We are announcing the first stable release of mruby 3.1 series - mruby 3.1.0.
Describes the new features and changes in mruby 3.1.
The main changes in mruby 3.1 are also described in doc/mruby3.1.md.
CRuby3.0 compatible keyword arguments are introduced. Keyword arguments are basically separated from ordinal arguments.
R-assignment by single-line pattern matching Ruby:Feature#15921Some configuration macros are available:
MRB_WORDBOX_NO_FLOAT_TRUNCATE: by default, float values are packed in the word if possible, but define this macro to allocate float values in the heap.MRB_USE_RO_DATA_P_ETEXT: define this macro if _etext is available on your platform.MRB_NO_DEFAULT_RO_DATA_P: define this macro to avoid using predefined mrb_ro_data_p() functionWe have added several new build configurations in the build_config directory.
cross-mingw-winetest.rbcross-mingw.rbnintendo_switch.rbserenity.rbminimal: minimal configurationhost-f32: compiles with mrb_float as 32 bit floathost-nofloat: compiles with no float configurationandroid_arm64_v8a.rb: renamed from android_arm64-v8a.rbArray#productArray#repeated_combinationArray#repeated_permutationKernel#__ENCODING__Random.bytesRandom#bytesString#centermrbgems/mruby-pack now supports M directive (Q encoding)mrbgems/mruby-pack now supports X directive (back-up by bytes)mrbgems/mruby-pack now supports @ directive (absolute position)mrbgems/mruby-pack now supports w directive (BER compression)mruby-config now supports --cc and --ld options.OP_ prefix from mruby -v code dump output.OP_EXT{1,2,3} by mrbc with --no-ext-ops option.c to mrb_get_args() for receive Class/Module.Kernel#printf (mruby-sprintf) Format specifiers %a and %A are removed.Kernel#puts (mruby-print) Now expand Array arguments.Due to improvements in the binary format, mruby binaries are no longer backward compatible.
To run the mruby binaries on mruby 3.1, recompile with the mruby 3.1 mrbc.
RITE_VM_VER to 0300 (means mruby 3.0 or after).RITE_BINARY_FORMAT_VER to 0300.mruby3.0 removed OP_EXT1, OP_EXT2, OP_EXT3 for operand extension. But the operand size limitations was too tight for real-world application.
mruby3.1 reintroduces those extension instructions.
mruby3.1 removed following instructions.
OP_LOADL16OP_LOADSYM16OP_STRING16OP_LAMBDA16OP_BLOCK16OP_METHOD16OP_EXEC16Those instructions are no longer needed by reintroduction of extension instructions.
OP_SENDVOP_SENDVBThose instructions for method calls with variable number of arguments are no longer needed. They are covered by OP_SEND instruction with n=15.
mruby3.1 introduces following new instructions.
OP_GETIDX: takes 1 operands R[a][a+1]OP_SETIDX: takes 1 operands R[a][a+1]=R[a+2]OP_SSEND: takes 3 operands a=self.b(c...); see OP_SENDOP_SSENDB: takes 3 operands a=self.b(c...){...}; see OP_SENDOP_SYMBOL: takes 2 operands R[a] = intern(Pool[b])OP_GETIDX and OP_SETIDXExecute obj[int] and obj[int] = value respectively, where obj is string|array|hash.
OP_SSEND and OP_SSENDBThey are similar to OP_SEND and OP_SENDB respectively. They initialize the R[a] by self first so that we can skip one OP_LOADSELF instruction for each call.
OP_SYMBOLExtracts the character string placed in the pool as a symbol.
OP_SEND and OP_SENDBMethod calling instructions are unified. Now OP_SEND and OP_SENDB (method call with a block) can support both splat arguments and keyword arguments as well.
The brief description of the instructions:
OP_SEND |
BBB | R[a] = R[a].call(Syms[b],R[a+1..n],R[a+n+1],R[a+n+2]..nk) c=n|nk<<4 |
OP_SENDB |
BBB | R[a] = R[a].call(Syms[b],R[a+1..n],R[a+n+1..nk],R[a+n+2..nk],&R[a+n+2*nk+2]) c=n|nk<<4 |
Operand C specifies the number of arguments. Lower 4 bits (n) represents the number of ordinal arguments, and higher 4 bits (nk) represents the number of keyword arguments.
When n == 15, the method takes arguments packed in an array. When nk == 15, the method takes keyword arguments are packed in a hash.
OP_ARYPUSHNow takes 2 operands and pushes multiple entries to an array.
MRB_WORD_BOXING now packs floating point numbers in the word, if the size of mrb_float is equal or smaller than the size of mrb_int by default.
If the size of mrb_float and mrb_int are same, the last 2 bits in the mrb_float are trimmed and used as flags. If you need full precision, you need to define MRB_WORDBOX_NO_FLOAT_TRUNCATE as described above.
Previous NaN boxing packs values in NaN representation, but pointer retrievals are far more frequent than floating point number references. So we add constant offset to NaN representation to clear higher bits of pointer representation. This representation is called “Favor Pointer” NaN Boxing.
Also, previous NaN boxing limit the size of mrb_int to 4 bytes (32 bits) to fit in NaN values. Now we allocate integer values in the heap, if the value does not fit in the 32 bit range, just like we did in Word Boxing.
The code generator was updated to reduce the number of instructions, e.g.
a = 2 * 5
will be interpreted as
a = 10
In addition, we have improved peephole optimizations, for example:
GETIV R4 :@foo
MOVE R1 R4
to
GETIV R1 :@foo
String#hash now use FNV1a algorithmFor better and faster hash values.
method_added hooks on method definitions; #2339time_zonename 26340a88Module.instance_eval bug #5528M packing bug bfe2bd49codegen_error() #5603mrb_ary_shift_m initialization bug 27d1e013super #5628Following CVEs do not cause problems in this release. They are fixed in the later release.
We have done 1,076 commits to 287 files with 17,188 additions and 15,549 deletions since mruby 3.0.0. For more detail of the updates, see Commit Log.
Thanks to all the contributors who have worked on bug fixes and improvements in the release of mruby 3.1.0.