root / tcg / README @ f24cb33e
History | View | Annotate | Download (12.5 kB)
1 |
Tiny Code Generator - Fabrice Bellard. |
---|---|
2 |
|
3 |
1) Introduction |
4 |
|
5 |
TCG (Tiny Code Generator) began as a generic backend for a C |
6 |
compiler. It was simplified to be used in QEMU. It also has its roots |
7 |
in the QOP code generator written by Paul Brook. |
8 |
|
9 |
2) Definitions |
10 |
|
11 |
The TCG "target" is the architecture for which we generate the |
12 |
code. It is of course not the same as the "target" of QEMU which is |
13 |
the emulated architecture. As TCG started as a generic C backend used |
14 |
for cross compiling, it is assumed that the TCG target is different |
15 |
from the host, although it is never the case for QEMU. |
16 |
|
17 |
A TCG "function" corresponds to a QEMU Translated Block (TB). |
18 |
|
19 |
A TCG "temporary" is a variable only live in a basic |
20 |
block. Temporaries are allocated explicitly in each function. |
21 |
|
22 |
A TCG "local temporary" is a variable only live in a function. Local |
23 |
temporaries are allocated explicitly in each function. |
24 |
|
25 |
A TCG "global" is a variable which is live in all the functions |
26 |
(equivalent of a C global variable). They are defined before the |
27 |
functions defined. A TCG global can be a memory location (e.g. a QEMU |
28 |
CPU register), a fixed host register (e.g. the QEMU CPU state pointer) |
29 |
or a memory location which is stored in a register outside QEMU TBs |
30 |
(not implemented yet). |
31 |
|
32 |
A TCG "basic block" corresponds to a list of instructions terminated |
33 |
by a branch instruction. |
34 |
|
35 |
3) Intermediate representation |
36 |
|
37 |
3.1) Introduction |
38 |
|
39 |
TCG instructions operate on variables which are temporaries, local |
40 |
temporaries or globals. TCG instructions and variables are strongly |
41 |
typed. Two types are supported: 32 bit integers and 64 bit |
42 |
integers. Pointers are defined as an alias to 32 bit or 64 bit |
43 |
integers depending on the TCG target word size. |
44 |
|
45 |
Each instruction has a fixed number of output variable operands, input |
46 |
variable operands and always constant operands. |
47 |
|
48 |
The notable exception is the call instruction which has a variable |
49 |
number of outputs and inputs. |
50 |
|
51 |
In the textual form, output operands usually come first, followed by |
52 |
input operands, followed by constant operands. The output type is |
53 |
included in the instruction name. Constants are prefixed with a '$'. |
54 |
|
55 |
add_i32 t0, t1, t2 (t0 <- t1 + t2) |
56 |
|
57 |
3.2) Assumptions |
58 |
|
59 |
* Basic blocks |
60 |
|
61 |
- Basic blocks end after branches (e.g. brcond_i32 instruction), |
62 |
goto_tb and exit_tb instructions. |
63 |
- Basic blocks end before legacy dyngen operations. |
64 |
- Basic blocks start after the end of a previous basic block, at a |
65 |
set_label instruction or after a legacy dyngen operation. |
66 |
|
67 |
After the end of a basic block, the content of temporaries is |
68 |
destroyed, but local temporaries and globals are preserved. |
69 |
|
70 |
* Floating point types are not supported yet |
71 |
|
72 |
* Pointers: depending on the TCG target, pointer size is 32 bit or 64 |
73 |
bit. The type TCG_TYPE_PTR is an alias to TCG_TYPE_I32 or |
74 |
TCG_TYPE_I64. |
75 |
|
76 |
* Helpers: |
77 |
|
78 |
Using the tcg_gen_helper_x_y it is possible to call any function |
79 |
taking i32, i64 or pointer types. Before calling an helper, all |
80 |
globals are stored at their canonical location and it is assumed that |
81 |
the function can modify them. In the future, function modifiers will |
82 |
be allowed to tell that the helper does not read or write some globals. |
83 |
|
84 |
On some TCG targets (e.g. x86), several calling conventions are |
85 |
supported. |
86 |
|
87 |
* Branches: |
88 |
|
89 |
Use the instruction 'br' to jump to a label. Use 'jmp' to jump to an |
90 |
explicit address. Conditional branches can only jump to labels. |
91 |
|
92 |
3.3) Code Optimizations |
93 |
|
94 |
When generating instructions, you can count on at least the following |
95 |
optimizations: |
96 |
|
97 |
- Single instructions are simplified, e.g. |
98 |
|
99 |
and_i32 t0, t0, $0xffffffff |
100 |
|
101 |
is suppressed. |
102 |
|
103 |
- A liveness analysis is done at the basic block level. The |
104 |
information is used to suppress moves from a dead variable to |
105 |
another one. It is also used to remove instructions which compute |
106 |
dead results. The later is especially useful for condition code |
107 |
optimization in QEMU. |
108 |
|
109 |
In the following example: |
110 |
|
111 |
add_i32 t0, t1, t2 |
112 |
add_i32 t0, t0, $1 |
113 |
mov_i32 t0, $1 |
114 |
|
115 |
only the last instruction is kept. |
116 |
|
117 |
3.4) Instruction Reference |
118 |
|
119 |
********* Function call |
120 |
|
121 |
* call <ret> <params> ptr |
122 |
|
123 |
call function 'ptr' (pointer type) |
124 |
|
125 |
<ret> optional 32 bit or 64 bit return value |
126 |
<params> optional 32 bit or 64 bit parameters |
127 |
|
128 |
********* Jumps/Labels |
129 |
|
130 |
* jmp t0 |
131 |
|
132 |
Absolute jump to address t0 (pointer type). |
133 |
|
134 |
* set_label $label |
135 |
|
136 |
Define label 'label' at the current program point. |
137 |
|
138 |
* br $label |
139 |
|
140 |
Jump to label. |
141 |
|
142 |
* brcond_i32/i64 cond, t0, t1, label |
143 |
|
144 |
Conditional jump if t0 cond t1 is true. cond can be: |
145 |
TCG_COND_EQ |
146 |
TCG_COND_NE |
147 |
TCG_COND_LT /* signed */ |
148 |
TCG_COND_GE /* signed */ |
149 |
TCG_COND_LE /* signed */ |
150 |
TCG_COND_GT /* signed */ |
151 |
TCG_COND_LTU /* unsigned */ |
152 |
TCG_COND_GEU /* unsigned */ |
153 |
TCG_COND_LEU /* unsigned */ |
154 |
TCG_COND_GTU /* unsigned */ |
155 |
|
156 |
********* Arithmetic |
157 |
|
158 |
* add_i32/i64 t0, t1, t2 |
159 |
|
160 |
t0=t1+t2 |
161 |
|
162 |
* sub_i32/i64 t0, t1, t2 |
163 |
|
164 |
t0=t1-t2 |
165 |
|
166 |
* neg_i32/i64 t0, t1 |
167 |
|
168 |
t0=-t1 (two's complement) |
169 |
|
170 |
* mul_i32/i64 t0, t1, t2 |
171 |
|
172 |
t0=t1*t2 |
173 |
|
174 |
* div_i32/i64 t0, t1, t2 |
175 |
|
176 |
t0=t1/t2 (signed). Undefined behavior if division by zero or overflow. |
177 |
|
178 |
* divu_i32/i64 t0, t1, t2 |
179 |
|
180 |
t0=t1/t2 (unsigned). Undefined behavior if division by zero. |
181 |
|
182 |
* rem_i32/i64 t0, t1, t2 |
183 |
|
184 |
t0=t1%t2 (signed). Undefined behavior if division by zero or overflow. |
185 |
|
186 |
* remu_i32/i64 t0, t1, t2 |
187 |
|
188 |
t0=t1%t2 (unsigned). Undefined behavior if division by zero. |
189 |
|
190 |
********* Logical |
191 |
|
192 |
* and_i32/i64 t0, t1, t2 |
193 |
|
194 |
t0=t1&t2 |
195 |
|
196 |
* or_i32/i64 t0, t1, t2 |
197 |
|
198 |
t0=t1|t2 |
199 |
|
200 |
* xor_i32/i64 t0, t1, t2 |
201 |
|
202 |
t0=t1^t2 |
203 |
|
204 |
* not_i32/i64 t0, t1 |
205 |
|
206 |
t0=~t1 |
207 |
|
208 |
* andc_i32/i64 t0, t1, t2 |
209 |
|
210 |
t0=t1&~t2 |
211 |
|
212 |
* eqv_i32/i64 t0, t1, t2 |
213 |
|
214 |
t0=~(t1^t2) |
215 |
|
216 |
* nand_i32/i64 t0, t1, t2 |
217 |
|
218 |
t0=~(t1&t2) |
219 |
|
220 |
* nor_i32/i64 t0, t1, t2 |
221 |
|
222 |
t0=~(t1|t2) |
223 |
|
224 |
* orc_i32/i64 t0, t1, t2 |
225 |
|
226 |
t0=t1|~t2 |
227 |
|
228 |
********* Shifts |
229 |
|
230 |
* shl_i32/i64 t0, t1, t2 |
231 |
|
232 |
t0=t1 << t2. Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
233 |
|
234 |
* shr_i32/i64 t0, t1, t2 |
235 |
|
236 |
t0=t1 >> t2 (unsigned). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
237 |
|
238 |
* sar_i32/i64 t0, t1, t2 |
239 |
|
240 |
t0=t1 >> t2 (signed). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
241 |
|
242 |
********* Misc |
243 |
|
244 |
* mov_i32/i64 t0, t1 |
245 |
|
246 |
t0 = t1 |
247 |
|
248 |
Move t1 to t0 (both operands must have the same type). |
249 |
|
250 |
* ext8s_i32/i64 t0, t1 |
251 |
ext8u_i32/i64 t0, t1 |
252 |
ext16s_i32/i64 t0, t1 |
253 |
ext16u_i32/i64 t0, t1 |
254 |
ext32s_i64 t0, t1 |
255 |
ext32u_i64 t0, t1 |
256 |
|
257 |
8, 16 or 32 bit sign/zero extension (both operands must have the same type) |
258 |
|
259 |
* bswap16_i32 t0, t1 |
260 |
|
261 |
16 bit byte swap on a 32 bit value. The two high order bytes must be set |
262 |
to zero. |
263 |
|
264 |
* bswap_i32 t0, t1 |
265 |
|
266 |
32 bit byte swap |
267 |
|
268 |
* bswap_i64 t0, t1 |
269 |
|
270 |
64 bit byte swap |
271 |
|
272 |
* discard_i32/i64 t0 |
273 |
|
274 |
Indicate that the value of t0 won't be used later. It is useful to |
275 |
force dead code elimination. |
276 |
|
277 |
********* Type conversions |
278 |
|
279 |
* ext_i32_i64 t0, t1 |
280 |
Convert t1 (32 bit) to t0 (64 bit) and does sign extension |
281 |
|
282 |
* extu_i32_i64 t0, t1 |
283 |
Convert t1 (32 bit) to t0 (64 bit) and does zero extension |
284 |
|
285 |
* trunc_i64_i32 t0, t1 |
286 |
Truncate t1 (64 bit) to t0 (32 bit) |
287 |
|
288 |
* concat_i32_i64 t0, t1, t2 |
289 |
Construct t0 (64-bit) taking the low half from t1 (32 bit) and the high half |
290 |
from t2 (32 bit). |
291 |
|
292 |
* concat32_i64 t0, t1, t2 |
293 |
Construct t0 (64-bit) taking the low half from t1 (64 bit) and the high half |
294 |
from t2 (64 bit). |
295 |
|
296 |
********* Load/Store |
297 |
|
298 |
* ld_i32/i64 t0, t1, offset |
299 |
ld8s_i32/i64 t0, t1, offset |
300 |
ld8u_i32/i64 t0, t1, offset |
301 |
ld16s_i32/i64 t0, t1, offset |
302 |
ld16u_i32/i64 t0, t1, offset |
303 |
ld32s_i64 t0, t1, offset |
304 |
ld32u_i64 t0, t1, offset |
305 |
|
306 |
t0 = read(t1 + offset) |
307 |
Load 8, 16, 32 or 64 bits with or without sign extension from host memory. |
308 |
offset must be a constant. |
309 |
|
310 |
* st_i32/i64 t0, t1, offset |
311 |
st8_i32/i64 t0, t1, offset |
312 |
st16_i32/i64 t0, t1, offset |
313 |
st32_i64 t0, t1, offset |
314 |
|
315 |
write(t0, t1 + offset) |
316 |
Write 8, 16, 32 or 64 bits to host memory. |
317 |
|
318 |
********* QEMU specific operations |
319 |
|
320 |
* tb_exit t0 |
321 |
|
322 |
Exit the current TB and return the value t0 (word type). |
323 |
|
324 |
* goto_tb index |
325 |
|
326 |
Exit the current TB and jump to the TB index 'index' (constant) if the |
327 |
current TB was linked to this TB. Otherwise execute the next |
328 |
instructions. |
329 |
|
330 |
* qemu_ld_i32/i64 t0, t1, flags |
331 |
qemu_ld8u_i32/i64 t0, t1, flags |
332 |
qemu_ld8s_i32/i64 t0, t1, flags |
333 |
qemu_ld16u_i32/i64 t0, t1, flags |
334 |
qemu_ld16s_i32/i64 t0, t1, flags |
335 |
qemu_ld32u_i64 t0, t1, flags |
336 |
qemu_ld32s_i64 t0, t1, flags |
337 |
|
338 |
Load data at the QEMU CPU address t1 into t0. t1 has the QEMU CPU |
339 |
address type. 'flags' contains the QEMU memory index (selects user or |
340 |
kernel access) for example. |
341 |
|
342 |
* qemu_st_i32/i64 t0, t1, flags |
343 |
qemu_st8_i32/i64 t0, t1, flags |
344 |
qemu_st16_i32/i64 t0, t1, flags |
345 |
qemu_st32_i64 t0, t1, flags |
346 |
|
347 |
Store the data t0 at the QEMU CPU Address t1. t1 has the QEMU CPU |
348 |
address type. 'flags' contains the QEMU memory index (selects user or |
349 |
kernel access) for example. |
350 |
|
351 |
Note 1: Some shortcuts are defined when the last operand is known to be |
352 |
a constant (e.g. addi for add, movi for mov). |
353 |
|
354 |
Note 2: When using TCG, the opcodes must never be generated directly |
355 |
as some of them may not be available as "real" opcodes. Always use the |
356 |
function tcg_gen_xxx(args). |
357 |
|
358 |
4) Backend |
359 |
|
360 |
tcg-target.h contains the target specific definitions. tcg-target.c |
361 |
contains the target specific code. |
362 |
|
363 |
4.1) Assumptions |
364 |
|
365 |
The target word size (TCG_TARGET_REG_BITS) is expected to be 32 bit or |
366 |
64 bit. It is expected that the pointer has the same size as the word. |
367 |
|
368 |
On a 32 bit target, all 64 bit operations are converted to 32 bits. A |
369 |
few specific operations must be implemented to allow it (see add2_i32, |
370 |
sub2_i32, brcond2_i32). |
371 |
|
372 |
Floating point operations are not supported in this version. A |
373 |
previous incarnation of the code generator had full support of them, |
374 |
but it is better to concentrate on integer operations first. |
375 |
|
376 |
On a 64 bit target, no assumption is made in TCG about the storage of |
377 |
the 32 bit values in 64 bit registers. |
378 |
|
379 |
4.2) Constraints |
380 |
|
381 |
GCC like constraints are used to define the constraints of every |
382 |
instruction. Memory constraints are not supported in this |
383 |
version. Aliases are specified in the input operands as for GCC. |
384 |
|
385 |
A target can define specific register or constant constraints. If an |
386 |
operation uses a constant input constraint which does not allow all |
387 |
constants, it must also accept registers in order to have a fallback. |
388 |
|
389 |
The movi_i32 and movi_i64 operations must accept any constants. |
390 |
|
391 |
The mov_i32 and mov_i64 operations must accept any registers of the |
392 |
same type. |
393 |
|
394 |
The ld/st instructions must accept signed 32 bit constant offsets. It |
395 |
can be implemented by reserving a specific register to compute the |
396 |
address if the offset is too big. |
397 |
|
398 |
The ld/st instructions must accept any destination (ld) or source (st) |
399 |
register. |
400 |
|
401 |
4.3) Function call assumptions |
402 |
|
403 |
- The only supported types for parameters and return value are: 32 and |
404 |
64 bit integers and pointer. |
405 |
- The stack grows downwards. |
406 |
- The first N parameters are passed in registers. |
407 |
- The next parameters are passed on the stack by storing them as words. |
408 |
- Some registers are clobbered during the call. |
409 |
- The function can return 0 or 1 value in registers. On a 32 bit |
410 |
target, functions must be able to return 2 values in registers for |
411 |
64 bit return type. |
412 |
|
413 |
5) Migration from dyngen to TCG |
414 |
|
415 |
TCG is backward compatible with QEMU "dyngen" operations. It means |
416 |
that TCG instructions can be freely mixed with dyngen operations. It |
417 |
is expected that QEMU targets will be progressively fully converted to |
418 |
TCG. Once a target is fully converted to TCG, it will be possible |
419 |
to apply more optimizations because more registers will be free for |
420 |
the generated code. |
421 |
|
422 |
The exception model is the same as the dyngen one. |
423 |
|
424 |
6) Recommended coding rules for best performance |
425 |
|
426 |
- Use globals to represent the parts of the QEMU CPU state which are |
427 |
often modified, e.g. the integer registers and the condition |
428 |
codes. TCG will be able to use host registers to store them. |
429 |
|
430 |
- Avoid globals stored in fixed registers. They must be used only to |
431 |
store the pointer to the CPU state and possibly to store a pointer |
432 |
to a register window. The other uses are to ensure backward |
433 |
compatibility with dyngen during the porting a new target to TCG. |
434 |
|
435 |
- Use temporaries. Use local temporaries only when really needed, |
436 |
e.g. when you need to use a value after a jump. Local temporaries |
437 |
introduce a performance hit in the current TCG implementation: their |
438 |
content is saved to memory at end of each basic block. |
439 |
|
440 |
- Free temporaries and local temporaries when they are no longer used |
441 |
(tcg_temp_free). Since tcg_const_x() also creates a temporary, you |
442 |
should free it after it is used. Freeing temporaries does not yield |
443 |
a better generated code, but it reduces the memory usage of TCG and |
444 |
the speed of the translation. |
445 |
|
446 |
- Don't hesitate to use helpers for complicated or seldom used target |
447 |
intructions. There is little performance advantage in using TCG to |
448 |
implement target instructions taking more than about twenty TCG |
449 |
instructions. |
450 |
|
451 |
- Use the 'discard' instruction if you know that TCG won't be able to |
452 |
prove that a given global is "dead" at a given program point. The |
453 |
x86 target uses it to improve the condition codes optimisation. |