Add a comment about cache coherency and retranslation, so that people
developping new targets based on existing ones are warned of the issue.
Acked-by: Edgar E. Iglesias <edgar.iglesias@gmail.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Improve constant loading in two ways:
- On all ARM versions, it's possible to load 0xffffff00 = -0x100 using
the mvn rd, #0. Fix the conditions.
- On <= ARMv6 versions, where movw and movt are not available, load the
constants using mov and orr with rotations depending on the constant
to load. This is very useful for example to load constants where the
low byte is 0. This reduce the generated code size by about 7%.
Also fix the coding style at the same time.
Cc: Andrzej Zaborowski <balrog@zabor.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Due to a typo, qemu_st64 doesn't properly byteswap the 32-bit low word of
a 64 bit word before saving it. This patch fixes that.
Acked-by: Andrzej Zaborowski <balrogg@gmail.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
QEMU uses code retranslation to restore the CPU state when an exception
happens. For it to work the retranslation must not modify the generated
code. This is what is currently implemented in ARM TCG.
However on CPU that don't have icache/dcache/memory synchronised like
ARM, this requirement is stronger and code retranslation must not modify
the generated code "atomically", as the cache line might be flushed
at any moment (interrupt, exception, task switching), even if not
triggered by QEMU. The probability for this to happen is very low, and
depends on cache size and associativiy, machine load, interrupts, so the
symptoms are might happen randomly.
This requirement is currently not followed in tcg/arm, for the
load/store code, which basically has the following structure:
1) tlb access code is written
2) conditional fast path code is written
3) branch is written with a temporary target
4) slow path code is written
5) branch target is updated
The cache lines corresponding to the retranslated code is not flushed
after code retranslation as the generated code is supposed to be the
same. However if the cache line corresponding to the branch instruction
is flushed between step 3 and 5, and is not flushed again before the
code is executed again, the branch target is wrong. In the guest, the
symptoms are MMU page fault at a random addresses, which leads to
kernel page fault or segmentation faults.
The patch fixes this issue by avoiding writing the branch target until
it is known, that is by writing only the branch instruction first, and
later only the offset.
This fixes booting linux guests on ARM hosts (tested: arm, i386, mips,
mipsel, sh4, sparc).
Acked-by: Edgar E. Iglesias <edgar.iglesias@gmail.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Both tcg_target_init and tcg_target_qemu_prologue
are unused outside of tcg.c.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Mirror tcg_out_movi in having a TYPE parameter. This allows x86_64
to perform the move at the proper width, which may elide a REX prefix.
Introduce a TCG_TYPE_REG enumerator to represent the "native width"
of the host register, and to distinguish the usage from "pointer data"
as represented by the existing TCG_TYPE_PTR.
Update all targets to match.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The beginning of the register allocation order list on the TCG arm
target matches the list of clobbered registers. This means that when an
helper is called, there is almost always clobbered registers that have
to be spilled.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
64-bit arguments should be aligned on an even register as specified
by the "Procedure Call Standard for the ARM Architecture".
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
addr_reg, data_reg and data_reg2 can't be register r0 or r1 du to the
constraints. Don't check if they equals these registers.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
On big endian targets, data arguments of qemu_ld/st ops have to be
byte swapped. Two temporary registers are needed for qemu_st to do
the bswap. r0 and r1 are used in system mode, do the same in user
mode, which implies reworking the constraints.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
While it make sense to pass a conditional argument to tcg_out_*()
functions as the ARM architecture allows that, it doesn't make sense
for qemu_ld/st functions. These functions use comparison instructions
and conditional execution already, so it is not possible to use a
second level of conditional execution.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Add an bswap16 and bswap32 ops, either using the rev and rev16
instructions on ARMv6+ or shifts and logical operations on previous
ARM versions. In both cases the result use less instructions than
the pure TCG version.
These ops are also needed by the qemu_ld/st functions.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Add an ext16u op, either using the uxth instruction on ARMv6+ or two
shifts on previous ARM versions. In both cases the result use the same
number or less instructions than the pure TCG version.
Also move all sign extension code to separate functions, so that they
can be reused in other parts of the code.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Use a set of variables to define the allowed ARM instructions, depending
on the __ARM_ARCH_*__ GCC defines.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The TCG ARM backends uses integer values to refer to both immediate
values and register number. This makes the code difficult to read.
The patch below replaces all (if I haven't miss any ;-) integer values
representing register number by TCG_REG_* enum values.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Instead of writing very compact code, declare all registers that are
clobbered or reserved one by one. This makes the code easier to read.
Also declare all the 16 registers to TCG, and mark pc as reserved.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
There is no need to save the LR register (r14) before a call to a
subroutine. According to the "Procedure Call Standard for the ARM
Architecture", it is the job of the callee to save this register.
Moreover, this register is already saved in the prologue/epilogue.
This patch removes the disabled SAVE_LR code, as there is no need to
reenable later.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Commit 86feb1c860
did not change all occurrences of INDEX_op_qemu_ld32u
for tcg/arm.
Please note that I could not test this patch
(I have currently no arm system available).
Cc: Richard Henderson <rth@twiddle.net>
Cc: Aurelien Jarno <aurelien@aurel32.net>
Signed-off-by: Stefan Weil <weil@mail.berlios.de>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Some targets (e.g. Alpha and MIPS64) need to keep 32-bit operands
sign-extended in 64-bit registers (regardless of the "real" sign
of the operand). For that, we need to be able to distinguish
between a 32-bit load with a 32-bit result and a 32-bit load with
a given extension to a 64-bit result. This distinction already
exists for the ld* loads, but not the qemu_ld* loads.
Reserve qemu_ld32u for 64-bit outputs and introduce qemu_ld32 for
32-bit outputs. Adjust all code generators to match.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Give the enumeration formed from tcg-opc.h a name: TCGOpcode.
Use that enumeration type instead of "int" whereever appropriate.
Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
There is no need to save r7, it is used to store the address
of the env structure and is not modified by GCC.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Some targets like ARM would benefit to use 32-bit helpers for
div/rem/divu/remu.
Create a #define for div2 so that targets can select between
div, div2 and helper implementation. Use the helper version if none
of the #define are present.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Since commit 6113d6d316 QEMU crashes
on ARM hosts. This is not a bug of this commit, but a latent bug
revealed by this commit.
The TCG code is called through a procedure call using the prologue
and epilogue code. This code does not save and restore enough registers.
The "Procedure Call Standard for the ARM Architecture" says:
A subroutine must preserve the contents of the registers r4-r8, r10,
r11 and SP (and r9 in PCS variants that designate r9 as v6).
The current code only saves and restores r9 to r11, and misses r4 to
r8. The patch fixes that by saving r4 to r12. Theoretically there is
no need to save and restore r12, but an even number of registers have
to be saved as per EABI.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
When restoring register values, increase the stack register for skipped
values.
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Signed-off-by: Andrzej Zaborowski <andrew.zaborowski@intel.com>
This patch uses sxtb for ext8s_i32 and sxth for ext16s_i32 in ARM back-end.
Signed-off-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
the encode_imm function in tcg/arm/tcg-target.c lacks shift declaration.
Laurent
Signed-off-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Signed-off-by: Andrzej Zaborowski <andrew.zaborowski@intel.com>
this patch handles all possible constants for immediate operand of ALU ops.
I'm not very satisfied by the implementation.
Laurent
Signed-off-by: Laurent Desnogues <laurent.desnogues@gmail.com>
Signed-off-by: Andrzej Zaborowski <andrew.zaborowski@intel.com>
