1 Tiny Code Generator - Fabrice Bellard.
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.
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.
17 A TCG "function" corresponds to a QEMU Translated Block (TB).
19 A TCG "temporary" is a variable only live in a basic
20 block. Temporaries are allocated explicitly in each function.
22 A TCG "local temporary" is a variable only live in a function. Local
23 temporaries are allocated explicitly in each function.
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).
32 A TCG "basic block" corresponds to a list of instructions terminated
33 by a branch instruction.
35 3) Intermediate representation
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.
45 Each instruction has a fixed number of output variable operands, input
46 variable operands and always constant operands.
48 The notable exception is the call instruction which has a variable
49 number of outputs and inputs.
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 '$'.
55 add_i32 t0, t1, t2 (t0 <- t1 + t2)
61 - Basic blocks end after branches (e.g. brcond_i32 instruction),
62 goto_tb and exit_tb instructions.
63 - Basic blocks start after the end of a previous basic block, or at a
64 set_label instruction.
66 After the end of a basic block, the content of temporaries is
67 destroyed, but local temporaries and globals are preserved.
69 * Floating point types are not supported yet
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
77 Using the tcg_gen_helper_x_y it is possible to call any function
78 taking i32, i64 or pointer types. By default, before calling a helper,
79 all globals are stored at their canonical location and it is assumed
80 that the function can modify them. This can be overridden by the
81 TCG_CALL_CONST function modifier. By default, the helper is allowed to
82 modify the CPU state or raise an exception. This can be overridden by
83 the TCG_CALL_PURE function modifier, in which case the call to the
84 function is removed if the return value is not used.
86 On some TCG targets (e.g. x86), several calling conventions are
91 Use the instruction 'br' to jump to a label. Use 'jmp' to jump to an
92 explicit address. Conditional branches can only jump to labels.
94 3.3) Code Optimizations
96 When generating instructions, you can count on at least the following
99 - Single instructions are simplified, e.g.
101 and_i32 t0, t0, $0xffffffff
105 - A liveness analysis is done at the basic block level. The
106 information is used to suppress moves from a dead variable to
107 another one. It is also used to remove instructions which compute
108 dead results. The later is especially useful for condition code
109 optimization in QEMU.
111 In the following example:
117 only the last instruction is kept.
119 3.4) Instruction Reference
121 ********* Function call
123 * call <ret> <params> ptr
125 call function 'ptr' (pointer type)
127 <ret> optional 32 bit or 64 bit return value
128 <params> optional 32 bit or 64 bit parameters
130 ********* Jumps/Labels
134 Absolute jump to address t0 (pointer type).
138 Define label 'label' at the current program point.
144 * brcond_i32/i64 t0, t1, cond, label
146 Conditional jump if t0 cond t1 is true. cond can be:
149 TCG_COND_LT /* signed */
150 TCG_COND_GE /* signed */
151 TCG_COND_LE /* signed */
152 TCG_COND_GT /* signed */
153 TCG_COND_LTU /* unsigned */
154 TCG_COND_GEU /* unsigned */
155 TCG_COND_LEU /* unsigned */
156 TCG_COND_GTU /* unsigned */
160 * add_i32/i64 t0, t1, t2
164 * sub_i32/i64 t0, t1, t2
170 t0=-t1 (two's complement)
172 * mul_i32/i64 t0, t1, t2
176 * div_i32/i64 t0, t1, t2
178 t0=t1/t2 (signed). Undefined behavior if division by zero or overflow.
180 * divu_i32/i64 t0, t1, t2
182 t0=t1/t2 (unsigned). Undefined behavior if division by zero.
184 * rem_i32/i64 t0, t1, t2
186 t0=t1%t2 (signed). Undefined behavior if division by zero or overflow.
188 * remu_i32/i64 t0, t1, t2
190 t0=t1%t2 (unsigned). Undefined behavior if division by zero.
194 * and_i32/i64 t0, t1, t2
198 * or_i32/i64 t0, t1, t2
202 * xor_i32/i64 t0, t1, t2
210 * andc_i32/i64 t0, t1, t2
214 * eqv_i32/i64 t0, t1, t2
216 t0=~(t1^t2), or equivalently, t0=t1^~t2
218 * nand_i32/i64 t0, t1, t2
222 * nor_i32/i64 t0, t1, t2
226 * orc_i32/i64 t0, t1, t2
230 ********* Shifts/Rotates
232 * shl_i32/i64 t0, t1, t2
234 t0=t1 << t2. Undefined behavior if t2 < 0 or t2 >= 32 (resp 64)
236 * shr_i32/i64 t0, t1, t2
238 t0=t1 >> t2 (unsigned). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64)
240 * sar_i32/i64 t0, t1, t2
242 t0=t1 >> t2 (signed). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64)
244 * rotl_i32/i64 t0, t1, t2
246 Rotation of t2 bits to the left. Undefined behavior if t2 < 0 or t2 >= 32 (resp 64)
248 * rotr_i32/i64 t0, t1, t2
250 Rotation of t2 bits to the right. Undefined behavior if t2 < 0 or t2 >= 32 (resp 64)
258 Move t1 to t0 (both operands must have the same type).
260 * ext8s_i32/i64 t0, t1
262 ext16s_i32/i64 t0, t1
263 ext16u_i32/i64 t0, t1
267 8, 16 or 32 bit sign/zero extension (both operands must have the same type)
269 * bswap16_i32/i64 t0, t1
271 16 bit byte swap on a 32/64 bit value. It assumes that the two/six high order
272 bytes are set to zero.
274 * bswap32_i32/i64 t0, t1
276 32 bit byte swap on a 32/64 bit value. With a 64 bit value, it assumes that
277 the four high order bytes are set to zero.
285 Indicate that the value of t0 won't be used later. It is useful to
286 force dead code elimination.
288 * deposit_i32/i64 dest, t1, t2, pos, len
290 Deposit T2 as a bitfield into T1, placing the result in DEST.
291 The bitfield is described by POS/LEN, which are immediate values:
293 LEN - the length of the bitfield
294 POS - the position of the first bit, counting from the LSB
296 For example, pos=8, len=4 indicates a 4-bit field at bit 8.
297 This operation would be equivalent to
299 dest = (t1 & ~0x0f00) | ((t2 << 8) & 0x0f00)
302 ********* Conditional moves
304 * setcond_i32/i64 dest, t1, t2, cond
308 Set DEST to 1 if (T1 cond T2) is true, otherwise set to 0.
310 * movcond_i32/i64 dest, c1, c2, v1, v2, cond
312 dest = (c1 cond c2 ? v1 : v2)
314 Set DEST to V1 if (C1 cond C2) is true, otherwise set to V2.
316 ********* Type conversions
319 Convert t1 (32 bit) to t0 (64 bit) and does sign extension
321 * extu_i32_i64 t0, t1
322 Convert t1 (32 bit) to t0 (64 bit) and does zero extension
324 * trunc_i64_i32 t0, t1
325 Truncate t1 (64 bit) to t0 (32 bit)
327 * concat_i32_i64 t0, t1, t2
328 Construct t0 (64-bit) taking the low half from t1 (32 bit) and the high half
331 * concat32_i64 t0, t1, t2
332 Construct t0 (64-bit) taking the low half from t1 (64 bit) and the high half
337 * ld_i32/i64 t0, t1, offset
338 ld8s_i32/i64 t0, t1, offset
339 ld8u_i32/i64 t0, t1, offset
340 ld16s_i32/i64 t0, t1, offset
341 ld16u_i32/i64 t0, t1, offset
342 ld32s_i64 t0, t1, offset
343 ld32u_i64 t0, t1, offset
345 t0 = read(t1 + offset)
346 Load 8, 16, 32 or 64 bits with or without sign extension from host memory.
347 offset must be a constant.
349 * st_i32/i64 t0, t1, offset
350 st8_i32/i64 t0, t1, offset
351 st16_i32/i64 t0, t1, offset
352 st32_i64 t0, t1, offset
354 write(t0, t1 + offset)
355 Write 8, 16, 32 or 64 bits to host memory.
357 ********* 64-bit target on 32-bit host support
359 The following opcodes are internal to TCG. Thus they are to be implemented by
360 32-bit host code generators, but are not to be emitted by guest translators.
361 They are emitted as needed by inline functions within "tcg-op.h".
363 * brcond2_i32 t0_low, t0_high, t1_low, t1_high, cond, label
365 Similar to brcond, except that the 64-bit values T0 and T1
366 are formed from two 32-bit arguments.
368 * add2_i32 t0_low, t0_high, t1_low, t1_high, t2_low, t2_high
369 * sub2_i32 t0_low, t0_high, t1_low, t1_high, t2_low, t2_high
371 Similar to add/sub, except that the 64-bit inputs T1 and T2 are
372 formed from two 32-bit arguments, and the 64-bit output T0
373 is returned in two 32-bit outputs.
375 * mulu2_i32 t0_low, t0_high, t1, t2
377 Similar to mul, except two 32-bit (unsigned) inputs T1 and T2 yielding
378 the full 64-bit product T0. The later is returned in two 32-bit outputs.
380 * setcond2_i32 dest, t1_low, t1_high, t2_low, t2_high, cond
382 Similar to setcond, except that the 64-bit values T1 and T2 are
383 formed from two 32-bit arguments. The result is a 32-bit value.
385 ********* QEMU specific operations
389 Exit the current TB and return the value t0 (word type).
393 Exit the current TB and jump to the TB index 'index' (constant) if the
394 current TB was linked to this TB. Otherwise execute the next
395 instructions. Only indices 0 and 1 are valid and tcg_gen_goto_tb may be issued
396 at most once with each slot index per TB.
398 * qemu_ld8u t0, t1, flags
399 qemu_ld8s t0, t1, flags
400 qemu_ld16u t0, t1, flags
401 qemu_ld16s t0, t1, flags
402 qemu_ld32 t0, t1, flags
403 qemu_ld32u t0, t1, flags
404 qemu_ld32s t0, t1, flags
405 qemu_ld64 t0, t1, flags
407 Load data at the QEMU CPU address t1 into t0. t1 has the QEMU CPU address
408 type. 'flags' contains the QEMU memory index (selects user or kernel access)
411 Note that "qemu_ld32" implies a 32-bit result, while "qemu_ld32u" and
412 "qemu_ld32s" imply a 64-bit result appropriately extended from 32 bits.
414 * qemu_st8 t0, t1, flags
415 qemu_st16 t0, t1, flags
416 qemu_st32 t0, t1, flags
417 qemu_st64 t0, t1, flags
419 Store the data t0 at the QEMU CPU Address t1. t1 has the QEMU CPU
420 address type. 'flags' contains the QEMU memory index (selects user or
421 kernel access) for example.
423 Note 1: Some shortcuts are defined when the last operand is known to be
424 a constant (e.g. addi for add, movi for mov).
426 Note 2: When using TCG, the opcodes must never be generated directly
427 as some of them may not be available as "real" opcodes. Always use the
428 function tcg_gen_xxx(args).
432 tcg-target.h contains the target specific definitions. tcg-target.c
433 contains the target specific code.
437 The target word size (TCG_TARGET_REG_BITS) is expected to be 32 bit or
438 64 bit. It is expected that the pointer has the same size as the word.
440 On a 32 bit target, all 64 bit operations are converted to 32 bits. A
441 few specific operations must be implemented to allow it (see add2_i32,
442 sub2_i32, brcond2_i32).
444 Floating point operations are not supported in this version. A
445 previous incarnation of the code generator had full support of them,
446 but it is better to concentrate on integer operations first.
448 On a 64 bit target, no assumption is made in TCG about the storage of
449 the 32 bit values in 64 bit registers.
453 GCC like constraints are used to define the constraints of every
454 instruction. Memory constraints are not supported in this
455 version. Aliases are specified in the input operands as for GCC.
457 The same register may be used for both an input and an output, even when
458 they are not explicitly aliased. If an op expands to multiple target
459 instructions then care must be taken to avoid clobbering input values.
460 GCC style "early clobber" outputs are not currently supported.
462 A target can define specific register or constant constraints. If an
463 operation uses a constant input constraint which does not allow all
464 constants, it must also accept registers in order to have a fallback.
466 The movi_i32 and movi_i64 operations must accept any constants.
468 The mov_i32 and mov_i64 operations must accept any registers of the
471 The ld/st instructions must accept signed 32 bit constant offsets. It
472 can be implemented by reserving a specific register to compute the
473 address if the offset is too big.
475 The ld/st instructions must accept any destination (ld) or source (st)
478 4.3) Function call assumptions
480 - The only supported types for parameters and return value are: 32 and
481 64 bit integers and pointer.
482 - The stack grows downwards.
483 - The first N parameters are passed in registers.
484 - The next parameters are passed on the stack by storing them as words.
485 - Some registers are clobbered during the call.
486 - The function can return 0 or 1 value in registers. On a 32 bit
487 target, functions must be able to return 2 values in registers for
490 5) Recommended coding rules for best performance
492 - Use globals to represent the parts of the QEMU CPU state which are
493 often modified, e.g. the integer registers and the condition
494 codes. TCG will be able to use host registers to store them.
496 - Avoid globals stored in fixed registers. They must be used only to
497 store the pointer to the CPU state and possibly to store a pointer
498 to a register window.
500 - Use temporaries. Use local temporaries only when really needed,
501 e.g. when you need to use a value after a jump. Local temporaries
502 introduce a performance hit in the current TCG implementation: their
503 content is saved to memory at end of each basic block.
505 - Free temporaries and local temporaries when they are no longer used
506 (tcg_temp_free). Since tcg_const_x() also creates a temporary, you
507 should free it after it is used. Freeing temporaries does not yield
508 a better generated code, but it reduces the memory usage of TCG and
509 the speed of the translation.
511 - Don't hesitate to use helpers for complicated or seldom used target
512 instructions. There is little performance advantage in using TCG to
513 implement target instructions taking more than about twenty TCG
514 instructions. Note that this rule of thumb is more applicable to
515 helpers doing complex logic or arithmetic, where the C compiler has
516 scope to do a good job of optimisation; it is less relevant where
517 the instruction is mostly doing loads and stores, and in those cases
518 inline TCG may still be faster for longer sequences.
520 - The hard limit on the number of TCG instructions you can generate
521 per target instruction is set by MAX_OP_PER_INSTR in exec-all.h --
522 you cannot exceed this without risking a buffer overrun.
524 - Use the 'discard' instruction if you know that TCG won't be able to
525 prove that a given global is "dead" at a given program point. The
526 x86 target uses it to improve the condition codes optimisation.