Although the original motivation given for the workaround[1] is correct,
nlr.o and nlrthumb.o are linked with a small enough distance that the
problem does not occur, and the workaround isn't necessary. The distance
between the b instruction and its target (nlr_push_tail) is just 64
bytes[2], well within the ±2046 byte range addressable by an
unconditional branch instruction in Thumb mode.
The workaround induces a relocation in the text section (textrel), which
isn't supported everywhere, notably not on musl-libc[3], where it causes
a crash on start-up. With the workaround removed, micropython works on an
ARMv5T Linux system built with musl-libc.
This commit changes nlrthumb.c to use a direct jump by default, but
leaves the long jump workaround as an option for those cases where it's
actually needed.
[1]: commit dd376a239d
Author: Damien George <damien.p.george@gmail.com>
Date: Fri Sep 1 15:25:29 2017 +1000
py/nlrthumb: Get working again on standard Thumb arch (ie not Thumb2).
"b" on Thumb might not be long enough for the jump to nlr_push_tail so
it must be done indirectly.
[2]: Excerpt from objdump -d micropython:
000095c4 <nlr_push_tail>:
95c4: b510 push {r4, lr}
95c6: 0004 movs r4, r0
95c8: f02d fd42 bl 37050 <mp_thread_get_state>
95cc: 6943 ldr r3, [r0, #20]
95ce: 6023 str r3, [r4, #0]
95d0: 6144 str r4, [r0, #20]
95d2: 2000 movs r0, #0
95d4: bd10 pop {r4, pc}
000095d6 <nlr_pop>:
95d6: b510 push {r4, lr}
95d8: f02d fd3a bl 37050 <mp_thread_get_state>
95dc: 6943 ldr r3, [r0, #20]
95de: 681b ldr r3, [r3, #0]
95e0: 6143 str r3, [r0, #20]
95e2: bd10 pop {r4, pc}
000095e4 <nlr_push>:
95e4: 60c4 str r4, [r0, #12]
95e6: 6105 str r5, [r0, #16]
95e8: 6146 str r6, [r0, #20]
95ea: 6187 str r7, [r0, #24]
95ec: 4641 mov r1, r8
95ee: 61c1 str r1, [r0, #28]
95f0: 4649 mov r1, r9
95f2: 6201 str r1, [r0, #32]
95f4: 4651 mov r1, sl
95f6: 6241 str r1, [r0, #36] @ 0x24
95f8: 4659 mov r1, fp
95fa: 6281 str r1, [r0, #40] @ 0x28
95fc: 4669 mov r1, sp
95fe: 62c1 str r1, [r0, #44] @ 0x2c
9600: 4671 mov r1, lr
9602: 6081 str r1, [r0, #8]
9604: e7de b.n 95c4 <nlr_push_tail>
[3]: https://www.openwall.com/lists/musl/2020/09/25/4
Signed-off-by: J. Neuschäfer <j.ne@posteo.net>
Newer versions of gcc (14 and up) have more sophisticated dead-code
detection, and the asm clobbers list needs to contain "memory" to inform
the compiler that the asm code actually does something.
Tested that adding this "memory" line does not change the generated code on
ARM Thumb2, x86-64 and Xtensa targets (using gcc 13.2).
Fixes issue #14115.
Signed-off-by: Damien George <damien@micropython.org>
This fixes compiling for older architectures (e.g. armv5tej).
According to [1], the limit of R0-R7 for the STR and LDR instructions is
tied to the Thumb instruction set and not any specific processor
architectures.
[1]: http://www.keil.com/support/man/docs/armasm/armasm_dom1361289906890.htm
These s16-s21 registers are used by gcc so need to be saved. Future
versions of gcc (beyond v9.1.0), or other compilers, may eventually need
additional registers saved/restored.
See issue #4844.
By adding __builtin_unreachable() at the end of nlr_push, we're
essentially telling the compiler that this function will never return.
When GCC LTO is in use, this means that any time nlr_push() is called
(which is often), the compiler thinks this function will never return
and thus eliminates all code following the call.
Note: I've added a 'return 0' for older GCC versions like 4.6 which
complain about not returning anything (which doesn't make sense in a
naked function). Newer GCC versions (tested 4.8, 5.4 and some others)
don't complain about this.
Each NLR implementation (Thumb, x86, x64, xtensa, setjmp) duplicates a lot
of the NLR code, specifically that dealing with pushing and popping the NLR
pointer to maintain the linked-list of NLR buffers. This patch factors all
of that code out of the specific implementations into generic functions in
nlr.c, along with a helper macro in nlr.h. This eliminates duplicated
code.
If MICROPY_NLR_SETJMP is not enabled and the machine is auto-detected then
nlr.h now defines some convenience macros for the individual NLR
implementations to use (eg MICROPY_NLR_THUMB). This keeps nlr.h and the
implementation in sync, and also makes the nlr_buf_t struct easier to read.
A function with a naked attribute must only contain basic inline asm
statements and no C code.
For nlr_push this means removing the "return 0" statement. But for some
gcc versions this induces a compiler warning so the __builtin_unreachable()
line needs to be added.
For nlr_jump, this function contains a combination of C code and inline asm
so cannot be naked.
This reverts commit 6a3a742a6c.
The above commit has number of faults starting from the motivation down
to the actual implementation.
1. Faulty implementation.
The original code contained functions like:
NORETURN void nlr_jump(void *val) {
nlr_buf_t **top_ptr = &MP_STATE_THREAD(nlr_top);
nlr_buf_t *top = *top_ptr;
...
__asm volatile (
"mov %0, %%edx \n" // %edx points to nlr_buf
"mov 28(%%edx), %%esi \n" // load saved %esi
"mov 24(%%edx), %%edi \n" // load saved %edi
"mov 20(%%edx), %%ebx \n" // load saved %ebx
"mov 16(%%edx), %%esp \n" // load saved %esp
"mov 12(%%edx), %%ebp \n" // load saved %ebp
"mov 8(%%edx), %%eax \n" // load saved %eip
"mov %%eax, (%%esp) \n" // store saved %eip to stack
"xor %%eax, %%eax \n" // clear return register
"inc %%al \n" // increase to make 1, non-local return
"ret \n" // return
: // output operands
: "r"(top) // input operands
: // clobbered registers
);
}
Which clearly stated that C-level variable should be a parameter of the
assembly, whcih then moved it into correct register.
Whereas now it's:
NORETURN void nlr_jump_tail(nlr_buf_t *top) {
(void)top;
__asm volatile (
"mov 28(%edx), %esi \n" // load saved %esi
"mov 24(%edx), %edi \n" // load saved %edi
"mov 20(%edx), %ebx \n" // load saved %ebx
"mov 16(%edx), %esp \n" // load saved %esp
"mov 12(%edx), %ebp \n" // load saved %ebp
"mov 8(%edx), %eax \n" // load saved %eip
"mov %eax, (%esp) \n" // store saved %eip to stack
"xor %eax, %eax \n" // clear return register
"inc %al \n" // increase to make 1, non-local return
"ret \n" // return
);
for (;;); // needed to silence compiler warning
}
Which just tries to perform operations on a completely random register (edx
in this case). The outcome is the expected: saving the pure random luck of
the compiler putting the right value in the random register above, there's
a crash.
2. Non-critical assessment.
The original commit message says "There is a small overhead introduced
(typically 1 machine instruction)". That machine instruction is a call
if a compiler doesn't perform tail optimization (happens regularly), and
it's 1 instruction only with the broken code shown above, fixing it
requires adding more. With inefficiencies already presented in the NLR
code, the overhead becomes "considerable" (several times more than 1%),
not "small".
The commit message also says "This eliminates duplicated code.". An
obvious way to eliminate duplication would be to factor out common code
to macros, not introduce overhead and breakage like above.
3. Faulty motivation.
All this started with a report of warnings/errors happening for a niche
compiler. It could have been solved in one the direct ways: a) fixing it
just for affected compiler(s); b) rewriting it in proper assembly (like
it was before BTW); c) by not doing anything at all, MICROPY_NLR_SETJMP
exists exactly to address minor-impact cases like thar (where a) or b) are
not applicable). Instead, a backwards "solution" was put forward, leading
to all the issues above.
The best action thus appears to be revert and rework, not trying to work
around what went haywire in the first place.
Each NLR implementation (Thumb, x86, x64, xtensa, setjmp) duplicates a lot
of the NLR code, specifically that dealing with pushing and popping the NLR
pointer to maintain the linked-list of NLR buffers. This patch factors all
of that code out of the specific implementations into generic functions in
nlr.c. This eliminates duplicated code.
The factoring also allows to make the machine-specific NLR code pure
assembler code, thus allowing nlrthumb.c to use naked function attributes
in the correct way (naked functions can only have basic inline assembler
code in them).
There is a small overhead introduced (typically 1 machine instruction)
because now the generic nlr_jump() must call nlr_jump_tail() rather than
them being one combined function.
This patch introduces the MICROPY_ENABLE_PYSTACK option (disabled by
default) which enables a "Python stack" that allows to allocate and free
memory in a scoped, or Last-In-First-Out (LIFO) way, similar to alloca().
A new memory allocation API is introduced along with this Py-stack. It
includes both "local" and "nonlocal" LIFO allocation. Local allocation is
intended to be equivalent to using alloca(), whereby the same function must
free the memory. Nonlocal allocation is where another function may free
the memory, so long as it's still LIFO.
Follow-up patches will convert all uses of alloca() and VLA to the new
scoped allocation API. The old behaviour (using alloca()) will still be
available, but when MICROPY_ENABLE_PYSTACK is enabled then alloca() is no
longer required or used.
The benefits of enabling this option are (or will be once subsequent
patches are made to convert alloca()/VLA):
- Toolchains without alloca() can use this feature to obtain correct and
efficient scoped memory allocation (compared to using the heap instead
of alloca(), which is slower).
- Even if alloca() is available, enabling the Py-stack gives slightly more
efficient use of stack space when calling nested Python functions, due to
the way that compilers implement alloca().
- Enabling the Py-stack with the stackless mode allows for even more
efficient stack usage, as well as retaining high performance (because the
heap is no longer used to build and destroy stackless code states).
- With Py-stack and stackless enabled, Python-calling-Python is no longer
recursive in the C mp_execute_bytecode function.
The micropython.pystack_use() function is included to measure usage of the
Python stack.
Header files that are considered internal to the py core and should not
normally be included directly are:
py/nlr.h - internal nlr configuration and declarations
py/bc0.h - contains bytecode macro definitions
py/runtime0.h - contains basic runtime enums
Instead, the top-level header files to include are one of:
py/obj.h - includes runtime0.h and defines everything to use the
mp_obj_t type
py/runtime.h - includes mpstate.h and hence nlr.h, obj.h, runtime0.h,
and defines everything to use the general runtime support functions
Additional, specific headers (eg py/objlist.h) can be included if needed.
Now only the bits that really need to be written in assembler are written
in it, otherwise C is used. This means that the assembler code no longer
needs to know about the global state structure which makes it much easier
to maintain.
J. Neuschäfer