Safe signals slowdown #10222
Comments
From @nwc10Perl 5.8.0 introduced "safe" signals, which changed signal handling to have int TAINT_NOT; was changed to check for and dispatch pending signals between ops, by changing #ifndef PERL_ASYNC_CHECK to this: #ifndef PERL_MICRO #ifndef PERL_ASYNC_CHECK (PERL_ASYNC_CHECK was originally added for MacOS classic. Classic is now gone) PERL_ASYNC_CHECK is also present in the restart loops of various IO ops. At the time this was believed to cause a slowdown, and IIRC a figure given (Change PERL_ASYNC_CHECK to NOOP, and force unsafe signals. Quite a few Anyway, on IRC Reini Urban was investigating unrolling the OP loop, as part of IIRC some time ago Leo was working on JITting the parrot, and took the So I tried removing PERL_ASYNC_CHECK from the oploop, and putting into The last one might be overkill, but originally I didn't have POP_BLOCK, ######## because POB_BLOCK has already restored PL_curcop before Perl_leave_scope() So, with this, signal dispatch is now (effectively) at C<;>, C<}> and any But, all tests pass, and perlbench says: $ ./perlbench-run /home/nclark/Sandpit/snap5.9.x-v5.11.5-98-g243b158/bin/perl5.11.5-{vanilla,unsafe,safefast} /home/nclark/Sandpit/snap5.9.x-v5.11.5-98-g243b158/bin/perl5.11.5-{vanilla,safefast,unsafe} B) perl-5.11.5 C) perl-5.11.5 D) perl-5.11.5 E) perl-5.11.5 F) perl-5.11.5 A B C D E F AVERAGE 100 103 103 100 104 104 ^ unmodified blead ^ So it looks like this approach recovers *all* the speed lost by safe Nicholas Clark Inline Patchdiff --git a/cop.h b/cop.h
index 13ce794..0faf0b0 100644
--- a/cop.h
+++ b/cop.h
@@ -563,12 +563,13 @@ struct block {
(long)cxstack_ix, PL_block_type[CxTYPE(cx)]); )
/* Exit a block (RETURN and LAST). */
-#define POPBLOCK(cx,pm) cx = &cxstack[cxstack_ix--], \
- newsp = PL_stack_base + cx->blk_oldsp, \
- PL_curcop = cx->blk_oldcop, \
- PL_markstack_ptr = PL_markstack + cx->blk_oldmarksp, \
- PL_scopestack_ix = cx->blk_oldscopesp, \
- pm = cx->blk_oldpm, \
+#define POPBLOCK(cx,pm) cx = &cxstack[cxstack_ix--]; \
+ PERL_ASYNC_CHECK(); \
+ newsp = PL_stack_base + cx->blk_oldsp; \
+ PL_curcop = cx->blk_oldcop; \
+ PL_markstack_ptr = PL_markstack + cx->blk_oldmarksp; \
+ PL_scopestack_ix = cx->blk_oldscopesp; \
+ pm = cx->blk_oldpm; \
gimme = cx->blk_gimme; \
DEBUG_SCOPE("POPBLOCK"); \
DEBUG_l( PerlIO_printf(Perl_debug_log, "Leaving block %ld, type %s\n", \
diff --git a/dump.c b/dump.c
index bc1ba58..d1fa26e 100644
--- a/dump.c
+++ b/dump.c
@@ -2026,7 +2026,6 @@ Perl_runops_debug(pTHX)
DEBUG_l(Perl_deb(aTHX_ "Entering new RUNOPS level\n"));
do {
- PERL_ASYNC_CHECK();
if (PL_debug) {
if (PL_watchaddr && (*PL_watchaddr != PL_watchok))
PerlIO_printf(Perl_debug_log,
diff --git a/pp_ctl.c b/pp_ctl.c
index 742bc3d..9f8a6d5 100644
--- a/pp_ctl.c
+++ b/pp_ctl.c
@@ -264,6 +264,9 @@ PP(pp_substcont)
register REGEXP * const rx = cx->sb_rx;
SV *nsv = NULL;
REGEXP *old = PM_GETRE(pm);
+
+ PERL_ASYNC_CHECK();
+
if(old != rx) {
if(old)
ReREFCNT_dec(old);
@@ -1865,6 +1868,8 @@ PP(pp_dbstate)
PL_stack_sp = PL_stack_base + cxstack[cxstack_ix].blk_oldsp;
FREETMPS;
+ PERL_ASYNC_CHECK();
+
if (PL_op->op_flags & OPf_SPECIAL /* breakpoint */
|| SvIV(PL_DBsingle) || SvIV(PL_DBsignal) || SvIV(PL_DBtrace))
{
@@ -2647,6 +2652,8 @@ PP(pp_goto)
else
label = cPVOP->op_pv;
+ PERL_ASYNC_CHECK();
+
if (label && *label) {
OP *gotoprobe = NULL;
bool leaving_eval = FALSE;
diff --git a/pp_hot.c b/pp_hot.c
index 3371e88..9f70147 100644
--- a/pp_hot.c
+++ b/pp_hot.c
@@ -52,6 +52,7 @@ PP(pp_nextstate)
TAINT_NOT; /* Each statement is presumed innocent */
PL_stack_sp = PL_stack_base + cxstack[cxstack_ix].blk_oldsp;
FREETMPS;
+ PERL_ASYNC_CHECK();
return NORMAL;
}
@@ -98,6 +99,7 @@ PP(pp_gv)
PP(pp_and)
{
dVAR; dSP;
+ PERL_ASYNC_CHECK();
if (!SvTRUE(TOPs))
RETURN;
else {
@@ -203,6 +205,7 @@ PP(pp_sassign)
PP(pp_cond_expr)
{
dVAR; dSP;
+ PERL_ASYNC_CHECK();
if (SvTRUEx(POPs))
RETURNOP(cLOGOP->op_other);
else
@@ -416,6 +419,7 @@ PP(pp_preinc)
PP(pp_or)
{
dVAR; dSP;
+ PERL_ASYNC_CHECK();
if (SvTRUE(TOPs))
RETURN;
else {
@@ -434,6 +438,7 @@ PP(pp_defined)
const bool is_dor = (op_type == OP_DOR || op_type == OP_DORASSIGN);
if (is_dor) {
+ PERL_ASYNC_CHECK();
sv = TOPs;
if (!sv || !SvANY(sv)) {
if (op_type == OP_DOR)
@@ -2071,6 +2076,8 @@ PP(pp_subst)
#endif
SV *nsv = NULL;
+ PERL_ASYNC_CHECK();
+
/* known replacement string? */
register SV *dstr = (pm->op_pmflags & PMf_CONST) ? POPs : NULL;
if (PL_op->op_flags & OPf_STACKED)
diff --git a/run.c b/run.c
index be280ee..20c711a 100644
--- a/run.c
+++ b/run.c
@@ -38,7 +38,6 @@ Perl_runops_standard(pTHX)
{
dVAR;
while ((PL_op = CALL_FPTR(PL_op->op_ppaddr)(aTHX))) {
- PERL_ASYNC_CHECK();
}
TAINT_NOT;
diff --git a/scope.c b/scope.c
index ed4c835..b9b50bf 100644
--- a/scope.c
+++ b/scope.c
@@ -1113,6 +1113,8 @@ Perl_leave_scope(pTHX_ I32 base)
}
PL_tainted = was;
+
+ PERL_ASYNC_CHECK();
}
void |
From @rurbanNicholas Clark (via RT) schrieb:
BTW: Just for reference. Here is my updated runops jitter x86 for Great patch, Nick! |
From @rurban/* Jit.xs #include <EXTERN.h> #ifndef _WIN32 #define T_CHARARR static unsigned char /* threaded: not-threaded: #if (defined(__i386__) || defined(_M_IX86)) && defined(USE_ITHREADS) /* after calling Perl_despatch_signals, restore my_perl into ebx and push for next epilog after final Perl_despatch_signals /* my_perl already on stack, but force 16-alignment for stack */ # define PROLOG x86thr_prolog #endif /* prolog: T_CHARARR x86_prolog[] = {0x55,0x89,0xe5}; /* just save ebp,esp */ # define PROLOG x86_prolog /* only after io funcs? rumor says after each and every op */ /* All ops are unrolled in execution order for the CPU cache, For now only implemented for x86 with certain hardcoded my_perl offsets. /* quirky pass 1: need code size to allocate string. #define PUSHc(what) memcpy(code,what,sizeof(what)); code += sizeof(what) /* pass 2: jit */ /*I_ASSERT(size == (code - c));*/ code = c; /* gdb: disassemble code code+200 */ MODULE=Jit PACKAGE=Jit PROTOTYPES: DISABLE BOOT: |
|
The RT System itself - Status changed from 'new' to 'open' |
From @hvdsNicholas Clark (via RT) <perlbug-followup@perl.org> wrote: This is what I get for blead (gcc 4.4.2 with -O6, no threads) without and (Adapted from output generated by <http://crypt.org/hv/perl/disfunc>.) It is quite distressing that it doesn't avoid redundant reload of The patch at the end of this email could be applied over yours to give :AVERAGE 100 103 103 100 104 104 That's awesome. :) Do you address map and grep? I believe they can run a long time without It might also trip up modules that use perl ops but not flow control, :So it looks like this approach recovers *all* the speed lost by safe Offhand, I'd expect it to reduce the overhead by 1-2 orders of (decimal) I think you might still notice it on very simple loop bodies: Hugo Inline Patch--- run.c.old 2010-03-11 19:06:10.000000000 +0000
+++ run.c 2010-03-11 22:05:44.000000000 +0000
@@ -37,7 +37,8 @@
Perl_runops_standard(pTHX)
{
dVAR;
- while ((PL_op = CALL_FPTR(PL_op->op_ppaddr)(aTHX))) {
+ register OP* op = PL_op;
+ while ((PL_op = op = CALL_FPTR(op->op_ppaddr)(aTHX))) {
}
TAINT_NOT; |
From [email protected]On Mar 10, 2010, at 3:23 AM, Nicholas Clark (via RT) wrote:
Not entirely. MacPerl may be gone, but Lamp (Lamp ain't Mac POSIX) Josh |
From @nwc10On Thu, Mar 11, 2010 at 10:26:43PM +0000, hv@crypt.org wrote:
Mmm, yes
I believe so, as I explicitly looked for all ops that do flow control
I'm having a hard time too.
Aha. Right yes, that would make sense.
Right, let's try: A) perl-5.11.5 B) perl-5.11.5 C) perl-5.11.5 D) perl-5.11.5 E) perl-5.11.5 F) perl-5.11.5 G) perl-5.11.5 H) perl-5.11.5 I) perl-5.11.5 J) perl-5.11.5 A B C D E F G H I J AVERAGE 100 104 105 100 105 104 101 100 107 106 Yes, it makes it faster, but it's within the noise level. http://perl5.git.perl.org/perl.git/shortlog/refs/heads/nicholas/HvFILL-shrink which reduces struct XPVHV by one IV, but that doesn't speed it up further Nicholas Clark |
From @rurban2010/3/11 <hv@crypt.org>:
This mov PL_op, %eax should not be needed.
This test neither.
Thanks. I have this for my jitted x86 no-threads, prolog: call: epilog:
Ah, this is where the additional PL_op in eax is coming from. |
From @nwc10On Thu, Mar 11, 2010 at 10:26:43PM +0000, hv@crypt.org wrote:
"Applied" (by hand) as 339aac2. Nicholas Clark |
|
@nwc10 - Status changed from 'open' to 'resolved' |
From @rurban2010/4/15 Nicholas Clark <nick@ccl4.org>:
Oh! Did you time it? I am working on a version which passes down the local op as arg down |
From @nwc10On Thu, Apr 15, 2010 at 04:53:52PM +0200, Reini Urban wrote:
perlbench suggested '103', compared to '100' without. (Bigger is better). Nicholas Clark |
Migrated from rt.perl.org#73480 (status was 'resolved')
Searchable as RT73480$
The text was updated successfully, but these errors were encountered: