructions causing instant jumps to the end.  Thanks to NIR being
224 *     structured, we can't represent arbitrary jumps to various points in the
225 *     program which is what an early return in the callee would have to turn
226 *     into when we inline it into the caller.  Instead, we require returns to
227 *     be lowered which lets us just copy+paste the callee directly into the
228 *     caller.
229 *
230 *  3. nir_inline_functions(shader)
231 *
232 *     This does the actual function inlining and the resulting shader will
233 *     contain no call instructions.
234 *
235 *  4. nir_opt_deref(shader)
236 *
237 *     Most functions contain pointer parameters where the result of a deref
238 *     instruction is passed in as a parameter, loaded via a load_param
239 *     intrinsic, and then turned back into a deref via a cast.  Function
240 *     inlining will get rid of the load_param but we are still left with a
241 *     cast.  Running nir_opt_deref gets rid of the intermediate cast and
242 *     results in a whole deref chain again.  This is currently required by a
243 *     number of optimizations and lowering passes at least for certain
244 *     variable modes.
245 *
246 *  5. Loop over the functions and delete all but the main entrypoint.
247 *
248 *     In the Intel Vulkan driver this looks like this:
249 *
250 *        foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
251 *           if (func != entry_point)
252 *              exec_node_remove(&func->node);
253 *        }
254 *        assert(exec_list_length(&nir->functions) == 1);
255 *
256 *    While nir_inline_functions does get rid of all call instructions, it
257 *    doesn't get rid of any functions because it doesn't know what the "root
258 *    function" is.  Instead, it's up to the individual driver to know how to
259 *    decide on a root function and delete the rest.  With SPIR-V,
260 *    spirv_to_nir returns the root function and so we can just use == whereas
261 *    with GL, you may have to look for a function named "main".
262 *
263 *  6. nir_lower_variable_initializers(shader, ~nir_var_function_temp)
264 *
265 *     Lowering constant initializers on inputs, outputs, global variables,
266 *     etc. requires that we know the main entrypoint so that we know where to
267 *     initialize them.  Otherwise, we would have to assume that anything
268 *     could be a main entrypoint and initialize them at the start of every
269 *     function but that would clearly be wrong if any of those functions were
270 *     ever called within another function.  Simply requiring a single-
271 *     entrypoint function shader is the best way to make it well-defined.
272 */
273bool
274nir_inline_functions(nir_shader *shader)
275{
276   struct set *inlined = _mesa_pointer_set_create(NULL);
277   bool progress = false;
278
279   nir_foreach_function(function, shader) {
280      if (function->impl)
281         progress = inline_function_impl(function->impl, inlined) || progress;
282   }
283
284   _mesa_set_destroy(inlined, NULL);
285
286   return progress;
287}
288