qemu/tcg
Emilio G. Cota e8feb96fcc tcg: introduce regions to split code_gen_buffer
This is groundwork for supporting multiple TCG contexts.

The naive solution here is to split code_gen_buffer statically
among the TCG threads; this however results in poor utilization
if translation needs are different across TCG threads.

What we do here is to add an extra layer of indirection, assigning
regions that act just like pages do in virtual memory allocation.
(BTW if you are wondering about the chosen naming, I did not want
to use blocks or pages because those are already heavily used in QEMU).

We use a global lock to serialize allocations as well as statistics
reporting (we now export the size of the used code_gen_buffer with
tcg_code_size()). Note that for the allocator we could just use
a counter and atomic_inc; however, that would complicate the gathering
of tcg_code_size()-like stats. So given that the region operations are
not a fast path, a lock seems the most reasonable choice.

The effectiveness of this approach is clear after seeing some numbers.
I used the bootup+shutdown of debian-arm with '-tb-size 80' as a benchmark.
Note that I'm evaluating this after enabling per-thread TCG (which
is done by a subsequent commit).

* -smp 1, 1 region (entire buffer):
    qemu: flush code_size=83885014 nb_tbs=154739 avg_tb_size=357
    qemu: flush code_size=83884902 nb_tbs=153136 avg_tb_size=363
    qemu: flush code_size=83885014 nb_tbs=152777 avg_tb_size=364
    qemu: flush code_size=83884950 nb_tbs=150057 avg_tb_size=373
    qemu: flush code_size=83884998 nb_tbs=150234 avg_tb_size=373
    qemu: flush code_size=83885014 nb_tbs=154009 avg_tb_size=360
    qemu: flush code_size=83885014 nb_tbs=151007 avg_tb_size=370
    qemu: flush code_size=83885014 nb_tbs=151816 avg_tb_size=367

That is, 8 flushes.

* -smp 8, 32 regions (80/32 MB per region) [i.e. this patch]:

    qemu: flush code_size=76328008 nb_tbs=141040 avg_tb_size=356
    qemu: flush code_size=75366534 nb_tbs=138000 avg_tb_size=361
    qemu: flush code_size=76864546 nb_tbs=140653 avg_tb_size=361
    qemu: flush code_size=76309084 nb_tbs=135945 avg_tb_size=375
    qemu: flush code_size=74581856 nb_tbs=132909 avg_tb_size=375
    qemu: flush code_size=73927256 nb_tbs=135616 avg_tb_size=360
    qemu: flush code_size=78629426 nb_tbs=142896 avg_tb_size=365
    qemu: flush code_size=76667052 nb_tbs=138508 avg_tb_size=368

Again, 8 flushes. Note how buffer utilization is not 100%, but it
is close. Smaller region sizes would yield higher utilization,
but we want region allocation to be rare (it acquires a lock), so
we do not want to go too small.

* -smp 8, static partitioning of 8 regions (10 MB per region):
    qemu: flush code_size=21936504 nb_tbs=40570 avg_tb_size=354
    qemu: flush code_size=11472174 nb_tbs=20633 avg_tb_size=370
    qemu: flush code_size=11603976 nb_tbs=21059 avg_tb_size=365
    qemu: flush code_size=23254872 nb_tbs=41243 avg_tb_size=377
    qemu: flush code_size=28289496 nb_tbs=52057 avg_tb_size=358
    qemu: flush code_size=43605160 nb_tbs=78896 avg_tb_size=367
    qemu: flush code_size=45166552 nb_tbs=82158 avg_tb_size=364
    qemu: flush code_size=63289640 nb_tbs=116494 avg_tb_size=358
    qemu: flush code_size=51389960 nb_tbs=93937 avg_tb_size=362
    qemu: flush code_size=59665928 nb_tbs=107063 avg_tb_size=372
    qemu: flush code_size=38380824 nb_tbs=68597 avg_tb_size=374
    qemu: flush code_size=44884568 nb_tbs=79901 avg_tb_size=376
    qemu: flush code_size=50782632 nb_tbs=90681 avg_tb_size=374
    qemu: flush code_size=39848888 nb_tbs=71433 avg_tb_size=372
    qemu: flush code_size=64708840 nb_tbs=119052 avg_tb_size=359
    qemu: flush code_size=49830008 nb_tbs=90992 avg_tb_size=362
    qemu: flush code_size=68372408 nb_tbs=123442 avg_tb_size=368
    qemu: flush code_size=33555560 nb_tbs=59514 avg_tb_size=378
    qemu: flush code_size=44748344 nb_tbs=80974 avg_tb_size=367
    qemu: flush code_size=37104248 nb_tbs=67609 avg_tb_size=364

That is, 20 flushes. Note how a static partitioning approach uses
the code buffer poorly, leading to many unnecessary flushes.

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
2017-10-24 13:53:42 -07:00
..
aarch64 tcg/aarch64: Fully convert tcg_target_op_def 2017-09-17 06:52:19 -07:00
arm tcg/arm: Fully convert tcg_target_op_def 2017-09-17 06:52:19 -07:00
i386 tcg/i386: constify tcg_target_callee_save_regs 2017-10-10 07:37:10 -07:00
mips tcg/mips: delete commented out extern keyword. 2017-10-10 09:45:01 -07:00
ppc tcg/ppc: Fully convert tcg_target_op_def 2017-09-17 06:52:19 -07:00
s390 tcg: Remove tcg_regset_set32 2017-09-17 06:52:19 -07:00
sparc tcg/sparc: Fully convert tcg_target_op_def 2017-09-17 06:52:19 -07:00
tci tcg: Remove tcg_regset_set32 2017-09-17 06:52:19 -07:00
LICENSE
optimize.c tcg: allocate optimizer temps with tcg_malloc 2017-10-24 13:53:42 -07:00
README docs: fix broken paths to docs/devel/atomics.txt 2017-07-31 13:12:47 +03:00
tcg-common.c exec: extract exec/tb-context.h 2016-05-19 16:42:29 +02:00
tcg-ldst.inc.c tcg: Rearrange ldst label tracking 2017-09-07 11:57:35 -07:00
tcg-mo.h tcg: move TCG_MO/BAR types into own file 2017-02-24 10:32:45 +00:00
tcg-op.c tcg: define tcg_init_ctx and make tcg_ctx a pointer 2017-10-24 13:53:42 -07:00
tcg-op.h tcg: Remove TCGV_EQUAL* 2017-10-24 21:50:15 +02:00
tcg-opc.h tcg: Introduce goto_ptr opcode and tcg_gen_lookup_and_goto_ptr 2017-06-05 09:25:42 -07:00
tcg-pool.inc.c tcg: Infrastructure for managing constant pools 2017-09-07 11:57:35 -07:00
tcg.c tcg: introduce regions to split code_gen_buffer 2017-10-24 13:53:42 -07:00
tcg.h tcg: introduce regions to split code_gen_buffer 2017-10-24 13:53:42 -07:00
tci.c tci: move tci_regs to tcg_qemu_tb_exec's stack 2017-10-10 07:37:10 -07:00
TODO tcg/TODO: remove setcond 2010-04-01 22:00:41 +02:00

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 afterwise.
- 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).

* 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.

********* 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 TCGMemOp 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.

*********

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.