also has a significant regression introduced in it. The default compiler
not only works, but also is noticably faster. Ant takes care of the
selection of the compiler automatically based on the JDK level, so the
following patch should make things better all around.
Patch from Igor:
The 2 attached patches allow to avoid wrapping of array indexes to Double object
when Interpreter knows that the index is an integer number. It speed up array
benchmark by 5-10%
array_access.diff adds to ScriptRuntime getStrIdElem and setStrIdElem to get/set
properties which known to be strings plus it modifies NativeArray to use these methods.
interpreter.diff contains the Interpreter modifications to call get/setElem for
integer or string properties when the property type is known for sure.
errors, etc.''):
We now ReportStatementTooLarge only if
- a jump offset overflows 32 bits, signed;
- there are 2**32 or more span dependencies in a script;
- a backpatch chain link is more than (2**30 - 1) bytecodes long;
- a source note's distance from the last note, or from script main entry
point, is > 0x7fffff bytes.
Narrative of the patch, by file:
- js.c
The js_SrcNoteName array of const char * is now a js_SrcNoteSpec array of
"specifiers", structs that include a const char *name member. Also, due to
span-dependent jumps at the ends of basic blocks where the decompiler knows
the basic block length, but not the jump format, we need an offset operand
for SRC_COND, SRC_IF_ELSE, and SRC_WHILE (to tell the distance from the
branch bytecode after the condition expression to the span-dependent jump).
- jsarena.[ch]
JS arenas are used mainly for last-in-first-out allocation with _en masse_
release to the malloc pool (or, optionally, to a private freelist). But
the code generator needs to allocate and grow (by doubling, to avoid O(n^2)
growth) allocations that hold bytecode, source notes, and span-dependency
records. This exception to LIFO allocation works by claiming an entire
arena from the pool and realloc'ing it, as soon as the allocation size
reaches the pool's default arena size. Call such an allocation a "large
single allocation".
This patch adds a new arena API, JS_ArenaFreeAllocation, which can be used
to free a large single allocation. If called with an allocation that's not
a large single allocation, it will nevertheless attempt to retract the arena
containing that allocation, if the allocation is last within its arena.
Thus JS_ArenaFreeAllocation adds a non-LIFO "free" special case to match the
non-LIFO "grow" special case already implemented under JS_ARENA_GROW for
large single allocations.
The code generator still benefits via this extension to arenas, over purely
manual malloc/realloc/free, by virtue of _en masse_ free (JS_ARENA_RELEASE
after code generation has completed, successfully or not).
To avoid searching for the previous arena, in order to update its next
member upon reallocation of the arena containing a large single allocation,
the oversized arena has a back-pointer to that next member stored (but not
as allocable space within the arena) in a (JSArena **) footer at its end.
- jscntxt.c
I've observed for many scripts that the bytes of source notes and bytecode
are of comparable lengths, but only now am I fixing the default arena size
for cx->notePool to match the size for cx->codePool (1024 instead of 256).
- jsemit.c
Span-dependent instructions in JS bytecode consist of the jump (JOF_JUMP)
and switch (JOF_LOOKUPSWITCH, JOF_TABLESWITCH) format opcodes, subdivided
into unconditional (gotos and gosubs), and conditional jumps or branches
(which pop a value, test it, and jump depending on its value). Most jumps
have just one immediate operand, a signed offset from the jump opcode's pc
to the target bytecode. The lookup and table switch opcodes may contain
many jump offsets.
This patch adds "X" counterparts to the opcodes/formats (X is suffixed, btw,
to prefer JSOP_ORX and thereby to avoid colliding on the JSOP_XOR name for
the extended form of the JSOP_OR branch opcode). The unextended or short
formats have 16-bit signed immediate offset operands, the extended or long
formats have 32-bit signed immediates. The span-dependency problem consists
of selecting as few long instructions as possible, or about as few -- since
jumps can span other jumps, extending one jump may cause another to need to
be extended.
Most JS scripts are short, so need no extended jumps. We optimize for this
case by generating short jumps until we know a long jump is needed. After
that point, we keep generating short jumps, but each jump's 16-bit immediate
offset operand is actually an unsigned index into cg->spanDeps, an array of
JSSpanDep structs. Each struct tells the top offset in the script of the
opcode, the "before" offset of the jump (which will be the same as top for
simplex jumps, but which will index further into the bytecode array for a
non-initial jump offset in a lookup or table switch), the after "offset"
adjusted during span-dependent instruction selection (initially the same
value as the "before" offset), and the jump target (more below).
Since we generate cg->spanDeps lazily, from within js_SetJumpOffset, we must
ensure that all bytecode generated so far can be inspected to discover where
the jump offset immediate operands lie within CG_CODE(cg). But the bonus is
that we generate span-dependency records sorted by their offsets, so we can
binary-search when trying to find a JSSpanDep for a given bytecode offset,
or the nearest JSSpanDep at or above a given pc.
To avoid limiting scripts to 64K jumps, if the cg->spanDeps index overflows
65534, we store SPANDEP_INDEX_HUGE in the jump's immediate operand. This
tells us that we need to binary-search for the cg->spanDeps entry by the
jump opcode's bytecode offset (sd->before).
Jump targets need to be maintained in a data structure that lets us look
up an already-known target by its address (jumps may have a common target),
and that also lets us update the addresses (script-relative, a.k.a. absolute
offsets) of targets that come after a jump target (for when a jump below
that target needs to be extended). We use an AVL tree, implemented using
recursion, but with some tricky optimizations to its height-balancing code
(see http://www.enteract.com/~bradapp/ftp/src/libs/C++/AvlTrees.html).
A final wrinkle: backpatch chains are linked by jump-to-jump offsets with
positive sign, even though they link "backward" (i.e., toward lower bytecode
address). We don't want to waste space and search time in the AVL tree for
such temporary backpatch deltas, so we use a single-bit wildcard scheme to
tag true JSJumpTarget pointers and encode untagged, signed (positive) deltas
in JSSpanDep.target pointers, depending on whether the JSSpanDep has a known
target, or is still awaiting backpatching.
Note that backpatch chains would present a problem for BuildSpanDepTable,
which inspects bytecode to build cg->spanDeps on demand, when the first
short jump offset overflows. To solve this temporary problem, we emit a
proxy bytecode (JSOP_BACKPATCH; JSOP_BACKPATCH_PUSH for jumps that push a
result on the interpreter's stack, namely JSOP_GOSUB; or JSOP_BACKPATCH_POP
for branch ops) whose nuses/ndefs counts help keep the stack balanced, but
whose opcode format distinguishes its backpatch delta immediate operand from
a normal jump offset.
The cg->spanDeps array and JSJumpTarget structs are allocated from the
cx->tempPool arena-pool. This created a LIFO vs. non-LIFO conflict: there
were two places under the TOK_SWITCH case in js_EmitTree that used tempPool
to allocate and release a chunk of memory, during whose lifetime JSSpanDep
and/or JSJumpTarget structs might also be allocated from tempPool -- the
ensuing release would prove disastrous. These bitmap and table temporaries
are now allocated from the malloc heap.
- jsinterp.c
Straightforward cloning and JUMP => JUMPX mutating of the jump and switch
format bytecode cases.
- jsobj.c
Silence warnings about %p used without (void *) casts.
- jsopcode.c
Massive and scary decompiler whackage to cope with extended jumps, using
source note offsets to help find jumps whose format (short or long) can't
be discovered from properties of prior instructions in the script.
One cute hack here: long || and && expressions are broken up to wrap before
the 80th column, with the operator at the end of each non-terminal line.
- jsopcode.h, jsopcode.tbl
The new extended jump opcodes, formats, and fundamental parameterization
macros. Also, more comments.
- jsparse.c
Random and probably only aesthetic fix to avoid decorating a foo[i]++ or
--foo[i] parse tree node with JSOP_SETCALL, wrongly (only foo(i)++ or
--foo(i), or the other post- or prefix form operator, should have such an
opcode decoration on its parse tree).
- jsscript.h
Random macro naming sanity: use trailing _ rather than leading _ for macro
local variables in order to avoid invading the standard C global namespace.
