LONG MUL can be optimzied into left shift if either operand is a power
of two. Conditions "IS_SIGNED_32BIT()" and "is_power_of_two()" are used
to filter out invalid candidates. However, there exists one exception,
i.e. -2147483648(that is 0xffff,ffff,8000,0000). See the stand-alone
case[1].
Assume "a = 3; b = -2147483648;". The expected result of "a * b" is one
negative value. However, it would be optimized to "a << 31", which is
positive.
This trigger condition is refined.
1) For x86 implementation, another check for positive numbers is added.
Note that LONG type, i.e. zend_long, is defined as int32_t for x86 arch
and int64_t for x64 arch. This optimization only accepts values which
can be represented by int32_t type as default. See IS_SIGNED_32BIlT(),
2) For AArch64, we employ helper function zend_long_is_power_of_two()
since values of int64_t type are used.
Overflow detection for left shifting is added in this patch as well.
Note 1: bit helper functions are arch-independent and we move them into
zend_jit_internals.h.
Note 2: two test cases are added. Test case mul_003.phpt is used to
check the trigger condition and mul_004.phpt is designed to check
overflow detection.
Note 3: overflow detection for x86 is not implemented yet as I think
anotehr temporay register besides R0 is needed. Hence mul_004.phpt would
fail on x86 machine.
If we can use R1 as tmp_reg, the code can be updated as below.
```
| GET_ZVAL_LVAL result_reg, op1_addr
if (may_overflow) {
use_ovf_flag = 0;
/* Compare 'op' and '((op << n) >> n)' for overflow.
* Flag: jne -> overflow. je -> no overflow.
*/
tmp_reg = ZREG_R1
| mov Ra(tmp_reg), Ra(result_reg)
| shl Ra(tmp_reg), floor_log2(Z_LVAL_P(Z_ZV(op2_addr)))
| sar Ra(tmp_reg), floor_log2(Z_LVAL_P(Z_ZV(op2_addr)))
| cmp Ra(tmp_reg), Ra(result_reg)
}
| shl Ra(result_reg), floor_log2(Z_LVAL_P(Z_ZV(op2_addr)))
```
[1]. https://godbolt.org/z/1vKbfv8oG
Change-Id: Ie90e1d4e7c8b94a0c8f61386dfe650fa2c6879a1
1. Pre-allocated bytes are missing in function
zend_jit_assign_const_stub().
2. 'w' register should be used for macro SET_ZVAL_TYPE_INFO.
3. 'w' register should be used to load "num_args" in function
zend_jit_do_fcall().
4. Remove the local path name in test case recv_002.phpt
5. One option is disabled temporarily in [1] and several test cases
would fail, e.g. shift_right_003.phpt. I suppose new execution paths are
touched. We will support them in the near future.
[1] https://github.com/php/php-src/commit/63d673d
SUMMARY
We implemented a prototype of PHP JIT/arm64. Briefly speaking,
1. build system
Changes to the build system are made so that PHP JIT can be successfully
built and run on ARM-based machine.
Major change lies in file zend_jit_arm64.dasc, where the handler for
each opcode is generated into machine code. Note that this file is just
copied from zend_jit_x86.dasc and the *unimplemented* parts are
substitued with 'brk' instruction for future work.
2. registers
AArch64 registers are defined in file zend_jit_arm64.h. From our
perspectives, the register usage is quite different from the x86
implementation due to the different ABI, number of registers and
addressing modes.
We had many confusions on this part, and will discuss it in details in
the final section.
3. opcodes
Several opcodes are partially supported, including INIT_FCALL, DO_UCALL,
DO_ICALL, RETURN, ADD, PRE_INC, JMP, QM_ASSIGN, etc. Hence, simple use
scenarios such as user function call, loops, addition with integer and
floating point numbers can be supported.
18 micro test cases are added under 'ext/opcache/tests/jit/arm64/'. Note
that majority of these test cases are design for functional JIT, and
cases 'hot_func_*.phpt' and 'loop_002.phpt' can trigger tracing JIT.
4. test
Our local test environment is an ARM-based server with Ubuntu 20.04 and
GCC-10. Note that both HYBRID and CALL VM modes are supported. We
suggest running the JIT test cases using the following command. Out of
all 130 test cases, 66 cases can be passed currently.
```
$ make test TESTS='-d opcache.jit=1203 ext/opcache/tests/jit/'
```
DETAILS
1. I-cache flush
Instruction cache must be flushed for the JIT-ed code on AArch64. See
macro JIT_CACHE_FLUSH in file 'zend_jit_internal.h'.
2. Disassembler
Add initialization and jump target parse operations for AArch64 backed.
See the updates in file 'zend_jit_disasm.c'.
3. redzone
Enable redzone for AArch64. See the update in zend_vm_opcodes.h.
Redzone is designated to prevent 'vm_stack_data' from being optimized
out by compilers. It's worth noting that this 16-byte redzone might be
reused as temporary use(treated as extra stack space) for HYBRID mode.
4. stack space reservation
The definitions of HYBRID_SPAD, SPAD and NR_SPAD are a bit tricky for
x86/64.
In AArch64, HYBRID_SPAD and SPAD are both defined as 16. These 16 bytes
are pre-allocated for tempoerary usage along the exuection of JIT-ed
code. Take line 4185 in file zend_jit_arm64.dasc as an example. NR_SPAD
is defined as 48, out of which 32 bytes to save FP/IP/LR registers.
Note that we choose to always reserve HYBRID_SPAD bytes in HYBRID mode,
no matter whether redzone is used or not, for the sake of safety.
5. stack alignment
In AArch64 the stack pointer should be 16-byte aligned. Since shadow
stack is used for JIT, it's easy to guarantee the stack alignment, via
simply moving SP with an offset like 16 or a multiple of 16. That's why
NR_SPAD is defined as 48 and we use 32 of them to save FP/IP/LR
registers which only occupies 24 bytes.
6. global registers
x27 and x28 are reserved as global registers. See the updates in file
zend_jit_vm_helpers.c
7. function prologue for CALL mode
Two callee-saved registers x27 and x28 should saved in function
zend_jit_prologue() in file zend_jit_arm64.dasc. Besides the LR, i.e.
x30, should also be saved since runtime C helper functions(such as
zend_jit_find_func_helper) might be invoked along the execution of
JIT-ed code.
8. regset
Minor changes are done to regset operations particularly for AArch64.
See the updates in file zend_jit_internal.h.
REGISTER USAGE
In this section, we will first talk about our understanding on register
usage and then demonstrate our design.
1. Register usage for HYBRID/CALL modes
Registers are used similarly between HYBRID mode and CALL mode.
One difference is how FP and IP are saved. In HYBRID mode, they are
assigned to global registers, while in CALL mode they are saved/restored
on the VM stack explicitly in prologue/epilogue.
The other difference is that LR register should also be saved/restored
in CALL mode since JIT-ed code are invoked as normal functions.
2. Register usage for functional/tracing JIT
The way registers are used differs a lot between functional JIT and
tracing JIT.
For functional JIT, runtime C code (e.g. helper functions) would be
invoked along the execution of JIT-ed code. As the operands for *most*
opcodes are accessed via the stack slot, i.e. FP + offset. Hence there
is no need to save/restore local(caller-saved) registers before/after
invoking runtime C code.
Exception lies in Phi node and registers might be allocated for these
nodes. Currently I don't fully understand the reason, why registers are
allocated for Phi functions, because I suppose for different versions of
SSA variables at the Phi function, their postions on the stack slot
should be identical(in other words, access via the stack slot is enough
and there is no need to allocate registers).
For tracing JIT, runtime information are recorded for traces(before the
JIT compilation), and the data types and control flows are concrete as
well. Hence it's would be faster to conduct operations and computations
via registers rather than stack slots(as functional JIT does) for these
collected hot paths. Besides, runtime C code can be invoked for tracing
JIT, however this only happends for deoptimization and all registers are
saved to stack in advance.
3. Candidates for register allocator
1) opcode candidates
Function zend_jit_opline_supports_reg() determines the candidate opcodes
which can use CPU registers.
2) register candidates
Registers in set "ZEND_REGSET_FP + ZEND_REGSET_GP - ZEND_REGSET_FIXED -
ZEND_REGSET_PRESERVED" are available for register allocator.
Note that registers from ZEND_REGSET_FIXED are reserved for special
purpose, such as the stack pointer, and they are excluded from register
allocation process.
Note that registers from ZEND_REGSET_PRESERVED are callee-saved based on
the ABI and it's safe to not use them either.
4. Temporary registers
Temporary registers are needed by some opcodes to save intermediate
computation results.
1) Functions zend_jit_get_def_scratch_regset() and
zend_jit_get_scratch_regset() return which registers might be clobbered
by some opcodes. Hence register allocator would spill these scratch
registers if necessary when encountering these opcodes.
2) Macro ZEND_REGSET_LOW_PRIORITY denotes a set of registers which would
be allocated with low priority, and these registers can be used as
temporary usage to avoid conflicts to its best.
5. Compared to the x86 implementation, in JIT/arm64
1) Called-saved FP registers are included into ZEND_REGSET_PRESERVED for
AArch64.
2) We follow the logic of function zend_jit_opline_supports_reg().
3) We reserve 4 GPRs and 2 FPRs out from register allocator and use them
as temporary registers in particular. Note that these 6 registers are
included in set ZEND_REGSET_FIXED.
Since they are reserved, may-clobbered registers can be removed for most
opcodes except for function calls. Besides, low-priority registers are
defined as empty since all candidate registers are of the same priority.
See the updates in function zend_jit_get_scratch_regset() and macro
ZEND_REGSET_LOW_PRIORITY.
6. Why we reserve registers for temporary usage?
1) Addressing mode in AArch64 needs more temporary registers.
The addressing mode is different from x86 and tempory registers might be
*always* needed for most opcodes. For instance, an immediate must be
first moved into one register before storing into memory in AArch64,
whereas in x86 this immediate can be stored directly.
2) There are more registers in AArch64.
Compared to the solution in JIT/x86(that is, temporary registers are
reserved on demand, i.e. different registers for different opcodes under
different conditions), our solution seems a coarse-granularity and
brute-force solution, and the execution performance might be downgraded
to some extent since the number of candidate registers used for
allocation becomes less.
We suppose the performance loss might be acceptable since there are more
registers in AArch64.
3) Based on my understanding, scratch registers defined in x86 are
excluded from candidates for register allocator with *low possibility*,
and it can still allocate these registers. Special handling should be
conducted, such as checking 'reg != ZREG_R0'.
Hence, as we see it, it's simpler to reserve some temporary registers
exclusively. See the updates in function zend_jit_math_long_long() for
instance. TMP1 can be used directly without checking.
Co-Developed-by: Nick Gasson <Nick.Gasson@arm.com>
Literal compaction was incorrectly assuming that literals with
the same base literal and the same number of related literals
would be equal. Maybe that was the case historically, but at
least it isn't true in PHP 8, where FETCH_CONSTANT and INIT_METHOD
have distinct literals at the second position.
Fix this by making the cache key a concatenation of all literals,
rather than just the base literal. We still distinguish the number
of related literals based on a bias added to the string hash.
Make sure that the previous opline is part of the same block,
otherwise it may be non-dominating.
The test case does not fail on PHP-7.4, but I think the general
problem can appear on 7.4 as well, so I'm applying the patch to
that branch.
As zend_update_parent_ce() only runs later, the parent reference
may still point to the original class entry rather than the
persisted one. Memory held by the original class entry may have
already been deallocated. Avoid use-after-free by explicitly
fetching the persisted parent CE.
* PHP-8.0:
Fix DCE of FREE of COALESCE
Note that this is a non-trivial merge, as opt/coalesce.phpt
regresses with this change. I'll have to see if it can be
improved.
When encountering the following SSA graph:
BB1:
#2.T1 [string] = COALESCE #1.CV0($str) [null, string] BB2
BB2:
#5.T1 [string] = QM_ASSIGN string("")
BB3:
#7.X1 [string] = Phi(#2.X1 [string], #5.X1 [string])
FREE #7.T1 [string]
We would currently determine that #7, #5 are dead, and eliminate
the FREE and QM_ASSIGN. However, we cannot eliminate #2, as
COALESCE is also responsible for control flow.
Fix this my marking all non-CV phis as live to start with. This
can be relaxed to check the kind of the source instruction, but
I couldn't immediately come up with a case where it would be
useful.
Some user opcode handler actually gets called when JIT is used, so do not disable JIT even if these user opcode handlers were registered, just ignore them.
ZEND_EXIT can help Swoole to detect exit/die operation and exit from the current coroutine and record exit_status correctly.
ZEND_*_SILENCE can help Swoole to switch EG(error_reporting) to keep the right behavior when using the error control operator.
So we ignore ZEND_EXIT, ZEND_BEGIN_SILENCE, ZEND_END_SILENCE in JIT check here.
Closes GH-6640.
This handles the degenerate case where SCCP replaced the value in
the RETURN opcode with a constant, but the VERIFY_RETURN is still
there. We can still apply the same optimization, just don't need
to adjust the use list in this case.
The result is still sub-optimal in that a dead QM_ASSIGN is left
behind.
Even if we don't know the exact method being called, include it
in the call graph with the is_prototype flag. In particular, we
can still make use of return types from prototype methods, as
PHP 8 makes LSP violations a hard error.
Most other places are adjusted to skip calls with !is_prototype.
Maybe some of them would be fine, but ignoring them is conservative.
Even if an explicit return type is given, we might still infer
a more narrow one based on return statements. We shouldn't
pessimize this just because a type has been declared.
Dmitry Stogov