Change mpz_realloc to FLINT_MPZ_REALLOC(_TIGHT)

* FLINT_MPZ_REALLOC only reallocates if necessary. Should be used in most functions. * FLINT_MPZ_REALLOC_TIGHT always reallocates, with a minimum number of MPZ_MIN_ALLOC limbs. Should be used in higher-level functions if it may be the case that it can improve the performance by using less memory, like routines as fmpz_bit_unpack. Also remove, now redundant, FLINT_MPZ_REALLOC in gmpcompat.h.
flintlib · May 17, 2024 · 3cdc0eb · 3cdc0eb
1 parent cc60cac
commit 3cdc0eb
Show file tree

Hide file tree

Showing 24 changed files with 62 additions and 129 deletions.
diff --git a/src/arf/get.c b/src/arf/get.c
@@ -179,14 +179,12 @@ arf_get_fmpz_2exp(fmpz_t man, fmpz_t exp, const arf_t x)
         else
         {
             mpz_ptr z = _fmpz_promote(man);
-
-            if (z->_mp_alloc < xn)
-                mpz_realloc(z, xn);
+            mp_ptr zptr = FLINT_MPZ_REALLOC(z, xn);
 
             if (shift == 0)
-                flint_mpn_copyi(z->_mp_d, xptr, xn);
+                flint_mpn_copyi(zptr, xptr, xn);
             else
-                mpn_rshift(z->_mp_d, xptr, xn, shift);
+                mpn_rshift(zptr, xptr, xn, shift);
 
             /* top limb cannot be zero */
             z->_mp_size = ARF_SGNBIT(x) ? -xn : xn;
@@ -321,10 +319,7 @@ arf_get_fmpz(fmpz_t z, const arf_t x, arf_rnd_t rnd)
     zn = (exp + (rnd != ARF_RND_DOWN) + FLINT_BITS - 1) / FLINT_BITS;
 
     zz = _fmpz_promote(z);
-    if (zz->_mp_alloc < zn)
-        mpz_realloc(zz, zn);
-
-    zp = zz->_mp_d;
+    zp = FLINT_MPZ_REALLOC(zz, zn);
 
     if (rnd == ARF_RND_DOWN)
     {

diff --git a/src/fft/fermat_to_mpz.c b/src/fft/fermat_to_mpz.c
@@ -15,24 +15,22 @@
 void fermat_to_mpz(mpz_t m, mp_limb_t * i, mp_size_t limbs)
 {
     mp_limb_signed_t hi;
+    mp_ptr mptr = FLINT_MPZ_REALLOC(m, limbs + 1);
 
-    if (m->_mp_alloc < limbs + 1)
-        mpz_realloc(m, limbs + 1);
-
-    flint_mpn_copyi(m->_mp_d, i, limbs + 1);
+    flint_mpn_copyi(mptr, i, limbs + 1);
     hi = i[limbs];
     if (hi < WORD(0))
     {
-        mpn_neg(m->_mp_d, m->_mp_d, limbs + 1);
+        mpn_neg(mptr, mptr, limbs + 1);
         m->_mp_size = limbs + 1;
-        while ((m->_mp_size) && (!m->_mp_d[m->_mp_size - 1]))
+        while ((m->_mp_size) && (!mptr[m->_mp_size - 1]))
             m->_mp_size--;
         m->_mp_size = -m->_mp_size;
     }
     else
     {
         m->_mp_size = limbs + 1;
-        while ((m->_mp_size) && (!m->_mp_d[m->_mp_size - 1]))
+        while ((m->_mp_size) && (!mptr[m->_mp_size - 1]))
             m->_mp_size--;
     }
 }
diff --git a/src/fft_small/fmpz_poly_mul.c b/src/fft_small/fmpz_poly_mul.c
@@ -200,10 +200,9 @@ void fmpz_neg_ui_array(fmpz_t out, const ulong * in, slong in_len)
     else
     {
         mpz_ptr mpz = _fmpz_promote(out);
-        if (mpz->_mp_alloc < size)
-            mpz_realloc2(mpz, FLINT_BITS * size);
+        mp_ptr mp = FLINT_MPZ_REALLOC(mpz, size);
         mpz->_mp_size = -size;
-        flint_mpn_copyi(mpz->_mp_d, in, size);
+        flint_mpn_copyi(mp, in, size);
     }
 }
 

diff --git a/src/flint.h.in b/src/flint.h.in
@@ -484,7 +484,10 @@ typedef fmpq fmpq_t[1];
 
 /* The minimum number of limbs that the fmpz collector should spit out */
 #define MPZ_MIN_ALLOC 2
-#define MPZ_REALLOC(ptr, sz) ((int) (sz) > ((ptr)->_mp_alloc) ? (mp_ptr) mpz_realloc(ptr, sz) : ((ptr)->_mp_d))
+
+/* NOTE: Never allocate less than MPZ_MIN_ALLOC */
+#define FLINT_MPZ_REALLOC(ptr, sz) ((int) (sz) > ((ptr)->_mp_alloc) ? (mp_ptr) mpz_realloc(ptr, sz) : (ptr)->_mp_d)
+#define FLINT_MPZ_REALLOC_TIGHT(ptr, sz) ((mp_ptr) mpz_realloc(ptr, FLINT_MAX(sz, MPZ_MIN_ALLOC)))
 
 #ifdef __cplusplus
 }

diff --git a/src/fmpz/addmul.c b/src/fmpz/addmul.c
@@ -16,12 +16,6 @@
 #include "ulong_extras.h"
 #include "fmpz.h"
 
-#define MPZ_FIT_SIZE(z, nlimbs) \
-    do { \
-        if (z->_mp_alloc < nlimbs) \
-            _mpz_realloc(z, nlimbs); \
-    } while (0)
-
 /* Will not get called with x or y small. */
 void
 _flint_mpz_addmul_large(mpz_ptr z, mpz_srcptr x, mpz_srcptr y, int negate)
@@ -122,8 +116,7 @@ _flint_mpz_addmul_large(mpz_ptr z, mpz_srcptr x, mpz_srcptr y, int negate)
 
     tn -= (top == 0);
     alloc = FLINT_MAX(tn, zn) + 1;
-    MPZ_FIT_SIZE(z, alloc);
-    zd = z->_mp_d;
+    zd = FLINT_MPZ_REALLOC(z, alloc);
 
     if (sgn >= 0)
     {

diff --git a/src/fmpz/aors_ui.c b/src/fmpz/aors_ui.c
@@ -69,10 +69,7 @@ fmpz_add_ui(fmpz_t res, const fmpz_t x, ulong y)
         else
             rp = _fmpz_promote_val(res);
 
-        if (rp->_mp_alloc < xn + 1)
-            _mpz_realloc(rp, xn + 1);
-
-        rd = rp->_mp_d;
+        rd = FLINT_MPZ_REALLOC(rp, xn + 1);
         xd = xp->_mp_d;
 
         if (xn_signed >= 0) /* positive + nonnegative */
@@ -182,10 +179,7 @@ fmpz_sub_ui(fmpz_t res, const fmpz_t x, ulong y)
         else
             rp = _fmpz_promote_val(res);
 
-        if (rp->_mp_alloc < xn + 1)
-            _mpz_realloc(rp, xn + 1);
-
-        rd = rp->_mp_d;
+        rd = FLINT_MPZ_REALLOC(rp, xn + 1);
         xd = xp->_mp_d;
 
         if (xn_signed <= 0) /* positive + nonnegative */

diff --git a/src/fmpz/bit_unpack.c b/src/fmpz/bit_unpack.c
@@ -77,9 +77,7 @@ fmpz_bit_unpack(fmpz_t coeff, nn_srcptr arr, flint_bitcnt_t shift,
         b = bits % FLINT_BITS;
 
         /* allocate space for l limbs only */
-        if (mcoeff->_mp_alloc < (int) l)
-            mpz_realloc(mcoeff, l);
-        p = mcoeff->_mp_d;
+        p = FLINT_MPZ_REALLOC_TIGHT(mcoeff, l);
 
         /* shift in l limbs */
         if (shift)
@@ -176,9 +174,7 @@ fmpz_bit_unpack_unsigned(fmpz_t coeff, nn_srcptr arr,
         b = bits % FLINT_BITS;
 
         /* allocate space for l limbs only */
-        if (mcoeff->_mp_alloc < (int) l)
-            mpz_realloc(mcoeff, l);
-        p = mcoeff->_mp_d;
+        p = FLINT_MPZ_REALLOC_TIGHT(mcoeff, l);
 
         /* shift in l limbs */
         if (shift)

diff --git a/src/fmpz/fdiv.c b/src/fmpz/fdiv.c
@@ -202,8 +202,7 @@ void _mpz_tdiv_qr_preinvn(mpz_ptr q, mpz_ptr r,
 
     nn_ptr qp, rp, ap, dp, tp, sp;
 
-    if ((ulong) r->_mp_alloc < usize1 + nm)
-        mpz_realloc(r, usize1 + nm);
+    rp = FLINT_MPZ_REALLOC(r, usize1 + nm);
 
     if (usize1 < usize2) /* special case preinv code can't deal with */
     {
@@ -213,13 +212,9 @@ void _mpz_tdiv_qr_preinvn(mpz_ptr q, mpz_ptr r,
         return;
     }
 
-    if ((ulong) q->_mp_alloc < qsize + nm)
-        mpz_realloc(q, qsize + nm);
-
     dp = d->_mp_d;
     ap = a->_mp_d;
-    qp = q->_mp_d;
-    rp = r->_mp_d;
+    qp = FLINT_MPZ_REALLOC(q, qsize + nm);
 
     TMP_START;
     if ((r == d || q == d) && !nm) /* we have alias with d */

diff --git a/src/fmpz/init2.c b/src/fmpz/init2.c
@@ -20,8 +20,7 @@ fmpz_init2(fmpz_t f, ulong limbs)
     {
         mpz_ptr mf = _fmpz_new_mpz();
         *f = PTR_TO_COEFF(mf);
-        if ((int) limbs > mf->_mp_alloc)
-            mpz_realloc(mf, limbs);
+        FLINT_MPZ_REALLOC(mf, limbs);
     }
     else
     {

diff --git a/src/fmpz/mul.c b/src/fmpz/mul.c
@@ -48,9 +48,7 @@ flint_mpz_mul(mpz_ptr z, mpz_srcptr x, mpz_srcptr y)
     }
 
     zn = xn + yn;
-    if (z->_mp_alloc < zn)
-        _mpz_realloc(z, zn);
-    zd = z->_mp_d;
+    zd = FLINT_MPZ_REALLOC(z, zn);
     /* Important: read after possibly resizing z, so that the
        pointers are valid in case of aliasing. */
     xd = x->_mp_d;

diff --git a/src/fmpz/mul_2exp.c b/src/fmpz/mul_2exp.c
@@ -43,23 +43,8 @@ fmpz_mul_2exp(fmpz_t f, const fmpz_t g, ulong exp)
         ulong * limbs;
         mpz_ptr mf;
 
-        /* Ensure enough limbs are allocated for f */
-        if (!COEFF_IS_MPZ(*f))
-        {
-            /* TODO: Initialize the new mpz with alloc limbs instead of
-             * reallocating them. */
-            mf = _fmpz_new_mpz();
-            *f = PTR_TO_COEFF(mf);
-            if (mf->_mp_alloc < alloc)
-                mpz_realloc(mf, alloc);
-        }
-        else
-        {
-            mf = COEFF_TO_PTR(*f);
-            if (mf->_mp_alloc < alloc)
-                mpz_realloc(mf, alloc);
-        }
-        limbs = mf->_mp_d;
+        mf = _fmpz_promote(f);
+        limbs = FLINT_MPZ_REALLOC(mf, alloc);
         mf->_mp_size = (c1 > 0) ? alloc : -alloc;
         memset(limbs, 0, sizeof(ulong) * alloc);
 

diff --git a/src/fmpz/primorial.c b/src/fmpz/primorial.c
@@ -136,6 +136,7 @@ fmpz_primorial(fmpz_t res, ulong n)
     slong len, pi;
     ulong bits;
     mpz_ptr mres;
+    mp_ptr rp;
     const ulong * primes;
 
     if (n <= LARGEST_ULONG_PRIMORIAL)
@@ -153,9 +154,8 @@ fmpz_primorial(fmpz_t res, ulong n)
     bits = FLINT_BIT_COUNT(primes[pi - 1]);
 
     mres = _fmpz_promote(res);
-    if (mres->_mp_alloc < (int) ((pi * bits) / FLINT_BITS) + 1)
-        mpz_realloc(mres, (pi * bits) / FLINT_BITS + 1);
+    rp = FLINT_MPZ_REALLOC(mres, (pi * bits) / FLINT_BITS + 1);
 
-    len = mpn_prod_limbs(mres->_mp_d, primes, pi, bits);
+    len = mpn_prod_limbs(rp, primes, pi, bits);
     mres->_mp_size = len;
 }
diff --git a/src/fmpz/set.c b/src/fmpz/set.c
@@ -187,11 +187,10 @@ fmpz_set_signed_uiuiui(fmpz_t r, ulong hi, ulong mid, ulong lo)
     else
     {
         mpz_ptr z = _fmpz_promote(r);
-        if (z->_mp_alloc < 3)
-            mpz_realloc(z, 3);
-        z->_mp_d[0] = lo;
-        z->_mp_d[1] = mid;
-        z->_mp_d[2] = hi;
+        mp_ptr zp = FLINT_MPZ_REALLOC(z, 3);
+        zp[0] = lo;
+        zp[1] = mid;
+        zp[2] = hi;
         z->_mp_size = negate ? -3 : 3;
     }
 }
@@ -217,24 +216,23 @@ void fmpz_set_ui_array(fmpz_t out, const ulong * in, slong in_len)
     else
     {
         mpz_ptr mpz = _fmpz_promote(out);
-        if (mpz->_mp_alloc < size)
-            mpz_realloc(mpz, size);
+        mp_ptr mp = FLINT_MPZ_REALLOC(mpz, size);
         mpz->_mp_size = size;
-        flint_mpn_copyi(mpz->_mp_d, in, size);
+        flint_mpn_copyi(mp, in, size);
     }
 }
 
 void fmpz_set_mpn_large(fmpz_t z, nn_srcptr src, slong n, int negative)
 {
     mpz_ptr zz;
+    mp_ptr zp;
     slong i;
-    zz = _fmpz_promote(z);
 
-    if (zz->_mp_alloc < n)
-        mpz_realloc(zz, n);
+    zz = _fmpz_promote(z);
+    zp = FLINT_MPZ_REALLOC(zz, n);
 
     for (i = 0; i < n; i++)
-        zz->_mp_d[i] = src[i];
+        zp[i] = src[i];
 
     zz->_mp_size = negative ? -n : n;
 }
diff --git a/src/fmpz_extras/lshift_mpn.c b/src/fmpz_extras/lshift_mpn.c
@@ -26,10 +26,7 @@ fmpz_lshift_mpn(fmpz_t z, nn_srcptr d, slong dn, int sgnbit, flint_bitcnt_t shif
     shift_bits = shift % FLINT_BITS;
     zn = dn + shift_limbs + (shift_bits != 0);
 
-    if (zmpz->_mp_alloc < zn)
-        mpz_realloc(zmpz, zn);
-
-    zp = zmpz->_mp_d;
+    zp = FLINT_MPZ_REALLOC(zmpz, zn);
     flint_mpn_zero(zp, shift_limbs);
 
     if (shift_bits == 0)

diff --git a/src/fmpz_factor/ecm.c b/src/fmpz_factor/ecm.c
@@ -98,11 +98,11 @@ fmpz_factor_ecm(fmpz_t f, ulong curves, ulong B1, ulong B2,
     fmpz_sub_ui(nm8, n_in, 8);
 
     ret = 0;
+    /* FIXME: Wait to promote f until after stage 2 precomputations? */
     fac = _fmpz_promote(f);
     {
         int alloc = fmpz_size(n_in);
-        if (fac->_mp_alloc < alloc)
-            mpz_realloc(fac, alloc);
+        FLINT_MPZ_REALLOC(fac, alloc);
     }
 
     /************************ STAGE I PRECOMPUTATIONS ************************/

diff --git a/src/fmpz_factor/factor_pp1.c b/src/fmpz_factor/factor_pp1.c
@@ -335,16 +335,14 @@ int fmpz_factor_pp1(fmpz_t fac, const fmpz_t n_in, ulong B1, ulong B2sqrt, ulong
         for (i = 0; (ulong) i < num_roots; i++)
         {
             mpz_ptr m = _fmpz_promote(roots + i);
-            if (m->_mp_alloc < nn)
-                mpz_realloc(m, nn);
+            FLINT_MPZ_REALLOC(m, nn);
         }
 
         roots2 = _fmpz_vec_init(num_roots);
         for (i = 0; (ulong) i < num_roots; i++)
         {
             mpz_ptr m = _fmpz_promote(roots2 + i);
-            if (m->_mp_alloc < nn)
-                mpz_realloc(m, nn);
+            FLINT_MPZ_REALLOC(m, nn);
         }
 
         evals = _fmpz_vec_init(num_roots);
@@ -553,9 +551,8 @@ int fmpz_factor_pp1(fmpz_t fac, const fmpz_t n_in, ulong B1, ulong B2sqrt, ulong
     if (ret)
     {
         mpz_ptr fm = _fmpz_promote(fac);
-        if (fm->_mp_alloc < r)
-            mpz_realloc(fm, r);
-        flint_mpn_copyi(fm->_mp_d, factor, r);
+        mp_ptr fp = FLINT_MPZ_REALLOC(fm, r);
+        flint_mpn_copyi(fp, factor, r);
         fm->_mp_size = r;
         _fmpz_demote_val(fac);
     }

diff --git a/src/fmpz_factor/pollard_brent.c b/src/fmpz_factor/pollard_brent.c
@@ -72,8 +72,7 @@ fmpz_factor_pollard_brent(fmpz_t p_factor, flint_rand_t state, fmpz_t n_in,
     flint_mpn_preinvn(ninv, n, n_size);
 
     fac = _fmpz_promote(p_factor);
-    if (fac->_mp_alloc < (int) n_size)
-        mpz_realloc(fac, n_size);
+    FLINT_MPZ_REALLOC(fac, n_size);
     fac->_mp_size = n_size;
 
     while (max_tries--)

diff --git a/src/fmpz_factor/pollard_brent_single.c b/src/fmpz_factor/pollard_brent_single.c
@@ -224,8 +224,7 @@ fmpz_factor_pollard_brent_single(fmpz_t p_factor, fmpz_t n_in, fmpz_t yi,
     flint_mpn_preinvn(ninv, n, n_size);
 
     fac = _fmpz_promote(p_factor);
-    if (fac->_mp_alloc < (int) n_size)
-        mpz_realloc(fac, n_size);
+    FLINT_MPZ_REALLOC(fac, n_size);
     fac->_mp_size = n_size;
 
     mpn_zero(a, n_size);