2 * Copyright © 2015 Intel Corporation
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:
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
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
25 #include "nir_builder.h"
26 #include "nir_control_flow.h"
29 void nir_inline_function_impl(struct nir_builder
*b
,
30 const nir_function_impl
*impl
,
33 nir_function_impl
*copy
= nir_function_impl_clone(b
->shader
, impl
);
35 /* Insert a nop at the cursor so we can keep track of where things are as
36 * we add/remove stuff from the CFG.
38 nir_intrinsic_instr
*nop
=
39 nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_nop
);
40 nir_builder_instr_insert(b
, &nop
->instr
);
42 exec_list_append(&b
->impl
->locals
, ©
->locals
);
43 exec_list_append(&b
->impl
->registers
, ©
->registers
);
45 nir_foreach_block(block
, copy
) {
46 nir_foreach_instr_safe(instr
, block
) {
47 /* Returns have to be lowered for this to work */
48 assert(instr
->type
!= nir_instr_type_jump
||
49 nir_instr_as_jump(instr
)->type
!= nir_jump_return
);
51 if (instr
->type
!= nir_instr_type_intrinsic
)
54 nir_intrinsic_instr
*load
= nir_instr_as_intrinsic(instr
);
55 if (load
->intrinsic
!= nir_intrinsic_load_param
)
58 unsigned param_idx
= nir_intrinsic_param_idx(load
);
59 assert(param_idx
< impl
->function
->num_params
);
60 assert(load
->dest
.is_ssa
);
61 nir_ssa_def_rewrite_uses(&load
->dest
.ssa
,
62 nir_src_for_ssa(params
[param_idx
]));
64 /* Remove any left-over load_param intrinsics because they're soon
65 * to be in another function and therefore no longer valid.
67 nir_instr_remove(&load
->instr
);
71 /* Pluck the body out of the function and place it here */
73 nir_cf_list_extract(&body
, ©
->body
);
74 nir_cf_reinsert(&body
, nir_before_instr(&nop
->instr
));
76 b
->cursor
= nir_instr_remove(&nop
->instr
);
79 static bool inline_function_impl(nir_function_impl
*impl
, struct set
*inlined
);
82 inline_functions_block(nir_block
*block
, nir_builder
*b
,
85 bool progress
= false;
86 /* This is tricky. We're iterating over instructions in a block but, as
87 * we go, the block and its instruction list are being split into
88 * pieces. However, this *should* be safe since foreach_safe always
89 * stashes the next thing in the iteration. That next thing will
90 * properly get moved to the next block when it gets split, and we
91 * continue iterating there.
93 nir_foreach_instr_safe(instr
, block
) {
94 if (instr
->type
!= nir_instr_type_call
)
99 nir_call_instr
*call
= nir_instr_as_call(instr
);
100 assert(call
->callee
->impl
);
102 /* Make sure that the function we're calling is already inlined */
103 inline_function_impl(call
->callee
->impl
, inlined
);
105 b
->cursor
= nir_instr_remove(&call
->instr
);
107 /* Rewrite all of the uses of the callee's parameters to use the call
108 * instructions sources. In order to ensure that the "load" happens
109 * here and not later (for register sources), we make sure to convert it
110 * to an SSA value first.
112 const unsigned num_params
= call
->num_params
;
113 NIR_VLA(nir_ssa_def
*, params
, num_params
);
114 for (unsigned i
= 0; i
< num_params
; i
++) {
115 params
[i
] = nir_ssa_for_src(b
, call
->params
[i
],
116 call
->callee
->params
[i
].num_components
);
119 nir_inline_function_impl(b
, call
->callee
->impl
, params
);
126 inline_function_impl(nir_function_impl
*impl
, struct set
*inlined
)
128 if (_mesa_set_search(inlined
, impl
))
129 return false; /* Already inlined */
132 nir_builder_init(&b
, impl
);
134 bool progress
= false;
135 nir_foreach_block_safe(block
, impl
) {
136 progress
|= inline_functions_block(block
, &b
, inlined
);
140 /* SSA and register indices are completely messed up now */
141 nir_index_ssa_defs(impl
);
142 nir_index_local_regs(impl
);
144 nir_metadata_preserve(impl
, nir_metadata_none
);
147 impl
->valid_metadata
&= ~nir_metadata_not_properly_reset
;
151 _mesa_set_add(inlined
, impl
);
156 /** A pass to inline all functions in a shader into their callers
158 * For most use-cases, function inlining is a multi-step process. The general
159 * pattern employed by SPIR-V consumers and others is as follows:
161 * 1. nir_lower_variable_initializers(shader, nir_var_function_temp)
163 * This is needed because local variables from the callee are simply added
164 * to the locals list for the caller and the information about where the
165 * constant initializer logically happens is lost. If the callee is
166 * called in a loop, this can cause the variable to go from being
167 * initialized once per loop iteration to being initialized once at the
168 * top of the caller and values to persist from one invocation of the
169 * callee to the next. The simple solution to this problem is to get rid
170 * of constant initializers before function inlining.
172 * 2. nir_lower_returns(shader)
174 * nir_inline_functions assumes that all functions end "naturally" by
175 * execution reaching the end of the function without any return
176 * instructions causing instant jumps to the end. Thanks to NIR being
177 * structured, we can't represent arbitrary jumps to various points in the
178 * program which is what an early return in the callee would have to turn
179 * into when we inline it into the caller. Instead, we require returns to
180 * be lowered which lets us just copy+paste the callee directly into the
183 * 3. nir_inline_functions(shader)
185 * This does the actual function inlining and the resulting shader will
186 * contain no call instructions.
188 * 4. nir_opt_deref(shader)
190 * Most functions contain pointer parameters where the result of a deref
191 * instruction is passed in as a parameter, loaded via a load_param
192 * intrinsic, and then turned back into a deref via a cast. Function
193 * inlining will get rid of the load_param but we are still left with a
194 * cast. Running nir_opt_deref gets rid of the intermediate cast and
195 * results in a whole deref chain again. This is currently required by a
196 * number of optimizations and lowering passes at least for certain
199 * 5. Loop over the functions and delete all but the main entrypoint.
201 * In the Intel Vulkan driver this looks like this:
203 * foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
204 * if (func != entry_point)
205 * exec_node_remove(&func->node);
207 * assert(exec_list_length(&nir->functions) == 1);
209 * While nir_inline_functions does get rid of all call instructions, it
210 * doesn't get rid of any functions because it doesn't know what the "root
211 * function" is. Instead, it's up to the individual driver to know how to
212 * decide on a root function and delete the rest. With SPIR-V,
213 * spirv_to_nir returns the root function and so we can just use == whereas
214 * with GL, you may have to look for a function named "main".
216 * 6. nir_lower_variable_initializers(shader, ~nir_var_function_temp)
218 * Lowering constant initializers on inputs, outputs, global variables,
219 * etc. requires that we know the main entrypoint so that we know where to
220 * initialize them. Otherwise, we would have to assume that anything
221 * could be a main entrypoint and initialize them at the start of every
222 * function but that would clearly be wrong if any of those functions were
223 * ever called within another function. Simply requiring a single-
224 * entrypoint function shader is the best way to make it well-defined.
227 nir_inline_functions(nir_shader
*shader
)
229 struct set
*inlined
= _mesa_pointer_set_create(NULL
);
230 bool progress
= false;
232 nir_foreach_function(function
, shader
) {
234 progress
= inline_function_impl(function
->impl
, inlined
) || progress
;
237 _mesa_set_destroy(inlined
, NULL
);