| 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 given |
| 20 |
function. Temporaries are allocated explicitly in each function. |
| 21 |
|
| 22 |
A TCG "global" is a variable which is live in all the functions. They |
| 23 |
are defined before the functions defined. A TCG global can be a memory |
| 24 |
location (e.g. a QEMU CPU register), a fixed host register (e.g. the |
| 25 |
QEMU CPU state pointer) or a memory location which is stored in a |
| 26 |
register outside QEMU TBs (not implemented yet). |
| 27 |
|
| 28 |
A TCG "basic block" corresponds to a list of instructions terminated |
| 29 |
by a branch instruction. |
| 30 |
|
| 31 |
3) Intermediate representation |
| 32 |
|
| 33 |
3.1) Introduction |
| 34 |
|
| 35 |
TCG instructions operate on variables which are temporaries or |
| 36 |
globals. TCG instructions and variables are strongly typed. Two types |
| 37 |
are supported: 32 bit integers and 64 bit integers. Pointers are |
| 38 |
defined as an alias to 32 bit or 64 bit integers depending on the TCG |
| 39 |
target word size. |
| 40 |
|
| 41 |
Each instruction has a fixed number of output variable operands, input |
| 42 |
variable operands and always constant operands. |
| 43 |
|
| 44 |
The notable exception is the call instruction which has a variable |
| 45 |
number of outputs and inputs. |
| 46 |
|
| 47 |
In the textual form, output operands come first, followed by input |
| 48 |
operands, followed by constant operands. The output type is included |
| 49 |
in the instruction name. Constants are prefixed with a '$'. |
| 50 |
|
| 51 |
add_i32 t0, t1, t2 (t0 <- t1 + t2) |
| 52 |
|
| 53 |
sub_i64 t2, t3, $4 (t2 <- t3 - 4) |
| 54 |
|
| 55 |
3.2) Assumptions |
| 56 |
|
| 57 |
* Basic blocks |
| 58 |
|
| 59 |
- Basic blocks end after branches (e.g. brcond_i32 instruction), |
| 60 |
goto_tb and exit_tb instructions. |
| 61 |
- Basic blocks end before legacy dyngen operations. |
| 62 |
- Basic blocks start after the end of a previous basic block, at a |
| 63 |
set_label instruction or after a legacy dyngen operation. |
| 64 |
|
| 65 |
After the end of a basic block, temporaries at destroyed and globals |
| 66 |
are stored at their initial storage (register or memory place |
| 67 |
depending on their declarations). |
| 68 |
|
| 69 |
* Floating point types are not supported yet |
| 70 |
|
| 71 |
* Pointers: depending on the TCG target, pointer size is 32 bit or 64 |
| 72 |
bit. The type TCG_TYPE_PTR is an alias to TCG_TYPE_I32 or |
| 73 |
TCG_TYPE_I64. |
| 74 |
|
| 75 |
* Helpers: |
| 76 |
|
| 77 |
Using the tcg_gen_helper_x_y it is possible to call any function |
| 78 |
taking i32, i64 or pointer types types. Before calling an helper, all |
| 79 |
globals are stored at their canonical location and it is assumed that |
| 80 |
the function can modify them. In the future, function modifiers will |
| 81 |
be allowed to tell that the helper does not read or write some globals. |
| 82 |
|
| 83 |
On some TCG targets (e.g. x86), several calling conventions are |
| 84 |
supported. |
| 85 |
|
| 86 |
* Branches: |
| 87 |
|
| 88 |
Use the instruction 'br' to jump to a label. Use 'jmp' to jump to an |
| 89 |
explicit address. Conditional branches can only jump to labels. |
| 90 |
|
| 91 |
3.3) Code Optimizations |
| 92 |
|
| 93 |
When generating instructions, you can count on at least the following |
| 94 |
optimizations: |
| 95 |
|
| 96 |
- Single instructions are simplified, e.g. |
| 97 |
|
| 98 |
and_i32 t0, t0, $0xffffffff |
| 99 |
|
| 100 |
is suppressed. |
| 101 |
|
| 102 |
- A liveness analysis is done at the basic block level. The |
| 103 |
information is used to suppress moves from a dead temporary to |
| 104 |
another one. It is also used to remove instructions which compute |
| 105 |
dead results. The later is especially useful for condition code |
| 106 |
optimization in QEMU. |
| 107 |
|
| 108 |
In the following example: |
| 109 |
|
| 110 |
add_i32 t0, t1, t2 |
| 111 |
add_i32 t0, t0, $1 |
| 112 |
mov_i32 t0, $1 |
| 113 |
|
| 114 |
only the last instruction is kept. |
| 115 |
|
| 116 |
- A macro system is supported (may get closer to function inlining |
| 117 |
some day). It is useful if the liveness analysis is likely to prove |
| 118 |
that some results of a computation are indeed not useful. With the |
| 119 |
macro system, the user can provide several alternative |
| 120 |
implementations which are used depending on the used results. It is |
| 121 |
especially useful for condition code optimization in QEMU. |
| 122 |
|
| 123 |
Here is an example: |
| 124 |
|
| 125 |
macro_2 t0, t1, $1 |
| 126 |
mov_i32 t0, $0x1234 |
| 127 |
|
| 128 |
The macro identified by the ID "$1" normally returns the values t0 |
| 129 |
and t1. Suppose its implementation is: |
| 130 |
|
| 131 |
macro_start |
| 132 |
brcond_i32 t2, $0, $TCG_COND_EQ, $1 |
| 133 |
mov_i32 t0, $2 |
| 134 |
br $2 |
| 135 |
set_label $1 |
| 136 |
mov_i32 t0, $3 |
| 137 |
set_label $2 |
| 138 |
add_i32 t1, t3, t4 |
| 139 |
macro_end |
| 140 |
|
| 141 |
If t0 is not used after the macro, the user can provide a simpler |
| 142 |
implementation: |
| 143 |
|
| 144 |
macro_start |
| 145 |
add_i32 t1, t2, t4 |
| 146 |
macro_end |
| 147 |
|
| 148 |
TCG automatically chooses the right implementation depending on |
| 149 |
which macro outputs are used after it. |
| 150 |
|
| 151 |
Note that if TCG did more expensive optimizations, macros would be |
| 152 |
less useful. In the previous example a macro is useful because the |
| 153 |
liveness analysis is done on each basic block separately. Hence TCG |
| 154 |
cannot remove the code computing 't0' even if it is not used after |
| 155 |
the first macro implementation. |
| 156 |
|
| 157 |
3.4) Instruction Reference |
| 158 |
|
| 159 |
********* Function call |
| 160 |
|
| 161 |
* call <ret> <params> ptr |
| 162 |
|
| 163 |
call function 'ptr' (pointer type) |
| 164 |
|
| 165 |
<ret> optional 32 bit or 64 bit return value |
| 166 |
<params> optional 32 bit or 64 bit parameters |
| 167 |
|
| 168 |
********* Jumps/Labels |
| 169 |
|
| 170 |
* jmp t0 |
| 171 |
|
| 172 |
Absolute jump to address t0 (pointer type). |
| 173 |
|
| 174 |
* set_label $label |
| 175 |
|
| 176 |
Define label 'label' at the current program point. |
| 177 |
|
| 178 |
* br $label |
| 179 |
|
| 180 |
Jump to label. |
| 181 |
|
| 182 |
* brcond_i32/i64 cond, t0, t1, label |
| 183 |
|
| 184 |
Conditional jump if t0 cond t1 is true. cond can be: |
| 185 |
TCG_COND_EQ |
| 186 |
TCG_COND_NE |
| 187 |
TCG_COND_LT /* signed */ |
| 188 |
TCG_COND_GE /* signed */ |
| 189 |
TCG_COND_LE /* signed */ |
| 190 |
TCG_COND_GT /* signed */ |
| 191 |
TCG_COND_LTU /* unsigned */ |
| 192 |
TCG_COND_GEU /* unsigned */ |
| 193 |
TCG_COND_LEU /* unsigned */ |
| 194 |
TCG_COND_GTU /* unsigned */ |
| 195 |
|
| 196 |
********* Arithmetic |
| 197 |
|
| 198 |
* add_i32/i64 t0, t1, t2 |
| 199 |
|
| 200 |
t0=t1+t2 |
| 201 |
|
| 202 |
* sub_i32/i64 t0, t1, t2 |
| 203 |
|
| 204 |
t0=t1-t2 |
| 205 |
|
| 206 |
* mul_i32/i64 t0, t1, t2 |
| 207 |
|
| 208 |
t0=t1*t2 |
| 209 |
|
| 210 |
* div_i32/i64 t0, t1, t2 |
| 211 |
|
| 212 |
t0=t1/t2 (signed). Undefined behavior if division by zero or overflow. |
| 213 |
|
| 214 |
* divu_i32/i64 t0, t1, t2 |
| 215 |
|
| 216 |
t0=t1/t2 (unsigned). Undefined behavior if division by zero. |
| 217 |
|
| 218 |
* rem_i32/i64 t0, t1, t2 |
| 219 |
|
| 220 |
t0=t1%t2 (signed). Undefined behavior if division by zero or overflow. |
| 221 |
|
| 222 |
* remu_i32/i64 t0, t1, t2 |
| 223 |
|
| 224 |
t0=t1%t2 (unsigned). Undefined behavior if division by zero. |
| 225 |
|
| 226 |
********* Logical |
| 227 |
|
| 228 |
* and_i32/i64 t0, t1, t2 |
| 229 |
|
| 230 |
t0=t1&t2 |
| 231 |
|
| 232 |
* or_i32/i64 t0, t1, t2 |
| 233 |
|
| 234 |
t0=t1|t2 |
| 235 |
|
| 236 |
* xor_i32/i64 t0, t1, t2 |
| 237 |
|
| 238 |
t0=t1^t2 |
| 239 |
|
| 240 |
********* Shifts |
| 241 |
|
| 242 |
* shl_i32/i64 t0, t1, t2 |
| 243 |
|
| 244 |
t0=t1 << t2. Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
| 245 |
|
| 246 |
* shr_i32/i64 t0, t1, t2 |
| 247 |
|
| 248 |
t0=t1 >> t2 (unsigned). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
| 249 |
|
| 250 |
* sar_i32/i64 t0, t1, t2 |
| 251 |
|
| 252 |
t0=t1 >> t2 (signed). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
| 253 |
|
| 254 |
********* Misc |
| 255 |
|
| 256 |
* mov_i32/i64 t0, t1 |
| 257 |
|
| 258 |
t0 = t1 |
| 259 |
|
| 260 |
Move t1 to t0 (both operands must have the same type). |
| 261 |
|
| 262 |
* ext8s_i32/i64 t0, t1 |
| 263 |
ext16s_i32/i64 t0, t1 |
| 264 |
ext32s_i64 t0, t1 |
| 265 |
|
| 266 |
8, 16 or 32 bit sign extension (both operands must have the same type) |
| 267 |
|
| 268 |
* bswap16_i32 t0, t1 |
| 269 |
|
| 270 |
16 bit byte swap on a 32 bit value. The two high order bytes must be set |
| 271 |
to zero. |
| 272 |
|
| 273 |
* bswap_i32 t0, t1 |
| 274 |
|
| 275 |
32 bit byte swap |
| 276 |
|
| 277 |
* bswap_i64 t0, t1 |
| 278 |
|
| 279 |
64 bit byte swap |
| 280 |
|
| 281 |
* discard_i32/i64 t0 |
| 282 |
|
| 283 |
Indicate that the value of t0 won't be used later. It is useful to |
| 284 |
force dead code elimination. |
| 285 |
|
| 286 |
********* Type conversions |
| 287 |
|
| 288 |
* ext_i32_i64 t0, t1 |
| 289 |
Convert t1 (32 bit) to t0 (64 bit) and does sign extension |
| 290 |
|
| 291 |
* extu_i32_i64 t0, t1 |
| 292 |
Convert t1 (32 bit) to t0 (64 bit) and does zero extension |
| 293 |
|
| 294 |
* trunc_i64_i32 t0, t1 |
| 295 |
Truncate t1 (64 bit) to t0 (32 bit) |
| 296 |
|
| 297 |
********* Load/Store |
| 298 |
|
| 299 |
* ld_i32/i64 t0, t1, offset |
| 300 |
ld8s_i32/i64 t0, t1, offset |
| 301 |
ld8u_i32/i64 t0, t1, offset |
| 302 |
ld16s_i32/i64 t0, t1, offset |
| 303 |
ld16u_i32/i64 t0, t1, offset |
| 304 |
ld32s_i64 t0, t1, offset |
| 305 |
ld32u_i64 t0, t1, offset |
| 306 |
|
| 307 |
t0 = read(t1 + offset) |
| 308 |
Load 8, 16, 32 or 64 bits with or without sign extension from host memory. |
| 309 |
offset must be a constant. |
| 310 |
|
| 311 |
* st_i32/i64 t0, t1, offset |
| 312 |
st8_i32/i64 t0, t1, offset |
| 313 |
st16_i32/i64 t0, t1, offset |
| 314 |
st32_i64 t0, t1, offset |
| 315 |
|
| 316 |
write(t0, t1 + offset) |
| 317 |
Write 8, 16, 32 or 64 bits to host memory. |
| 318 |
|
| 319 |
********* QEMU specific operations |
| 320 |
|
| 321 |
* tb_exit t0 |
| 322 |
|
| 323 |
Exit the current TB and return the value t0 (word type). |
| 324 |
|
| 325 |
* goto_tb index |
| 326 |
|
| 327 |
Exit the current TB and jump to the TB index 'index' (constant) if the |
| 328 |
current TB was linked to this TB. Otherwise execute the next |
| 329 |
instructions. |
| 330 |
|
| 331 |
* qemu_ld_i32/i64 t0, t1, flags |
| 332 |
qemu_ld8u_i32/i64 t0, t1, flags |
| 333 |
qemu_ld8s_i32/i64 t0, t1, flags |
| 334 |
qemu_ld16u_i32/i64 t0, t1, flags |
| 335 |
qemu_ld16s_i32/i64 t0, t1, flags |
| 336 |
qemu_ld32u_i64 t0, t1, flags |
| 337 |
qemu_ld32s_i64 t0, t1, flags |
| 338 |
|
| 339 |
Load data at the QEMU CPU address t1 into t0. t1 has the QEMU CPU |
| 340 |
address type. 'flags' contains the QEMU memory index (selects user or |
| 341 |
kernel access) for example. |
| 342 |
|
| 343 |
* qemu_st_i32/i64 t0, t1, flags |
| 344 |
qemu_st8_i32/i64 t0, t1, flags |
| 345 |
qemu_st16_i32/i64 t0, t1, flags |
| 346 |
qemu_st32_i64 t0, t1, flags |
| 347 |
|
| 348 |
Store the data t0 at the QEMU CPU Address t1. t1 has the QEMU CPU |
| 349 |
address type. 'flags' contains the QEMU memory index (selects user or |
| 350 |
kernel access) for example. |
| 351 |
|
| 352 |
Note 1: Some shortcuts are defined when the last operand is known to be |
| 353 |
a constant (e.g. addi for add, movi for mov). |
| 354 |
|
| 355 |
Note 2: When using TCG, the opcodes must never be generated directly |
| 356 |
as some of them may not be available as "real" opcodes. Always use the |
| 357 |
function tcg_gen_xxx(args). |
| 358 |
|
| 359 |
4) Backend |
| 360 |
|
| 361 |
tcg-target.h contains the target specific definitions. tcg-target.c |
| 362 |
contains the target specific code. |
| 363 |
|
| 364 |
4.1) Assumptions |
| 365 |
|
| 366 |
The target word size (TCG_TARGET_REG_BITS) is expected to be 32 bit or |
| 367 |
64 bit. It is expected that the pointer has the same size as the word. |
| 368 |
|
| 369 |
On a 32 bit target, all 64 bit operations are converted to 32 bits. A |
| 370 |
few specific operations must be implemented to allow it (see add2_i32, |
| 371 |
sub2_i32, brcond2_i32). |
| 372 |
|
| 373 |
Floating point operations are not supported in this version. A |
| 374 |
previous incarnation of the code generator had full support of them, |
| 375 |
but it is better to concentrate on integer operations first. |
| 376 |
|
| 377 |
On a 64 bit target, no assumption is made in TCG about the storage of |
| 378 |
the 32 bit values in 64 bit registers. |
| 379 |
|
| 380 |
4.2) Constraints |
| 381 |
|
| 382 |
GCC like constraints are used to define the constraints of every |
| 383 |
instruction. Memory constraints are not supported in this |
| 384 |
version. Aliases are specified in the input operands as for GCC. |
| 385 |
|
| 386 |
A target can define specific register or constant constraints. If an |
| 387 |
operation uses a constant input constraint which does not allow all |
| 388 |
constants, it must also accept registers in order to have a fallback. |
| 389 |
|
| 390 |
The movi_i32 and movi_i64 operations must accept any constants. |
| 391 |
|
| 392 |
The mov_i32 and mov_i64 operations must accept any registers of the |
| 393 |
same type. |
| 394 |
|
| 395 |
The ld/st instructions must accept signed 32 bit constant offsets. It |
| 396 |
can be implemented by reserving a specific register to compute the |
| 397 |
address if the offset is too big. |
| 398 |
|
| 399 |
The ld/st instructions must accept any destination (ld) or source (st) |
| 400 |
register. |
| 401 |
|
| 402 |
4.3) Function call assumptions |
| 403 |
|
| 404 |
- The only supported types for parameters and return value are: 32 and |
| 405 |
64 bit integers and pointer. |
| 406 |
- The stack grows downwards. |
| 407 |
- The first N parameters are passed in registers. |
| 408 |
- The next parameters are passed on the stack by storing them as words. |
| 409 |
- Some registers are clobbered during the call. |
| 410 |
- The function can return 0 or 1 value in registers. On a 32 bit |
| 411 |
target, functions must be able to return 2 values in registers for |
| 412 |
64 bit return type. |
| 413 |
|
| 414 |
5) Migration from dyngen to TCG |
| 415 |
|
| 416 |
TCG is backward compatible with QEMU "dyngen" operations. It means |
| 417 |
that TCG instructions can be freely mixed with dyngen operations. It |
| 418 |
is expected that QEMU targets will be progressively fully converted to |
| 419 |
TCG. Once a target is fully converted to TCG, it will be possible |
| 420 |
to apply more optimizations because more registers will be free for |
| 421 |
the generated code. |
| 422 |
|
| 423 |
The exception model is the same as the dyngen one. |