i. improvements: ravel fntbl; check in old patch for lit2/lit4 i.; avx512 int table cmp; special code for two-column table. Sample time a cycle or two earlier in timedlock

Author: moon-chilled

jsrc/j.h

@@ -1586,6 +1586,7 @@ if(likely(!((I)jtinplace&JTWILLBEOPENED)))z=EPILOGNORET(z); RETF(z); \
 // -mavx or /arch:AVX should already generate VEX encoded for SSE instructions
 #define _mm256_zeroupperx(x)
 // this is faster than reusing another register as the source anyway, because it's not a recognised idiom, so we would have a false dependency on the other register
+#define _mm_setone_si128() _mm_cmpeq_epi32(_mm_setzero_si128(), _mm_setzero_si128())
 #define _mm256_setone_epi64() _mm256_cmpeq_epi64(_mm256_setzero_si256(), _mm256_setzero_si256())
 #define _mm256_setone_pd() _mm256_castsi256_pd(_mm256_setone_epi64())
 static inline __m256i LOADV32I(void *x) { return _mm256_loadu_si256(x); }
@@ -2186,6 +2187,7 @@ if(likely(type _i<3)){z=(I)&oneone; z=type _i>1?(I)_zzt:z; _zzt=type _i<1?(I*)z:
 #if (C_AVX2&&SY_64)
 #define PEXT(s,m) _pext_u64(s,m)
 #define PDEP(s,m) _pdep_u64(s,m)
+#define BZHI(s,i) _bzhi_u64(s,i)
 #else
 // #define PEXT(s,m) _pext_u32(s,m)
 // #define PDEP(s,m) _pdep_u32(s,m)
@@ -2311,6 +2313,7 @@ static inline UINT _clearfp(void){int r=fetestexcept(FE_ALL_EXCEPT);
 #define DPMULDE(x,y,z) ASSERT(!__builtin_smulll_overflow(x,y,&z),EVLIMIT)
 #define DPUMUL(x,y,z,h) {__int128 _t; _t=(__int128)(x)*(__int128)(y); z=(I)_t; h=(I)(_t>>64);}  // product in z and h
 #define DPUMULH(x,y,h) {__int128 _t; _t=(__int128)(x)*(__int128)(y); h=(I)(_t>>64);}  // high product in h
+#define DPUMULU(x,y,z,h) {__uint128_t _t=(__uint128_t)(x)*(__uint128_t)(y);z=(UI)_t;h=(UI)(_t>>64);}
 #endif
 #else // C_USEMULTINTRINSIC 0 - use standard-C version (64-bit)
 #define DPMULDECLS

jsrc/mt.c

@@ -189,8 +189,8 @@ fail:CLRFUTEXWT; R r;}  // error return, with our internal errorcode
 I jtpthread_mutex_timedlock(J jt,jtpthread_mutex_t *m,UI ns,I self){ //lock m, with a timeout of ns ns.  Largely the same as lock
  I r;  // internal return code in case of error
  if(unlikely(!casa((US*)&m->v,&(US){FREE},LOCK))){    //fast and common path: attempt to install LOCK in place of FREE; if so, we have acquired the lock
-  if(uncommon(m->owner==self)){if(unlikely(!m->recursive))R EVCONCURRENCY; m->ct++;R 0;} //handle deadlock and recursive cases
   struct jtimespec tgt=jtmtil(ns);
+  if(uncommon(m->owner==self)){if(unlikely(!m->recursive))R EVCONCURRENCY; m->ct++;R 0;} //handle deadlock and recursive cases
   sta(&jt->futexwt,&m->v); //ensure other threads know how to wake us up for systemlock
   while(xchga((US*)&m->v,WAIT)!=FREE){ //exit when _we_ successfully installed WAIT in place of FREE
    UI4 waitval=lda(&m->v); C breakb;  // get the serial number before we check.  Must be atomic; this is supposed to synchronise with writes to the same location via futexwt by wakeall