| 1 |
.. Copyright (C) 2014-2025 Free Software Foundation, Inc. |
| 2 |
Originally contributed by David Malcolm <[email protected]> |
| 3 |
|
| 4 |
This is free software: you can redistribute it and/or modify it |
| 5 |
under the terms of the GNU General Public License as published by |
| 6 |
the Free Software Foundation, either version 3 of the License, or |
| 7 |
(at your option) any later version. |
| 8 |
|
| 9 |
This program is distributed in the hope that it will be useful, but |
| 10 |
WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 12 |
General Public License for more details. |
| 13 |
|
| 14 |
You should have received a copy of the GNU General Public License |
| 15 |
along with this program. If not, see |
| 16 |
<https://www.gnu.org/licenses/>. |
| 17 |
|
| 18 |
.. default-domain:: cpp |
| 19 |
|
| 20 |
Tutorial part 2: Creating a trivial machine code function |
| 21 |
--------------------------------------------------------- |
| 22 |
|
| 23 |
Consider this C function: |
| 24 |
|
| 25 |
.. code-block:: c |
| 26 |
|
| 27 |
int square (int i) |
| 28 |
{ |
| 29 |
return i * i; |
| 30 |
} |
| 31 |
|
| 32 |
How can we construct this at run-time using libgccjit's C++ API? |
| 33 |
|
| 34 |
First we need to include the relevant header: |
| 35 |
|
| 36 |
.. code-block:: c++ |
| 37 |
|
| 38 |
#include <libgccjit++.h> |
| 39 |
|
| 40 |
All state associated with compilation is associated with a |
| 41 |
:type:`gccjit::context`, which is a thin C++ wrapper around the C API's |
| 42 |
:c:expr:`gcc_jit_context *`. |
| 43 |
|
| 44 |
Create one using :func:`gccjit::context::acquire`: |
| 45 |
|
| 46 |
.. code-block:: c++ |
| 47 |
|
| 48 |
gccjit::context ctxt; |
| 49 |
ctxt = gccjit::context::acquire (); |
| 50 |
|
| 51 |
The JIT library has a system of types. It is statically-typed: every |
| 52 |
expression is of a specific type, fixed at compile-time. In our example, |
| 53 |
all of the expressions are of the C `int` type, so let's obtain this from |
| 54 |
the context, as a :type:`gccjit::type`, using |
| 55 |
:func:`gccjit::context::get_type`: |
| 56 |
|
| 57 |
.. code-block:: c++ |
| 58 |
|
| 59 |
gccjit::type int_type = ctxt.get_type (GCC_JIT_TYPE_INT); |
| 60 |
|
| 61 |
:type:`gccjit::type` is an example of a "contextual" object: every |
| 62 |
entity in the API is associated with a :type:`gccjit::context`. |
| 63 |
|
| 64 |
Memory management is easy: all such "contextual" objects are automatically |
| 65 |
cleaned up for you when the context is released, using |
| 66 |
:func:`gccjit::context::release`: |
| 67 |
|
| 68 |
.. code-block:: c++ |
| 69 |
|
| 70 |
ctxt.release (); |
| 71 |
|
| 72 |
so you don't need to manually track and cleanup all objects, just the |
| 73 |
contexts. |
| 74 |
|
| 75 |
All of the C++ classes in the API are thin wrappers around pointers to |
| 76 |
types in the C API. |
| 77 |
|
| 78 |
The C++ class hierarchy within the ``gccjit`` namespace looks like this:: |
| 79 |
|
| 80 |
+- object |
| 81 |
+- location |
| 82 |
+- type |
| 83 |
+- struct |
| 84 |
+- field |
| 85 |
+- function |
| 86 |
+- block |
| 87 |
+- rvalue |
| 88 |
+- lvalue |
| 89 |
+- param |
| 90 |
|
| 91 |
One thing you can do with a :type:`gccjit::object` is |
| 92 |
to ask it for a human-readable description as a :type:`std::string`, using |
| 93 |
:func:`gccjit::object::get_debug_string`: |
| 94 |
|
| 95 |
.. code-block:: c++ |
| 96 |
|
| 97 |
printf ("obj: %s\n", obj.get_debug_string ().c_str ()); |
| 98 |
|
| 99 |
giving this text on stdout: |
| 100 |
|
| 101 |
.. code-block:: bash |
| 102 |
|
| 103 |
obj: int |
| 104 |
|
| 105 |
This is invaluable when debugging. |
| 106 |
|
| 107 |
Let's create the function. To do so, we first need to construct |
| 108 |
its single parameter, specifying its type and giving it a name, |
| 109 |
using :func:`gccjit::context::new_param`: |
| 110 |
|
| 111 |
.. code-block:: c++ |
| 112 |
|
| 113 |
gccjit::param param_i = ctxt.new_param (int_type, "i"); |
| 114 |
|
| 115 |
and we can then make a vector of all of the params of the function, |
| 116 |
in this case just one: |
| 117 |
|
| 118 |
.. code-block:: c++ |
| 119 |
|
| 120 |
std::vector<gccjit::param> params; |
| 121 |
params.push_back (param_i); |
| 122 |
|
| 123 |
Now we can create the function, using |
| 124 |
:cpp:func:`gccjit::context::new_function`: |
| 125 |
|
| 126 |
.. code-block:: c++ |
| 127 |
|
| 128 |
gccjit::function func = |
| 129 |
ctxt.new_function (GCC_JIT_FUNCTION_EXPORTED, |
| 130 |
int_type, |
| 131 |
"square", |
| 132 |
params, |
| 133 |
0); |
| 134 |
|
| 135 |
To define the code within the function, we must create basic blocks |
| 136 |
containing statements. |
| 137 |
|
| 138 |
Every basic block contains a list of statements, eventually terminated |
| 139 |
by a statement that either returns, or jumps to another basic block. |
| 140 |
|
| 141 |
Our function has no control-flow, so we just need one basic block: |
| 142 |
|
| 143 |
.. code-block:: c++ |
| 144 |
|
| 145 |
gccjit::block block = func.new_block (); |
| 146 |
|
| 147 |
Our basic block is relatively simple: it immediately terminates by |
| 148 |
returning the value of an expression. |
| 149 |
|
| 150 |
We can build the expression using :func:`gccjit::context::new_binary_op`: |
| 151 |
|
| 152 |
.. code-block:: c++ |
| 153 |
|
| 154 |
gccjit::rvalue expr = |
| 155 |
ctxt.new_binary_op ( |
| 156 |
GCC_JIT_BINARY_OP_MULT, int_type, |
| 157 |
param_i, param_i); |
| 158 |
|
| 159 |
A :type:`gccjit::rvalue` is another example of a |
| 160 |
:type:`gccjit::object` subclass. As before, we can print it with |
| 161 |
:func:`gccjit::object::get_debug_string`. |
| 162 |
|
| 163 |
.. code-block:: c++ |
| 164 |
|
| 165 |
printf ("expr: %s\n", expr.get_debug_string ().c_str ()); |
| 166 |
|
| 167 |
giving this output: |
| 168 |
|
| 169 |
.. code-block:: bash |
| 170 |
|
| 171 |
expr: i * i |
| 172 |
|
| 173 |
Note that :type:`gccjit::rvalue` provides numerous overloaded operators |
| 174 |
which can be used to dramatically reduce the amount of typing needed. |
| 175 |
We can build the above binary operation more directly with this one-liner: |
| 176 |
|
| 177 |
.. code-block:: c++ |
| 178 |
|
| 179 |
gccjit::rvalue expr = param_i * param_i; |
| 180 |
|
| 181 |
Creating the expression in itself doesn't do anything; we have to add |
| 182 |
this expression to a statement within the block. In this case, we use it |
| 183 |
to build a return statement, which terminates the basic block: |
| 184 |
|
| 185 |
.. code-block:: c++ |
| 186 |
|
| 187 |
block.end_with_return (expr); |
| 188 |
|
| 189 |
OK, we've populated the context. We can now compile it using |
| 190 |
:func:`gccjit::context::compile`: |
| 191 |
|
| 192 |
.. code-block:: c++ |
| 193 |
|
| 194 |
gcc_jit_result *result; |
| 195 |
result = ctxt.compile (); |
| 196 |
|
| 197 |
and get a :c:expr:`gcc_jit_result *`. |
| 198 |
|
| 199 |
We can now use :c:func:`gcc_jit_result_get_code` to look up a specific |
| 200 |
machine code routine within the result, in this case, the function we |
| 201 |
created above. |
| 202 |
|
| 203 |
.. code-block:: c++ |
| 204 |
|
| 205 |
void *fn_ptr = gcc_jit_result_get_code (result, "square"); |
| 206 |
if (!fn_ptr) |
| 207 |
{ |
| 208 |
fprintf (stderr, "NULL fn_ptr"); |
| 209 |
goto error; |
| 210 |
} |
| 211 |
|
| 212 |
We can now cast the pointer to an appropriate function pointer type, and |
| 213 |
then call it: |
| 214 |
|
| 215 |
.. code-block:: c++ |
| 216 |
|
| 217 |
typedef int (*fn_type) (int); |
| 218 |
fn_type square = (fn_type)fn_ptr; |
| 219 |
printf ("result: %d", square (5)); |
| 220 |
|
| 221 |
.. code-block:: bash |
| 222 |
|
| 223 |
result: 25 |
| 224 |
|
| 225 |
|
| 226 |
Options |
| 227 |
******* |
| 228 |
|
| 229 |
To get more information on what's going on, you can set debugging flags |
| 230 |
on the context using :func:`gccjit::context::set_bool_option`. |
| 231 |
|
| 232 |
.. (I'm deliberately not mentioning |
| 233 |
:c:macro:`GCC_JIT_BOOL_OPTION_DUMP_INITIAL_TREE` here since I think |
| 234 |
it's probably more of use to implementors than to users) |
| 235 |
|
| 236 |
Setting :c:macro:`GCC_JIT_BOOL_OPTION_DUMP_INITIAL_GIMPLE` will dump a |
| 237 |
C-like representation to stderr when you compile (GCC's "GIMPLE" |
| 238 |
representation): |
| 239 |
|
| 240 |
.. code-block:: c++ |
| 241 |
|
| 242 |
ctxt.set_bool_option (GCC_JIT_BOOL_OPTION_DUMP_INITIAL_GIMPLE, 1); |
| 243 |
result = ctxt.compile (); |
| 244 |
|
| 245 |
.. code-block:: c |
| 246 |
|
| 247 |
square (signed int i) |
| 248 |
{ |
| 249 |
signed int D.260; |
| 250 |
|
| 251 |
entry: |
| 252 |
D.260 = i * i; |
| 253 |
return D.260; |
| 254 |
} |
| 255 |
|
| 256 |
We can see the generated machine code in assembler form (on stderr) by |
| 257 |
setting :c:macro:`GCC_JIT_BOOL_OPTION_DUMP_GENERATED_CODE` on the context |
| 258 |
before compiling: |
| 259 |
|
| 260 |
.. code-block:: c++ |
| 261 |
|
| 262 |
ctxt.set_bool_option (GCC_JIT_BOOL_OPTION_DUMP_GENERATED_CODE, 1); |
| 263 |
result = ctxt.compile (); |
| 264 |
|
| 265 |
.. code-block:: gas |
| 266 |
|
| 267 |
.file "fake.c" |
| 268 |
.text |
| 269 |
.globl square |
| 270 |
.type square, @function |
| 271 |
square: |
| 272 |
.LFB6: |
| 273 |
.cfi_startproc |
| 274 |
pushq %rbp |
| 275 |
.cfi_def_cfa_offset 16 |
| 276 |
.cfi_offset 6, -16 |
| 277 |
movq %rsp, %rbp |
| 278 |
.cfi_def_cfa_register 6 |
| 279 |
movl %edi, -4(%rbp) |
| 280 |
.L14: |
| 281 |
movl -4(%rbp), %eax |
| 282 |
imull -4(%rbp), %eax |
| 283 |
popq %rbp |
| 284 |
.cfi_def_cfa 7, 8 |
| 285 |
ret |
| 286 |
.cfi_endproc |
| 287 |
.LFE6: |
| 288 |
.size square, .-square |
| 289 |
.ident "GCC: (GNU) 4.9.0 20131023 (Red Hat 0.2)" |
| 290 |
.section .note.GNU-stack,"",@progbits |
| 291 |
|
| 292 |
By default, no optimizations are performed, the equivalent of GCC's |
| 293 |
`-O0` option. We can turn things up to e.g. `-O3` by calling |
| 294 |
:func:`gccjit::context::set_int_option` with |
| 295 |
:c:macro:`GCC_JIT_INT_OPTION_OPTIMIZATION_LEVEL`: |
| 296 |
|
| 297 |
.. code-block:: c++ |
| 298 |
|
| 299 |
ctxt.set_int_option (GCC_JIT_INT_OPTION_OPTIMIZATION_LEVEL, 3); |
| 300 |
|
| 301 |
.. code-block:: gas |
| 302 |
|
| 303 |
.file "fake.c" |
| 304 |
.text |
| 305 |
.p2align 4,,15 |
| 306 |
.globl square |
| 307 |
.type square, @function |
| 308 |
square: |
| 309 |
.LFB7: |
| 310 |
.cfi_startproc |
| 311 |
.L16: |
| 312 |
movl %edi, %eax |
| 313 |
imull %edi, %eax |
| 314 |
ret |
| 315 |
.cfi_endproc |
| 316 |
.LFE7: |
| 317 |
.size square, .-square |
| 318 |
.ident "GCC: (GNU) 4.9.0 20131023 (Red Hat 0.2)" |
| 319 |
.section .note.GNU-stack,"",@progbits |
| 320 |
|
| 321 |
Naturally this has only a small effect on such a trivial function. |
| 322 |
|
| 323 |
|
| 324 |
Full example |
| 325 |
************ |
| 326 |
|
| 327 |
Here's what the above looks like as a complete program: |
| 328 |
|
| 329 |
.. literalinclude:: ../../examples/tut02-square.cc |
| 330 |
:lines: 1- |
| 331 |
:language: c++ |
| 332 |
|
| 333 |
Building and running it: |
| 334 |
|
| 335 |
.. code-block:: console |
| 336 |
|
| 337 |
$ gcc \ |
| 338 |
tut02-square.cc \ |
| 339 |
-o tut02-square \ |
| 340 |
-lgccjit |
| 341 |
|
| 342 |
# Run the built program: |
| 343 |
$ ./tut02-square |
| 344 |
result: 25 |