root / tcg / README @ affa3264
History | View | Annotate | Download (12.1 kB)
1 | c896fe29 | bellard | Tiny Code Generator - Fabrice Bellard. |
---|---|---|---|
2 | c896fe29 | bellard | |
3 | c896fe29 | bellard | 1) Introduction |
4 | c896fe29 | bellard | |
5 | c896fe29 | bellard | TCG (Tiny Code Generator) began as a generic backend for a C |
6 | c896fe29 | bellard | compiler. It was simplified to be used in QEMU. It also has its roots |
7 | c896fe29 | bellard | in the QOP code generator written by Paul Brook. |
8 | c896fe29 | bellard | |
9 | c896fe29 | bellard | 2) Definitions |
10 | c896fe29 | bellard | |
11 | c896fe29 | bellard | The TCG "target" is the architecture for which we generate the |
12 | c896fe29 | bellard | code. It is of course not the same as the "target" of QEMU which is |
13 | c896fe29 | bellard | the emulated architecture. As TCG started as a generic C backend used |
14 | c896fe29 | bellard | for cross compiling, it is assumed that the TCG target is different |
15 | c896fe29 | bellard | from the host, although it is never the case for QEMU. |
16 | c896fe29 | bellard | |
17 | c896fe29 | bellard | A TCG "function" corresponds to a QEMU Translated Block (TB). |
18 | c896fe29 | bellard | |
19 | 0a6b7b78 | bellard | A TCG "temporary" is a variable only live in a basic |
20 | 0a6b7b78 | bellard | block. Temporaries are allocated explicitly in each function. |
21 | c896fe29 | bellard | |
22 | 0a6b7b78 | bellard | A TCG "local temporary" is a variable only live in a function. Local |
23 | 0a6b7b78 | bellard | temporaries are allocated explicitly in each function. |
24 | 0a6b7b78 | bellard | |
25 | 0a6b7b78 | bellard | A TCG "global" is a variable which is live in all the functions |
26 | 0a6b7b78 | bellard | (equivalent of a C global variable). They are defined before the |
27 | 0a6b7b78 | bellard | functions defined. A TCG global can be a memory location (e.g. a QEMU |
28 | 0a6b7b78 | bellard | CPU register), a fixed host register (e.g. the QEMU CPU state pointer) |
29 | 0a6b7b78 | bellard | or a memory location which is stored in a register outside QEMU TBs |
30 | 0a6b7b78 | bellard | (not implemented yet). |
31 | c896fe29 | bellard | |
32 | c896fe29 | bellard | A TCG "basic block" corresponds to a list of instructions terminated |
33 | c896fe29 | bellard | by a branch instruction. |
34 | c896fe29 | bellard | |
35 | c896fe29 | bellard | 3) Intermediate representation |
36 | c896fe29 | bellard | |
37 | c896fe29 | bellard | 3.1) Introduction |
38 | c896fe29 | bellard | |
39 | 0a6b7b78 | bellard | TCG instructions operate on variables which are temporaries, local |
40 | 0a6b7b78 | bellard | temporaries or globals. TCG instructions and variables are strongly |
41 | 0a6b7b78 | bellard | typed. Two types are supported: 32 bit integers and 64 bit |
42 | 0a6b7b78 | bellard | integers. Pointers are defined as an alias to 32 bit or 64 bit |
43 | 0a6b7b78 | bellard | integers depending on the TCG target word size. |
44 | c896fe29 | bellard | |
45 | c896fe29 | bellard | Each instruction has a fixed number of output variable operands, input |
46 | c896fe29 | bellard | variable operands and always constant operands. |
47 | c896fe29 | bellard | |
48 | c896fe29 | bellard | The notable exception is the call instruction which has a variable |
49 | c896fe29 | bellard | number of outputs and inputs. |
50 | c896fe29 | bellard | |
51 | 0a6b7b78 | bellard | In the textual form, output operands usually come first, followed by |
52 | 0a6b7b78 | bellard | input operands, followed by constant operands. The output type is |
53 | 0a6b7b78 | bellard | included in the instruction name. Constants are prefixed with a '$'. |
54 | c896fe29 | bellard | |
55 | c896fe29 | bellard | add_i32 t0, t1, t2 (t0 <- t1 + t2) |
56 | c896fe29 | bellard | |
57 | c896fe29 | bellard | 3.2) Assumptions |
58 | c896fe29 | bellard | |
59 | c896fe29 | bellard | * Basic blocks |
60 | c896fe29 | bellard | |
61 | c896fe29 | bellard | - Basic blocks end after branches (e.g. brcond_i32 instruction), |
62 | c896fe29 | bellard | goto_tb and exit_tb instructions. |
63 | c896fe29 | bellard | - Basic blocks end before legacy dyngen operations. |
64 | c896fe29 | bellard | - Basic blocks start after the end of a previous basic block, at a |
65 | c896fe29 | bellard | set_label instruction or after a legacy dyngen operation. |
66 | c896fe29 | bellard | |
67 | 0a6b7b78 | bellard | After the end of a basic block, the content of temporaries is |
68 | 0a6b7b78 | bellard | destroyed, but local temporaries and globals are preserved. |
69 | c896fe29 | bellard | |
70 | c896fe29 | bellard | * Floating point types are not supported yet |
71 | c896fe29 | bellard | |
72 | c896fe29 | bellard | * Pointers: depending on the TCG target, pointer size is 32 bit or 64 |
73 | c896fe29 | bellard | bit. The type TCG_TYPE_PTR is an alias to TCG_TYPE_I32 or |
74 | c896fe29 | bellard | TCG_TYPE_I64. |
75 | c896fe29 | bellard | |
76 | c896fe29 | bellard | * Helpers: |
77 | c896fe29 | bellard | |
78 | c896fe29 | bellard | Using the tcg_gen_helper_x_y it is possible to call any function |
79 | 811d4cf4 | balrog | taking i32, i64 or pointer types. Before calling an helper, all |
80 | c896fe29 | bellard | globals are stored at their canonical location and it is assumed that |
81 | c896fe29 | bellard | the function can modify them. In the future, function modifiers will |
82 | c896fe29 | bellard | be allowed to tell that the helper does not read or write some globals. |
83 | c896fe29 | bellard | |
84 | c896fe29 | bellard | On some TCG targets (e.g. x86), several calling conventions are |
85 | c896fe29 | bellard | supported. |
86 | c896fe29 | bellard | |
87 | c896fe29 | bellard | * Branches: |
88 | c896fe29 | bellard | |
89 | c896fe29 | bellard | Use the instruction 'br' to jump to a label. Use 'jmp' to jump to an |
90 | c896fe29 | bellard | explicit address. Conditional branches can only jump to labels. |
91 | c896fe29 | bellard | |
92 | c896fe29 | bellard | 3.3) Code Optimizations |
93 | c896fe29 | bellard | |
94 | c896fe29 | bellard | When generating instructions, you can count on at least the following |
95 | c896fe29 | bellard | optimizations: |
96 | c896fe29 | bellard | |
97 | c896fe29 | bellard | - Single instructions are simplified, e.g. |
98 | c896fe29 | bellard | |
99 | c896fe29 | bellard | and_i32 t0, t0, $0xffffffff |
100 | c896fe29 | bellard | |
101 | c896fe29 | bellard | is suppressed. |
102 | c896fe29 | bellard | |
103 | c896fe29 | bellard | - A liveness analysis is done at the basic block level. The |
104 | 0a6b7b78 | bellard | information is used to suppress moves from a dead variable to |
105 | c896fe29 | bellard | another one. It is also used to remove instructions which compute |
106 | c896fe29 | bellard | dead results. The later is especially useful for condition code |
107 | 9804c8e2 | bellard | optimization in QEMU. |
108 | c896fe29 | bellard | |
109 | c896fe29 | bellard | In the following example: |
110 | c896fe29 | bellard | |
111 | c896fe29 | bellard | add_i32 t0, t1, t2 |
112 | c896fe29 | bellard | add_i32 t0, t0, $1 |
113 | c896fe29 | bellard | mov_i32 t0, $1 |
114 | c896fe29 | bellard | |
115 | c896fe29 | bellard | only the last instruction is kept. |
116 | c896fe29 | bellard | |
117 | c896fe29 | bellard | 3.4) Instruction Reference |
118 | c896fe29 | bellard | |
119 | c896fe29 | bellard | ********* Function call |
120 | c896fe29 | bellard | |
121 | c896fe29 | bellard | * call <ret> <params> ptr |
122 | c896fe29 | bellard | |
123 | c896fe29 | bellard | call function 'ptr' (pointer type) |
124 | c896fe29 | bellard | |
125 | c896fe29 | bellard | <ret> optional 32 bit or 64 bit return value |
126 | c896fe29 | bellard | <params> optional 32 bit or 64 bit parameters |
127 | c896fe29 | bellard | |
128 | c896fe29 | bellard | ********* Jumps/Labels |
129 | c896fe29 | bellard | |
130 | c896fe29 | bellard | * jmp t0 |
131 | c896fe29 | bellard | |
132 | c896fe29 | bellard | Absolute jump to address t0 (pointer type). |
133 | c896fe29 | bellard | |
134 | c896fe29 | bellard | * set_label $label |
135 | c896fe29 | bellard | |
136 | c896fe29 | bellard | Define label 'label' at the current program point. |
137 | c896fe29 | bellard | |
138 | c896fe29 | bellard | * br $label |
139 | c896fe29 | bellard | |
140 | c896fe29 | bellard | Jump to label. |
141 | c896fe29 | bellard | |
142 | c896fe29 | bellard | * brcond_i32/i64 cond, t0, t1, label |
143 | c896fe29 | bellard | |
144 | c896fe29 | bellard | Conditional jump if t0 cond t1 is true. cond can be: |
145 | c896fe29 | bellard | TCG_COND_EQ |
146 | c896fe29 | bellard | TCG_COND_NE |
147 | c896fe29 | bellard | TCG_COND_LT /* signed */ |
148 | c896fe29 | bellard | TCG_COND_GE /* signed */ |
149 | c896fe29 | bellard | TCG_COND_LE /* signed */ |
150 | c896fe29 | bellard | TCG_COND_GT /* signed */ |
151 | c896fe29 | bellard | TCG_COND_LTU /* unsigned */ |
152 | c896fe29 | bellard | TCG_COND_GEU /* unsigned */ |
153 | c896fe29 | bellard | TCG_COND_LEU /* unsigned */ |
154 | c896fe29 | bellard | TCG_COND_GTU /* unsigned */ |
155 | c896fe29 | bellard | |
156 | c896fe29 | bellard | ********* Arithmetic |
157 | c896fe29 | bellard | |
158 | c896fe29 | bellard | * add_i32/i64 t0, t1, t2 |
159 | c896fe29 | bellard | |
160 | c896fe29 | bellard | t0=t1+t2 |
161 | c896fe29 | bellard | |
162 | c896fe29 | bellard | * sub_i32/i64 t0, t1, t2 |
163 | c896fe29 | bellard | |
164 | c896fe29 | bellard | t0=t1-t2 |
165 | c896fe29 | bellard | |
166 | 390efc54 | pbrook | * neg_i32/i64 t0, t1 |
167 | 390efc54 | pbrook | |
168 | 390efc54 | pbrook | t0=-t1 (two's complement) |
169 | 390efc54 | pbrook | |
170 | c896fe29 | bellard | * mul_i32/i64 t0, t1, t2 |
171 | c896fe29 | bellard | |
172 | c896fe29 | bellard | t0=t1*t2 |
173 | c896fe29 | bellard | |
174 | c896fe29 | bellard | * div_i32/i64 t0, t1, t2 |
175 | c896fe29 | bellard | |
176 | c896fe29 | bellard | t0=t1/t2 (signed). Undefined behavior if division by zero or overflow. |
177 | c896fe29 | bellard | |
178 | c896fe29 | bellard | * divu_i32/i64 t0, t1, t2 |
179 | c896fe29 | bellard | |
180 | c896fe29 | bellard | t0=t1/t2 (unsigned). Undefined behavior if division by zero. |
181 | c896fe29 | bellard | |
182 | c896fe29 | bellard | * rem_i32/i64 t0, t1, t2 |
183 | c896fe29 | bellard | |
184 | c896fe29 | bellard | t0=t1%t2 (signed). Undefined behavior if division by zero or overflow. |
185 | c896fe29 | bellard | |
186 | c896fe29 | bellard | * remu_i32/i64 t0, t1, t2 |
187 | c896fe29 | bellard | |
188 | c896fe29 | bellard | t0=t1%t2 (unsigned). Undefined behavior if division by zero. |
189 | c896fe29 | bellard | |
190 | c896fe29 | bellard | ********* Logical |
191 | c896fe29 | bellard | |
192 | 5e85404a | aurel32 | * and_i32/i64 t0, t1, t2 |
193 | 5e85404a | aurel32 | |
194 | c896fe29 | bellard | t0=t1&t2 |
195 | c896fe29 | bellard | |
196 | c896fe29 | bellard | * or_i32/i64 t0, t1, t2 |
197 | c896fe29 | bellard | |
198 | c896fe29 | bellard | t0=t1|t2 |
199 | c896fe29 | bellard | |
200 | c896fe29 | bellard | * xor_i32/i64 t0, t1, t2 |
201 | c896fe29 | bellard | |
202 | c896fe29 | bellard | t0=t1^t2 |
203 | c896fe29 | bellard | |
204 | 0a6b7b78 | bellard | * not_i32/i64 t0, t1 |
205 | 0a6b7b78 | bellard | |
206 | 0a6b7b78 | bellard | t0=~t1 |
207 | 0a6b7b78 | bellard | |
208 | c896fe29 | bellard | ********* Shifts |
209 | c896fe29 | bellard | |
210 | c896fe29 | bellard | * shl_i32/i64 t0, t1, t2 |
211 | c896fe29 | bellard | |
212 | c896fe29 | bellard | t0=t1 << t2. Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
213 | c896fe29 | bellard | |
214 | c896fe29 | bellard | * shr_i32/i64 t0, t1, t2 |
215 | c896fe29 | bellard | |
216 | c896fe29 | bellard | t0=t1 >> t2 (unsigned). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
217 | c896fe29 | bellard | |
218 | c896fe29 | bellard | * sar_i32/i64 t0, t1, t2 |
219 | c896fe29 | bellard | |
220 | c896fe29 | bellard | t0=t1 >> t2 (signed). Undefined behavior if t2 < 0 or t2 >= 32 (resp 64) |
221 | c896fe29 | bellard | |
222 | c896fe29 | bellard | ********* Misc |
223 | c896fe29 | bellard | |
224 | c896fe29 | bellard | * mov_i32/i64 t0, t1 |
225 | c896fe29 | bellard | |
226 | c896fe29 | bellard | t0 = t1 |
227 | c896fe29 | bellard | |
228 | c896fe29 | bellard | Move t1 to t0 (both operands must have the same type). |
229 | c896fe29 | bellard | |
230 | c896fe29 | bellard | * ext8s_i32/i64 t0, t1 |
231 | 86831435 | pbrook | ext8u_i32/i64 t0, t1 |
232 | c896fe29 | bellard | ext16s_i32/i64 t0, t1 |
233 | 86831435 | pbrook | ext16u_i32/i64 t0, t1 |
234 | c896fe29 | bellard | ext32s_i64 t0, t1 |
235 | 86831435 | pbrook | ext32u_i64 t0, t1 |
236 | c896fe29 | bellard | |
237 | 86831435 | pbrook | 8, 16 or 32 bit sign/zero extension (both operands must have the same type) |
238 | c896fe29 | bellard | |
239 | c896fe29 | bellard | * bswap16_i32 t0, t1 |
240 | c896fe29 | bellard | |
241 | c896fe29 | bellard | 16 bit byte swap on a 32 bit value. The two high order bytes must be set |
242 | c896fe29 | bellard | to zero. |
243 | c896fe29 | bellard | |
244 | c896fe29 | bellard | * bswap_i32 t0, t1 |
245 | c896fe29 | bellard | |
246 | c896fe29 | bellard | 32 bit byte swap |
247 | c896fe29 | bellard | |
248 | c896fe29 | bellard | * bswap_i64 t0, t1 |
249 | c896fe29 | bellard | |
250 | c896fe29 | bellard | 64 bit byte swap |
251 | c896fe29 | bellard | |
252 | 5ff9d6a4 | bellard | * discard_i32/i64 t0 |
253 | 5ff9d6a4 | bellard | |
254 | 5ff9d6a4 | bellard | Indicate that the value of t0 won't be used later. It is useful to |
255 | 5ff9d6a4 | bellard | force dead code elimination. |
256 | 5ff9d6a4 | bellard | |
257 | c896fe29 | bellard | ********* Type conversions |
258 | c896fe29 | bellard | |
259 | c896fe29 | bellard | * ext_i32_i64 t0, t1 |
260 | c896fe29 | bellard | Convert t1 (32 bit) to t0 (64 bit) and does sign extension |
261 | c896fe29 | bellard | |
262 | c896fe29 | bellard | * extu_i32_i64 t0, t1 |
263 | c896fe29 | bellard | Convert t1 (32 bit) to t0 (64 bit) and does zero extension |
264 | c896fe29 | bellard | |
265 | c896fe29 | bellard | * trunc_i64_i32 t0, t1 |
266 | c896fe29 | bellard | Truncate t1 (64 bit) to t0 (32 bit) |
267 | c896fe29 | bellard | |
268 | c896fe29 | bellard | ********* Load/Store |
269 | c896fe29 | bellard | |
270 | c896fe29 | bellard | * ld_i32/i64 t0, t1, offset |
271 | c896fe29 | bellard | ld8s_i32/i64 t0, t1, offset |
272 | c896fe29 | bellard | ld8u_i32/i64 t0, t1, offset |
273 | c896fe29 | bellard | ld16s_i32/i64 t0, t1, offset |
274 | c896fe29 | bellard | ld16u_i32/i64 t0, t1, offset |
275 | c896fe29 | bellard | ld32s_i64 t0, t1, offset |
276 | c896fe29 | bellard | ld32u_i64 t0, t1, offset |
277 | c896fe29 | bellard | |
278 | c896fe29 | bellard | t0 = read(t1 + offset) |
279 | c896fe29 | bellard | Load 8, 16, 32 or 64 bits with or without sign extension from host memory. |
280 | c896fe29 | bellard | offset must be a constant. |
281 | c896fe29 | bellard | |
282 | c896fe29 | bellard | * st_i32/i64 t0, t1, offset |
283 | c896fe29 | bellard | st8_i32/i64 t0, t1, offset |
284 | c896fe29 | bellard | st16_i32/i64 t0, t1, offset |
285 | c896fe29 | bellard | st32_i64 t0, t1, offset |
286 | c896fe29 | bellard | |
287 | c896fe29 | bellard | write(t0, t1 + offset) |
288 | c896fe29 | bellard | Write 8, 16, 32 or 64 bits to host memory. |
289 | c896fe29 | bellard | |
290 | c896fe29 | bellard | ********* QEMU specific operations |
291 | c896fe29 | bellard | |
292 | c896fe29 | bellard | * tb_exit t0 |
293 | c896fe29 | bellard | |
294 | c896fe29 | bellard | Exit the current TB and return the value t0 (word type). |
295 | c896fe29 | bellard | |
296 | c896fe29 | bellard | * goto_tb index |
297 | c896fe29 | bellard | |
298 | c896fe29 | bellard | Exit the current TB and jump to the TB index 'index' (constant) if the |
299 | c896fe29 | bellard | current TB was linked to this TB. Otherwise execute the next |
300 | c896fe29 | bellard | instructions. |
301 | c896fe29 | bellard | |
302 | c896fe29 | bellard | * qemu_ld_i32/i64 t0, t1, flags |
303 | c896fe29 | bellard | qemu_ld8u_i32/i64 t0, t1, flags |
304 | c896fe29 | bellard | qemu_ld8s_i32/i64 t0, t1, flags |
305 | c896fe29 | bellard | qemu_ld16u_i32/i64 t0, t1, flags |
306 | c896fe29 | bellard | qemu_ld16s_i32/i64 t0, t1, flags |
307 | c896fe29 | bellard | qemu_ld32u_i64 t0, t1, flags |
308 | c896fe29 | bellard | qemu_ld32s_i64 t0, t1, flags |
309 | c896fe29 | bellard | |
310 | c896fe29 | bellard | Load data at the QEMU CPU address t1 into t0. t1 has the QEMU CPU |
311 | c896fe29 | bellard | address type. 'flags' contains the QEMU memory index (selects user or |
312 | c896fe29 | bellard | kernel access) for example. |
313 | c896fe29 | bellard | |
314 | c896fe29 | bellard | * qemu_st_i32/i64 t0, t1, flags |
315 | c896fe29 | bellard | qemu_st8_i32/i64 t0, t1, flags |
316 | c896fe29 | bellard | qemu_st16_i32/i64 t0, t1, flags |
317 | c896fe29 | bellard | qemu_st32_i64 t0, t1, flags |
318 | c896fe29 | bellard | |
319 | c896fe29 | bellard | Store the data t0 at the QEMU CPU Address t1. t1 has the QEMU CPU |
320 | c896fe29 | bellard | address type. 'flags' contains the QEMU memory index (selects user or |
321 | c896fe29 | bellard | kernel access) for example. |
322 | c896fe29 | bellard | |
323 | c896fe29 | bellard | Note 1: Some shortcuts are defined when the last operand is known to be |
324 | c896fe29 | bellard | a constant (e.g. addi for add, movi for mov). |
325 | c896fe29 | bellard | |
326 | c896fe29 | bellard | Note 2: When using TCG, the opcodes must never be generated directly |
327 | c896fe29 | bellard | as some of them may not be available as "real" opcodes. Always use the |
328 | c896fe29 | bellard | function tcg_gen_xxx(args). |
329 | c896fe29 | bellard | |
330 | c896fe29 | bellard | 4) Backend |
331 | c896fe29 | bellard | |
332 | c896fe29 | bellard | tcg-target.h contains the target specific definitions. tcg-target.c |
333 | c896fe29 | bellard | contains the target specific code. |
334 | c896fe29 | bellard | |
335 | c896fe29 | bellard | 4.1) Assumptions |
336 | c896fe29 | bellard | |
337 | c896fe29 | bellard | The target word size (TCG_TARGET_REG_BITS) is expected to be 32 bit or |
338 | c896fe29 | bellard | 64 bit. It is expected that the pointer has the same size as the word. |
339 | c896fe29 | bellard | |
340 | c896fe29 | bellard | On a 32 bit target, all 64 bit operations are converted to 32 bits. A |
341 | c896fe29 | bellard | few specific operations must be implemented to allow it (see add2_i32, |
342 | c896fe29 | bellard | sub2_i32, brcond2_i32). |
343 | c896fe29 | bellard | |
344 | c896fe29 | bellard | Floating point operations are not supported in this version. A |
345 | c896fe29 | bellard | previous incarnation of the code generator had full support of them, |
346 | c896fe29 | bellard | but it is better to concentrate on integer operations first. |
347 | c896fe29 | bellard | |
348 | c896fe29 | bellard | On a 64 bit target, no assumption is made in TCG about the storage of |
349 | c896fe29 | bellard | the 32 bit values in 64 bit registers. |
350 | c896fe29 | bellard | |
351 | c896fe29 | bellard | 4.2) Constraints |
352 | c896fe29 | bellard | |
353 | c896fe29 | bellard | GCC like constraints are used to define the constraints of every |
354 | c896fe29 | bellard | instruction. Memory constraints are not supported in this |
355 | c896fe29 | bellard | version. Aliases are specified in the input operands as for GCC. |
356 | c896fe29 | bellard | |
357 | c896fe29 | bellard | A target can define specific register or constant constraints. If an |
358 | c896fe29 | bellard | operation uses a constant input constraint which does not allow all |
359 | c896fe29 | bellard | constants, it must also accept registers in order to have a fallback. |
360 | c896fe29 | bellard | |
361 | c896fe29 | bellard | The movi_i32 and movi_i64 operations must accept any constants. |
362 | c896fe29 | bellard | |
363 | c896fe29 | bellard | The mov_i32 and mov_i64 operations must accept any registers of the |
364 | c896fe29 | bellard | same type. |
365 | c896fe29 | bellard | |
366 | c896fe29 | bellard | The ld/st instructions must accept signed 32 bit constant offsets. It |
367 | c896fe29 | bellard | can be implemented by reserving a specific register to compute the |
368 | c896fe29 | bellard | address if the offset is too big. |
369 | c896fe29 | bellard | |
370 | c896fe29 | bellard | The ld/st instructions must accept any destination (ld) or source (st) |
371 | c896fe29 | bellard | register. |
372 | c896fe29 | bellard | |
373 | c896fe29 | bellard | 4.3) Function call assumptions |
374 | c896fe29 | bellard | |
375 | c896fe29 | bellard | - The only supported types for parameters and return value are: 32 and |
376 | c896fe29 | bellard | 64 bit integers and pointer. |
377 | c896fe29 | bellard | - The stack grows downwards. |
378 | c896fe29 | bellard | - The first N parameters are passed in registers. |
379 | c896fe29 | bellard | - The next parameters are passed on the stack by storing them as words. |
380 | c896fe29 | bellard | - Some registers are clobbered during the call. |
381 | c896fe29 | bellard | - The function can return 0 or 1 value in registers. On a 32 bit |
382 | c896fe29 | bellard | target, functions must be able to return 2 values in registers for |
383 | c896fe29 | bellard | 64 bit return type. |
384 | c896fe29 | bellard | |
385 | c896fe29 | bellard | 5) Migration from dyngen to TCG |
386 | c896fe29 | bellard | |
387 | c896fe29 | bellard | TCG is backward compatible with QEMU "dyngen" operations. It means |
388 | c896fe29 | bellard | that TCG instructions can be freely mixed with dyngen operations. It |
389 | c896fe29 | bellard | is expected that QEMU targets will be progressively fully converted to |
390 | 9804c8e2 | bellard | TCG. Once a target is fully converted to TCG, it will be possible |
391 | c896fe29 | bellard | to apply more optimizations because more registers will be free for |
392 | c896fe29 | bellard | the generated code. |
393 | c896fe29 | bellard | |
394 | c896fe29 | bellard | The exception model is the same as the dyngen one. |
395 | 0a6b7b78 | bellard | |
396 | 0a6b7b78 | bellard | 6) Recommended coding rules for best performance |
397 | 0a6b7b78 | bellard | |
398 | 0a6b7b78 | bellard | - Use globals to represent the parts of the QEMU CPU state which are |
399 | 0a6b7b78 | bellard | often modified, e.g. the integer registers and the condition |
400 | 0a6b7b78 | bellard | codes. TCG will be able to use host registers to store them. |
401 | 0a6b7b78 | bellard | |
402 | 0a6b7b78 | bellard | - Avoid globals stored in fixed registers. They must be used only to |
403 | 0a6b7b78 | bellard | store the pointer to the CPU state and possibly to store a pointer |
404 | 0a6b7b78 | bellard | to a register window. The other uses are to ensure backward |
405 | 0a6b7b78 | bellard | compatibility with dyngen during the porting a new target to TCG. |
406 | 0a6b7b78 | bellard | |
407 | 0a6b7b78 | bellard | - Use temporaries. Use local temporaries only when really needed, |
408 | 0a6b7b78 | bellard | e.g. when you need to use a value after a jump. Local temporaries |
409 | 0a6b7b78 | bellard | introduce a performance hit in the current TCG implementation: their |
410 | 0a6b7b78 | bellard | content is saved to memory at end of each basic block. |
411 | 0a6b7b78 | bellard | |
412 | 0a6b7b78 | bellard | - Free temporaries and local temporaries when they are no longer used |
413 | 0a6b7b78 | bellard | (tcg_temp_free). Since tcg_const_x() also creates a temporary, you |
414 | 0a6b7b78 | bellard | should free it after it is used. Freeing temporaries does not yield |
415 | 0a6b7b78 | bellard | a better generated code, but it reduces the memory usage of TCG and |
416 | 0a6b7b78 | bellard | the speed of the translation. |
417 | 0a6b7b78 | bellard | |
418 | 0a6b7b78 | bellard | - Don't hesitate to use helpers for complicated or seldom used target |
419 | 0a6b7b78 | bellard | intructions. There is little performance advantage in using TCG to |
420 | 0a6b7b78 | bellard | implement target instructions taking more than about twenty TCG |
421 | 0a6b7b78 | bellard | instructions. |
422 | 0a6b7b78 | bellard | |
423 | 0a6b7b78 | bellard | - Use the 'discard' instruction if you know that TCG won't be able to |
424 | 0a6b7b78 | bellard | prove that a given global is "dead" at a given program point. The |
425 | 0a6b7b78 | bellard | x86 target uses it to improve the condition codes optimisation. |