FUNCTION static Word_t j__udy1LCountSM(
const Pjp_t Pjp, // top of Judy (sub)SM.
const Word_t Index, // count at or above this Index.
const Pjpm_t Pjpm) // for returning error info.
{
Pjbl_t Pjbl; // Pjp->jp_Addr masked and cast to types:
Pjbb_t Pjbb;
Pjbu_t Pjbu;
Pjll_t Pjll; // a Judy lower-level linear leaf.
Word_t digit; // next digit to decode from Index.
long jpnum; // JP number in a branch (base 0).
int offset; // index ordinal within a leaf, base 0.
Word_t pop1; // total population of an expanse.
Word_t pop1above; // to return.
// Common code to check Decode bits in a JP against the equivalent portion of
// Index; XOR together, then mask bits of interest; must be all 0:
//
// Note: Why does this code only assert() compliance rather than actively
// checking for outliers? Its because Index is supposed to be valid, hence
// always match any Dcd bits traversed.
//
// Note: This assertion turns out to be always true for cState = 3 on 32-bit
// and 7 on 64-bit, but its harmless, probably removed by the compiler.
#define CHECKDCD(Pjp,cState) \
assert(! JU_DCDNOTMATCHINDEX(Index, Pjp, cState))
// Common code to prepare to handle a root-level or lower-level branch:
// Extract a state-dependent digit from Index in a "constant" way, obtain the
// total population for the branch in a state-dependent way, and then branch to
// common code for multiple cases:
//
// For root-level branches, the state is always cJU_ROOTSTATE, and the
// population is received in Pjpm->jpm_Pop0.
//
// Note: The total population is only needed in cases where the common code
// "counts up" instead of down to minimize cache line fills. However, its
// available cheaply, and its better to do it with a constant shift (constant
// state value) instead of a variable shift later "when needed".
#define PREPB_ROOT(Pjp,Next) \
digit = JU_DIGITATSTATE(Index, cJU_ROOTSTATE); \
pop1 = (Pjpm->jpm_Pop0) + 1; \
goto Next
#define PREPB(Pjp,cState,Next) \
digit = JU_DIGITATSTATE(Index, cState); \
pop1 = JU_JPBRANCH_POP0(Pjp, (cState)) + 1; \
goto Next
// SWITCH ON JP TYPE:
//
// WARNING: For run-time efficiency the following cases replicate code with
// varying constants, rather than using common code with variable values!
switch (JU_JPTYPE(Pjp))
{
// ----------------------------------------------------------------------------
// ROOT-STATE LEAF that starts with a Pop0 word; just count within the leaf:
case cJU_LEAFW:
{
Pjlw_t Pjlw = P_JLW(Pjp->jp_Addr); // first word of leaf.
assert((Pjpm->jpm_Pop0) + 1 == Pjlw[0] + 1); // sent correctly.
offset = j__udySearchLeafW(Pjlw + 1, Pjpm->jpm_Pop0 + 1, Index);
assert(offset >= 0); // Index must exist.
assert(offset < (Pjpm->jpm_Pop0) + 1); // Index be in range.
return((Pjpm->jpm_Pop0) + 1 - offset); // INCLUSIVE of Index.
}
// ----------------------------------------------------------------------------
// LINEAR BRANCH; count populations in JPs in the JBL ABOVE the next digit in
// Index, and recurse for the next digit in Index:
//
// Note: There are no null JPs in a JBL; watch out for pop1 == 0.
//
// Note: A JBL should always fit in one cache line => no need to count up
// versus down to save cache line fills. (PREPB() sets pop1 for no reason.)
case cJU_JPBRANCH_L2: CHECKDCD(Pjp, 2); PREPB(Pjp, 2, BranchL);
case cJU_JPBRANCH_L3: CHECKDCD(Pjp, 3); PREPB(Pjp, 3, BranchL);
#ifdef JU_64BIT
case cJU_JPBRANCH_L4: CHECKDCD(Pjp, 4); PREPB(Pjp, 4, BranchL);
case cJU_JPBRANCH_L5: CHECKDCD(Pjp, 5); PREPB(Pjp, 5, BranchL);
case cJU_JPBRANCH_L6: CHECKDCD(Pjp, 6); PREPB(Pjp, 6, BranchL);
case cJU_JPBRANCH_L7: CHECKDCD(Pjp, 7); PREPB(Pjp, 7, BranchL);
#endif
case cJU_JPBRANCH_L: PREPB_ROOT(Pjp, BranchL);
// Common code (state-independent) for all cases of linear branches:
BranchL:
Pjbl = P_JBL(Pjp->jp_Addr);
jpnum = Pjbl->jbl_NumJPs; // above last JP.
pop1above = 0;
while (digit < (Pjbl->jbl_Expanse[--jpnum])) // still ABOVE digit.
{
if ((pop1 = j__udyJPPop1((Pjbl->jbl_jp) + jpnum)) == cJU_ALLONES)
{
JU_SET_ERRNO_NONNULL(Pjpm, JU_ERRNO_CORRUPT);
return(C_JERR);
}
pop1above += pop1;
assert(jpnum > 0); // should find digit.
}
assert(digit == (Pjbl->jbl_Expanse[jpnum])); // should find digit.
pop1 = j__udy1LCountSM((Pjbl->jbl_jp) + jpnum, Index, Pjpm);
if (pop1 == C_JERR) return(C_JERR); // pass error up.
assert(pop1above + pop1);
return(pop1above + pop1);
// ----------------------------------------------------------------------------
// BITMAP BRANCH; count populations in JPs in the JBB ABOVE the next digit in
// Index, and recurse for the next digit in Index:
//
// Note: There are no null JPs in a JBB; watch out for pop1 == 0.
case cJU_JPBRANCH_B2: CHECKDCD(Pjp, 2); PREPB(Pjp, 2, BranchB);
case cJU_JPBRANCH_B3: CHECKDCD(Pjp, 3); PREPB(Pjp, 3, BranchB);
#ifdef JU_64BIT
case cJU_JPBRANCH_B4: CHECKDCD(Pjp, 4); PREPB(Pjp, 4, BranchB);
case cJU_JPBRANCH_B5: CHECKDCD(Pjp, 5); PREPB(Pjp, 5, BranchB);
case cJU_JPBRANCH_B6: CHECKDCD(Pjp, 6); PREPB(Pjp, 6, BranchB);
case cJU_JPBRANCH_B7: CHECKDCD(Pjp, 7); PREPB(Pjp, 7, BranchB);
#endif
case cJU_JPBRANCH_B: PREPB_ROOT(Pjp, BranchB);
// Common code (state-independent) for all cases of bitmap branches:
BranchB:
{
intptr_t subexp; // for stepping through layer 1 (subexpanses).
intptr_t findsub; // subexpanse containing Index (digit).
Word_t findbit; // bit representing Index (digit).
Word_t lowermask; // bits for indexes at or below Index.
Word_t jpcount; // JPs in a subexpanse.
Word_t clbelow; // cache lines below digits cache line.
Word_t clabove; // cache lines above digits cache line.
Pjbb = P_JBB(Pjp->jp_Addr);
findsub = digit / cJU_BITSPERSUBEXPB;
findbit = digit % cJU_BITSPERSUBEXPB;
lowermask = JU_MASKLOWERINC(JU_BITPOSMASKB(findbit));
clbelow = clabove = 0; // initial/default => always downward.
assert(JU_BITMAPTESTB(Pjbb, digit)); // digit must have a JP.
assert(findsub < cJU_NUMSUBEXPB); // falls in expected range.
// Shorthand for one subexpanse in a bitmap and for one JP in a bitmap branch:
//
// Note: BMPJP0 exists separately to support assertions.
#define BMPJP0(Subexp) (P_JP(JU_JBB_PJP(Pjbb, Subexp)))
#define BMPJP(Subexp,JPnum) (BMPJP0(Subexp) + (JPnum))
#ifndef NOSMARTJBB // enable to turn off smart code for comparison purposes.
// FIGURE OUT WHICH DIRECTION CAUSES FEWER CACHE LINE FILLS; adding the pop1s
// in JPs above Indexs JP, or subtracting the pop1s in JPs below Indexs JP.
//
// This is tricky because, while each set bit in the bitmap represents a JP,
// the JPs are scattered over cJU_NUMSUBEXPB subexpanses, each of which can
// contain JPs packed into multiple cache lines, and this code must visit every
// JP either BELOW or ABOVE the JP for Index.
//
// Number of cache lines required to hold a linear list of the given number of
// JPs, assuming the first JP is at the start of a cache line or the JPs in
// jpcount fit wholly within a single cache line, which is ensured by
// JudyMalloc():
#define CLPERJPS(jpcount) \
((((jpcount) * cJU_WORDSPERJP) + cJU_WORDSPERCL - 1) / cJU_WORDSPERCL)
// Count cache lines below/above for each subexpanse:
for (subexp = 0; subexp < cJU_NUMSUBEXPB; ++subexp)
{
jpcount = j__udyCountBitsB(JU_JBB_BITMAP(Pjbb, subexp));
// When at the subexpanse containing Index (digit), add cache lines
// below/above appropriately, excluding the cache line containing the JP for
// Index itself:
if (subexp < findsub) clbelow += CLPERJPS(jpcount);
else if (subexp > findsub) clabove += CLPERJPS(jpcount);
else // (subexp == findsub)
{
Word_t clfind; // cache line containing Index (digit).
clfind = CLPERJPS(j__udyCountBitsB(
JU_JBB_BITMAP(Pjbb, subexp) & lowermask));
assert(clfind > 0); // digit itself should have 1 CL.
clbelow += clfind - 1;
clabove += CLPERJPS(jpcount) - clfind;
}
}
#endif // ! NOSMARTJBB
// Note: Its impossible to get through the following "if" without setting
// jpnum -- see some of the assertions below -- but gcc -Wall doesnt know
// this, so preset jpnum to make it happy:
jpnum = 0;
// COUNT POPULATION FOR A BITMAP BRANCH, in whichever direction should result
// in fewer cache line fills:
//
// Note: If the remainder of Index is zero, pop1above is the pop1 of the
// entire expanse and theres no point in recursing to lower levels; but this
// should be so rare that its not worth checking for;
// Judy1Count()/JudyLCount() never even calls the motor for Index == 0 (all
// bytes).
// COUNT UPWARD, subtracting each "below or at" JPs pop1 from the whole
// expanses pop1:
//
// Note: If this causes clbelow + 1 cache line fills including JPs cache
// line, thats OK; at worst this is the same as clabove.
if (clbelow < clabove)
{
#ifdef SMARTMETRICS
++jbb_upward;
#endif
pop1above = pop1; // subtract JPs at/below Index.
// Count JPs for which to accrue pop1s in this subexpanse:
//
// TBD: If JU_JBB_BITMAP is cJU_FULLBITMAPB, dont bother counting.
for (subexp = 0; subexp <= findsub; ++subexp)
{
jpcount = j__udyCountBitsB((subexp < findsub) ?
JU_JBB_BITMAP(Pjbb, subexp) :
JU_JBB_BITMAP(Pjbb, subexp) & lowermask);
// should always find findbit:
assert((subexp < findsub) || jpcount);
// Subtract pop1s from JPs BELOW OR AT Index (digit):
//
// Note: The pop1 for Indexs JP itself is partially added back later at a
// lower state.
//
// Note: An empty subexpanse (jpcount == 0) is handled "for free".
//
// Note: Must be null JP subexp pointer in empty subexpanse and non-empty in
// non-empty subexpanse:
assert( jpcount || (BMPJP0(subexp) == (Pjp_t) NULL));
assert((! jpcount) || (BMPJP0(subexp) != (Pjp_t) NULL));
for (jpnum = 0; jpnum < jpcount; ++jpnum)
{
if ((pop1 = j__udyJPPop1(BMPJP(subexp, jpnum)))
== cJU_ALLONES)
{
JU_SET_ERRNO_NONNULL(Pjpm, JU_ERRNO_CORRUPT);
return(C_JERR);
}
pop1above -= pop1;
}
jpnum = jpcount - 1; // make correct for digit.
}
}
// COUNT DOWNWARD, adding each "above" JPs pop1:
else
{
long jpcountbf; // below findbit, inclusive.
#ifdef SMARTMETRICS
++jbb_downward;
#endif
pop1above = 0; // add JPs above Index.
jpcountbf = 0; // until subexp == findsub.
// Count JPs for which to accrue pop1s in this subexpanse:
//
// This is more complicated than counting upward because the scan of digits
// subexpanse must count ALL JPs, to know where to START counting down, and
// ALSO note the offset of digits JP to know where to STOP counting down.
for (subexp = cJU_NUMSUBEXPB - 1; subexp >= findsub; --subexp)
{
jpcount = j__udyCountBitsB(JU_JBB_BITMAP(Pjbb, subexp));
// should always find findbit:
assert((subexp > findsub) || jpcount);
if (! jpcount) continue; // empty subexpanse, save time.
// Count JPs below digit, inclusive:
if (subexp == findsub)
{
jpcountbf = j__udyCountBitsB(JU_JBB_BITMAP(Pjbb, subexp)
& lowermask);
}
// should always find findbit:
assert((subexp > findsub) || jpcountbf);
assert(jpcount >= jpcountbf); // proper relationship.
// Add pop1s from JPs ABOVE Index (digit):
// no null JP subexp pointers:
assert(BMPJP0(subexp) != (Pjp_t) NULL);
for (jpnum = jpcount - 1; jpnum >= jpcountbf; --jpnum)
{
if ((pop1 = j__udyJPPop1(BMPJP(subexp, jpnum)))
== cJU_ALLONES)
{
JU_SET_ERRNO_NONNULL(Pjpm, JU_ERRNO_CORRUPT);
return(C_JERR);
}
pop1above += pop1;
}
// jpnum is now correct for digit.
}
} // else.
// Return the net population ABOVE the digits JP at this state (in this JBB)
// plus the population AT OR ABOVE Index in the SM under the digits JP:
pop1 = j__udy1LCountSM(BMPJP(findsub, jpnum), Index, Pjpm);
if (pop1 == C_JERR) return(C_JERR); // pass error up.
assert(pop1above + pop1);
return(pop1above + pop1);
} // case.
// ----------------------------------------------------------------------------
// UNCOMPRESSED BRANCH; count populations in JPs in the JBU ABOVE the next
// digit in Index, and recurse for the next digit in Index:
//
// Note: If the remainder of Index is zero, pop1above is the pop1 of the
// entire expanse and theres no point in recursing to lower levels; but this
// should be so rare that its not worth checking for;
// Judy1Count()/JudyLCount() never even calls the motor for Index == 0 (all
// bytes).
case cJU_JPBRANCH_U2: CHECKDCD(Pjp, 2); PREPB(Pjp, 2, BranchU);
case cJU_JPBRANCH_U3: CHECKDCD(Pjp, 3); PREPB(Pjp, 3, BranchU);
#ifdef JU_64BIT
case cJU_JPBRANCH_U4: CHECKDCD(Pjp, 4); PREPB(Pjp, 4, BranchU);
case cJU_JPBRANCH_U5: CHECKDCD(Pjp, 5); PREPB(Pjp, 5, BranchU);
case cJU_JPBRANCH_U6: CHECKDCD(Pjp, 6); PREPB(Pjp, 6, BranchU);
case cJU_JPBRANCH_U7: CHECKDCD(Pjp, 7); PREPB(Pjp, 7, BranchU);
#endif
case cJU_JPBRANCH_U: PREPB_ROOT(Pjp, BranchU);
// Common code (state-independent) for all cases of uncompressed branches:
BranchU:
Pjbu = P_JBU(Pjp->jp_Addr);
#ifndef NOSMARTJBU // enable to turn off smart code for comparison purposes.
// FIGURE OUT WHICH WAY CAUSES FEWER CACHE LINE FILLS; adding the JPs above
// Indexs JP, or subtracting the JPs below Indexs JP.
//
// COUNT UPWARD, subtracting the pop1 of each JP BELOW OR AT Index, from the
// whole expanses pop1:
if (digit < (cJU_BRANCHUNUMJPS / 2))
{
pop1above = pop1; // subtract JPs below Index.
#ifdef SMARTMETRICS
++jbu_upward;
#endif
for (jpnum = 0; jpnum <= digit; ++jpnum)
{
if ((Pjbu->jbu_jp[jpnum].jp_Type) <= cJU_JPNULLMAX)
continue; // shortcut, save a function call.
if ((pop1 = j__udyJPPop1(Pjbu->jbu_jp + jpnum))
== cJU_ALLONES)
{
JU_SET_ERRNO_NONNULL(Pjpm, JU_ERRNO_CORRUPT);
return(C_JERR);
}
pop1above -= pop1;
}
}
// COUNT DOWNWARD, simply adding the pop1 of each JP ABOVE Index:
else
#endif // NOSMARTJBU
{
assert(digit < cJU_BRANCHUNUMJPS);
#ifdef SMARTMETRICS
++jbu_downward;
#endif
pop1above = 0; // add JPs above Index.
for (jpnum = cJU_BRANCHUNUMJPS - 1; jpnum > digit; --jpnum)
{
if ((Pjbu->jbu_jp[jpnum].jp_Type) <= cJU_JPNULLMAX)
continue; // shortcut, save a function call.
if ((pop1 = j__udyJPPop1(Pjbu->jbu_jp + jpnum))
== cJU_ALLONES)
{
JU_SET_ERRNO_NONNULL(Pjpm, JU_ERRNO_CORRUPT);
return(C_JERR);
}
pop1above += pop1;
}
}
if ((pop1 = j__udy1LCountSM(Pjbu->jbu_jp + digit, Index, Pjpm))
== C_JERR) return(C_JERR); // pass error up.
assert(pop1above + pop1);
return(pop1above + pop1);
// ----------------------------------------------------------------------------
// LEAF COUNT MACROS:
//
// LEAF*ABOVE() are common code for different JP types (linear leaves, bitmap
// leaves, and immediates) and different leaf Index Sizes, which result in
// calling different leaf search functions. Linear leaves get the leaf address
// from jp_Addr and the Population from jp_DcdPopO, while immediates use Pjp
// itself as the leaf address and get Population from jp_Type.
#define LEAFLABOVE(Func) \
Pjll = P_JLL(Pjp->jp_Addr); \
pop1 = JU_JPLEAF_POP0(Pjp) + 1; \
LEAFABOVE(Func, Pjll, pop1)
#define LEAFB1ABOVE(Func) LEAFLABOVE(Func) // different Func, otherwise same.
#ifdef JUDY1
#define IMMABOVE(Func,Pop1) \
Pjll = (Pjll_t) Pjp; \
LEAFABOVE(Func, Pjll, Pop1)
#else
// Note: For JudyL immediates with >= 2 Indexes, the index bytes are in a
// different place than for Judy1:
#define IMMABOVE(Func,Pop1) \
LEAFABOVE(Func, (Pjll_t) (Pjp->jp_LIndex), Pop1)
#endif
// For all leaf types, the population AT OR ABOVE is the total pop1 less the
// offset of Index; and Index should always be found:
#define LEAFABOVE(Func,Pjll,Pop1) \
offset = Func(Pjll, Pop1, Index); \
assert(offset >= 0); \
assert(offset < (Pop1)); \
return((Pop1) - offset)
// IMMABOVE_01 handles the special case of an immediate JP with 1 index, which
// the search functions arent used for anyway:
//
// The target Index should be the one in this Immediate, in which case the
// count above (inclusive) is always 1.
#define IMMABOVE_01 \
assert((JU_JPDCDPOP0(Pjp)) == JU_TRIMTODCDSIZE(Index)); \
return(1)
// ----------------------------------------------------------------------------
// LINEAR LEAF; search the leaf for Index; size is computed from jp_Type:
#if (defined(JUDYL) || (! defined(JU_64BIT)))
case cJU_JPLEAF1: LEAFLABOVE(j__udySearchLeaf1);
#endif
case cJU_JPLEAF2: LEAFLABOVE(j__udySearchLeaf2);
case cJU_JPLEAF3: LEAFLABOVE(j__udySearchLeaf3);
#ifdef JU_64BIT
case cJU_JPLEAF4: LEAFLABOVE(j__udySearchLeaf4);
case cJU_JPLEAF5: LEAFLABOVE(j__udySearchLeaf5);
case cJU_JPLEAF6: LEAFLABOVE(j__udySearchLeaf6);
case cJU_JPLEAF7: LEAFLABOVE(j__udySearchLeaf7);
#endif
// ----------------------------------------------------------------------------
// BITMAP LEAF; search the leaf for Index:
//
// Since the bitmap describes Indexes digitally rather than linearly, this is
// not really a search, but just a count.
case cJU_JPLEAF_B1: LEAFB1ABOVE(j__udyCountLeafB1);
#ifdef JUDY1
// ----------------------------------------------------------------------------
// FULL POPULATION:
//
// Return the count of Indexes AT OR ABOVE Index, which is the total population
// of the expanse (a constant) less the value of the undecoded digit remaining
// in Index (its base-0 offset in the expanse), which yields an inclusive count
// above.
//
// TBD: This only supports a 1-byte full expanse. Should this extract a
// stored value for pop0 and possibly more LSBs of Index, to handle larger full
// expanses?
case cJ1_JPFULLPOPU1:
return(cJU_JPFULLPOPU1_POP0 + 1 - JU_DIGITATSTATE(Index, 1));
#endif
// ----------------------------------------------------------------------------
// IMMEDIATE:
case cJU_JPIMMED_1_01: IMMABOVE_01;
case cJU_JPIMMED_2_01: IMMABOVE_01;
case cJU_JPIMMED_3_01: IMMABOVE_01;
#ifdef JU_64BIT
case cJU_JPIMMED_4_01: IMMABOVE_01;
case cJU_JPIMMED_5_01: IMMABOVE_01;
case cJU_JPIMMED_6_01: IMMABOVE_01;
case cJU_JPIMMED_7_01: IMMABOVE_01;
#endif
case cJU_JPIMMED_1_02: IMMABOVE(j__udySearchLeaf1, 2);
case cJU_JPIMMED_1_03: IMMABOVE(j__udySearchLeaf1, 3);
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_1_04: IMMABOVE(j__udySearchLeaf1, 4);
case cJU_JPIMMED_1_05: IMMABOVE(j__udySearchLeaf1, 5);
case cJU_JPIMMED_1_06: IMMABOVE(j__udySearchLeaf1, 6);
case cJU_JPIMMED_1_07: IMMABOVE(j__udySearchLeaf1, 7);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_1_08: IMMABOVE(j__udySearchLeaf1, 8);
case cJ1_JPIMMED_1_09: IMMABOVE(j__udySearchLeaf1, 9);
case cJ1_JPIMMED_1_10: IMMABOVE(j__udySearchLeaf1, 10);
case cJ1_JPIMMED_1_11: IMMABOVE(j__udySearchLeaf1, 11);
case cJ1_JPIMMED_1_12: IMMABOVE(j__udySearchLeaf1, 12);
case cJ1_JPIMMED_1_13: IMMABOVE(j__udySearchLeaf1, 13);
case cJ1_JPIMMED_1_14: IMMABOVE(j__udySearchLeaf1, 14);
case cJ1_JPIMMED_1_15: IMMABOVE(j__udySearchLeaf1, 15);
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_2_02: IMMABOVE(j__udySearchLeaf2, 2);
case cJU_JPIMMED_2_03: IMMABOVE(j__udySearchLeaf2, 3);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_2_04: IMMABOVE(j__udySearchLeaf2, 4);
case cJ1_JPIMMED_2_05: IMMABOVE(j__udySearchLeaf2, 5);
case cJ1_JPIMMED_2_06: IMMABOVE(j__udySearchLeaf2, 6);
case cJ1_JPIMMED_2_07: IMMABOVE(j__udySearchLeaf2, 7);
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_3_02: IMMABOVE(j__udySearchLeaf3, 2);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_3_03: IMMABOVE(j__udySearchLeaf3, 3);
case cJ1_JPIMMED_3_04: IMMABOVE(j__udySearchLeaf3, 4);
case cJ1_JPIMMED_3_05: IMMABOVE(j__udySearchLeaf3, 5);
case cJ1_JPIMMED_4_02: IMMABOVE(j__udySearchLeaf4, 2);
case cJ1_JPIMMED_4_03: IMMABOVE(j__udySearchLeaf4, 3);
case cJ1_JPIMMED_5_02: IMMABOVE(j__udySearchLeaf5, 2);
case cJ1_JPIMMED_5_03: IMMABOVE(j__udySearchLeaf5, 3);
case cJ1_JPIMMED_6_02: IMMABOVE(j__udySearchLeaf6, 2);
case cJ1_JPIMMED_7_02: IMMABOVE(j__udySearchLeaf7, 2);
#endif
// ----------------------------------------------------------------------------
// OTHER CASES:
default: JU_SET_ERRNO_NONNULL(Pjpm, JU_ERRNO_CORRUPT); return(C_JERR);
} // switch on JP type
/*NOTREACHED*/
} // j__udy1LCountSM()
FUNCTION static Word_t j__udyGetMemActive(
Pjp_t Pjp) // top of subtree.
{
Word_t offset; // in a branch.
Word_t Bytes = 0; // actual bytes used at this level.
Word_t IdxSz; // bytes per index in leaves
switch (JU_JPTYPE(Pjp))
{
case cJU_JPBRANCH_L2:
case cJU_JPBRANCH_L3:
#ifdef JU_64BIT
case cJU_JPBRANCH_L4:
case cJU_JPBRANCH_L5:
case cJU_JPBRANCH_L6:
case cJU_JPBRANCH_L7:
#endif
case cJU_JPBRANCH_L:
{
Pjbl_t Pjbl = P_JBL(Pjp->jp_Addr);
for (offset = 0; offset < (Pjbl->jbl_NumJPs); ++offset)
Bytes += j__udyGetMemActive((Pjbl->jbl_jp) + offset);
return(Bytes + sizeof(jbl_t));
}
case cJU_JPBRANCH_B2:
case cJU_JPBRANCH_B3:
#ifdef JU_64BIT
case cJU_JPBRANCH_B4:
case cJU_JPBRANCH_B5:
case cJU_JPBRANCH_B6:
case cJU_JPBRANCH_B7:
#endif
case cJU_JPBRANCH_B:
{
Word_t subexp;
Word_t jpcount;
Pjbb_t Pjbb = P_JBB(Pjp->jp_Addr);
for (subexp = 0; subexp < cJU_NUMSUBEXPB; ++subexp)
{
jpcount = j__udyCountBitsB(JU_JBB_BITMAP(Pjbb, subexp));
Bytes += jpcount * sizeof(jp_t);
for (offset = 0; offset < jpcount; ++offset)
{
Bytes += j__udyGetMemActive(P_JP(JU_JBB_PJP(Pjbb, subexp))
+ offset);
}
}
return(Bytes + sizeof(jbb_t));
}
case cJU_JPBRANCH_U2:
case cJU_JPBRANCH_U3:
#ifdef JU_64BIT
case cJU_JPBRANCH_U4:
case cJU_JPBRANCH_U5:
case cJU_JPBRANCH_U6:
case cJU_JPBRANCH_U7:
#endif
case cJU_JPBRANCH_U:
{
Pjbu_t Pjbu = P_JBU(Pjp->jp_Addr);
for (offset = 0; offset < cJU_BRANCHUNUMJPS; ++offset)
{
if (((Pjbu->jbu_jp[offset].jp_Type) >= cJU_JPNULL1)
&& ((Pjbu->jbu_jp[offset].jp_Type) <= cJU_JPNULLMAX))
{
continue; // skip null JP to save time.
}
Bytes += j__udyGetMemActive(Pjbu->jbu_jp + offset);
}
return(Bytes + sizeof(jbu_t));
}
// -- Cases below here terminate and do not recurse. --
#if (defined(JUDYL) || (! defined(JU_64BIT)))
case cJU_JPLEAF1: IdxSz = 1; goto LeafWords;
#endif
case cJU_JPLEAF2: IdxSz = 2; goto LeafWords;
case cJU_JPLEAF3: IdxSz = 3; goto LeafWords;
#ifdef JU_64BIT
case cJU_JPLEAF4: IdxSz = 4; goto LeafWords;
case cJU_JPLEAF5: IdxSz = 5; goto LeafWords;
case cJU_JPLEAF6: IdxSz = 6; goto LeafWords;
case cJU_JPLEAF7: IdxSz = 7; goto LeafWords;
#endif
LeafWords:
#ifdef JUDY1
return(IdxSz * (JU_JPLEAF_POP0(Pjp) + 1));
#else
return((IdxSz + sizeof(Word_t))
* (JU_JPLEAF_POP0(Pjp) + 1));
#endif
case cJU_JPLEAF_B1:
{
#ifdef JUDY1
return(sizeof(jlb_t));
#else
Bytes = (JU_JPLEAF_POP0(Pjp) + 1) * sizeof(Word_t);
return(Bytes + sizeof(jlb_t));
#endif
}
JUDY1CODE(case cJ1_JPFULLPOPU1: return(0);)
#ifdef JUDY1
#define J__Mpy 0
#else
#define J__Mpy sizeof(Word_t)
#endif
case cJU_JPIMMED_1_01: return(0);
case cJU_JPIMMED_2_01: return(0);
case cJU_JPIMMED_3_01: return(0);
#ifdef JU_64BIT
case cJU_JPIMMED_4_01: return(0);
case cJU_JPIMMED_5_01: return(0);
case cJU_JPIMMED_6_01: return(0);
case cJU_JPIMMED_7_01: return(0);
#endif
case cJU_JPIMMED_1_02: return(J__Mpy * 2);
case cJU_JPIMMED_1_03: return(J__Mpy * 3);
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_1_04: return(J__Mpy * 4);
case cJU_JPIMMED_1_05: return(J__Mpy * 5);
case cJU_JPIMMED_1_06: return(J__Mpy * 6);
case cJU_JPIMMED_1_07: return(J__Mpy * 7);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_1_08: return(0);
case cJ1_JPIMMED_1_09: return(0);
case cJ1_JPIMMED_1_10: return(0);
case cJ1_JPIMMED_1_11: return(0);
case cJ1_JPIMMED_1_12: return(0);
case cJ1_JPIMMED_1_13: return(0);
case cJ1_JPIMMED_1_14: return(0);
case cJ1_JPIMMED_1_15: return(0);
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_2_02: return(J__Mpy * 2);
case cJU_JPIMMED_2_03: return(J__Mpy * 3);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_2_04: return(0);
case cJ1_JPIMMED_2_05: return(0);
case cJ1_JPIMMED_2_06: return(0);
case cJ1_JPIMMED_2_07: return(0);
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_3_02: return(J__Mpy * 2);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_3_03: return(0);
case cJ1_JPIMMED_3_04: return(0);
case cJ1_JPIMMED_3_05: return(0);
case cJ1_JPIMMED_4_02: return(0);
case cJ1_JPIMMED_4_03: return(0);
case cJ1_JPIMMED_5_02: return(0);
case cJ1_JPIMMED_5_03: return(0);
case cJ1_JPIMMED_6_02: return(0);
case cJ1_JPIMMED_7_02: return(0);
#endif
} // switch (JU_JPTYPE(Pjp))
FUNCTION int j__udyCreateBranchB(
Pjp_t Pjp, // Build JPs from this place
Pjp_t PJPs, // Array of JPs to put into Bitmap branch
uint8_t Exp[], // Array of expanses to put into bitmap
Word_t ExpCnt, // Number of above JPs and Expanses
Pvoid_t Pjpm)
{
Pjbb_t PjbbRaw; // pointer to bitmap branch.
Pjbb_t Pjbb;
Word_t ii, jj; // Temps
uint8_t CurrSubExp; // Current sub expanse for BM
// This assertion says the number of populated subexpanses is not too large.
// This function is only called when a BranchL overflows to a BranchB or when a
// cascade occurs, meaning a leaf overflows. Either way ExpCnt cant be very
// large, in fact a lot smaller than cJU_BRANCHBMAXJPS. (Otherwise a BranchU
// would be used.) Popping this assertion means something (unspecified) has
// gone very wrong, or else Judys design criteria have changed, although in
// fact there should be no HARM in creating a BranchB with higher actual
// fanout.
assert(ExpCnt <= cJU_BRANCHBMAXJPS);
// Get memory for a Bitmap branch
PjbbRaw = j__udyAllocJBB(Pjpm);
if (PjbbRaw == (Pjbb_t) NULL) return(-1);
Pjbb = P_JBB(PjbbRaw);
// Get 1st "sub" expanse (0..7) of bitmap branch
CurrSubExp = Exp[0] / cJU_BITSPERSUBEXPB;
// Index thru all 1 byte sized expanses:
for (jj = ii = 0; ii <= ExpCnt; ii++)
{
Word_t SubExp; // Cannot be a uint8_t
// Make sure we cover the last one
if (ii == ExpCnt)
{
SubExp = cJU_ALLONES; // Force last one
}
else
{
// Calculate the "sub" expanse of the byte expanse
SubExp = Exp[ii] / cJU_BITSPERSUBEXPB; // Bits 5..7.
// Set the bit that represents the expanse in Exp[]
JU_JBB_BITMAP(Pjbb, SubExp) |= JU_BITPOSMASKB(Exp[ii]);
}
// Check if a new "sub" expanse range needed
if (SubExp != CurrSubExp)
{
// Get number of JPs in this sub expanse
Word_t NumJP = ii - jj;
Pjp_t PjpRaw;
Pjp_t Pjp;
PjpRaw = j__udyAllocJBBJP(NumJP, Pjpm);
Pjp = P_JP(PjpRaw);
if (PjpRaw == (Pjp_t) NULL) // out of memory.
{
// Free any previous allocations:
while(CurrSubExp--)
{
NumJP = j__udyCountBitsB(JU_JBB_BITMAP(Pjbb,
CurrSubExp));
if (NumJP)
{
j__udyFreeJBBJP(JU_JBB_PJP(Pjbb,
CurrSubExp), NumJP, Pjpm);
}
}
j__udyFreeJBB(PjbbRaw, Pjpm);
return(-1);
}
// Place the array of JPs in bitmap branch:
JU_JBB_PJP(Pjbb, CurrSubExp) = PjpRaw;
// Copy the JPs to new leaf:
JU_COPYMEM(Pjp, PJPs + jj, NumJP);
// On to the next bitmap branch "sub" expanse:
jj = ii;
CurrSubExp = SubExp;
}
} // for each 1-byte expanse
// Pass back some of the JP to the new Bitmap branch:
Pjp->jp_Addr = (Word_t) PjbbRaw;
return(1);
} // j__udyCreateBranchB()
FUNCTION int Judy1PrevEmpty
#else
FUNCTION int Judy1NextEmpty
#endif
#else
#ifdef JUDYPREV
FUNCTION int JudyLPrevEmpty
#else
FUNCTION int JudyLNextEmpty
#endif
#endif
(
Pcvoid_t PArray, // Judy array to search.
Word_t * PIndex, // starting point and result.
PJError_t PJError // optional, for returning error info.
)
{
Word_t Index; // fast copy, in a register.
Pjp_t Pjp; // current JP.
Pjbl_t Pjbl; // Pjp->jp_Addr masked and cast to types:
Pjbb_t Pjbb;
Pjbu_t Pjbu;
Pjlb_t Pjlb;
PWord_t Pword; // alternate name for use by GET* macros.
Word_t digit; // next digit to decode from Index.
Word_t digits; // current state in SM = digits left to decode.
Word_t pop0; // in a leaf.
Word_t pop0mask; // precalculated to avoid variable shifts.
long offset; // within a branch or leaf (can be large).
int subexp; // subexpanse in a bitmap branch.
BITMAPB_t bitposmaskB; // bit in bitmap for bitmap branch.
BITMAPL_t bitposmaskL; // bit in bitmap for bitmap leaf.
Word_t possfullJP1; // JP types for possibly full subexpanses:
Word_t possfullJP2;
Word_t possfullJP3;
// ----------------------------------------------------------------------------
// M A C R O S
//
// These are intended to make the code a bit more readable and less redundant.
// CHECK FOR NULL JP:
//
// TBD: In principle this can be reduced (here and in other *.c files) to just
// the latter clause since no Type should ever be below cJU_JPNULL1, but in
// fact some root pointer types can be lower, so for safety do both checks.
#define JPNULL(Type) (((Type) >= cJU_JPNULL1) && ((Type) <= cJU_JPNULLMAX))
// CHECK FOR A FULL JP:
//
// Given a JP, indicate if it is fully populated. Use digits, pop0mask, and
// possfullJP1..3 in the context.
//
// This is a difficult problem because it requires checking the Pop0 bits for
// all-ones, but the number of bytes depends on the JP type, which is not
// directly related to the parent branchs type or level -- the JPs child
// could be under a narrow pointer (hence not full). The simple answer
// requires switching on or otherwise calculating the JP type, which could be
// slow. Instead, in SMPREPB* precalculate pop0mask and also record in
// possfullJP1..3 the child JP (branch) types that could possibly be full (one
// level down), and use them here. For level-2 branches (with digits == 2),
// the test for a full child depends on Judy1/JudyL.
//
// Note: This cannot be applied to the JP in a JPM because it doesnt have
// enough pop0 digits.
//
// TBD: JPFULL_BRANCH diligently checks for BranchL or BranchB, where neither
// of those can ever be full as it turns out. Could just check for a BranchU
// at the right level. Also, pop0mask might be overkill, its not used much,
// so perhaps just call cJU_POP0MASK(digits - 1) here?
//
// First, JPFULL_BRANCH checks for a full expanse for a JP whose child can be a
// branch, that is, a JP in a branch at level 3 or higher:
#define JPFULL_BRANCH(Pjp) \
((((JU_JPDCDPOP0(Pjp) ^ cJU_ALLONES) & pop0mask) == 0) \
&& ((JU_JPTYPE(Pjp) == possfullJP1) \
|| (JU_JPTYPE(Pjp) == possfullJP2) \
|| (JU_JPTYPE(Pjp) == possfullJP3)))
#ifdef JUDY1
#define JPFULL(Pjp) \
((digits == 2) ? \
(JU_JPTYPE(Pjp) == cJ1_JPFULLPOPU1) : JPFULL_BRANCH(Pjp))
#else
#define JPFULL(Pjp) \
((digits == 2) ? \
(JU_JPTYPE(Pjp) == cJU_JPLEAF_B1) \
&& (((JU_JPDCDPOP0(Pjp) & cJU_POP0MASK(1)) == cJU_POP0MASK(1))) : \
JPFULL_BRANCH(Pjp))
#endif
// RETURN SUCCESS:
//
// This hides the need to set *PIndex back to the local value of Index -- use a
// local value for faster operation. Note that the callers *PIndex is ALWAYS
// modified upon success, at least decremented/incremented.
#define RET_SUCCESS { *PIndex = Index; return(1); }
// RETURN A CORRUPTION:
#define RET_CORRUPT { JU_SET_ERRNO(PJError, JU_ERRNO_CORRUPT); return(JERRI); }
// SEARCH A BITMAP BRANCH:
//
// This is a weak analog of j__udySearchLeaf*() for bitmap branches. Return
// the actual or next-left position, base 0, of Digit in a BITMAPB_t bitmap
// (subexpanse of a full bitmap), also given a Bitposmask for Digit. The
// position is the offset within the set bits.
//
// Unlike j__udySearchLeaf*(), the offset is not returned bit-complemented if
// Digits bit is unset, because the caller can check the bitmap themselves to
// determine that. Also, if Digits bit is unset, the returned offset is to
// the next-left JP or index (including -1), not to the "ideal" position for
// the index = next-right JP or index.
//
// Shortcut and skip calling j__udyCountBitsB() if the bitmap is full, in which
// case (Digit % cJU_BITSPERSUBEXPB) itself is the base-0 offset.
#define SEARCHBITMAPB(Bitmap,Digit,Bitposmask) \
(((Bitmap) == cJU_FULLBITMAPB) ? (Digit % cJU_BITSPERSUBEXPB) : \
j__udyCountBitsB((Bitmap) & JU_MASKLOWERINC(Bitposmask)) - 1)
#ifdef JUDYPREV
// Equivalent to search for the highest offset in Bitmap, that is, one less
// than the number of bits set:
#define SEARCHBITMAPMAXB(Bitmap) \
(((Bitmap) == cJU_FULLBITMAPB) ? cJU_BITSPERSUBEXPB - 1 : \
j__udyCountBitsB(Bitmap) - 1)
#endif
// CHECK DECODE BYTES:
//
// Check Decode bytes in a JP against the equivalent portion of Index. If they
// dont match, Index is outside the subexpanse of a narrow pointer, hence is
// empty.
#define CHECKDCD(cDigits) \
if (JU_DCDNOTMATCHINDEX(Index, Pjp, cDigits)) RET_SUCCESS
// REVISE REMAINDER OF INDEX:
//
// Put one digit in place in Index and clear/set the lower digits, if any, so
// the resulting Index is at the start/end of an expanse, or just clear/set the
// least digits.
//
// Actually, to make simple use of JU_LEASTBYTESMASK, first clear/set all least
// digits of Index including the digit to be overridden, then set the value of
// that one digit. If Digits == 1 the first operation is redundant, but either
// very fast or even removed by the optimizer.
#define CLEARLEASTDIGITS(Digits) Index &= ~JU_LEASTBYTESMASK(Digits)
#define SETLEASTDIGITS( Digits) Index |= JU_LEASTBYTESMASK(Digits)
#define CLEARLEASTDIGITS_D(Digit,Digits) \
{ \
CLEARLEASTDIGITS(Digits); \
JU_SETDIGIT(Index, Digit, Digits); \
}
#define SETLEASTDIGITS_D(Digit,Digits) \
{ \
SETLEASTDIGITS(Digits); \
JU_SETDIGIT(Index, Digit, Digits); \
}
// SET REMAINDER OF INDEX AND THEN RETURN OR CONTINUE:
#define SET_AND_RETURN(OpLeastDigits,Digit,Digits) \
{ \
OpLeastDigits(Digit, Digits); \
RET_SUCCESS; \
}
#define SET_AND_CONTINUE(OpLeastDigits,Digit,Digits) \
{ \
OpLeastDigits(Digit, Digits); \
goto SMGetContinue; \
}
// PREPARE TO HANDLE A LEAFW OR JP BRANCH IN THE STATE MACHINE:
//
// Extract a state-dependent digit from Index in a "constant" way, then jump to
// common code for multiple cases.
//
// TBD: Should this macro do more, such as preparing variable-shift masks for
// use in CLEARLEASTDIGITS and SETLEASTDIGITS?
#define SMPREPB(cDigits,Next,PossFullJP1,PossFullJP2,PossFullJP3) \
digits = (cDigits); \
digit = JU_DIGITATSTATE(Index, cDigits); \
pop0mask = cJU_POP0MASK((cDigits) - 1); /* for branchs JPs */ \
possfullJP1 = (PossFullJP1); \
possfullJP2 = (PossFullJP2); \
possfullJP3 = (PossFullJP3); \
goto Next
// Variations for specific-level branches and for shorthands:
//
// Note: SMPREPB2 need not initialize possfullJP* because JPFULL does not use
// them for digits == 2, but gcc -Wall isnt quite smart enough to see this, so
// waste a bit of time and space to get rid of the warning:
#define SMPREPB2(Next) \
digits = 2; \
digit = JU_DIGITATSTATE(Index, 2); \
pop0mask = cJU_POP0MASK(1); /* for branchs JPs */ \
possfullJP1 = possfullJP2 = possfullJP3 = 0; \
goto Next
#define SMPREPB3(Next) SMPREPB(3, Next, cJU_JPBRANCH_L2, \
cJU_JPBRANCH_B2, \
cJU_JPBRANCH_U2)
#ifndef JU_64BIT
#define SMPREPBL(Next) SMPREPB(cJU_ROOTSTATE, Next, cJU_JPBRANCH_L3, \
cJU_JPBRANCH_B3, \
cJU_JPBRANCH_U3)
#else
#define SMPREPB4(Next) SMPREPB(4, Next, cJU_JPBRANCH_L3, \
cJU_JPBRANCH_B3, \
cJU_JPBRANCH_U3)
#define SMPREPB5(Next) SMPREPB(5, Next, cJU_JPBRANCH_L4, \
cJU_JPBRANCH_B4, \
cJU_JPBRANCH_U4)
#define SMPREPB6(Next) SMPREPB(6, Next, cJU_JPBRANCH_L5, \
cJU_JPBRANCH_B5, \
cJU_JPBRANCH_U5)
#define SMPREPB7(Next) SMPREPB(7, Next, cJU_JPBRANCH_L6, \
cJU_JPBRANCH_B6, \
cJU_JPBRANCH_U6)
#define SMPREPBL(Next) SMPREPB(cJU_ROOTSTATE, Next, cJU_JPBRANCH_L7, \
cJU_JPBRANCH_B7, \
cJU_JPBRANCH_U7)
#endif
// RESTART AFTER SECONDARY DEAD END:
//
// Set Index to the first/last index in the branch or leaf subexpanse and start
// over at the top of the tree.
#ifdef JUDYPREV
#define SMRESTART(Digits) { CLEARLEASTDIGITS(Digits); goto SMGetRestart; }
#else
#define SMRESTART(Digits) { SETLEASTDIGITS( Digits); goto SMGetRestart; }
#endif
// CHECK EDGE OF LEAFS EXPANSE:
//
// Given the LSBs of the lowest/highest valid index in a leaf (or equivalently
// in an immediate JP), the level (index size) of the leaf, and the full index
// to return (as Index in the context) already set to the full index matching
// the lowest/highest one, determine if there is an empty index in the leafs
// expanse below/above the lowest/highest index, which is true if the
// lowest/highest index is not at the "edge" of the leafs expanse based on its
// LSBs. If so, return Index decremented/incremented; otherwise restart at the
// top of the tree.
//
// Note: In many cases Index is already at the right spot and calling
// SMRESTART instead of just going directly to SMGetRestart is a bit of
// overkill.
//
// Note: Variable shift occurs if Digits is not a constant.
#ifdef JUDYPREV
#define LEAF_EDGE(MinIndex,Digits) \
{ \
if (MinIndex) { --Index; RET_SUCCESS; } \
SMRESTART(Digits); \
}
#else
#define LEAF_EDGE(MaxIndex,Digits) \
{ \
if ((MaxIndex) != JU_LEASTBYTES(cJU_ALLONES, Digits)) \
{ ++Index; RET_SUCCESS; } \
SMRESTART(Digits); \
}
#endif
// Same as above except Index is not already set to match the lowest/highest
// index, so do that before decrementing/incrementing it:
#ifdef JUDYPREV
#define LEAF_EDGE_SET(MinIndex,Digits) \
{ \
if (MinIndex) \
{ JU_SETDIGITS(Index, MinIndex, Digits); --Index; RET_SUCCESS; } \
SMRESTART(Digits); \
}
#else
#define LEAF_EDGE_SET(MaxIndex,Digits) \
{ \
if ((MaxIndex) != JU_LEASTBYTES(cJU_ALLONES, Digits)) \
{ JU_SETDIGITS(Index, MaxIndex, Digits); ++Index; RET_SUCCESS; } \
SMRESTART(Digits); \
}
#endif
// FIND A HOLE (EMPTY INDEX) IN AN IMMEDIATE OR LEAF:
//
// Given an index location in a leaf (or equivalently an immediate JP) known to
// contain a usable hole (an empty index less/greater than Index), and the LSBs
// of a minimum/maximum index to locate, find the previous/next empty index and
// return it.
//
// Note: "Even" index sizes (1,2,4[,8] bytes) have corresponding native C
// types; "odd" index sizes dont, but they are not represented here because
// they are handled completely differently; see elsewhere.
#ifdef JUDYPREV
#define LEAF_HOLE_EVEN(cDigits,Pjll,IndexLSB) \
{ \
while (*(Pjll) > (IndexLSB)) --(Pjll); /* too high */ \
if (*(Pjll) < (IndexLSB)) RET_SUCCESS /* Index is empty */ \
while (*(--(Pjll)) == --(IndexLSB)) /* null, find a hole */;\
JU_SETDIGITS(Index, IndexLSB, cDigits); \
RET_SUCCESS; \
}
#else
#define LEAF_HOLE_EVEN(cDigits,Pjll,IndexLSB) \
{ \
while (*(Pjll) < (IndexLSB)) ++(Pjll); /* too low */ \
if (*(Pjll) > (IndexLSB)) RET_SUCCESS /* Index is empty */ \
while (*(++(Pjll)) == ++(IndexLSB)) /* null, find a hole */;\
JU_SETDIGITS(Index, IndexLSB, cDigits); \
RET_SUCCESS; \
}
#endif
// SEARCH FOR AN EMPTY INDEX IN AN IMMEDIATE OR LEAF:
//
// Given a pointer to the first index in a leaf (or equivalently an immediate
// JP), the population of the leaf, and a first empty Index to find (inclusive,
// as Index in the context), where Index is known to fall within the expanse of
// the leaf to search, efficiently find the previous/next empty index in the
// leaf, if any. For simplicity the following overview is stated in terms of
// Judy*NextEmpty() only, but the same concepts apply symmetrically for
// Judy*PrevEmpty(). Also, in each case the comparisons are for the LSBs of
// Index and leaf indexes, according to the leafs level.
//
// 1. If Index is GREATER than the last (highest) index in the leaf
// (maxindex), return success, Index is empty. (Remember, Index is known
// to be in the leafs expanse.)
//
// 2. If Index is EQUAL to maxindex: If maxindex is not at the edge of the
// leafs expanse, increment Index and return success, there is an empty
// Index one higher than any in the leaf; otherwise restart with Index
// reset to the upper edge of the leafs expanse. Note: This might cause
// an extra cache line fill, but this is OK for repeatedly-called search
// code, and it saves CPU time.
//
// 3. If Index is LESS than maxindex, check for "dense to end of leaf":
// Subtract Index from maxindex, and back up that many slots in the leaf.
// If the resulting offset is not before the start of the leaf then compare
// the index at this offset (baseindex) with Index:
//
// 3a. If GREATER, the leaf must be corrupt, since indexes are sorted and
// there are no duplicates.
//
// 3b. If EQUAL, the leaf is "dense" from Index to maxindex, meaning there is
// no reason to search it. "Slide right" to the high end of the leaf
// (modify Index to maxindex) and continue with step 2 above.
//
// 3c. If LESS, continue with step 4.
//
// 4. If the offset based on maxindex minus Index falls BEFORE the start of
// the leaf, or if, per 3c above, baseindex is LESS than Index, the leaf is
// guaranteed "not dense to the end" and a usable empty Index must exist.
// This supports a more efficient search loop. Start at the FIRST index in
// the leaf, or one BEYOND baseindex, respectively, and search the leaf as
// follows, comparing each current index (currindex) with Index:
//
// 4a. If LESS, keep going to next index. Note: This is certain to terminate
// because maxindex is known to be greater than Index, hence the loop can
// be small and fast.
//
// 4b. If EQUAL, loop and increment Index until finding currindex greater than
// Index, and return success with the modified Index.
//
// 4c. If GREATER, return success, Index (unmodified) is empty.
//
// Note: These are macros rather than functions for speed.
#ifdef JUDYPREV
#define JSLE_EVEN(Addr,Pop0,cDigits,LeafType) \
{ \
LeafType * PjllLSB = (LeafType *) (Addr); \
LeafType IndexLSB = Index; /* auto-masking */ \
\
/* Index before or at start of leaf: */ \
\
if (*PjllLSB >= IndexLSB) /* no need to search */ \
{ \
if (*PjllLSB > IndexLSB) RET_SUCCESS; /* Index empty */ \
LEAF_EDGE(*PjllLSB, cDigits); \
} \
\
/* Index in or after leaf: */ \
\
offset = IndexLSB - *PjllLSB; /* tentative offset */ \
if (offset <= (Pop0)) /* can check density */ \
{ \
PjllLSB += offset; /* move to slot */ \
\
if (*PjllLSB <= IndexLSB) /* dense or corrupt */ \
{ \
if (*PjllLSB == IndexLSB) /* dense, check edge */ \
LEAF_EDGE_SET(PjllLSB[-offset], cDigits); \
RET_CORRUPT; \
} \
--PjllLSB; /* not dense, start at previous */ \
} \
else PjllLSB = ((LeafType *) (Addr)) + (Pop0); /* start at max */ \
\
LEAF_HOLE_EVEN(cDigits, PjllLSB, IndexLSB); \
}
// JSLE_ODD is completely different from JSLE_EVEN because its important to
// minimize copying odd indexes to compare them (see 4.14). Furthermore, a
// very complex version (4.17, but abandoned before fully debugged) that
// avoided calling j__udySearchLeaf*() ran twice as fast as 4.14, but still
// half as fast as SearchValid. Doug suggested that to minimize complexity and
// share common code we should use j__udySearchLeaf*() for the initial search
// to establish if Index is empty, which should be common. If Index is valid
// in a leaf or immediate indexes, odds are good that an empty Index is nearby,
// so for simplicity just use a *COPY* function to linearly search the
// remainder.
//
// TBD: Pathological case? Average performance should be good, but worst-case
// might suffer. When Search says the initial Index is valid, so a linear
// copy-and-compare is begun, if the caller builds fairly large leaves with
// dense clusters AND frequently does a SearchEmpty at one end of such a
// cluster, performance wont be very good. Might a dense-check help? This
// means checking offset against the index at offset, and then against the
// first/last index in the leaf. We doubt the pathological case will appear
// much in real applications because they will probably alternate SearchValid
// and SearchEmpty calls.
#define JSLE_ODD(cDigits,Pjll,Pop0,Search,Copy) \
{ \
Word_t IndexLSB; /* least bytes only */ \
Word_t IndexFound; /* in leaf */ \
\
if ((offset = Search(Pjll, (Pop0) + 1, Index)) < 0) \
RET_SUCCESS; /* Index is empty */ \
\
IndexLSB = JU_LEASTBYTES(Index, cDigits); \
offset *= (cDigits); \
\
while ((offset -= (cDigits)) >= 0) \
{ /* skip until empty or start */ \
Copy(IndexFound, ((uint8_t *) (Pjll)) + offset); \
if (IndexFound != (--IndexLSB)) /* found an empty */ \
{ JU_SETDIGITS(Index, IndexLSB, cDigits); RET_SUCCESS; }\
} \
LEAF_EDGE_SET(IndexLSB, cDigits); \
}
#else // JUDYNEXT
#define JSLE_EVEN(Addr,Pop0,cDigits,LeafType) \
{ \
LeafType * PjllLSB = ((LeafType *) (Addr)) + (Pop0); \
LeafType IndexLSB = Index; /* auto-masking */ \
\
/* Index at or after end of leaf: */ \
\
if (*PjllLSB <= IndexLSB) /* no need to search */ \
{ \
if (*PjllLSB < IndexLSB) RET_SUCCESS; /* Index empty */\
LEAF_EDGE(*PjllLSB, cDigits); \
} \
\
/* Index before or in leaf: */ \
\
offset = *PjllLSB - IndexLSB; /* tentative offset */ \
if (offset <= (Pop0)) /* can check density */ \
{ \
PjllLSB -= offset; /* move to slot */ \
\
if (*PjllLSB >= IndexLSB) /* dense or corrupt */ \
{ \
if (*PjllLSB == IndexLSB) /* dense, check edge */ \
LEAF_EDGE_SET(PjllLSB[offset], cDigits); \
RET_CORRUPT; \
} \
++PjllLSB; /* not dense, start at next */ \
} \
else PjllLSB = (LeafType *) (Addr); /* start at minimum */ \
\
LEAF_HOLE_EVEN(cDigits, PjllLSB, IndexLSB); \
}
#define JSLE_ODD(cDigits,Pjll,Pop0,Search,Copy) \
{ \
Word_t IndexLSB; /* least bytes only */ \
Word_t IndexFound; /* in leaf */ \
int offsetmax; /* in bytes */ \
\
if ((offset = Search(Pjll, (Pop0) + 1, Index)) < 0) \
RET_SUCCESS; /* Index is empty */ \
\
IndexLSB = JU_LEASTBYTES(Index, cDigits); \
offset *= (cDigits); \
offsetmax = (Pop0) * (cDigits); /* single multiply */ \
\
while ((offset += (cDigits)) <= offsetmax) \
{ /* skip until empty or end */ \
Copy(IndexFound, ((uint8_t *) (Pjll)) + offset); \
if (IndexFound != (++IndexLSB)) /* found an empty */ \
{ JU_SETDIGITS(Index, IndexLSB, cDigits); RET_SUCCESS; } \
} \
LEAF_EDGE_SET(IndexLSB, cDigits); \
}
#endif // JUDYNEXT
// Note: Immediate indexes never fill a single index group, so for odd index
// sizes, save time by calling JSLE_ODD_IMM instead of JSLE_ODD.
#define j__udySearchLeafEmpty1(Addr,Pop0) \
JSLE_EVEN(Addr, Pop0, 1, uint8_t)
#define j__udySearchLeafEmpty2(Addr,Pop0) \
JSLE_EVEN(Addr, Pop0, 2, uint16_t)
#define j__udySearchLeafEmpty3(Addr,Pop0) \
JSLE_ODD(3, Addr, Pop0, j__udySearchLeaf3, JU_COPY3_PINDEX_TO_LONG)
#ifndef JU_64BIT
#define j__udySearchLeafEmptyL(Addr,Pop0) \
JSLE_EVEN(Addr, Pop0, 4, Word_t)
#else
#define j__udySearchLeafEmpty4(Addr,Pop0) \
JSLE_EVEN(Addr, Pop0, 4, uint32_t)
#define j__udySearchLeafEmpty5(Addr,Pop0) \
JSLE_ODD(5, Addr, Pop0, j__udySearchLeaf5, JU_COPY5_PINDEX_TO_LONG)
#define j__udySearchLeafEmpty6(Addr,Pop0) \
JSLE_ODD(6, Addr, Pop0, j__udySearchLeaf6, JU_COPY6_PINDEX_TO_LONG)
#define j__udySearchLeafEmpty7(Addr,Pop0) \
JSLE_ODD(7, Addr, Pop0, j__udySearchLeaf7, JU_COPY7_PINDEX_TO_LONG)
#define j__udySearchLeafEmptyL(Addr,Pop0) \
JSLE_EVEN(Addr, Pop0, 8, Word_t)
#endif // JU_64BIT
// ----------------------------------------------------------------------------
// START OF CODE:
//
// CHECK FOR SHORTCUTS:
//
// Error out if PIndex is null.
if (PIndex == (PWord_t) NULL)
{
JU_SET_ERRNO(PJError, JU_ERRNO_NULLPINDEX);
return(JERRI);
}
Index = *PIndex; // fast local copy.
// Set and pre-decrement/increment Index, watching for underflow/overflow:
//
// An out-of-bounds Index means failure: No previous/next empty index.
SMGetRestart: // return here with revised Index.
#ifdef JUDYPREV
if (Index-- == 0) return(0);
#else
if (++Index == 0) return(0);
#endif
// An empty array with an in-bounds (not underflowed/overflowed) Index means
// success:
//
// Note: This check is redundant after restarting at SMGetRestart, but should
// take insignificant time.
if (PArray == (Pvoid_t) NULL) RET_SUCCESS;
// ----------------------------------------------------------------------------
// ROOT-LEVEL LEAF that starts with a Pop0 word; just look within the leaf:
//
// If Index is not in the leaf, return success; otherwise return the first
// empty Index, if any, below/above where it would belong.
if (JU_LEAFW_POP0(PArray) < cJU_LEAFW_MAXPOP1) // must be a LEAFW
{
Pjlw_t Pjlw = P_JLW(PArray); // first word of leaf.
pop0 = Pjlw[0];
#ifdef JUDY1
if (pop0 == 0) // special case.
{
#ifdef JUDYPREV
if ((Index != Pjlw[1]) || (Index-- != 0)) RET_SUCCESS;
#else
if ((Index != Pjlw[1]) || (++Index != 0)) RET_SUCCESS;
#endif
return(0); // no previous/next empty index.
}
#endif // JUDY1
j__udySearchLeafEmptyL(Pjlw + 1, pop0);
// No return -- thanks ALAN
}
else
// ----------------------------------------------------------------------------
// HANDLE JRP Branch:
//
// For JRP branches, traverse the JPM; handle LEAFW
// directly; but look for the most common cases first.
{
Pjpm_t Pjpm = P_JPM(PArray);
Pjp = &(Pjpm->jpm_JP);
// goto SMGetContinue;
}
// ============================================================================
// STATE MACHINE -- GET INDEX:
//
// Search for Index (already decremented/incremented so as to be an inclusive
// search). If not found (empty index), return success. Otherwise do a
// previous/next search, and if successful modify Index to the empty index
// found. See function header comments.
//
// ENTRY: Pjp points to next JP to interpret, whose Decode bytes have not yet
// been checked.
//
// Note: Check Decode bytes at the start of each loop, not after looking up a
// new JP, so its easy to do constant shifts/masks.
//
// EXIT: Return, or branch to SMGetRestart with modified Index, or branch to
// SMGetContinue with a modified Pjp, as described elsewhere.
//
// WARNING: For run-time efficiency the following cases replicate code with
// varying constants, rather than using common code with variable values!
SMGetContinue: // return here for next branch/leaf.
#ifdef TRACEJPSE
JudyPrintJP(Pjp, "sf", __LINE__);
#endif
switch (JU_JPTYPE(Pjp))
{
// ----------------------------------------------------------------------------
// LINEAR BRANCH:
//
// Check Decode bytes, if any, in the current JP, then search for a JP for the
// next digit in Index.
case cJU_JPBRANCH_L2:
CHECKDCD(2);
SMPREPB2(SMBranchL);
case cJU_JPBRANCH_L3:
CHECKDCD(3);
SMPREPB3(SMBranchL);
#ifdef JU_64BIT
case cJU_JPBRANCH_L4:
CHECKDCD(4);
SMPREPB4(SMBranchL);
case cJU_JPBRANCH_L5:
CHECKDCD(5);
SMPREPB5(SMBranchL);
case cJU_JPBRANCH_L6:
CHECKDCD(6);
SMPREPB6(SMBranchL);
case cJU_JPBRANCH_L7:
CHECKDCD(7);
SMPREPB7(SMBranchL);
#endif
case cJU_JPBRANCH_L:
SMPREPBL(SMBranchL);
// Common code (state-independent) for all cases of linear branches:
SMBranchL:
Pjbl = P_JBL(Pjp->jp_Addr);
// First, check if Indexs expanse (digit) is below/above the first/last
// populated expanse in the BranchL, in which case Index is empty; otherwise
// find the offset of the lowest/highest populated expanse at or above/below
// digit, if any:
//
// Note: The for-loop is guaranteed to exit eventually because the first/last
// expanse is known to be a terminator.
//
// Note: Cannot use j__udySearchLeaf*Empty1() here because it only applies to
// leaves and does not know about partial versus full JPs, unlike the use of
// j__udySearchLeaf1() for BranchLs in SearchValid code. Also, since linear
// leaf expanse lists are small, dont waste time calling j__udySearchLeaf1(),
// just scan the expanse list.
#ifdef JUDYPREV
if ((Pjbl->jbl_Expanse[0]) > digit) RET_SUCCESS;
for (offset = (Pjbl->jbl_NumJPs) - 1; /* null */; --offset)
#else
if ((Pjbl->jbl_Expanse[(Pjbl->jbl_NumJPs) - 1]) < digit)
RET_SUCCESS;
for (offset = 0; /* null */; ++offset)
#endif
{
// Too low/high, keep going; or too high/low, meaning the loop passed a hole
// and the initial Index is empty:
#ifdef JUDYPREV
if ((Pjbl->jbl_Expanse[offset]) > digit) continue;
if ((Pjbl->jbl_Expanse[offset]) < digit) RET_SUCCESS;
#else
if ((Pjbl->jbl_Expanse[offset]) < digit) continue;
if ((Pjbl->jbl_Expanse[offset]) > digit) RET_SUCCESS;
#endif
// Found expanse matching digit; if its not full, traverse through it:
if (! JPFULL((Pjbl->jbl_jp) + offset))
{
Pjp = (Pjbl->jbl_jp) + offset;
goto SMGetContinue;
}
// Common code: While searching for a lower/higher hole or a non-full JP, upon
// finding a lower/higher hole, adjust Index using the revised digit and
// return; or upon finding a consecutive lower/higher expanse, if the expanses
// JP is non-full, modify Index and traverse through the JP:
#define BRANCHL_CHECK(OpIncDec,OpLeastDigits,Digit,Digits) \
{ \
if ((Pjbl->jbl_Expanse[offset]) != OpIncDec digit) \
SET_AND_RETURN(OpLeastDigits, Digit, Digits); \
\
if (! JPFULL((Pjbl->jbl_jp) + offset)) \
{ \
Pjp = (Pjbl->jbl_jp) + offset; \
SET_AND_CONTINUE(OpLeastDigits, Digit, Digits); \
} \
}
// BranchL primary dead end: Expanse matching Index/digit is full (rare except
// for dense/sequential indexes):
//
// Search for a lower/higher hole, a non-full JP, or the end of the expanse
// list, while decrementing/incrementing digit.
#ifdef JUDYPREV
while (--offset >= 0)
BRANCHL_CHECK(--, SETLEASTDIGITS_D, digit, digits)
#else
while (++offset < Pjbl->jbl_NumJPs)
BRANCHL_CHECK(++, CLEARLEASTDIGITS_D, digit, digits)
#endif
// Passed end of BranchL expanse list after finding a matching but full
// expanse:
//
// Digit now matches the lowest/highest expanse, which is a full expanse; if
// digit is at the end of BranchLs expanse (no hole before/after), break out
// of the loop; otherwise modify Index to the next lower/higher digit and
// return success:
#ifdef JUDYPREV
if (digit == 0) break;
--digit;
SET_AND_RETURN(SETLEASTDIGITS_D, digit, digits);
#else
if (digit == JU_LEASTBYTES(cJU_ALLONES, 1)) break;
++digit;
SET_AND_RETURN(CLEARLEASTDIGITS_D, digit, digits);
#endif
} // for-loop
// BranchL secondary dead end, no non-full previous/next JP:
SMRESTART(digits);
// ----------------------------------------------------------------------------
// BITMAP BRANCH:
//
// Check Decode bytes, if any, in the current JP, then search for a JP for the
// next digit in Index.
case cJU_JPBRANCH_B2:
CHECKDCD(2);
SMPREPB2(SMBranchB);
case cJU_JPBRANCH_B3:
CHECKDCD(3);
SMPREPB3(SMBranchB);
#ifdef JU_64BIT
case cJU_JPBRANCH_B4:
CHECKDCD(4);
SMPREPB4(SMBranchB);
case cJU_JPBRANCH_B5:
CHECKDCD(5);
SMPREPB5(SMBranchB);
case cJU_JPBRANCH_B6:
CHECKDCD(6);
SMPREPB6(SMBranchB);
case cJU_JPBRANCH_B7:
CHECKDCD(7);
SMPREPB7(SMBranchB);
#endif
case cJU_JPBRANCH_B:
SMPREPBL(SMBranchB);
// Common code (state-independent) for all cases of bitmap branches:
SMBranchB:
Pjbb = P_JBB(Pjp->jp_Addr);
// Locate the digits JP in the subexpanse list, if present:
subexp = digit / cJU_BITSPERSUBEXPB;
assert(subexp < cJU_NUMSUBEXPB); // falls in expected range.
bitposmaskB = JU_BITPOSMASKB(digit);
// Absent JP = no JP matches current digit in Index:
// if (! JU_BITMAPTESTB(Pjbb, digit)) // slower.
if (! (JU_JBB_BITMAP(Pjbb, subexp) & bitposmaskB)) // faster.
RET_SUCCESS;
// Non-full JP matches current digit in Index:
//
// Iterate to the subsidiary non-full JP.
offset = SEARCHBITMAPB(JU_JBB_BITMAP(Pjbb, subexp), digit,
bitposmaskB);
// not negative since at least one bit is set:
assert(offset >= 0);
assert(offset < (int) cJU_BITSPERSUBEXPB);
// Watch for null JP subarray pointer with non-null bitmap (a corruption):
if ((Pjp = P_JP(JU_JBB_PJP(Pjbb, subexp)))
== (Pjp_t) NULL) RET_CORRUPT;
Pjp += offset;
if (! JPFULL(Pjp)) goto SMGetContinue;
// BranchB primary dead end:
//
// Upon hitting a full JP in a BranchB for the next digit in Index, search
// sideways for a previous/next absent JP (unset bit) or non-full JP (set bit
// with non-full JP); first in the current bitmap subexpanse, then in
// lower/higher subexpanses. Upon entry, Pjp points to a known-unusable JP,
// ready to decrement/increment.
//
// Note: The preceding code is separate from this loop because Index does not
// need revising (see SET_AND_*()) if the initial index is an empty index.
//
// TBD: For speed, shift bitposmaskB instead of using JU_BITMAPTESTB or
// JU_BITPOSMASKB, but this shift has knowledge of bit order that really should
// be encapsulated in a header file.
#define BRANCHB_CHECKBIT(OpLeastDigits) \
if (! (JU_JBB_BITMAP(Pjbb, subexp) & bitposmaskB)) /* absent JP */ \
SET_AND_RETURN(OpLeastDigits, digit, digits)
#define BRANCHB_CHECKJPFULL(OpLeastDigits) \
if (! JPFULL(Pjp)) \
SET_AND_CONTINUE(OpLeastDigits, digit, digits)
#define BRANCHB_STARTSUBEXP(OpLeastDigits) \
if (! JU_JBB_BITMAP(Pjbb, subexp)) /* empty subexpanse, shortcut */ \
SET_AND_RETURN(OpLeastDigits, digit, digits) \
if ((Pjp = P_JP(JU_JBB_PJP(Pjbb, subexp))) == (Pjp_t) NULL) RET_CORRUPT
#ifdef JUDYPREV
--digit; // skip initial digit.
bitposmaskB >>= 1; // see TBD above.
BranchBNextSubexp: // return here to check next bitmap subexpanse.
while (bitposmaskB) // more bits to check in subexp.
{
BRANCHB_CHECKBIT(SETLEASTDIGITS_D);
--Pjp; // previous in subarray.
BRANCHB_CHECKJPFULL(SETLEASTDIGITS_D);
assert(digit >= 0);
--digit;
bitposmaskB >>= 1;
}
if (subexp-- > 0) // more subexpanses.
{
BRANCHB_STARTSUBEXP(SETLEASTDIGITS_D);
Pjp += SEARCHBITMAPMAXB(JU_JBB_BITMAP(Pjbb, subexp)) + 1;
bitposmaskB = (1U << (cJU_BITSPERSUBEXPB - 1));
goto BranchBNextSubexp;
}
#else // JUDYNEXT
++digit; // skip initial digit.
bitposmaskB <<= 1; // note: BITMAPB_t.
BranchBNextSubexp: // return here to check next bitmap subexpanse.
while (bitposmaskB) // more bits to check in subexp.
{
BRANCHB_CHECKBIT(CLEARLEASTDIGITS_D);
++Pjp; // previous in subarray.
BRANCHB_CHECKJPFULL(CLEARLEASTDIGITS_D);
assert(digit < cJU_SUBEXPPERSTATE);
++digit;
bitposmaskB <<= 1; // note: BITMAPB_t.
}
if (++subexp < cJU_NUMSUBEXPB) // more subexpanses.
{
BRANCHB_STARTSUBEXP(CLEARLEASTDIGITS_D);
--Pjp; // pre-decrement.
bitposmaskB = 1;
goto BranchBNextSubexp;
}
#endif // JUDYNEXT
// BranchB secondary dead end, no non-full previous/next JP:
SMRESTART(digits);
// ----------------------------------------------------------------------------
// UNCOMPRESSED BRANCH:
//
// Check Decode bytes, if any, in the current JP, then search for a JP for the
// next digit in Index.
case cJU_JPBRANCH_U2:
CHECKDCD(2);
SMPREPB2(SMBranchU);
case cJU_JPBRANCH_U3:
CHECKDCD(3);
SMPREPB3(SMBranchU);
#ifdef JU_64BIT
case cJU_JPBRANCH_U4:
CHECKDCD(4);
SMPREPB4(SMBranchU);
case cJU_JPBRANCH_U5:
CHECKDCD(5);
SMPREPB5(SMBranchU);
case cJU_JPBRANCH_U6:
CHECKDCD(6);
SMPREPB6(SMBranchU);
case cJU_JPBRANCH_U7:
CHECKDCD(7);
SMPREPB7(SMBranchU);
#endif
case cJU_JPBRANCH_U:
SMPREPBL(SMBranchU);
// Common code (state-independent) for all cases of uncompressed branches:
SMBranchU:
Pjbu = P_JBU(Pjp->jp_Addr);
Pjp = (Pjbu->jbu_jp) + digit;
// Absent JP = null JP for current digit in Index:
if (JPNULL(JU_JPTYPE(Pjp))) RET_SUCCESS;
// Non-full JP matches current digit in Index:
//
// Iterate to the subsidiary JP.
if (! JPFULL(Pjp)) goto SMGetContinue;
// BranchU primary dead end:
//
// Upon hitting a full JP in a BranchU for the next digit in Index, search
// sideways for a previous/next null or non-full JP. BRANCHU_CHECKJP() is
// shorthand for common code.
//
// Note: The preceding code is separate from this loop because Index does not
// need revising (see SET_AND_*()) if the initial index is an empty index.
#define BRANCHU_CHECKJP(OpIncDec,OpLeastDigits) \
{ \
OpIncDec Pjp; \
\
if (JPNULL(JU_JPTYPE(Pjp))) \
SET_AND_RETURN(OpLeastDigits, digit, digits) \
\
if (! JPFULL(Pjp)) \
SET_AND_CONTINUE(OpLeastDigits, digit, digits) \
}
#ifdef JUDYPREV
while (digit-- > 0)
BRANCHU_CHECKJP(--, SETLEASTDIGITS_D);
#else
while (++digit < cJU_BRANCHUNUMJPS)
BRANCHU_CHECKJP(++, CLEARLEASTDIGITS_D);
#endif
// BranchU secondary dead end, no non-full previous/next JP:
SMRESTART(digits);
// ----------------------------------------------------------------------------
// LINEAR LEAF:
//
// Check Decode bytes, if any, in the current JP, then search the leaf for the
// previous/next empty index starting at Index. Primary leaf dead end is
// hidden within j__udySearchLeaf*Empty*(). In case of secondary leaf dead
// end, restart at the top of the tree.
//
// Note: Pword is the name known to GET*; think of it as Pjlw.
#define SMLEAFL(cDigits,Func) \
Pword = (PWord_t) P_JLW(Pjp->jp_Addr); \
pop0 = JU_JPLEAF_POP0(Pjp); \
Func(Pword, pop0)
#if (defined(JUDYL) || (! defined(JU_64BIT)))
case cJU_JPLEAF1:
CHECKDCD(1);
SMLEAFL(1, j__udySearchLeafEmpty1);
#endif
case cJU_JPLEAF2:
CHECKDCD(2);
SMLEAFL(2, j__udySearchLeafEmpty2);
case cJU_JPLEAF3:
CHECKDCD(3);
SMLEAFL(3, j__udySearchLeafEmpty3);
#ifdef JU_64BIT
case cJU_JPLEAF4:
CHECKDCD(4);
SMLEAFL(4, j__udySearchLeafEmpty4);
case cJU_JPLEAF5:
CHECKDCD(5);
SMLEAFL(5, j__udySearchLeafEmpty5);
case cJU_JPLEAF6:
CHECKDCD(6);
SMLEAFL(6, j__udySearchLeafEmpty6);
case cJU_JPLEAF7:
CHECKDCD(7);
SMLEAFL(7, j__udySearchLeafEmpty7);
#endif
// ----------------------------------------------------------------------------
// BITMAP LEAF:
//
// Check Decode bytes, if any, in the current JP, then search the leaf for the
// previous/next empty index starting at Index.
case cJU_JPLEAF_B1:
CHECKDCD(1);
Pjlb = P_JLB(Pjp->jp_Addr);
digit = JU_DIGITATSTATE(Index, 1);
subexp = digit / cJU_BITSPERSUBEXPL;
bitposmaskL = JU_BITPOSMASKL(digit);
assert(subexp < cJU_NUMSUBEXPL); // falls in expected range.
// Absent index = no index matches current digit in Index:
// if (! JU_BITMAPTESTL(Pjlb, digit)) // slower.
if (! (JU_JLB_BITMAP(Pjlb, subexp) & bitposmaskL)) // faster.
RET_SUCCESS;
// LeafB1 primary dead end:
//
// Upon hitting a valid (non-empty) index in a LeafB1 for the last digit in
// Index, search sideways for a previous/next absent index, first in the
// current bitmap subexpanse, then in lower/higher subexpanses.
// LEAFB1_CHECKBIT() is shorthand for common code to handle one bit in one
// bitmap subexpanse.
//
// Note: The preceding code is separate from this loop because Index does not
// need revising (see SET_AND_*()) if the initial index is an empty index.
//
// TBD: For speed, shift bitposmaskL instead of using JU_BITMAPTESTL or
// JU_BITPOSMASKL, but this shift has knowledge of bit order that really should
// be encapsulated in a header file.
#define LEAFB1_CHECKBIT(OpLeastDigits) \
if (! (JU_JLB_BITMAP(Pjlb, subexp) & bitposmaskL)) \
SET_AND_RETURN(OpLeastDigits, digit, 1)
#define LEAFB1_STARTSUBEXP(OpLeastDigits) \
if (! JU_JLB_BITMAP(Pjlb, subexp)) /* empty subexp */ \
SET_AND_RETURN(OpLeastDigits, digit, 1)
#ifdef JUDYPREV
--digit; // skip initial digit.
bitposmaskL >>= 1; // see TBD above.
LeafB1NextSubexp: // return here to check next bitmap subexpanse.
while (bitposmaskL) // more bits to check in subexp.
{
LEAFB1_CHECKBIT(SETLEASTDIGITS_D);
assert(digit >= 0);
--digit;
bitposmaskL >>= 1;
}
if (subexp-- > 0) // more subexpanses.
{
LEAFB1_STARTSUBEXP(SETLEASTDIGITS_D);
#ifdef JU_64BIT
bitposmaskL = (1ULL << (cJU_BITSPERSUBEXPL - 1));
#else
bitposmaskL = (1UL << (cJU_BITSPERSUBEXPL - 1));
#endif
goto LeafB1NextSubexp;
}
#else // JUDYNEXT
++digit; // skip initial digit.
bitposmaskL <<= 1; // note: BITMAPL_t.
LeafB1NextSubexp: // return here to check next bitmap subexpanse.
while (bitposmaskL) // more bits to check in subexp.
{
LEAFB1_CHECKBIT(CLEARLEASTDIGITS_D);
assert(digit < cJU_SUBEXPPERSTATE);
++digit;
bitposmaskL <<= 1; // note: BITMAPL_t.
}
if (++subexp < cJU_NUMSUBEXPL) // more subexpanses.
{
LEAFB1_STARTSUBEXP(CLEARLEASTDIGITS_D);
bitposmaskL = 1;
goto LeafB1NextSubexp;
}
#endif // JUDYNEXT
// LeafB1 secondary dead end, no empty index:
SMRESTART(1);
#ifdef JUDY1
// ----------------------------------------------------------------------------
// FULL POPULATION:
//
// If the Decode bytes do not match, Index is empty (without modification);
// otherwise restart.
case cJ1_JPFULLPOPU1:
CHECKDCD(1);
SMRESTART(1);
#endif
// ----------------------------------------------------------------------------
// IMMEDIATE:
//
// Pop1 = 1 Immediate JPs:
//
// If Index is not in the immediate JP, return success; otherwise check if
// there is an empty index below/above the immediate JPs index, and if so,
// return success with modified Index, else restart.
//
// Note: Doug says its fast enough to calculate the index size (digits) in
// the following; no need to set it separately for each case.
case cJU_JPIMMED_1_01:
case cJU_JPIMMED_2_01:
case cJU_JPIMMED_3_01:
#ifdef JU_64BIT
case cJU_JPIMMED_4_01:
case cJU_JPIMMED_5_01:
case cJU_JPIMMED_6_01:
case cJU_JPIMMED_7_01:
#endif
if (JU_JPDCDPOP0(Pjp) != JU_TRIMTODCDSIZE(Index)) RET_SUCCESS;
digits = JU_JPTYPE(Pjp) - cJU_JPIMMED_1_01 + 1;
LEAF_EDGE(JU_LEASTBYTES(JU_JPDCDPOP0(Pjp), digits), digits);
// Immediate JPs with Pop1 > 1:
#define IMM_MULTI(Func,BaseJPType) \
JUDY1CODE(Pword = (PWord_t) (Pjp->jp_1Index);) \
JUDYLCODE(Pword = (PWord_t) (Pjp->jp_LIndex);) \
Func(Pword, JU_JPTYPE(Pjp) - (BaseJPType) + 1)
case cJU_JPIMMED_1_02:
case cJU_JPIMMED_1_03:
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_1_04:
case cJU_JPIMMED_1_05:
case cJU_JPIMMED_1_06:
case cJU_JPIMMED_1_07:
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_1_08:
case cJ1_JPIMMED_1_09:
case cJ1_JPIMMED_1_10:
case cJ1_JPIMMED_1_11:
case cJ1_JPIMMED_1_12:
case cJ1_JPIMMED_1_13:
case cJ1_JPIMMED_1_14:
case cJ1_JPIMMED_1_15:
#endif
IMM_MULTI(j__udySearchLeafEmpty1, cJU_JPIMMED_1_02);
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_2_02:
case cJU_JPIMMED_2_03:
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_2_04:
case cJ1_JPIMMED_2_05:
case cJ1_JPIMMED_2_06:
case cJ1_JPIMMED_2_07:
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
IMM_MULTI(j__udySearchLeafEmpty2, cJU_JPIMMED_2_02);
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
case cJU_JPIMMED_3_02:
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_3_03:
case cJ1_JPIMMED_3_04:
case cJ1_JPIMMED_3_05:
#endif
#if (defined(JUDY1) || defined(JU_64BIT))
IMM_MULTI(j__udySearchLeafEmpty3, cJU_JPIMMED_3_02);
#endif
#if (defined(JUDY1) && defined(JU_64BIT))
case cJ1_JPIMMED_4_02:
case cJ1_JPIMMED_4_03:
IMM_MULTI(j__udySearchLeafEmpty4, cJ1_JPIMMED_4_02);
case cJ1_JPIMMED_5_02:
case cJ1_JPIMMED_5_03:
IMM_MULTI(j__udySearchLeafEmpty5, cJ1_JPIMMED_5_02);
case cJ1_JPIMMED_6_02:
IMM_MULTI(j__udySearchLeafEmpty6, cJ1_JPIMMED_6_02);
case cJ1_JPIMMED_7_02:
IMM_MULTI(j__udySearchLeafEmpty7, cJ1_JPIMMED_7_02);
#endif
// ----------------------------------------------------------------------------
// INVALID JP TYPE:
default:
RET_CORRUPT;
} // SMGet switch.
} // Judy1PrevEmpty() / Judy1NextEmpty() / JudyLPrevEmpty() / JudyLNextEmpty()
FUNCTION int j__udyCreateBranchU(
Pjp_t Pjp,
Pvoid_t Pjpm)
{
jp_t JPNull;
Pjbu_t PjbuRaw;
Pjbu_t Pjbu;
Pjbb_t PjbbRaw;
Pjbb_t Pjbb;
Word_t ii, jj;
BITMAPB_t BitMap;
Pjp_t PDstJP;
#ifdef JU_STAGED_EXP
jbu_t BranchU; // Staged uncompressed branch
#else
// Allocate memory for a BranchU:
PjbuRaw = j__udyAllocJBU(Pjpm);
if (PjbuRaw == (Pjbu_t) NULL) return(-1);
Pjbu = P_JBU(PjbuRaw);
#endif
JU_JPSETADT(&JPNull, 0, 0, JU_JPTYPE(Pjp) - cJU_JPBRANCH_B2 + cJU_JPNULL1);
// Get the pointer to the BranchB:
PjbbRaw = (Pjbb_t) (Pjp->jp_Addr);
Pjbb = P_JBB(PjbbRaw);
// Set the pointer to the Uncompressed branch
#ifdef JU_STAGED_EXP
PDstJP = BranchU.jbu_jp;
#else
PDstJP = Pjbu->jbu_jp;
#endif
for (ii = 0; ii < cJU_NUMSUBEXPB; ii++)
{
Pjp_t PjpA;
Pjp_t PjpB;
PjpB = PjpA = P_JP(JU_JBB_PJP(Pjbb, ii));
// Get the bitmap for this subexpanse
BitMap = JU_JBB_BITMAP(Pjbb, ii);
// NULL empty subexpanses
if (BitMap == 0)
{
// But, fill with NULLs
for (jj = 0; jj < cJU_BITSPERSUBEXPB; jj++)
{
PDstJP[jj] = JPNull;
}
PDstJP += cJU_BITSPERSUBEXPB;
continue;
}
// Check if Uncompressed subexpanse
if (BitMap == cJU_FULLBITMAPB)
{
// Copy subexpanse to the Uncompressed branch intact
JU_COPYMEM(PDstJP, PjpA, cJU_BITSPERSUBEXPB);
// Bump to next subexpanse
PDstJP += cJU_BITSPERSUBEXPB;
// Set length of subexpanse
jj = cJU_BITSPERSUBEXPB;
}
else
{
for (jj = 0; jj < cJU_BITSPERSUBEXPB; jj++)
{
// Copy JP or NULLJP depending on bit
if (BitMap & 1) { *PDstJP = *PjpA++; }
else { *PDstJP = JPNull; }
PDstJP++; // advance to next JP
BitMap >>= 1;
}
jj = PjpA - PjpB;
}
// Free the subexpanse:
j__udyFreeJBBJP(JU_JBB_PJP(Pjbb, ii), jj, Pjpm);
} // for each JP in BranchU
#ifdef JU_STAGED_EXP
// Allocate memory for a BranchU:
PjbuRaw = j__udyAllocJBU(Pjpm);
if (PjbuRaw == (Pjbu_t) NULL) return(-1);
Pjbu = P_JBU(PjbuRaw);
// Copy staged branch to newly allocated branch:
//
// TBD: I think this code is broken.
*Pjbu = BranchU;
#endif // JU_STAGED_EXP
// Finally free the BranchB and put the BranchU in its place:
j__udyFreeJBB(PjbbRaw, Pjpm);
Pjp->jp_Addr = (Word_t) PjbuRaw;
Pjp->jp_Type += cJU_JPBRANCH_U - cJU_JPBRANCH_B;
return(1);
} // j__udyCreateBranchU()