github
/
duktape
mirror of https://github.com/svaarala/duktape.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

Merge pull request #370 from svaarala/executor-setjmp-external 

Move executor setjmp catchpoint to an outer wrapper function
pull/388/head
Sami Vaarala 9 years ago
parent
391053e7c2 31e727525b
commit
616557edef
17 changed files with 538 additions and 234 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 4
      RELEASES.rst
  2. 14
      config/config-options/DUK_USE_EXEC_FUN_LOCAL.yaml
  3. 2
      config/examples/low_memory.yaml
  4. 2
      config/examples/performance_sensitive.yaml
  5. 22
      doc/code-issues.rst
  6. 5
      src/duk_dblunion.h.in
  7. 433
      src/duk_js_executor.c
  8. 26
      tests/perf/test-add-fastint.js
  9. 26
      tests/perf/test-add-float.js
  10. 29
      tests/perf/test-add-nan-fastint.js
  11. 29
      tests/perf/test-add-nan.js
  12. 43
      tests/perf/test-assign-addto-nan.js
  13. 2
      tests/perf/test-assign-addto.js
  14. 2
      tests/perf/test-assign-addto.lua
  15. 2
      tests/perf/test-assign-addto.pl
  16. 1
      tests/perf/test-assign-addto.py
  17. 130
      tests/perf/test-assign-reg.pl

4
RELEASES.rst
View File

@ -1163,6 +1163,10 @@ Planned
assignment statements with an identifier left-hand-side value, especially assignment statements with an identifier left-hand-side value, especially
when the assignment is a top level expression like "x = y + z;" (GH-380) when the assignment is a top level expression like "x = y + z;" (GH-380)
* Internal performance improvement: split bytecode executor into an inner and
outer function, with the outer function containing a setjmp/longjmp catch
point and the inner function free of setjmp/longjmp (GH-369, GH-370)
2.0.0 (XXXX-XX-XX) 2.0.0 (XXXX-XX-XX)
------------------ ------------------

14
config/config-options/DUK_USE_EXEC_FUN_LOCAL.yaml
View File

@ -0,0 +1,14 @@
define: DUK_USE_EXEC_FUN_LOCAL
introduced: 1.4.0
default: false
tags:
- performance
- execution
description: >
Use a local variable "fun" pointing to the current function in the bytecode
dispatch loop instead of looking up the current function through "thr" every
time it is needed. On x64 performance is slightly better without a "fun"
local; on x86 performance is slightly better with one.
You should only tweak this if you're really interested in performance, and
should then do proper testing to see which value works better.

2
config/examples/low_memory.yaml
View File

@ -39,4 +39,4 @@ DUK_USE_STRTAB_CHAIN_SIZE: 128
# Consider to reduce code footprint at the expense of more erratic RAM usage # Consider to reduce code footprint at the expense of more erratic RAM usage
#DUK_USE_REFERENCE_COUNTING: false #DUK_USE_REFERENCE_COUNTING: false
#DUK_USE_DOUBLE_LINKED_LIST: false #DUK_USE_DOUBLE_LINKED_HEAP: false

2
config/examples/performance_sensitive.yaml
View File

@ -3,6 +3,7 @@
# You should choose the fastest setjmp/longjmp for your platform. # You should choose the fastest setjmp/longjmp for your platform.
DUK_USE_PACKED_TVAL: false # packed duk_tval slower in most cases
DUK_USE_FASTINT: true DUK_USE_FASTINT: true
DUK_USE_VALSTACK_UNSAFE: true DUK_USE_VALSTACK_UNSAFE: true
DUK_USE_FAST_REFCOUNT_DEFAULT: true DUK_USE_FAST_REFCOUNT_DEFAULT: true
@ -13,3 +14,4 @@ DUK_USE_JSON_DECNUMBER_FASTPATH: true
DUK_USE_JSON_EATWHITE_FASTPATH: true DUK_USE_JSON_EATWHITE_FASTPATH: true
DUK_USE_INTERRUPT_COUNTER: false DUK_USE_INTERRUPT_COUNTER: false
DUK_USE_DEBUGGER_SUPPORT: false DUK_USE_DEBUGGER_SUPPORT: false
#DUK_USE_EXEC_FUN_LOCAL: test both values, marginal benefit

22
doc/code-issues.rst
View File

@ -1167,9 +1167,26 @@ Optimizations may also cause odd situations, see e.g.:
* http://blog.sam.liddicott.com/2013/09/why-are-setjmp-volatile-hacks-still.html * http://blog.sam.liddicott.com/2013/09/why-are-setjmp-volatile-hacks-still.html
With Emscripten a function containing ``setjmp()`` executes much more slowly
than a function without it. For example, for the bytecode executor the speed
improvement of refactoring ``setjmp()`` out of the main executor function was
around 25%:
* https://github.com/svaarala/duktape/pull/370
Some compilers generate incorrect code with setjmp. Some workarounds may be
needed (e.g. optimizations may need to be disabled completely) for functions
containing a setjmp:
* https://github.com/svaarala/duktape/issues/369
To minimize the chances of the compiler handling setjmp/longjmp incorrectly, To minimize the chances of the compiler handling setjmp/longjmp incorrectly,
the cleanest approach would probable be to: the cleanest approach would probable be to:
* Minimize the size of functions containing a ``setjmp()``; use a wrapper
with just the ``setjmp()`` and an inner function with the rest of the
function when that's possible.
* Declare all variables used in the ``setjmp()`` non-zero return case (when * Declare all variables used in the ``setjmp()`` non-zero return case (when
called through ``longjmp()``) as volatile, so that we don't ever rely on called through ``longjmp()``) as volatile, so that we don't ever rely on
non-volatile variable values in that code path. non-volatile variable values in that code path.
@ -1195,7 +1212,7 @@ variables, e.g.::
/* ... */ /* ... */
(As of Duktape 1.1 this has not yet been done for all setjmp/longjmp (As of Duktape 1.3 this has not yet been done for all setjmp/longjmp
functions. Rather, volatile declarations have been added where they functions. Rather, volatile declarations have been added where they
seem to be needed in practice.) seem to be needed in practice.)
@ -1604,7 +1621,8 @@ arguments". The fix is to remove the comment from inside the macro::
Character values in char literals and strings, EBCDIC Character values in char literals and strings, EBCDIC
===================================================== =====================================================
**FIXME: under work** **FIXME: under work, while some other projects do support EBCDIC,
EBCDIC may not be a useful portability target for Duktape.**
Overview Overview
-------- --------

5
src/duk_dblunion.h.in
View File

@ -203,7 +203,7 @@ typedef union duk_double_union duk_double_union;
* *
* When packed duk_tval is used, the NaN space is used to store pointers * When packed duk_tval is used, the NaN space is used to store pointers
* and other tagged values in addition to NaNs. Actual NaNs are normalized * and other tagged values in addition to NaNs. Actual NaNs are normalized
* to a specific format. The macros below are used by the implementation * to a specific quiet NaN. The macros below are used by the implementation
* to check and normalize NaN values when they might be created. The macros * to check and normalize NaN values when they might be created. The macros
* are essentially NOPs when the non-packed duk_tval representation is used. * are essentially NOPs when the non-packed duk_tval representation is used.
* *
@ -211,7 +211,8 @@ typedef union duk_double_union duk_double_union;
* the packed duk_tval and works correctly for all NaNs except those that * the packed duk_tval and works correctly for all NaNs except those that
* begin with 0x7ff0. Since the 'normalized NaN' values used with packed * begin with 0x7ff0. Since the 'normalized NaN' values used with packed
* duk_tval begin with 0x7ff8, the partial check is reliable when packed * duk_tval begin with 0x7ff8, the partial check is reliable when packed
* duk_tval is used. * duk_tval is used. The 0x7ff8 prefix means the normalized NaN will be a
* quiet NaN regardless of its remaining lower bits.
* *
* The ME variant below is specifically for ARM byte order, which has the * The ME variant below is specifically for ARM byte order, which has the
* feature that while doubles have a mixed byte order (32107654), unsigned * feature that while doubles have a mixed byte order (32107654), unsigned

433
src/duk_js_executor.c
View File

@ -4,17 +4,25 @@
#include "duk_internal.h" #include "duk_internal.h"
/*
* Local declarations.
*/
DUK_LOCAL_DECL void duk__js_execute_bytecode_inner(duk_hthread *entry_thread, duk_size_t entry_callstack_top);
/* /*
* Arithmetic, binary, and logical helpers. * Arithmetic, binary, and logical helpers.
* *
* Note: there is no opcode for logical AND or logical OR; this is on * Note: there is no opcode for logical AND or logical OR; this is on
* purpose, because the evalution order semantics for them make such * purpose, because the evalution order semantics for them make such
* opcodes pretty pointless (short circuiting means they are most * opcodes pretty pointless: short circuiting means they are most
* comfortably implemented as jumps). However, a logical NOT opcode * comfortably implemented as jumps. However, a logical NOT opcode
* is useful. * is useful.
* *
* Note: careful with duk_tval pointers here: they are potentially * Note: careful with duk_tval pointers here: they are potentially
* invalidated by any DECREF and almost any API call. * invalidated by any DECREF and almost any API call. It's still
* preferable to work without making a copy but that's not always
* possible.
*/ */
DUK_LOCAL duk_double_t duk__compute_mod(duk_double_t d1, duk_double_t d2) { DUK_LOCAL duk_double_t duk__compute_mod(duk_double_t d1, duk_double_t d2) {
@ -393,7 +401,9 @@ DUK_LOCAL void duk__vm_bitwise_binary_op(duk_hthread *thr, duk_tval *tv_x, duk_t
#if defined(DUK_USE_FASTINT) #if defined(DUK_USE_FASTINT)
/* Result is always fastint compatible. */ /* Result is always fastint compatible. */
/* XXX: set 32-bit result (but must handle signed and unsigned) */ /* XXX: Set 32-bit result (but must then handle signed and
* unsigned results separately).
*/
fi3 = (duk_int64_t) i3; fi3 = (duk_int64_t) i3;
fastint_result_set: fastint_result_set:
@ -572,18 +582,15 @@ DUK_LOCAL void duk__vm_logical_not(duk_hthread *thr, duk_tval *tv_x, duk_tval *t
* Longjmp and other control flow transfer for the bytecode executor. * Longjmp and other control flow transfer for the bytecode executor.
* *
* The longjmp handler can handle all longjmp types: error, yield, and * The longjmp handler can handle all longjmp types: error, yield, and
* resume (pseudotypes are never actually thrown). Error policy for * resume (pseudotypes are never actually thrown).
* longjmp: should not ordinarily throw errors; if errors occur (e.g. *
* due to out-of-memory) they bubble outwards rather than being handled * Error policy for longjmp: should not ordinarily throw errors; if errors
* recursively. * occur (e.g. due to out-of-memory) they bubble outwards rather than being
* handled recursively.
*/ */
/* XXX: duk_api operations for cross-thread reg manipulation? */
/* XXX: post-condition: value stack must be correct; for ecmascript functions, clamped to 'nregs' */
#define DUK__LONGJMP_RESTART 0 /* state updated, restart bytecode execution */ #define DUK__LONGJMP_RESTART 0 /* state updated, restart bytecode execution */
#define DUK__LONGJMP_FINISHED 1 /* exit bytecode executor with return value */ #define DUK__LONGJMP_RETHROW 1 /* exit bytecode executor by rethrowing an error to caller */
#define DUK__LONGJMP_RETHROW 2 /* exit bytecode executor by rethrowing an error to caller */
#define DUK__RETHAND_RESTART 0 /* state updated, restart bytecode execution */ #define DUK__RETHAND_RESTART 0 /* state updated, restart bytecode execution */
#define DUK__RETHAND_FINISHED 1 /* exit bytecode execution with return value */ #define DUK__RETHAND_FINISHED 1 /* exit bytecode execution with return value */
@ -619,10 +626,10 @@ DUK_LOCAL void duk__reconfig_valstack_ecma_return(duk_hthread *thr, duk_size_t a
act = NULL; act = NULL;
(void) duk_valstack_resize_raw((duk_context *) thr, (void) duk_valstack_resize_raw((duk_context *) thr,
(thr->valstack_bottom - thr->valstack) + /* bottom of current func */ (thr->valstack_bottom - thr->valstack) + /* bottom of current func */
h_func->nregs + /* reg count */ h_func->nregs + /* reg count */
DUK_VALSTACK_INTERNAL_EXTRA, /* + spare */ DUK_VALSTACK_INTERNAL_EXTRA, /* + spare */
DUK_VSRESIZE_FLAG_SHRINK | /* flags */ DUK_VSRESIZE_FLAG_SHRINK | /* flags */
0 /* no compact */ | 0 /* no compact */ |
DUK_VSRESIZE_FLAG_THROW); DUK_VSRESIZE_FLAG_THROW);
@ -653,10 +660,10 @@ DUK_LOCAL void duk__reconfig_valstack_ecma_catcher(duk_hthread *thr, duk_size_t
cat = NULL; cat = NULL;
(void) duk_valstack_resize_raw((duk_context *) thr, (void) duk_valstack_resize_raw((duk_context *) thr,
(thr->valstack_bottom - thr->valstack) + /* bottom of current func */ (thr->valstack_bottom - thr->valstack) + /* bottom of current func */
h_func->nregs + /* reg count */ h_func->nregs + /* reg count */
DUK_VALSTACK_INTERNAL_EXTRA, /* + spare */ DUK_VALSTACK_INTERNAL_EXTRA, /* + spare */
DUK_VSRESIZE_FLAG_SHRINK | /* flags */ DUK_VSRESIZE_FLAG_SHRINK | /* flags */
0 /* no compact */ | 0 /* no compact */ |
DUK_VSRESIZE_FLAG_THROW); DUK_VSRESIZE_FLAG_THROW);
@ -1462,8 +1469,6 @@ DUK_LOCAL duk_small_uint_t duk__handle_return(duk_hthread *thr,
DUK_DD(DUK_DDPRINT("no calling activation, thread finishes (similar to yield)")); DUK_DD(DUK_DDPRINT("no calling activation, thread finishes (similar to yield)"));
/* Share yield longjmp handler. */
DUK_ASSERT(thr->resumer != NULL); DUK_ASSERT(thr->resumer != NULL);
DUK_ASSERT(thr->resumer->callstack_top >= 2); /* Ecmascript activation + Duktape.Thread.resume() activation */ DUK_ASSERT(thr->resumer->callstack_top >= 2); /* Ecmascript activation + Duktape.Thread.resume() activation */
DUK_ASSERT(DUK_ACT_GET_FUNC(thr->resumer->callstack + thr->resumer->callstack_top - 1) != NULL && DUK_ASSERT(DUK_ACT_GET_FUNC(thr->resumer->callstack + thr->resumer->callstack_top - 1) != NULL &&
@ -1477,6 +1482,7 @@ DUK_LOCAL duk_small_uint_t duk__handle_return(duk_hthread *thr,
resumer = thr->resumer; resumer = thr->resumer;
/* Share yield longjmp handler. */
DUK_ASSERT(thr->valstack_top - 1 >= thr->valstack_bottom); DUK_ASSERT(thr->valstack_top - 1 >= thr->valstack_bottom);
duk__handle_yield(thr, resumer, resumer->callstack_top - 2, thr->valstack_top - 1); duk__handle_yield(thr, resumer, resumer->callstack_top - 2, thr->valstack_top - 1);
@ -1804,6 +1810,7 @@ DUK_LOCAL duk_small_uint_t duk__executor_interrupt(duk_hthread *thr) {
* execute 1 instruction (after interrupt handler return), counter must * execute 1 instruction (after interrupt handler return), counter must
* be 0. * be 0.
*/ */
DUK_ASSERT(ctr >= 1);
thr->interrupt_init = ctr; thr->interrupt_init = ctr;
thr->interrupt_counter = ctr - 1; thr->interrupt_counter = ctr - 1;
DUK_HEAP_CLEAR_INTERRUPT_RUNNING(thr->heap); DUK_HEAP_CLEAR_INTERRUPT_RUNNING(thr->heap);
@ -1822,7 +1829,7 @@ DUK_LOCAL duk_small_uint_t duk__executor_interrupt(duk_hthread *thr) {
*/ */
#if defined(DUK_USE_DEBUGGER_SUPPORT) #if defined(DUK_USE_DEBUGGER_SUPPORT)
DUK_LOCAL void duk__executor_handle_debugger(duk_hthread *thr, duk_activation *act, duk_hcompiledfunction *fun) { DUK_LOCAL void duk__executor_recheck_debugger(duk_hthread *thr, duk_activation *act, duk_hcompiledfunction *fun) {
duk_heap *heap; duk_heap *heap;
duk_tval *tv_tmp; duk_tval *tv_tmp;
duk_hstring *filename; duk_hstring *filename;
@ -1958,15 +1965,15 @@ DUK_LOCAL void duk__executor_handle_debugger(duk_hthread *thr, duk_activation *a
* setjmp() jmpbuf. * setjmp() jmpbuf.
* *
* Ecmascript function calls and coroutine resumptions are handled * Ecmascript function calls and coroutine resumptions are handled
* internally without recursive C calls. Other function calls are * internally (by the outer executor function) without recursive C calls.
* handled using duk_handle_call(), increasing C recursion depth. * Other function calls are handled using duk_handle_call(), increasing
* * C recursion depth.
* There are many other tricky control flow situations, such as:
* *
* - Break and continue (fast and slow) * Abrupt completions (= long control tranfers) are handled either
* - Return (fast and slow) * directly by reconfiguring relevant stacks and restarting execution,
* - Error throwing * or via a longjmp. Longjmp-free handling is preferable for performance
* - Thread resume and yield * (especially Emscripten performance), and is used for: break, continue,
* and return.
* *
* For more detailed notes, see doc/execution.rst. * For more detailed notes, see doc/execution.rst.
* *
@ -1974,7 +1981,15 @@ DUK_LOCAL void duk__executor_handle_debugger(duk_hthread *thr, duk_activation *a
* and volatile. * and volatile.
*/ */
#define DUK__STRICT() (DUK_HOBJECT_HAS_STRICT(&(fun)->obj)) /* Presence of 'fun' is config based, there's a marginal performance
* difference and the best option is architecture dependent.
*/
#if defined(DUK_USE_EXEC_FUN_LOCAL)
#define DUK__FUN() fun
#else
#define DUK__FUN() ((duk_hcompiledfunction *) DUK_ACT_GET_FUNC((thr)->callstack + (thr)->callstack_top - 1))
#endif
#define DUK__STRICT() (DUK_HOBJECT_HAS_STRICT((duk_hobject *) DUK__FUN()))
#define DUK__REG(x) (*(thr->valstack_bottom + (x))) #define DUK__REG(x) (*(thr->valstack_bottom + (x)))
#define DUK__REGP(x) (thr->valstack_bottom + (x)) #define DUK__REGP(x) (thr->valstack_bottom + (x))
#define DUK__CONST(x) (*(consts + (x))) #define DUK__CONST(x) (*(consts + (x)))
@ -2004,51 +2019,17 @@ DUK_LOCAL void duk__executor_handle_debugger(duk_hthread *thr, duk_activation *a
thr->ptr_curr_pc = NULL; \ thr->ptr_curr_pc = NULL; \
} while (0) } while (0)
/* Outer executor with setjmp/longjmp handling. */
DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) { DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
/* Entry level info. Although these are assigned to before setjmp() /* Entry level info. */
* a 'volatile' seems to be needed. Note placement of "volatile" for duk_hthread *entry_thread;
* pointers. See doc/code-issues.rst for more discussion. duk_size_t entry_callstack_top;
*/ duk_int_t entry_call_recursion_depth;
duk_hthread * volatile entry_thread; /* volatile copy of exec_thr */ duk_jmpbuf *entry_jmpbuf_ptr;
volatile duk_size_t entry_callstack_top;
volatile duk_int_t entry_call_recursion_depth;
duk_jmpbuf * volatile entry_jmpbuf_ptr;
/* current PC, volatile because it is accessed by other functions
* through thr->ptr_to_curr_pc. Critical for performance. It would
* be safest to make this volatile, but that eliminates performance
* benefits. Aliasing guarantees should be enough though.
*/
duk_instr_t *curr_pc; /* stable */
/* "hot" variables for interpretation -- not volatile, value not guaranteed in setjmp error handling */
duk_hthread *thr; /* stable */
duk_hcompiledfunction *fun; /* stable */
duk_tval *consts; /* stable */
/* 'funcs' is quite rarely used, so no local for it */
/* "hot" temps for interpretation -- not volatile, value not guaranteed in setjmp error handling */
duk_uint_fast32_t ins; /* XXX: check performance impact on x64 between fast/non-fast variant */
/* jmpbuf */ duk_heap *heap;
duk_jmpbuf jmpbuf; duk_jmpbuf jmpbuf;
#ifdef DUK_USE_INTERRUPT_COUNTER
duk_int_t int_ctr;
#endif
#ifdef DUK_USE_ASSERTIONS
duk_size_t valstack_top_base; /* valstack top, should match before interpreting each op (no leftovers) */
#endif
/* XXX: document assumptions on setjmp and volatile variables
* (see duk_handle_call()).
*/
/*
* Preliminaries
*/
DUK_ASSERT(exec_thr != NULL); DUK_ASSERT(exec_thr != NULL);
DUK_ASSERT(exec_thr->heap != NULL); DUK_ASSERT(exec_thr->heap != NULL);
DUK_ASSERT(exec_thr->heap->curr_thread != NULL); DUK_ASSERT(exec_thr->heap->curr_thread != NULL);
@ -2057,111 +2038,111 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
DUK_ASSERT(DUK_ACT_GET_FUNC(exec_thr->callstack + exec_thr->callstack_top - 1) != NULL); DUK_ASSERT(DUK_ACT_GET_FUNC(exec_thr->callstack + exec_thr->callstack_top - 1) != NULL);
DUK_ASSERT(DUK_HOBJECT_IS_COMPILEDFUNCTION(DUK_ACT_GET_FUNC(exec_thr->callstack + exec_thr->callstack_top - 1))); DUK_ASSERT(DUK_HOBJECT_IS_COMPILEDFUNCTION(DUK_ACT_GET_FUNC(exec_thr->callstack + exec_thr->callstack_top - 1)));
entry_thread = exec_thr; /* volatile copy */ entry_thread = exec_thr;
thr = (duk_hthread *) entry_thread; heap = entry_thread->heap;
entry_callstack_top = thr->callstack_top; entry_callstack_top = entry_thread->callstack_top;
entry_call_recursion_depth = thr->heap->call_recursion_depth; entry_call_recursion_depth = entry_thread->heap->call_recursion_depth;
entry_jmpbuf_ptr = thr->heap->lj.jmpbuf_ptr; entry_jmpbuf_ptr = entry_thread->heap->lj.jmpbuf_ptr;
/* /*
* Setjmp catchpoint setup.
*
* Note: we currently assume that the setjmp() catchpoint is * Note: we currently assume that the setjmp() catchpoint is
* not re-entrant (longjmp() cannot be called more than once * not re-entrant (longjmp() cannot be called more than once
* for a single setjmp()). * for a single setjmp()).
*
* See doc/code-issues.rst for notes on variable assignment
* before and after setjmp().
*/ */
reset_setjmp_catchpoint: for (;;) {
heap->lj.jmpbuf_ptr = &jmpbuf;
DUK_ASSERT(heap->lj.jmpbuf_ptr != NULL);
DUK_ASSERT(thr != NULL); if (DUK_SETJMP(heap->lj.jmpbuf_ptr->jb) != 0) {
thr->heap->lj.jmpbuf_ptr = &jmpbuf; duk_small_uint_t lj_ret;
DUK_ASSERT(thr->heap->lj.jmpbuf_ptr != NULL);
if (DUK_SETJMP(thr->heap->lj.jmpbuf_ptr->jb) != 0) { DUK_DDD(DUK_DDDPRINT("longjmp caught by bytecode executor"));
/*
* Note: any local variables accessed here must have their value
* assigned *before* the setjmp() call, OR they must be declared
* volatile. Otherwise their value is not guaranteed to be correct.
*
* 'thr' might seem to be a risky variable because it is changed
* for yield and resume. However, yield and resume are handled
* using longjmp()s.
*/
duk_small_uint_t lj_ret; /* Longjmp callers are required to sync-and-null thr->ptr_curr_pc
duk_hthread *tmp_entry_thread; * before longjmp.
duk_size_t tmp_entry_callstack_top; */
DUK_ASSERT(heap->curr_thread != NULL);
DUK_ASSERT(heap->curr_thread->ptr_curr_pc == NULL);
DUK_DDD(DUK_DDDPRINT("longjmp caught by bytecode executor")); /* XXX: signalling the need to shrink check (only if unwound) */
/* Relookup 'thr': it's not volatile so its value is not /* Must be restored here to handle e.g. yields properly. */
* guaranteed. The heap->curr_thread value should always be heap->call_recursion_depth = entry_call_recursion_depth;
* valid here because longjmp callers don't switch threads,
* only the longjmp handler does that (even for RESUME and
* YIELD).
*/
DUK_ASSERT(entry_thread != NULL);
thr = entry_thread->heap->curr_thread;
/* Don't sync curr_pc when unwinding: with recursive executor /* Switch to caller's setjmp() catcher so that if an error occurs
* calls thr->ptr_curr_pc may be dangling. * during error handling, it is always propagated outwards instead
*/ * of causing an infinite loop in our own handler.
*/
heap->lj.jmpbuf_ptr = (duk_jmpbuf *) entry_jmpbuf_ptr;
/* XXX: signalling the need to shrink check (only if unwound) */ lj_ret = duk__handle_longjmp(heap->curr_thread, entry_thread, entry_callstack_top);
/* Must be restored here to handle e.g. yields properly. */ if (lj_ret == DUK__LONGJMP_RESTART) {
thr->heap->call_recursion_depth = entry_call_recursion_depth; /* Restart bytecode execution, possibly with a changed thread. */
;
} else {
/* Rethrow error to calling state. */
DUK_ASSERT(lj_ret == DUK__LONGJMP_RETHROW);
/* Switch to caller's setjmp() catcher so that if an error occurs /* Longjmp handling has restored jmpbuf_ptr. */
* during error handling, it is always propagated outwards instead DUK_ASSERT(heap->lj.jmpbuf_ptr == entry_jmpbuf_ptr);
* of causing an infinite loop in our own handler.
*/
DUK_DDD(DUK_DDDPRINT("restore jmpbuf_ptr: %p -> %p", /* Thread may have changed, e.g. YIELD converted to THROW. */
(void *) ((thr && thr->heap) ? thr->heap->lj.jmpbuf_ptr : NULL), duk_err_longjmp(heap->curr_thread);
(void *) entry_jmpbuf_ptr)); DUK_UNREACHABLE();
thr->heap->lj.jmpbuf_ptr = (duk_jmpbuf *) entry_jmpbuf_ptr; }
} else {
/* Execute bytecode until returned or longjmp(). */
duk__js_execute_bytecode_inner(entry_thread, entry_callstack_top);
tmp_entry_thread = (duk_hthread *) entry_thread; /* use temps to avoid GH-318 */ /* Successful return: restore jmpbuf and return to caller. */
tmp_entry_callstack_top = (duk_size_t) entry_callstack_top; heap->lj.jmpbuf_ptr = (duk_jmpbuf *) entry_jmpbuf_ptr;
lj_ret = duk__handle_longjmp(thr, tmp_entry_thread, tmp_entry_callstack_top); return;
}
}
if (lj_ret == DUK__LONGJMP_RESTART) { DUK_UNREACHABLE();
/* }
* Restart bytecode execution, possibly with a changed thread.
*/
thr = thr->heap->curr_thread;
goto reset_setjmp_catchpoint;
} else if (lj_ret == DUK__LONGJMP_RETHROW) {
/*
* Rethrow error to calling state.
*/
/* thread may have changed (e.g. YIELD converted to THROW) */ /* Inner executor, performance critical. */
thr = thr->heap->curr_thread; DUK_LOCAL DUK_NOINLINE void duk__js_execute_bytecode_inner(duk_hthread *entry_thread, duk_size_t entry_callstack_top) {
/* Current PC, accessed by other functions through thr->ptr_to_curr_pc.
* Critical for performance. It would be safest to make this volatile,
* but that eliminates performance benefits; aliasing guarantees
* should be enough though.
*/
duk_instr_t *curr_pc; /* bytecode has a stable pointer */
DUK_ASSERT(thr->heap->lj.jmpbuf_ptr == entry_jmpbuf_ptr); /* Hot variables for interpretation. Critical for performance,
* but must add sparingly to minimize register shuffling.
*/
duk_hthread *thr; /* stable */
duk_tval *consts; /* stable */
duk_uint_fast32_t ins;
/* 'funcs' is quite rarely used, so no local for it */
#if defined(DUK_USE_EXEC_FUN_LOCAL)
duk_hcompiledfunction *fun;
#else
/* 'fun' is quite rarely used, so no local for it */
#endif
duk_err_longjmp(thr); #ifdef DUK_USE_INTERRUPT_COUNTER
DUK_UNREACHABLE(); duk_int_t int_ctr;
} else { #endif
/*
* Return from bytecode executor with a return value.
*/
DUK_ASSERT(lj_ret == DUK__LONGJMP_FINISHED);
/* XXX: return assertions for valstack, callstack, catchstack */
DUK_ASSERT(thr->heap->lj.jmpbuf_ptr == entry_jmpbuf_ptr); #ifdef DUK_USE_ASSERTIONS
return; duk_size_t valstack_top_base; /* valstack top, should match before interpreting each op (no leftovers) */
} #endif
DUK_UNREACHABLE();
}
/* /*
* Restart execution by reloading thread state. * Restart execution by reloading thread state.
* *
* Note that 'thr' and any thread configuration may have changed, * Note that 'thr' and any thread configuration may have changed,
* so all local variables are suspect. * so all local variables are suspect and we need to reinitialize.
* *
* The number of local variables should be kept to a minimum: if * The number of local variables should be kept to a minimum: if
* the variables are spilled, they will need to be loaded from * the variables are spilled, they will need to be loaded from
@ -2170,13 +2151,38 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
* Any 'goto restart_execution;' code path in opcode dispatch must * Any 'goto restart_execution;' code path in opcode dispatch must
* ensure 'curr_pc' is synced back to act->curr_pc before the goto * ensure 'curr_pc' is synced back to act->curr_pc before the goto
* takes place. * takes place.
*
* The interpreter must be very careful with memory pointers, as
* many pointers are not guaranteed to be 'stable' and may be
* reallocated and relocated on-the-fly quite easily (e.g. by a
* memory allocation or a property access).
*
* The following are assumed to have stable pointers:
* - the current thread
* - the current function
* - the bytecode, constant table, inner function table of the
* current function (as they are a part of the function allocation)
*
* The following are assumed to have semi-stable pointers:
* - the current activation entry: stable as long as callstack
* is not changed (reallocated by growing or shrinking), or
* by any garbage collection invocation (through finalizers)
* - Note in particular that ANY DECREF can invalidate the
* activation pointer, so for the most part a fresh lookup
* is required
*
* The following are not assumed to have stable pointers at all:
* - the value stack (registers) of the current thread
* - the catch stack of the current thread
*
* See execution.rst for discussion.
*/ */
restart_execution: restart_execution:
/* Lookup current thread; use the volatile 'entry_thread' for this to /* Lookup current thread; use the stable 'entry_thread' for this to
* avoid clobber warnings. (Any valid, reachable 'thr' value would be * avoid clobber warnings. Any valid, reachable 'thr' value would be
* fine for this, so using 'entry_thread' is just to silence warnings.) * fine for this, so using 'entry_thread' is just to silence warnings.
*/ */
thr = entry_thread->heap->curr_thread; thr = entry_thread->heap->curr_thread;
DUK_ASSERT(thr != NULL); DUK_ASSERT(thr != NULL);
@ -2186,11 +2192,15 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
thr->ptr_curr_pc = &curr_pc; thr->ptr_curr_pc = &curr_pc;
/* Assume interrupt init/counter are properly initialized here. */ /* Relookup and initialize dispatch loop variables. Debugger check. */
/* assume that thr->valstack_bottom has been set-up before getting here */
{ {
duk_activation *act; duk_activation *act;
#if !defined(DUK_USE_EXEC_FUN_LOCAL)
duk_hcompiledfunction *fun;
#endif
/* Assume interrupt init/counter are properly initialized here. */
/* Assume that thr->valstack_bottom has been set-up before getting here. */
act = thr->callstack + thr->callstack_top - 1; act = thr->callstack + thr->callstack_top - 1;
fun = (duk_hcompiledfunction *) DUK_ACT_GET_FUNC(act); fun = (duk_hcompiledfunction *) DUK_ACT_GET_FUNC(act);
@ -2198,80 +2208,43 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
DUK_ASSERT(thr->valstack_top - thr->valstack_bottom == fun->nregs); DUK_ASSERT(thr->valstack_top - thr->valstack_bottom == fun->nregs);
consts = DUK_HCOMPILEDFUNCTION_GET_CONSTS_BASE(thr->heap, fun); consts = DUK_HCOMPILEDFUNCTION_GET_CONSTS_BASE(thr->heap, fun);
DUK_ASSERT(consts != NULL); DUK_ASSERT(consts != NULL);
}
#if defined(DUK_USE_DEBUGGER_SUPPORT) #if defined(DUK_USE_DEBUGGER_SUPPORT)
if (DUK_HEAP_IS_DEBUGGER_ATTACHED(thr->heap) && !thr->heap->dbg_processing) { if (DUK_HEAP_IS_DEBUGGER_ATTACHED(thr->heap) && !thr->heap->dbg_processing) {
duk_activation *act; thr->heap->dbg_processing = 1;
duk__executor_recheck_debugger(thr, act, fun);
thr->heap->dbg_processing = 1; act = thr->callstack + thr->callstack_top - 1; /* relookup after side effects */
act = thr->callstack + thr->callstack_top - 1; thr->heap->dbg_processing = 0;
duk__executor_handle_debugger(thr, act, fun); }
thr->heap->dbg_processing = 0;
}
#endif /* DUK_USE_DEBUGGER_SUPPORT */ #endif /* DUK_USE_DEBUGGER_SUPPORT */
/* XXX: shrink check flag? */ #ifdef DUK_USE_ASSERTIONS
valstack_top_base = (duk_size_t) (thr->valstack_top - thr->valstack);
/* #endif
* Bytecode interpreter.
*
* The interpreter must be very careful with memory pointers, as
* many pointers are not guaranteed to be 'stable' and may be
* reallocated and relocated on-the-fly quite easily (e.g. by a
* memory allocation or a property access).
*
* The following are assumed to have stable pointers:
* - the current thread
* - the current function
* - the bytecode, constant table, inner function table of the
* current function (as they are a part of the function allocation)
*
* The following are assumed to have semi-stable pointers:
* - the current activation entry: stable as long as callstack
* is not changed (reallocated by growing or shrinking), or
* by any garbage collection invocation (through finalizers)
* - Note in particular that ANY DECREF can invalidate the
* activation pointer
*
* The following are not assumed to have stable pointers at all:
* - the value stack (registers) of the current thread
* - the catch stack of the current thread
*
* See execution.rst for discussion.
*/
DUK_ASSERT(thr != NULL); /* Set up curr_pc for opcode dispatch. */
DUK_ASSERT(fun != NULL); curr_pc = act->curr_pc;
}
DUK_DD(DUK_DDPRINT("restarting execution, thr %p, act idx %ld, fun %p," DUK_DD(DUK_DDPRINT("restarting execution, thr %p, act idx %ld, fun %p,"
"consts %p, funcs %p, lev %ld, regbot %ld, regtop %ld, catchstack_top=%ld, " "consts %p, funcs %p, lev %ld, regbot %ld, regtop %ld, catchstack_top=%ld, "
"preventcount=%ld", "preventcount=%ld",
(void *) thr, (void *) thr,
(long) (thr->callstack_top - 1), (long) (thr->callstack_top - 1),
(void *) fun, (void *) DUK__FUN(),
(void *) DUK_HCOMPILEDFUNCTION_GET_CONSTS_BASE(thr->heap, fun), (void *) DUK_HCOMPILEDFUNCTION_GET_CONSTS_BASE(thr->heap, DUK__FUN()),
(void *) DUK_HCOMPILEDFUNCTION_GET_FUNCS_BASE(thr->heap, fun), (void *) DUK_HCOMPILEDFUNCTION_GET_FUNCS_BASE(thr->heap, DUK__FUN()),
(long) (thr->callstack_top - 1), (long) (thr->callstack_top - 1),
(long) (thr->valstack_bottom - thr->valstack), (long) (thr->valstack_bottom - thr->valstack),
(long) (thr->valstack_top - thr->valstack), (long) (thr->valstack_top - thr->valstack),
(long) thr->catchstack_top, (long) thr->catchstack_top,
(long) thr->callstack_preventcount)); (long) thr->callstack_preventcount));
#ifdef DUK_USE_ASSERTIONS /* Dispatch loop. */
valstack_top_base = (duk_size_t) (thr->valstack_top - thr->valstack);
#endif
/* Set up curr_pc for opcode dispatch. */
{
duk_activation *act;
act = thr->callstack + thr->callstack_top - 1;
curr_pc = act->curr_pc;
}
for (;;) { for (;;) {
DUK_ASSERT(thr->callstack_top >= 1); DUK_ASSERT(thr->callstack_top >= 1);
DUK_ASSERT(thr->valstack_top - thr->valstack_bottom == fun->nregs); DUK_ASSERT(thr->valstack_top - thr->valstack_bottom == DUK__FUN()->nregs);
DUK_ASSERT((duk_size_t) (thr->valstack_top - thr->valstack) == valstack_top_base); DUK_ASSERT((duk_size_t) (thr->valstack_top - thr->valstack) == valstack_top_base);
/* Executor interrupt counter check, used to implement breakpoints, /* Executor interrupt counter check, used to implement breakpoints,
@ -2317,6 +2290,9 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
} }
#endif /* DUK_USE_INTERRUPT_COUNTER */ #endif /* DUK_USE_INTERRUPT_COUNTER */
#if defined(DUK_USE_INTERRUPT_COUNTER) && defined(DUK_USE_DEBUG) #if defined(DUK_USE_INTERRUPT_COUNTER) && defined(DUK_USE_DEBUG)
/* For cross-checking during development: ensure dispatch count
* matches cumulative interrupt counter init value sums.
*/
thr->heap->inst_count_exec++; thr->heap->inst_count_exec++;
#endif #endif
@ -2324,17 +2300,17 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
{ {
duk_activation *act; duk_activation *act;
act = thr->callstack + thr->callstack_top - 1; act = thr->callstack + thr->callstack_top - 1;
DUK_ASSERT(curr_pc >= DUK_HCOMPILEDFUNCTION_GET_CODE_BASE(thr->heap, fun)); DUK_ASSERT(curr_pc >= DUK_HCOMPILEDFUNCTION_GET_CODE_BASE(thr->heap, DUK__FUN()));
DUK_ASSERT(curr_pc < DUK_HCOMPILEDFUNCTION_GET_CODE_END(thr->heap, fun)); DUK_ASSERT(curr_pc < DUK_HCOMPILEDFUNCTION_GET_CODE_END(thr->heap, DUK__FUN()));
DUK_UNREF(act); /* if debugging disabled */ DUK_UNREF(act); /* if debugging disabled */
DUK_DDD(DUK_DDDPRINT("executing bytecode: pc=%ld, ins=0x%08lx, op=%ld, valstack_top=%ld/%ld, nregs=%ld --> %!I", DUK_DDD(DUK_DDDPRINT("executing bytecode: pc=%ld, ins=0x%08lx, op=%ld, valstack_top=%ld/%ld, nregs=%ld --> %!I",
(long) (curr_pc - DUK_HCOMPILEDFUNCTION_GET_CODE_BASE(thr->heap, fun)), (long) (curr_pc - DUK_HCOMPILEDFUNCTION_GET_CODE_BASE(thr->heap, DUK__FUN())),
(unsigned long) *curr_pc, (unsigned long) *curr_pc,
(long) DUK_DEC_OP(*curr_pc), (long) DUK_DEC_OP(*curr_pc),
(long) (thr->valstack_top - thr->valstack), (long) (thr->valstack_top - thr->valstack),
(long) (thr->valstack_end - thr->valstack), (long) (thr->valstack_end - thr->valstack),
(long) (fun ? fun->nregs : -1), (long) (DUK__FUN() ? DUK__FUN()->nregs : -1),
(duk_instr_t) *curr_pc)); (duk_instr_t) *curr_pc));
} }
#endif #endif
@ -2894,6 +2870,7 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
case DUK_OP_CLOSURE: { case DUK_OP_CLOSURE: {
duk_context *ctx = (duk_context *) thr; duk_context *ctx = (duk_context *) thr;
duk_activation *act; duk_activation *act;
duk_hcompiledfunction *fun;
duk_small_uint_fast_t a = DUK_DEC_A(ins); duk_small_uint_fast_t a = DUK_DEC_A(ins);
duk_uint_fast_t bc = DUK_DEC_BC(ins); duk_uint_fast_t bc = DUK_DEC_BC(ins);
duk_hobject *fun_temp; duk_hobject *fun_temp;
@ -2903,10 +2880,13 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
*/ */
DUK_DDD(DUK_DDDPRINT("CLOSURE to target register %ld, fnum %ld (count %ld)", DUK_DDD(DUK_DDDPRINT("CLOSURE to target register %ld, fnum %ld (count %ld)",
(long) a, (long) bc, (long) DUK_HCOMPILEDFUNCTION_GET_FUNCS_COUNT(thr->heap, fun))); (long) a, (long) bc, (long) DUK_HCOMPILEDFUNCTION_GET_FUNCS_COUNT(thr->heap, DUK__FUN())));
DUK_ASSERT_DISABLE(bc >= 0); /* unsigned */ DUK_ASSERT_DISABLE(bc >= 0); /* unsigned */
DUK_ASSERT((duk_uint_t) bc < (duk_uint_t) DUK_HCOMPILEDFUNCTION_GET_FUNCS_COUNT(thr->heap, fun)); DUK_ASSERT((duk_uint_t) bc < (duk_uint_t) DUK_HCOMPILEDFUNCTION_GET_FUNCS_COUNT(thr->heap, DUK__FUN()));
act = thr->callstack + thr->callstack_top - 1;
fun = (duk_hcompiledfunction *) DUK_ACT_GET_FUNC(act);
fun_temp = DUK_HCOMPILEDFUNCTION_GET_FUNCS_BASE(thr->heap, fun)[bc]; fun_temp = DUK_HCOMPILEDFUNCTION_GET_FUNCS_BASE(thr->heap, fun)[bc];
DUK_ASSERT(fun_temp != NULL); DUK_ASSERT(fun_temp != NULL);
DUK_ASSERT(DUK_HOBJECT_IS_COMPILEDFUNCTION(fun_temp)); DUK_ASSERT(DUK_HOBJECT_IS_COMPILEDFUNCTION(fun_temp));
@ -2914,7 +2894,6 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
DUK_DDD(DUK_DDDPRINT("CLOSURE: function template is: %p -> %!O", DUK_DDD(DUK_DDDPRINT("CLOSURE: function template is: %p -> %!O",
(void *) fun_temp, (duk_heaphdr *) fun_temp)); (void *) fun_temp, (duk_heaphdr *) fun_temp));
act = thr->callstack + thr->callstack_top - 1;
if (act->lex_env == NULL) { if (act->lex_env == NULL) {
DUK_ASSERT(act->var_env == NULL); DUK_ASSERT(act->var_env == NULL);
duk_js_init_activation_environment_records_delayed(thr, act); duk_js_init_activation_environment_records_delayed(thr, act);
@ -3288,10 +3267,6 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
} }
DUK_ASSERT(ret_result == DUK__RETHAND_FINISHED); DUK_ASSERT(ret_result == DUK__RETHAND_FINISHED);
DUK_DDD(DUK_DDDPRINT("restore jmpbuf_ptr: %p -> %p",
(void *) ((thr && thr->heap) ? thr->heap->lj.jmpbuf_ptr : NULL),
(void *) entry_jmpbuf_ptr));
thr->heap->lj.jmpbuf_ptr = (duk_jmpbuf *) entry_jmpbuf_ptr;
DUK_DDD(DUK_DDDPRINT("exiting executor after RETURN handling")); DUK_DDD(DUK_DDDPRINT("exiting executor after RETURN handling"));
return; return;
} }
@ -3309,6 +3284,9 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
duk_hobject *obj_func; duk_hobject *obj_func;
duk_bool_t setup_rc; duk_bool_t setup_rc;
duk_idx_t num_stack_args; duk_idx_t num_stack_args;
#if !defined(DUK_USE_EXEC_FUN_LOCAL)
duk_hcompiledfunction *fun;
#endif
/* A -> flags /* A -> flags
* B -> base register for call (base -> func, base+1 -> this, base+2 -> arg1 ... base+2+N-1 -> argN) * B -> base register for call (base -> func, base+1 -> this, base+2 -> arg1 ... base+2+N-1 -> argN)
@ -3394,6 +3372,7 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
tv_func = DUK__REGP(idx); /* Relookup if relocated */ tv_func = DUK__REGP(idx); /* Relookup if relocated */
if (DUK_TVAL_IS_LIGHTFUNC(tv_func)) { if (DUK_TVAL_IS_LIGHTFUNC(tv_func)) {
call_flags = 0; /* not protected, respect reclimit, not constructor */ call_flags = 0; /* not protected, respect reclimit, not constructor */
/* There is no eval() special handling here: eval() is never /* There is no eval() special handling here: eval() is never
@ -3408,6 +3387,9 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
/* duk_js_call.c is required to restore the stack reserve /* duk_js_call.c is required to restore the stack reserve
* so we only need to reset the top. * so we only need to reset the top.
*/ */
#if !defined(DUK_USE_EXEC_FUN_LOCAL)
fun = DUK__FUN();
#endif
duk_set_top(ctx, (duk_idx_t) fun->nregs); duk_set_top(ctx, (duk_idx_t) fun->nregs);
/* No need to reinit setjmp() catchpoint, as call handling /* No need to reinit setjmp() catchpoint, as call handling
@ -3451,6 +3433,9 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
/* duk_js_call.c is required to restore the stack reserve /* duk_js_call.c is required to restore the stack reserve
* so we only need to reset the top. * so we only need to reset the top.
*/ */
#if !defined(DUK_USE_EXEC_FUN_LOCAL)
fun = DUK__FUN();
#endif
duk_set_top(ctx, (duk_idx_t) fun->nregs); duk_set_top(ctx, (duk_idx_t) fun->nregs);
/* No need to reinit setjmp() catchpoint, as call handling /* No need to reinit setjmp() catchpoint, as call handling
@ -4275,10 +4260,6 @@ DUK_INTERNAL void duk_js_execute_bytecode(duk_hthread *exec_thr) {
} }
DUK_ASSERT(ret_result == DUK__RETHAND_FINISHED); DUK_ASSERT(ret_result == DUK__RETHAND_FINISHED);
DUK_DDD(DUK_DDDPRINT("restore jmpbuf_ptr: %p -> %p",
(void *) ((thr && thr->heap) ? thr->heap->lj.jmpbuf_ptr : NULL),
(void *) entry_jmpbuf_ptr));
thr->heap->lj.jmpbuf_ptr = (duk_jmpbuf *) entry_jmpbuf_ptr;
DUK_DDD(DUK_DDDPRINT("exiting executor after ENDFIN and RETURN (pseudo) longjmp type")); DUK_DDD(DUK_DDDPRINT("exiting executor after ENDFIN and RETURN (pseudo) longjmp type"));
return; return;
} }

26
tests/perf/test-add-fastint.js
View File

@ -0,0 +1,26 @@
function test() {
var i;
var x = 123;
var y = 234;
var t;
for (i = 0; i < 1e6; i++) {
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
}
}
try {
test();
} catch (e) {
print(e.stack || e);
throw e;
}

26
tests/perf/test-add-float.js
View File

@ -0,0 +1,26 @@
function test() {
var i;
var x = 123.1;
var y = 234.2;
var t;
for (i = 0; i < 1e6; i++) {
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
}
}
try {
test();
} catch (e) {
print(e.stack || e);
throw e;
}

29
tests/perf/test-add-nan-fastint.js
View File

@ -0,0 +1,29 @@
/*
* Test addition when it involves a NaN, which matters on x86.
*/
function test() {
var i;
var x = 0 / 0;
var y = 234; // fastint
var t;
for (i = 0; i < 1e6; i++) {
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
}
}
try {
test();
} catch (e) {
print(e.stack || e);
throw e;
}

29
tests/perf/test-add-nan.js
View File

@ -0,0 +1,29 @@
/*
* Test addition when it involves a NaN, which matters on x86.
*/
function test() {
var i;
var x = 0 / 0;
var y = 234.5; // not fastint
var t;
for (i = 0; i < 1e6; i++) {
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
t = x + y;
}
}
try {
test();
} catch (e) {
print(e.stack || e);
throw e;
}

43
tests/perf/test-assign-addto-nan.js
View File

@ -0,0 +1,43 @@
/*
* Special variant of assign-addto test where the initial 't' is not
* initialized so it becomes NaN on first assignment and stays that
* way for the whole test.
*
* For x64 there's not much performance difference between 't' being
* a NaN, a fastint, or a double.
*
* But for x86 there's a roughly 10x difference between 't' being NaN
* and 't' being a double (and 20x difference between NaN and fastint).
* A similar difference occurs when duk_tval is packed or unpacked, so
* the difference is not caused by packed duk_tval and NaN normalization
* involved in that.
*/
function test() {
var i;
// With this initializer: x86 packed ~2,4 sec, x86 unpacked ~1,6sec, x64 unpacked ~1,5 sec
//var t = 1;
// With this initializer: x86 packed ~5,8 sec, x86 unpacked ~3,1 sec, x64 unpacked ~1,7 sec
//var t = 0.1;
// With this initializer: x86 packed ~33,1 sec (!), x86 unpacked ~30,0 sec (!), x64 unpacked ~1,7 sec
var t;
var a = 10;
for (i = 0; i < 1e7; i++) {
t += a; t += a; t += a; t += a; t += a;
t += a; t += a; t += a; t += a; t += a;
t += a; t += a; t += a; t += a; t += a;
t += a; t += a; t += a; t += a; t += a;
}
}
try {
test();
} catch (e) {
print(e.stack || e);
throw e;
}

2
tests/perf/test-assign-addto.js
View File

@ -1,6 +1,6 @@
function test() { function test() {
var i; var i;
var t; var t = 0;
var a = 10; var a = 10;
for (i = 0; i < 1e7; i++) { for (i = 0; i < 1e7; i++) {

2
tests/perf/test-assign-addto.lua
View File

@ -1,5 +1,5 @@
local function test() local function test()
local t local t = 0
local a = 10 local a = 10
for i=1,1e7 do for i=1,1e7 do

2
tests/perf/test-assign-addto.pl
View File

@ -1,6 +1,6 @@
sub test { sub test {
my $i; my $i;
my $t; my $t = 0;
my $a = 10; my $a = 10;
my $b = 20; my $b = 20;

1
tests/perf/test-assign-addto.py
View File

@ -1,4 +1,5 @@
def test(): def test():
t = 0
a = 10 a = 10
i = 0 i = 0

130
tests/perf/test-assign-reg.pl
View File

@ -0,0 +1,130 @@
sub test {
my $i;
my $t;
my $r0 = 123.0;
my $r1 = 123.0;
my $r2 = 123.0;
my $r3 = 123.0;
my $r4 = 123.0;
my $r5 = 123.0;
my $r6 = 123.0;
my $r7 = 123.0;
my $r8 = 123.0;
my $r9 = 123.0;
$i = 0;
for ($i = 0; $i < 1e7; $i++) {
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
$t = $r0;
$t = $r1;
$t = $r2;
$t = $r3;
$t = $r4;
$t = $r5;
$t = $r6;
$t = $r7;
$t = $r8;
$t = $r9;
}
}
test();
Write Preview
Loading…
Cancel
Save