Tiny Code Generator - Fabrice Bellard.

1) Introduction

TCG (Tiny Code Generator) began as a generic backend for a C

compiler. It was simplified to be used in QEMU. It also has its roots

in the QOP code generator written by Paul Brook.

2) Definitions

TCG receives RISC-like "TCG ops" and performs some optimizations on them,

including liveness analysis and trivial constant expression

evaluation. TCG ops are then implemented in the host CPU back end,

also known as the TCG "target".

The TCG "target" is the architecture for which we generate the

code. It is of course not the same as the "target" of QEMU which is

the emulated architecture. As TCG started as a generic C backend used

for cross compiling, it is assumed that the TCG target is different

from the host, although it is never the case for QEMU.

In this document, we use "guest" to specify what architecture we are

emulating; "target" always means the TCG target, the machine on which

we are running QEMU.

A TCG "function" corresponds to a QEMU Translated Block (TB).

A TCG "temporary" is a variable only live in a basic

block. Temporaries are allocated explicitly in each function.

A TCG "local temporary" is a variable only live in a function. Local

temporaries are allocated explicitly in each function.

A TCG "global" is a variable which is live in all the functions

(equivalent of a C global variable). They are defined before the

functions defined. A TCG global can be a memory location (e.g. a QEMU

CPU register), a fixed host register (e.g. the QEMU CPU state pointer)

or a memory location which is stored in a register outside QEMU TBs

(not implemented yet).

A TCG "basic block" corresponds to a list of instructions terminated

by a branch instruction.

An operation with "undefined behavior" may result in a crash.

An operation with "unspecified behavior" shall not crash. However,

the result may be one of several possibilities so may be considered

an "undefined result".

3) Intermediate representation

3.1) Introduction

TCG instructions operate on variables which are temporaries, local

temporaries or globals. TCG instructions and variables are strongly

typed. Two types are supported: 32 bit integers and 64 bit

integers. Pointers are defined as an alias to 32 bit or 64 bit

integers depending on the TCG target word size.

Each instruction has a fixed number of output variable operands, input

variable operands and always constant operands.

The notable exception is the call instruction which has a variable

number of outputs and inputs.

In the textual form, output operands usually come first, followed by

input operands, followed by constant operands. The output type is

included in the instruction name. Constants are prefixed with a '$'.

add_i32 t0, t1, t2 (t0 <- t1 + t2)

3.2) Assumptions

* Basic blocks

- Basic blocks end after branches (e.g. brcond_i32 instruction),

goto_tb and exit_tb instructions.

- Basic blocks start after the end of a previous basic block, or at a

set_label instruction.

After the end of a basic block, the content of temporaries is

destroyed, but local temporaries and globals are preserved.

* Floating point types are not supported yet

* Pointers: depending on the TCG target, pointer size is 32 bit or 64

bit. The type TCG_TYPE_PTR is an alias to TCG_TYPE_I32 or

TCG_TYPE_I64.

* Helpers:

Using the tcg_gen_helper_x_y it is possible to call any function

taking i32, i64 or pointer types. By default, before calling a helper,

all globals are stored at their canonical location and it is assumed

that the function can modify them. By default, the helper is allowed to

modify the CPU state or raise an exception.

This can be overridden using the following function modifiers:

- TCG_CALL_NO_READ_GLOBALS means that the helper does not read globals,

either directly or via an exception. They will not be saved to their

canonical locations before calling the helper.

- TCG_CALL_NO_WRITE_GLOBALS means that the helper does not modify any globals.

They will only be saved to their canonical location before calling helpers,

but they won't be reloaded afterwards.

- TCG_CALL_NO_SIDE_EFFECTS means that the call to the function is removed if

the return value is not used.

Note that TCG_CALL_NO_READ_GLOBALS implies TCG_CALL_NO_WRITE_GLOBALS.

On some TCG targets (e.g. x86), several calling conventions are

supported.

* Branches:

Use the instruction 'br' to jump to a label.

3.3) Code Optimizations

When generating instructions, you can count on at least the following

optimizations:

- Single instructions are simplified, e.g.

and_i32 t0, t0, $0xffffffff

is suppressed.

- A liveness analysis is done at the basic block level. The

information is used to suppress moves from a dead variable to

another one. It is also used to remove instructions which compute

dead results. The later is especially useful for condition code

optimization in QEMU.

In the following example:

add_i32 t0, t1, t2

add_i32 t0, t0, $1

mov_i32 t0, $1

only the last instruction is kept.

3.4) Instruction Reference

********* Function call

* call <ret> <params> ptr

call function 'ptr' (pointer type)

<ret> optional 32 bit or 64 bit return value

<params> optional 32 bit or 64 bit parameters

********* Jumps/Labels

* set_label $label

Define label 'label' at the current program point.

* br $label

Jump to label.

* brcond_i32/i64 t0, t1, cond, label

Conditional jump if t0 cond t1 is true. cond can be:

TCG_COND_EQ

TCG_COND_NE

TCG_COND_LT /* signed */

TCG_COND_GE /* signed */

TCG_COND_LE /* signed */

TCG_COND_GT /* signed */

TCG_COND_LTU /* unsigned */

TCG_COND_GEU /* unsigned */

TCG_COND_LEU /* unsigned */

TCG_COND_GTU /* unsigned */

********* Arithmetic

* add_i32/i64 t0, t1, t2

t0=t1+t2

* sub_i32/i64 t0, t1, t2

t0=t1-t2

* neg_i32/i64 t0, t1

t0=-t1 (two's complement)

* mul_i32/i64 t0, t1, t2

t0=t1*t2

* div_i32/i64 t0, t1, t2

t0=t1/t2 (signed). Undefined behavior if division by zero or overflow.

* divu_i32/i64 t0, t1, t2

t0=t1/t2 (unsigned). Undefined behavior if division by zero.

* rem_i32/i64 t0, t1, t2

t0=t1%t2 (signed). Undefined behavior if division by zero or overflow.

* remu_i32/i64 t0, t1, t2

t0=t1%t2 (unsigned). Undefined behavior if division by zero.

********* Logical

* and_i32/i64 t0, t1, t2

t0=t1&t2

* or_i32/i64 t0, t1, t2

t0=t1|t2

* xor_i32/i64 t0, t1, t2

t0=t1^t2

* not_i32/i64 t0, t1

t0=~t1

* andc_i32/i64 t0, t1, t2

t0=t1&~t2

* eqv_i32/i64 t0, t1, t2

t0=~(t1^t2), or equivalently, t0=t1^~t2

* nand_i32/i64 t0, t1, t2

t0=~(t1&t2)

* nor_i32/i64 t0, t1, t2

t0=~(t1|t2)

* orc_i32/i64 t0, t1, t2

t0=t1|~t2

* clz_i32/i64 t0, t1, t2

t0 = t1 ? clz(t1) : t2

* ctz_i32/i64 t0, t1, t2

t0 = t1 ? ctz(t1) : t2

********* Shifts/Rotates

* shl_i32/i64 t0, t1, t2

t0=t1 << t2. Unspecified behavior if t2 < 0 or t2 >= 32 (resp 64)

* shr_i32/i64 t0, t1, t2

t0=t1 >> t2 (unsigned). Unspecified behavior if t2 < 0 or t2 >= 32 (resp 64)

* sar_i32/i64 t0, t1, t2

t0=t1 >> t2 (signed). Unspecified behavior if t2 < 0 or t2 >= 32 (resp 64)

* rotl_i32/i64 t0, t1, t2

Rotation of t2 bits to the left.

Unspecified behavior if t2 < 0 or t2 >= 32 (resp 64)

* rotr_i32/i64 t0, t1, t2

Rotation of t2 bits to the right.

Unspecified behavior if t2 < 0 or t2 >= 32 (resp 64)

********* Misc

* mov_i32/i64 t0, t1

t0 = t1

Move t1 to t0 (both operands must have the same type).

* ext8s_i32/i64 t0, t1

ext8u_i32/i64 t0, t1

ext16s_i32/i64 t0, t1

ext16u_i32/i64 t0, t1

ext32s_i64 t0, t1

ext32u_i64 t0, t1

8, 16 or 32 bit sign/zero extension (both operands must have the same type)

* bswap16_i32/i64 t0, t1

16 bit byte swap on a 32/64 bit value. It assumes that the two/six high order

bytes are set to zero.

* bswap32_i32/i64 t0, t1

32 bit byte swap on a 32/64 bit value. With a 64 bit value, it assumes that

the four high order bytes are set to zero.

* bswap64_i64 t0, t1

64 bit byte swap

* discard_i32/i64 t0

Indicate that the value of t0 won't be used later. It is useful to

force dead code elimination.

* deposit_i32/i64 dest, t1, t2, pos, len

Deposit T2 as a bitfield into T1, placing the result in DEST.

The bitfield is described by POS/LEN, which are immediate values:

LEN - the length of the bitfield

POS - the position of the first bit, counting from the LSB

For example, "deposit_i32 dest, t1, t2, 8, 4" indicates a 4-bit field

at bit 8. This operation would be equivalent to

dest = (t1 & ~0x0f00) | ((t2 << 8) & 0x0f00)

* extract_i32/i64 dest, t1, pos, len

* sextract_i32/i64 dest, t1, pos, len

Extract a bitfield from T1, placing the result in DEST.

The bitfield is described by POS/LEN, which are immediate values,

as above for deposit. For extract_*, the result will be extended

to the left with zeros; for sextract_*, the result will be extended

to the left with copies of the bitfield sign bit at pos + len - 1.

For example, "sextract_i32 dest, t1, 8, 4" indicates a 4-bit field

at bit 8. This operation would be equivalent to

dest = (t1 << 20) >> 28

(using an arithmetic right shift).

* extract2_i32/i64 dest, t1, t2, pos

For N = {32,64}, extract an N-bit quantity from the concatenation

of t2:t1, beginning at pos. The tcg_gen_extract2_{i32,i64} expander

accepts 0 <= pos <= N as inputs. The backend code generator will

not see either 0 or N as inputs for these opcodes.

* extrl_i64_i32 t0, t1

For 64-bit hosts only, extract the low 32-bits of input T1 and place it

into 32-bit output T0. Depending on the host, this may be a simple move,

or may require additional canonicalization.

* extrh_i64_i32 t0, t1

For 64-bit hosts only, extract the high 32-bits of input T1 and place it

into 32-bit output T0. Depending on the host, this may be a simple shift,

or may require additional canonicalization.

********* Conditional moves

* setcond_i32/i64 dest, t1, t2, cond

dest = (t1 cond t2)

Set DEST to 1 if (T1 cond T2) is true, otherwise set to 0.

* movcond_i32/i64 dest, c1, c2, v1, v2, cond

dest = (c1 cond c2 ? v1 : v2)

Set DEST to V1 if (C1 cond C2) is true, otherwise set to V2.

********* Type conversions

* ext_i32_i64 t0, t1

Convert t1 (32 bit) to t0 (64 bit) and does sign extension

* extu_i32_i64 t0, t1

Convert t1 (32 bit) to t0 (64 bit) and does zero extension

* trunc_i64_i32 t0, t1

Truncate t1 (64 bit) to t0 (32 bit)

* concat_i32_i64 t0, t1, t2

Construct t0 (64-bit) taking the low half from t1 (32 bit) and the high half

from t2 (32 bit).

* concat32_i64 t0, t1, t2

Construct t0 (64-bit) taking the low half from t1 (64 bit) and the high half

from t2 (64 bit).

********* Load/Store

* ld_i32/i64 t0, t1, offset

ld8s_i32/i64 t0, t1, offset

ld8u_i32/i64 t0, t1, offset

ld16s_i32/i64 t0, t1, offset

ld16u_i32/i64 t0, t1, offset

ld32s_i64 t0, t1, offset

ld32u_i64 t0, t1, offset

t0 = read(t1 + offset)

Load 8, 16, 32 or 64 bits with or without sign extension from host memory.

offset must be a constant.

* st_i32/i64 t0, t1, offset

st8_i32/i64 t0, t1, offset

st16_i32/i64 t0, t1, offset

st32_i64 t0, t1, offset

write(t0, t1 + offset)

Write 8, 16, 32 or 64 bits to host memory.

All this opcodes assume that the pointed host memory doesn't correspond

to a global. In the latter case the behaviour is unpredictable.

********* Multiword arithmetic support

* add2_i32/i64 t0_low, t0_high, t1_low, t1_high, t2_low, t2_high

* sub2_i32/i64 t0_low, t0_high, t1_low, t1_high, t2_low, t2_high

Similar to add/sub, except that the double-word inputs T1 and T2 are

formed from two single-word arguments, and the double-word output T0

is returned in two single-word outputs.

* mulu2_i32/i64 t0_low, t0_high, t1, t2

Similar to mul, except two unsigned inputs T1 and T2 yielding the full

double-word product T0. The later is returned in two single-word outputs.

* muls2_i32/i64 t0_low, t0_high, t1, t2

Similar to mulu2, except the two inputs T1 and T2 are signed.

* mulsh_i32/i64 t0, t1, t2

* muluh_i32/i64 t0, t1, t2

Provide the high part of a signed or unsigned multiply, respectively.

If mulu2/muls2 are not provided by the backend, the tcg-op generator

can obtain the same results can be obtained by emitting a pair of

opcodes, mul+muluh/mulsh.

********* Memory Barrier support

* mb <$arg>

Generate a target memory barrier instruction to ensure memory ordering as being

enforced by a corresponding guest memory barrier instruction. The ordering

enforced by the backend may be stricter than the ordering required by the guest.

It cannot be weaker. This opcode takes a constant argument which is required to

generate the appropriate barrier instruction. The backend should take care to

emit the target barrier instruction only when necessary i.e., for SMP guests and

when MTTCG is enabled.

The guest translators should generate this opcode for all guest instructions

which have ordering side effects.

Please see docs/devel/atomics.txt for more information on memory barriers.

********* 64-bit guest on 32-bit host support

The following opcodes are internal to TCG. Thus they are to be implemented by

32-bit host code generators, but are not to be emitted by guest translators.

They are emitted as needed by inline functions within "tcg-op.h".

* brcond2_i32 t0_low, t0_high, t1_low, t1_high, cond, label

Similar to brcond, except that the 64-bit values T0 and T1

are formed from two 32-bit arguments.

* setcond2_i32 dest, t1_low, t1_high, t2_low, t2_high, cond

Similar to setcond, except that the 64-bit values T1 and T2 are

formed from two 32-bit arguments. The result is a 32-bit value.

********* QEMU specific operations

* exit_tb t0

Exit the current TB and return the value t0 (word type).

* goto_tb index

Exit the current TB and jump to the TB index 'index' (constant) if the

current TB was linked to this TB. Otherwise execute the next

instructions. Only indices 0 and 1 are valid and tcg_gen_goto_tb may be issued

at most once with each slot index per TB.

* lookup_and_goto_ptr tb_addr

Look up a TB address ('tb_addr') and jump to it if valid. If not valid,

jump to the TCG epilogue to go back to the exec loop.

This operation is optional. If the TCG backend does not implement the

goto_ptr opcode, emitting this op is equivalent to emitting exit_tb(0).

* qemu_ld_i32/i64 t0, t1, flags, memidx

* qemu_st_i32/i64 t0, t1, flags, memidx

Load data at the guest address t1 into t0, or store data in t0 at guest

address t1. The _i32/_i64 size applies to the size of the input/output

register t0 only. The address t1 is always sized according to the guest,

and the width of the memory operation is controlled by flags.

Both t0 and t1 may be split into little-endian ordered pairs of registers

if dealing with 64-bit quantities on a 32-bit host.

The memidx selects the qemu tlb index to use (e.g. user or kernel access).

The flags are the MemOp bits, selecting the sign, width, and endianness

of the memory access.

For a 32-bit host, qemu_ld/st_i64 is guaranteed to only be used with a

64-bit memory access specified in flags.

********* Host vector operations

All of the vector ops have two parameters, TCGOP_VECL & TCGOP_VECE.

The former specifies the length of the vector in log2 64-bit units; the

later specifies the length of the element (if applicable) in log2 8-bit units.

E.g. VECL=1 -> 64 << 1 -> v128, and VECE=2 -> 1 << 2 -> i32.

* mov_vec v0, v1

* ld_vec v0, t1

* st_vec v0, t1

Move, load and store.

* dup_vec v0, r1

Duplicate the low N bits of R1 into VECL/VECE copies across V0.

* dupi_vec v0, c

Similarly, for a constant.

Smaller values will be replicated to host register size by the expanders.

* dup2_vec v0, r1, r2

Duplicate r2:r1 into VECL/64 copies across V0. This opcode is

only present for 32-bit hosts.

* add_vec v0, v1, v2

v0 = v1 + v2, in elements across the vector.

* sub_vec v0, v1, v2

Similarly, v0 = v1 - v2.

* mul_vec v0, v1, v2

Similarly, v0 = v1 * v2.

* neg_vec v0, v1

Similarly, v0 = -v1.

* abs_vec v0, v1

Similarly, v0 = v1 < 0 ? -v1 : v1, in elements across the vector.

* smin_vec:

* umin_vec:

Similarly, v0 = MIN(v1, v2), for signed and unsigned element types.

* smax_vec:

* umax_vec:

Similarly, v0 = MAX(v1, v2), for signed and unsigned element types.

* ssadd_vec:

* sssub_vec:

* usadd_vec:

* ussub_vec:

Signed and unsigned saturating addition and subtraction. If the true

result is not representable within the element type, the element is

set to the minimum or maximum value for the type.

* and_vec v0, v1, v2

* or_vec v0, v1, v2

* xor_vec v0, v1, v2

* andc_vec v0, v1, v2

* orc_vec v0, v1, v2

* not_vec v0, v1

Similarly, logical operations with and without complement.

Note that VECE is unused.

* shli_vec v0, v1, i2

* shls_vec v0, v1, s2

Shift all elements from v1 by a scalar i2/s2. I.e.

for (i = 0; i < VECL/VECE; ++i) {

v0[i] = v1[i] << s2;

}

* shri_vec v0, v1, i2

* sari_vec v0, v1, i2

* shrs_vec v0, v1, s2

* sars_vec v0, v1, s2

Similarly for logical and arithmetic right shift.

* shlv_vec v0, v1, v2

Shift elements from v1 by elements from v2. I.e.

for (i = 0; i < VECL/VECE; ++i) {

v0[i] = v1[i] << v2[i];

}

* shrv_vec v0, v1, v2

* sarv_vec v0, v1, v2

Similarly for logical and arithmetic right shift.

* cmp_vec v0, v1, v2, cond

Compare vectors by element, storing -1 for true and 0 for false.

* bitsel_vec v0, v1, v2, v3

Bitwise select, v0 = (v2 & v1) | (v3 & ~v1), across the entire vector.

* cmpsel_vec v0, c1, c2, v3, v4, cond

Select elements based on comparison results:

for (i = 0; i < n; ++i) {

v0[i] = (c1[i] cond c2[i]) ? v3[i] : v4[i].

}

*********

Note 1: Some shortcuts are defined when the last operand is known to be

a constant (e.g. addi for add, movi for mov).

Note 2: When using TCG, the opcodes must never be generated directly

as some of them may not be available as "real" opcodes. Always use the

function tcg_gen_xxx(args).

4) Backend

tcg-target.h contains the target specific definitions. tcg-target.inc.c

contains the target specific code; it is #included by tcg/tcg.c, rather

than being a standalone C file.

4.1) Assumptions

The target word size (TCG_TARGET_REG_BITS) is expected to be 32 bit or

64 bit. It is expected that the pointer has the same size as the word.

On a 32 bit target, all 64 bit operations are converted to 32 bits. A

few specific operations must be implemented to allow it (see add2_i32,

sub2_i32, brcond2_i32).

On a 64 bit target, the values are transferred between 32 and 64-bit

registers using the following ops:

- trunc_shr_i64_i32

- ext_i32_i64

- extu_i32_i64

They ensure that the values are correctly truncated or extended when

moved from a 32-bit to a 64-bit register or vice-versa. Note that the

trunc_shr_i64_i32 is an optional op. It is not necessary to implement

it if all the following conditions are met:

- 64-bit registers can hold 32-bit values

- 32-bit values in a 64-bit register do not need to stay zero or

sign extended

- all 32-bit TCG ops ignore the high part of 64-bit registers

Floating point operations are not supported in this version. A

previous incarnation of the code generator had full support of them,

but it is better to concentrate on integer operations first.

4.2) Constraints

GCC like constraints are used to define the constraints of every

instruction. Memory constraints are not supported in this

version. Aliases are specified in the input operands as for GCC.

The same register may be used for both an input and an output, even when

they are not explicitly aliased. If an op expands to multiple target

instructions then care must be taken to avoid clobbering input values.

GCC style "early clobber" outputs are supported, with '&'.

A target can define specific register or constant constraints. If an

operation uses a constant input constraint which does not allow all

constants, it must also accept registers in order to have a fallback.

The constraint 'i' is defined generically to accept any constant.

The constraint 'r' is not defined generically, but is consistently

used by each backend to indicate all registers.

The movi_i32 and movi_i64 operations must accept any constants.

The mov_i32 and mov_i64 operations must accept any registers of the

same type.

The ld/st/sti instructions must accept signed 32 bit constant offsets.

This can be implemented by reserving a specific register in which to

compute the address if the offset is too big.

The ld/st instructions must accept any destination (ld) or source (st)

register.

The sti instruction may fail if it cannot store the given constant.

4.3) Function call assumptions

- The only supported types for parameters and return value are: 32 and

64 bit integers and pointer.

- The stack grows downwards.

- The first N parameters are passed in registers.

- The next parameters are passed on the stack by storing them as words.

- Some registers are clobbered during the call.

- The function can return 0 or 1 value in registers. On a 32 bit

target, functions must be able to return 2 values in registers for

64 bit return type.

5) Recommended coding rules for best performance

- Use globals to represent the parts of the QEMU CPU state which are

often modified, e.g. the integer registers and the condition

codes. TCG will be able to use host registers to store them.

- Avoid globals stored in fixed registers. They must be used only to

store the pointer to the CPU state and possibly to store a pointer

to a register window.

- Use temporaries. Use local temporaries only when really needed,

e.g. when you need to use a value after a jump. Local temporaries

introduce a performance hit in the current TCG implementation: their

content is saved to memory at end of each basic block.

- Free temporaries and local temporaries when they are no longer used

(tcg_temp_free). Since tcg_const_x() also creates a temporary, you

should free it after it is used. Freeing temporaries does not yield

a better generated code, but it reduces the memory usage of TCG and

the speed of the translation.

- Don't hesitate to use helpers for complicated or seldom used guest

instructions. There is little performance advantage in using TCG to

implement guest instructions taking more than about twenty TCG

instructions. Note that this rule of thumb is more applicable to

helpers doing complex logic or arithmetic, where the C compiler has

scope to do a good job of optimisation; it is less relevant where

the instruction is mostly doing loads and stores, and in those cases

inline TCG may still be faster for longer sequences.

- The hard limit on the number of TCG instructions you can generate

per guest instruction is set by MAX_OP_PER_INSTR in exec-all.h --

you cannot exceed this without risking a buffer overrun.

- Use the 'discard' instruction if you know that TCG won't be able to

prove that a given global is "dead" at a given program point. The

x86 guest uses it to improve the condition codes optimisation.