* include/lightning.h, include/lightning/jit_private.h,
lib/jit_names.c, lib/lightning.c: Add initial support
for the new jit_va_start, jit_va_arg, jit_va_arg_d, and
jit_va_end interfaces. The jit_va_start call is supposed
to return a va_list compatible pointer, but not yet
decided if it will be "declared" stdarg compatible,
as for now only x86 support has been added (and should
be compatible), but issues may arise on other backends.
* check/lightning.c: Add wrappers to call the new jit_va_*
interfaces.
* lib/jit_x86-cpu.c, lib/jit_x86.c: Implement the new
jit_va_* for x86.
* lib/jit_x86-sz.c: Add fields, but not yet fully updated,
as this is an intermediate commit.
* lib/jit_aarch64-sz.c, lib/jit_aarch64.c,
lib/jit_alpha-sz.c, lib/jit_alpha.c,
lib/jit_arm-sz.c, lib/jit_arm.c,
lib/jit_hppa-sz.c, lib/jit_hppa.c,
lib/jit_ia64-sz.c, lib/jit_ia64.c,
lib/jit_mips-sz.c, lib/jit_mips.c,
lib/jit_ppc-sz.c, lib/jit_ppc.c,
lib/jit_s390-sz.c, lib/jit_s390.c,
lib/jit_sparc-sz.c, lib/jit_sparc.c: Prepare for the
new jit_va_* interfaces. Not yet implemented, and will
cause an assertion if used.
* check/va_list.tst: Simple early test case, that works
on x86_64, x32, ix86, cygwin, and cygwin64.
* include/lightning.h, include/lightning/jit_private.h,
lib/jit_aarch64-cpu.c, lib/jit_aarch64.c,
lib/jit_alpha-cpu.c, lib/jit_alpha.c,
lib/jit_arm-cpu.c, lib/jit_arm.c,
lib/jit_hppa-cpu.c, lib/jit_hppa.c,
lib/jit_ia64-cpu.c, lib/jit_ia64.c,
lib/jit_mips-cpu.c, lib/jit_mips.c,
lib/jit_ppc-cpu.c, lib/jit_ppc.c,
lib/jit_s390-cpu.c, lib/jit_s390.c,
lib/jit_sparc-cpu.c, lib/jit_sparc.c,
lib/jit_x86-cpu.c, lib/jit_x86.c: Implement the new
jit_allocar(offs, size) interface, that receives
two integer registers arguments, allocates space
dynamically in the stack, returns the offset in
the first argument, and uses the second argument
for the size in bytes of the memory to be allocated.
* check/allocar.ok, check/allocar.tst: New files
implementing test cases for the new jit_allocar
interface.
* check/Makefile.am, check/lightning.c: Update for
the new test case and interface.
* doc/body.texi: Add documentation of the new
interface.
* include/lightning/jit_x86.h, lib/jit_x86-cpu.c,
lib/jit_x86-x87.c: No longer make st(7) available.
Need to keep one x87 slots empty to avoid exceptions.
This has the side effect of no longer needing the
hackish emms instruction before a function call.
* lib/lightning.c: Remove the jit_regno_patch bitfield
register fields before actual emit, as it is only really
used before emit, otherwise, on special conditions it
may consider live registers as dead during code emit.
* lib/jit_x86-cpu.c, lib/jit_x86-sse.c, lib/jit_x86-x87.c:
Correct encoding of ldxr* stxr* in the x32 abi. If the
displacement register is negative, it would generate
a 64 bit instruction with a 32 bit unsigned displacement.
* check/ranger.tst, check/ranger.ok: New files, implementing
a test case for negative loads and stores. This is range.tst
converted to use registers instead of immediate offsets.
check/Makefile.am: Update for the new test case.
* lib/jit_size.c: Preventively use at least 144 bytes
if JIT_INSTR_MAX is less than it. The logic is not
guaranteed to be 100% precise, it is mostly heuristics
to allocate a buffer with as close as possible size,
but a wrong value may cause code generation to write
past the end of the buffer.
* lib/lightning.c: Correct the reason the bug in
simplify_stxi was not triggered before, it was due to
incorrectly resetting the value->code field, what was
causing it to never properly optimize:
stxi Im0 Rb0 Rt0
ldxi Rt1 Rb1 Im1
when Rb0 == Rb1, Rt0 == Rt1 and Im0 == Im1
There was another possible issue, that has been also
addressed in this commit, that would be the case of
Rbn == Rtn, where no redundancy removal is possible.
* lib/lightning.c: Correct wrong check in simplify_stxi.
The test was incorrectly comparing the target register
and the displacement offset. This was a time bomb bug,
that would trigger in code like:
stxi Im0 Rb0 Rt0
stxi Im1 Rb1 Rt1
if Rb0 == Rb1 && Rt0 == Rt1 && Im0 == Rt1, that is,
the wrong check was Im0 == Rt1, instead of the supposed
Im0 == Imm1 (that was what the code mean't to do). It
was removing the second stxi assuming it was redundantly
generated; as that is not uncommon pattern on
translators generating jit.
* configure.ac, include/lightning/jit_private.h,
lib/jit_aarch64.c, lib/jit_alpha.c, lib/jit_arm.c,
lib/jit_disasm.c, lib/jit_hppa.c, lib/jit_ia64.c,
lib/jit_mips.c, lib/jit_ppc.c, lib/jit_print.c,
lib/jit_s390.c, lib/jit_sparc.c, lib/jit_x86.c: Add a new
--enable-devel-disassembler option, that should be used
during development, or lightning debug. This option
intermixes previous jit_print and jit_disassemble
output, making it easier to visualize what lightning
call was used, and what code was generated.
* include/lightning.h, lib/jit_aarch64.c,
lib/jit_alpha.c, lib/jit_arm-vfp.c, lib/jit_arm.c,
lib/jit_hppa.c, lib/jit_ia64.c, lib/jit_mips.c,
lib/jit_ppc.c, lib/jit_s390.c, lib/jit_sparc.c,
lib/jit_x86.c: Add the new jit_arg_register_p predicate.
The predicate is expected to be used to know if an
argument is in a register, what would need special
handling if code that can overwrite non callee save
registers is executed.
* check/carg.c: New test case to check consistency and
expected usage of jit_arg_register_p.
* check/Makefile.am: Update for new test case.
* include/lightning/jit_aarch64.h,
include/lightning/jit_alpha.h,
include/lightning/jit_arm.h,
include/lightning/jit_hppa.h,
include/lightning/jit_mips.h,
include/lightning/jit_ppc.h,
include/lightning/jit_s390.h,
include/lightning/jit_sparc.h,
include/lightning/jit_x86.h,
lib/jit_aarch64.c, lib/jit_alpha.c,
lib/jit_arm.c, lib/jit_hppa.c,
lib/jit_ia64.c, lib/jit_mips.c,
lib/jit_ppc.c, lib/jit_s390.c,
lib/jit_sparc.c, lib/jit_x86.c: Remove jit_arg_reg_p and
jit_arg_f_reg_p from a public header, and define it only
on port specific files where an integer offset is used
to qualify an argument identifier. Exported code expects
an opaque pointer (but of jit_node_t* type) to "qualify"
an argument identifier.
This patch, and the code review/simplification done during
it also corrected some bugs:
o Inconsistent jit_arg_d value of double argument after 3
integer arguments in arm for jit_functions; tested, C
functions were being properly called.
o Inconsistent use of getarg_{f,d} and putarg*_{f,d} on
s390 (32-bit) that happened to not have a proper test
case, as it would only happen for jit functions, and
tested, called C functions had proper arguments.
o Corrected a "last minute" correction that did not go
to the committed version, and would not compile on hppa,
due to bad _jit_putargi_d prototype definition.
* include/lightning.h, lib/jit_aarch64.c,
lib/jit_alpha.c, lib/jit_arm.c, lib/jit_hppa.c,
lib/jit_ia64.c, lib/jit_mips.c, lib/jit_ppc.c,
lib/jit_s390.c, lib/jit_sparc.c, lib/jit_x86.c:
Implement jit_putarg*. It works as a mix of jit_getarg*
and jit_pusharg*, in the way that the first argument is
a register or immediate, and the second is a pointer
returned by jit_arg*. The use of the interface is to change
values of arguments to the current jit function.
* check/put.ok, check/put.tst: New test cases exercising
the new jit_putarg* interface.
* check/Makefile.am, check/lightning.c: Update for the
new test case and interface.
* include/lightning.h, include/lightning/jit_private.h,
include/lightning/jit_s390x.h, lib/jit_disasm.c,
lib/jit_s390x-cpu.c, lib/jit_s390x-fpu.c, lib/jit_s390x-sz.c,
lib/jit_s390x.c, lib/jit_size.c, lib/lightning.c:
Add support for generating jit for s390 32 bit. This change
also removed %f15 from the list of temporaries fpr registers;
it was not being used, but if were, it would corrupt the
stack frame because the spill address would overwrite grp
offsets.
* include/lightning.h: Split jit_htonr in the new 3 interfaces
jit_htonr_us, jit_htonr_ui and jit_htonr_ul, the later only
available on 64 bit. The plain/untyped jit_htonr macro call
maps to the wordsize one.
* lib/jit_aarch64-cpu.c, lib/jit_aarch64-sz.c, lib/jit_aarch64.c,
lib/jit_alpha-cpu.c, lib/jit_alpha-sz.c, lib/jit_alpha.c,
lib/jit_arm-cpu.c, lib/jit_arm-sz.c, lib/jit_arm.c,
lib/jit_hppa-cpu.c, lib/jit_hppa-sz.c, lib/jit_hppa.c,
lib/jit_ia64-cpu.c, lib/jit_ia64-sz.c, lib/jit_ia64.c,
lib/jit_mips-cpu.c, lib/jit_mips-sz.c, lib/jit_mips.c,
lib/jit_ppc-cpu.c, lib/jit_ppc-sz.c, lib/jit_ppc.c,
lib/jit_s390x-cpu.c, lib/jit_s390x-sz.c, lib/jit_s390x.c,
lib/jit_sparc-cpu.c, lib/jit_sparc-sz.c, lib/jit_sparc.c,
lib/jit_x86-cpu.c, lib/jit_x86-sz.c, lib/jit_x86.c:
Update backends for the new jit_htonr*.
* check/lightning.c, lib/jit_names.c, lib/lightning.c:
Update for the new jit_htonr* interfaces.
* check/Makefile.am: Update for new test cases.
* check/hton.ok, check/hton.tst: New test cases.
* include/lightning/jit_private.h, include/lightning/jit_x86.h,
lib/jit_disasm.c, lib/jit_x86-cpu.c, lib/jit_x86-sse.c,
lib/jit_x86-sz.c, lib/jit_x86-x87.c, lib/jit_x86.c,
size.c: Implement support for the x32 abi. Built and
tested on Gentoo default/linux/amd64/13.0/x32 profile.
* lib/jit_arm.c: Call __clear_cache for every page.
This should only be required for older boards or
toolchain setup, but has been reported to be required
for lightning at some point.
* lib/jit_disasm.c: Change thumb or arm disassemble based on
jit code before disassembly.
* lib/jit_arm-cpu.c: Correct reversed arguments to LDRD and
STRD instructions, and correct checking for support of those.
* lib/jit_arm-swf.c: Correct wrong use of LDRD and STRD and
only use those if the register is even.
* check/check.arm.swf.sh, check/check.arm4.swf.sh: New files
to test LDRD and STRD, as well as the alternate code path
when those are not available, in the .arm4. test case.
* check/Makefile.am: Update for the new test cases.
* include/lightning/jit_private.h, lib/jit_aarch64.c,
lib/jit_alpha.c, lib/jit_arm.c, lib/jit_hppa.c,
lib/jit_ia64.c, lib/jit_mips.c, lib/jit_ppc.c,
lib/jit_s390x.c, lib/jit_sparc.c, lib/jit_x86.c:
Implement a private jit_flush call, that flushes
the cache, if applicable, aligning down to the
previous and up to the next page boundary.
Make jit_memcpy, jit_memmove, jit_data take const pointers to
allow jit_note to be used with a const string (e.g. a string
literal, __FILE__ or __func__). This is needed for GNU Smalltalk
to silence compiler warnings.
Sadly "const jit_pointer_t" is not the same as "typedef const void *"
so I introduced a new typedef for a const jit pointer. The other
option would be to replace jit_pointer_t with void*.
* lib/jit_aarch64.c, lib/jit_alpha.c, lib/jit_arm.c,
lib/jit_hppa.c, lib/jit_ia64.c, lib/jit_mips.c, lib/jit_ppc.c,
lib/jit_s390x.c, lib/jit_sparc.c, lib/jit_x86.c: Add an
assertion to all code generation "drivers" to ensure
_jitc->regarg is empty or in an expected state, after
translation of a lightning instruction to native code.
This change was a brute force test to find out other cases
of a temporary not being release (like was happening with
_bmsi and _bmci on x86), but no other case was found,
after running make check, with assertions enabled, on all
backends.
* lib/jit_ia64.c, lib/jit_ppc.c: Correct handling of function
descriptor when first prolog is a jit_tramp prolog. The
test case was using the same jit_context_t, so was not
triggering this condition.
* lib/jit_ppc-cpu.c: Properly handle jump displacements that
do not fit on 24 powerpc. This required changing from previous
"mtlr reg, blr" to "mtctr reg, bctr" to properly handle
the logic to "hide" function descriptors, but that would
also be required as the proper jit_jmpr when/if implementing
optimizations to leaf functions (was working with blr because
it is saved/reloaded in prolog/epilog).
* include/lightning.h, lib/lightning.c: Add three predicates
to query information about labels. jit_forward_p(label)
will return non zero if the label is "forward", that is
need a call to jit_link(label), jit_indirect_p(label)
that returns non zero if the label was created with the
jit_indirect() call, and jit_target_p(label) that will
return non zero if there is at least one jump patched
to land at that label.
* check/range.ok, check/range.tst: New test case designed
to catch incorrect code generation, usually due to incorrect
test of immediate size. The test checks a large amount of
encodings in "power of two" boundaries. This test exorcises
a significant amount of code paths that was previously not
tested.
* check/Makefile.am: Add range test to make check target.
* lib/jit_aarch64-cpu.c: Correct wrong address calculation
for stxi_c, stxi_s, stxi_i and stxi_l when the offset is
too large.
* lib/jit_mips-fpu.c: Correct wrong size test to check if
an immediate can be encoded in a float or double store.
* lib/jit_s390x-cpu.c: Correct inverted encoding to stxi_s
when the offset cannot be encoded, and fallbacks to an
alternate encoding in 2 instructions.
* check/alu_rsb.ok, check/alu_rsb.tst: New files implementing
tests for jit_rsb*.
* check/Makefile.am, check/lightning.c, include/lightning.h,
lib/jit_aarch64-cpu.c, lib/jit_aarch64-fpu.c, lib/jit_aarch64-sz.c,
lib/jit_aarch64.c, lib/jit_alpha-cpu.c, lib/jit_alpha-fpu.c,
lib/jit_alpha-sz.c, lib/jit_alpha.c, lib/jit_arm-cpu.c,
lib/jit_arm-swf.c, lib/jit_arm-sz.c, lib/jit_arm-vfp.c,
lib/jit_arm.c, lib/jit_hppa-cpu.c, lib/jit_hppa-fpu.c,
lib/jit_hppa-sz.c, lib/jit_hppa.c, lib/jit_ia64-cpu.c,
lib/jit_ia64-fpu.c, lib/jit_ia64-sz.c, lib/jit_ia64.c,
lib/jit_mips-cpu.c, lib/jit_mips-fpu.c, lib/jit_mips-sz.c,
lib/jit_mips.c, lib/jit_names.c, lib/jit_ppc-cpu.c,
lib/jit_ppc-fpu.c, lib/jit_ppc-sz.c, lib/jit_ppc.c,
lib/jit_s390x-cpu.c, lib/jit_s390x-fpu.c, lib/jit_s390x-sz.c,
lib/jit_s390x.c, lib/jit_sparc-cpu.c, lib/jit_sparc-fpu.c,
lib/jit_sparc-sz.c, lib/jit_sparc.c, lib/jit_x86-cpu.c,
lib/jit_x86-sse.c, lib/jit_x86-sz.c, lib/jit_x86-x87.c,
lib/jit_x86.c, lib/lightning.c: Implement jit_rsb*. This
was a missing lightning 1.x interface, that on most
backends is synthesized, but on a few backends (hppa and ia64),
it can generate better code as on those there is, or the
only instruction with an immediate is in "rsb" format
(left operand).
* lib/jit_names.c: New file with single definition of string
representation of lightning IR codes.
* size.c: Modified to append the code name in a C comment
after the maximum instruction size.
* lib/jit_print.c: Minor change to not duplicate jit_names.c
contents.
* lib/jit_aarch64-sz.c, lib/jit_alpha-sz.c, lib/jit_arm-sz.c,
lib/jit_hppa-sz.c, lib/jit_ia64-sz.c, lib/jit_mips-sz.c,
lib/jit_ppc-sz.c, lib/jit_s390x-sz.c, lib/jit_sparc-sz.c,
lib/jit_x86-sz.c: Rewritten to add string representation of
IR codes in a C comment.
* lib/jit_aarch64-cpu.c, lib/jit_alpha-cpu.c, lib/jit_arm-cpu.c,
lib/jit_hppa-cpu.c, lib/jit_mips-cpu.c, lib/jit_ppc-cpu.c,
lib/jit_sparc-cpu.c: Implement or correct the internal
nop(count) call that receives an argument that tells the
modulo bytes to align the code for the next instruction.
* include/lightning.h, lib/lightning.c, lib/jit_aarch64.c,
lib/jit_alpha.c, lib/jit_arm.c, lib/jit_hppa.c, lib/jit_ia64.c,
lib/jit_mips.c, lib/jit_ppc.c, lib/jit_s390x.c, lib/jit_sparc.c,
lib/jit_x86.c: Implement the new jit_align() call that receive
an argument, that tells the modulo, in bytes, to align the
next instruction. In most backends the only value that makes
a difference is a value that matches sizeof(void*), as all
other values usually are already automatically aligned in
labels, but not guaranteed to be aligned at word size bytes.
* check/align.ok, check/align.tst: New files, implementing
a simple test for the new jit_align() interface.
* check/Makefile.am, check/lightning.c, lib/jit_aarch64-sz.c,
lib/jit_alpha-sz.c, lib/jit_arm-sz.c, lib/jit_hppa-sz.c,
lib/jit_ia64-sz.c, lib/jit_mips-sz.c, lib/jit_ppc-sz.c,
lib/jit_print.c, lib/jit_s390x-sz.c, lib/jit_sparc-sz.c,
lib/jit_x86-sz.c: Update for the new jit_code_align code and
the jit_align() interface.
* include/lightning.h, include/lightning/jit_private.h,
lib/jit_aarch64-cpu.c, lib/jit_alpha-cpu.c, lib/jit_arm-cpu.c,
lib/jit_hppa-cpu.c, lib/jit_ia64-cpu.c, lib/jit_mips-cpu.c,
lib/jit_ppc-cpu.c, lib/jit_s390x-cpu.c, lib/jit_sparc-cpu.c,
lib/jit_x86-cpu.c, lib/lightning.c: Implement the new
jit_frame and jit_tramp interfaces, that allow writing
trampoline like calls, where a single dispatcher jit buffer
is written, and later other jit buffers are created, with
the same stack frame layout as the dispatcher. This is the
logic that GNU Smalltalk used in lightning 1.x, and is required
to make a sane port for lighting 2.x.
* jit_ia64-cpu.c: Implement support for jit_frame and jit_tramp,
and also correct wrong encoding for B4 instructions, that
implement jmpr, as well as correct reverse logic in _jmpr,
that was moving the branch register to the jump register,
and not vice-versa.
Also, if a stack frame is to be assumed, always assume it may
call a function with up to 8 arguments, regardless of the
hint frame argument.
* lib/jit_arm.c: Add a new must_align_p() interface to ensure
function prologs are always aligned. This condition was
previously always true, somewhat by accident, but with
jit_tramp it is not guaranteed.
* jit_ia64-cpu.c: lib/jit_ppc.c: Add minor special handling
required to implement jit_tramp, where a function descriptor
should not be added before a prolog, as jit_tramp means omit
prolog.
* check/lightning.c: Update test driver for the new interfaces.
* check/Makefile.am, check/tramp.tst, check/tramp.ok: Add
a simple test and example of the jit_frame and jit_tramp
usage implementing a simple Fibonacci function using a
simulation of an interpreter stack and how it would handle
state in language specific variables.
* doc/body.texi: Add documentation for jit_frame and
jit_tramp.
* lib/jit_aarch64.c, lib/jit_alpha.c, lib/jit_arm.c,
lib/jit_hppa.c, lib/jit_ia64.c, lib/jit_mips.c,
lib/jit_ppc.c, lib/jit_s390x.c, lib/jit_sparc.c,
lib/jit_x86.c, lib/lightning.c: Allow jit_jmpi on a
target that is not a node. This may lead to hard to
debug code generation, but is a required feature for
certain generators, like the ones that used lightning
1.2x. Note that previously, but not really well
documented, it was instructed to use:
jit_movi(rn, addr); jit_jmpr(rn);
but now, plain:
jit_patch_abs(jit_jmpi(), addr);
should also work.
* lib/jit_x86-sz.c: Generate information about instruction
lengths for more precise calculation of buffer size on
Windows x64. This change is specially important because
the maximum instruction length is larger than other
systems, what could cause an out of bounds write on
special conditions without this update.
