aboutsummaryrefslogtreecommitdiff
path: root/patches/gmp/5.1.1
diff options
context:
space:
mode:
Diffstat (limited to 'patches/gmp/5.1.1')
-rw-r--r--patches/gmp/5.1.1/100-fix-bulldozer-piledriver.patch21
-rw-r--r--patches/gmp/5.1.1/110-mpz_powm_ui.patch39
-rw-r--r--patches/gmp/5.1.1/120-fix-mpn_sbpi1_div_qr_sec.patch164
-rw-r--r--patches/gmp/5.1.1/130-do-not-clobber-f16-f18.patch73
4 files changed, 297 insertions, 0 deletions
diff --git a/patches/gmp/5.1.1/100-fix-bulldozer-piledriver.patch b/patches/gmp/5.1.1/100-fix-bulldozer-piledriver.patch
new file mode 100644
index 00000000..f6faca1a
--- /dev/null
+++ b/patches/gmp/5.1.1/100-fix-bulldozer-piledriver.patch
@@ -0,0 +1,21 @@
+
+# HG changeset patch
+# User Torbjorn Granlund <tege@gmplib.org>
+# Date 1368052461 -7200
+# Node ID 11fbd825bc45385d00b69c30bd7566acee11f0d2
+# Parent bd4c950486ab4c161ea69b91c25150719c3cb700
+Fix typo.
+
+diff -r bd4c950486ab -r 11fbd825bc45 mpn/x86_64/bd1/mul_1.asm
+--- a/mpn/x86_64/bd1/mul_1.asm Thu May 02 18:20:37 2013 +0200
++++ b/mpn/x86_64/bd1/mul_1.asm Thu May 09 00:34:21 2013 +0200
+@@ -53,7 +53,7 @@
+ IFDOS(` define(`v0', ``%r9'') ') dnl
+ IFDOS(` define(`r9', ``rdi'') ') dnl
+ IFDOS(` define(`n', ``%r8'') ') dnl
+-IFDOS(` define(`r8', ``r11'') ') dnl
++IFDOS(` define(`r8', ``rbx'') ') dnl
+
+ ASM_START()
+ TEXT
+
diff --git a/patches/gmp/5.1.1/110-mpz_powm_ui.patch b/patches/gmp/5.1.1/110-mpz_powm_ui.patch
new file mode 100644
index 00000000..f0d240c6
--- /dev/null
+++ b/patches/gmp/5.1.1/110-mpz_powm_ui.patch
@@ -0,0 +1,39 @@
+
+# HG changeset patch
+# User Torbjorn Granlund <tege@gmplib.org>
+# Date 1363174284 -3600
+# Node ID ec4d88674036804b26f22c6a2bfca6ae1e92d370
+# Parent e616ff715c34e112d0a5f7535d31ffe1194a5c7d
+(mod): Adhere to mpn_mu_div_qr's overlap requirements.
+
+diff -r e616ff715c34 -r ec4d88674036 mpz/powm_ui.c
+--- a/mpz/powm_ui.c Sun Feb 17 19:40:16 2013 +0100
++++ b/mpz/powm_ui.c Wed Mar 13 12:31:24 2013 +0100
+@@ -2,8 +2,8 @@
+
+ Contributed to the GNU project by Torbjorn Granlund.
+
+-Copyright 1991, 1993, 1994, 1996, 1997, 2000, 2001, 2002, 2005, 2008,
+-2009, 2011, 2012 Free Software Foundation, Inc.
++Copyright 1991, 1993, 1994, 1996, 1997, 2000, 2001, 2002, 2005, 2008, 2009,
++2011, 2012, 2013 Free Software Foundation, Inc.
+
+ This file is part of the GNU MP Library.
+
+@@ -70,9 +70,14 @@
+ }
+ else
+ {
++ /* We need to allocate separate remainder area, since mpn_mu_div_qr does
++ not handle overlap between the numerator and remainder areas.
++ FIXME: Make it handle such overlap. */
++ mp_ptr rp = TMP_ALLOC_LIMBS (dn);
+ mp_size_t itch = mpn_mu_div_qr_itch (nn, dn, 0);
+ mp_ptr scratch = TMP_ALLOC_LIMBS (itch);
+- mpn_mu_div_qr (qp, np, np, nn, dp, dn, scratch);
++ mpn_mu_div_qr (qp, rp, np, nn, dp, dn, scratch);
++ MPN_COPY (np, rp, dn);
+ }
+
+ TMP_FREE;
+
diff --git a/patches/gmp/5.1.1/120-fix-mpn_sbpi1_div_qr_sec.patch b/patches/gmp/5.1.1/120-fix-mpn_sbpi1_div_qr_sec.patch
new file mode 100644
index 00000000..13a50ad4
--- /dev/null
+++ b/patches/gmp/5.1.1/120-fix-mpn_sbpi1_div_qr_sec.patch
@@ -0,0 +1,164 @@
+
+# HG changeset patch
+# User Torbjorn Granlund <tege@gmplib.org>
+# Date 1373624469 -7200
+# Node ID a447c0c537891ed23edf180594a89616364ee633
+# Parent 6540e0b2925ead29f7158bb182e4fabfb9441433
+Partial rewrite.
+
+diff -r 6540e0b2925e -r a447c0c53789 mpn/generic/sbpi1_div_sec.c
+--- a/mpn/generic/sbpi1_div_sec.c Mon Jul 01 19:16:32 2013 +0200
++++ b/mpn/generic/sbpi1_div_sec.c Fri Jul 12 12:21:09 2013 +0200
+@@ -8,7 +8,7 @@
+ SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES. IN FACT, IT IS ALMOST
+ GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GNU MP RELEASE.
+
+-Copyright 2011, 2012 Free Software Foundation, Inc.
++Copyright 2011, 2012, 2013 Free Software Foundation, Inc.
+
+ This file is part of the GNU MP Library.
+
+@@ -29,6 +29,28 @@
+ #include "gmp-impl.h"
+ #include "longlong.h"
+
++/* This side-channel silent division algorithm reduces the partial remainder by
++ GMP_NUMB_BITS/2 bits at a time, compared to GMP_NUMB_BITS for the main
++ division algorithm. We do not insists on reducing by exactly
++ GMP_NUMB_BITS/2, but may leave a partial remainder that is D*B^i to 3D*B^i
++ too large (B is the limb base, D is the divisor, and i is the induction
++ variable); the subsequent step will handle the extra partial remainder bits.
++
++ WIth that partial remainder reduction, each step generates a quotient "half
++ limb". The outer loop generates two quotient half limbs, an upper (q1h) and
++ a lower (q0h) which are stored sparsely in separate limb arrays. These
++ arrays are added at the end; using separate arrays avoids data-dependent
++ carry propagation which could else pose a side-channel leakage problem.
++
++ The quotient half limbs may be between -3 to 0 from the accurate value
++ ("accurate" being the one which corresponds to a reduction to a principal
++ partial remainder). Too small quotient half limbs correspond to too large
++ remainders, which we reduce later, as described above.
++
++ In order to keep quotients from getting too big, corresponding to a negative
++ partial remainder, we use an inverse which is sligtly smaller than usually.
++*/
++
+ #if OPERATION_sbpi1_div_qr_sec
+ /* Needs (dn + 1) + (nn - dn) + (nn - dn) = 2nn - dn + 1 limbs at tp. */
+ #define FNAME mpn_sbpi1_div_qr_sec
+@@ -49,7 +71,7 @@
+ mp_limb_t dinv,
+ mp_ptr tp)
+ {
+- mp_limb_t nh, cy, q1h, q0h, dummy, h;
++ mp_limb_t nh, cy, q1h, q0h, dummy, cnd;
+ mp_size_t i;
+ mp_ptr hp;
+ #if OPERATION_sbpi1_div_qr_sec
+@@ -72,77 +94,69 @@
+ #endif
+ }
+
++ /* Decremenet inverse to keep quotient half limbs from being too large. */
++ dinv -= dinv != 0; /* FIXME: cmp-to-int */
++
+ /* Create a divisor copy shifted half a limb. */
+ hp = tp; /* (dn + 1) limbs */
+- cy = mpn_lshift (hp, dp, dn, GMP_NUMB_BITS / 2);
+- hp[dn] = dp[dn - 1] >> GMP_NUMB_BITS / 2;
++ hp[dn] = mpn_lshift (hp, dp, dn, GMP_NUMB_BITS / 2);
+
+ #if OPERATION_sbpi1_div_qr_sec
+ qlp = tp + (dn + 1); /* (nn - dn) limbs */
+ qhp = tp + (nn + 1); /* (nn - dn) limbs */
+ #endif
+
+- np += nn;
++ np += nn - dn;
++ nh = 0;
+
+- /* Main loop. Develop one full limb per iteration, but do it in two steps in
+- order to avoid conditionals. Quotient bits will be either correct or
+- underestimates. When a quotient is underestimated, the next quotient will
+- compensate, since quotients are to be added at consecutive weight distance
+- GMP_NUMB_BITS/2. We make two quotient arrays, each with GMP_NUMB_BITS/2+2
+- bits per entry. The arrays are added late after the loop. Separate
+- arrays avoid data-dependent carry propagation. */
+- nh = 0;
+ for (i = nn - dn - 1; i >= 0; i--)
+ {
+ np--;
+
+- nh = (nh << GMP_NUMB_BITS/2) + (np[0] >> GMP_NUMB_BITS/2);
++ nh = (nh << GMP_NUMB_BITS/2) + (np[dn] >> GMP_NUMB_BITS/2);
+ umul_ppmm (q1h, dummy, nh, dinv);
+ q1h += nh;
+ #if OPERATION_sbpi1_div_qr_sec
+ qhp[i] = q1h;
+ #endif
+- cy = mpn_submul_1 (np - dn, hp, dn + 1, q1h);
++ mpn_submul_1 (np, hp, dn + 1, q1h);
+
+- nh = np[0];
++ nh = np[dn];
+ umul_ppmm (q0h, dummy, nh, dinv);
+ q0h += nh;
+ #if OPERATION_sbpi1_div_qr_sec
+ qlp[i] = q0h;
+ #endif
+- cy = mpn_submul_1 (np - dn, dp, dn, q0h);
+-
+- nh -= cy;
++ nh -= mpn_submul_1 (np, dp, dn, q0h);
+ }
+
+- np[0] = nh;
+-
+- np -= dn;
+-
+ /* 1st adjustment depends on extra high remainder limb. */
+- h = np[dn];
++ cnd = nh != 0; /* FIXME: cmp-to-int */
+ #if OPERATION_sbpi1_div_qr_sec
+- qlp[0] += h;
++ qlp[0] += cnd;
+ #endif
+- h -= mpn_subcnd_n (np, np, dp, dn, h);
++ nh -= mpn_subcnd_n (np, np, dp, dn, cnd);
+
+- /* 2nd adjustment depends on remainder/divisor comparision as well as whether
++ /* 2nd adjustment depends on remainder/divisor comparison as well as whether
+ extra remainder limb was nullified by previous subtract. */
+ cy = mpn_sub_n (np, np, dp, dn);
+- cy = cy == h; /* FIXME: might leak on some archs */
++ cy = cy - nh;
+ #if OPERATION_sbpi1_div_qr_sec
+- qlp[0] += cy;
++ qlp[0] += 1 - cy;
+ #endif
+- mpn_addcnd_n (np, np, dp, dn, 1 - cy);
++ mpn_addcnd_n (np, np, dp, dn, cy);
+
++ /* 3rd adjustment depends on remainder/divisor comparison. */
++ cy = mpn_sub_n (np, np, dp, dn);
++#if OPERATION_sbpi1_div_qr_sec
++ qlp[0] += 1 - cy;
++#endif
++ mpn_addcnd_n (np, np, dp, dn, cy);
++
++#if OPERATION_sbpi1_div_qr_sec
+ /* Combine quotient halves into final quotient. */
+-#if OPERATION_sbpi1_div_qr_sec
+- qh = 0;
+- if (nn - dn != 0)
+- {
+- qh = mpn_lshift (qhp, qhp, nn - dn, GMP_NUMB_BITS/2);
+- qh += mpn_add_n (qp, qhp, qlp, nn - dn);
+- }
++ qh = mpn_lshift (qhp, qhp, nn - dn, GMP_NUMB_BITS/2);
++ qh += mpn_add_n (qp, qhp, qlp, nn - dn);
+
+ return qh;
+ #else
+
diff --git a/patches/gmp/5.1.1/130-do-not-clobber-f16-f18.patch b/patches/gmp/5.1.1/130-do-not-clobber-f16-f18.patch
new file mode 100644
index 00000000..9f9f470c
--- /dev/null
+++ b/patches/gmp/5.1.1/130-do-not-clobber-f16-f18.patch
@@ -0,0 +1,73 @@
+
+# HG changeset patch
+# User Torbjorn Granlund <tege@gmplib.org>
+# Date 1369216954 -7200
+# Node ID 394bdf8fdaee749660cc5680cc0636ea0242fae7
+# Parent 0771124b9f13b218a3f07f7dd7c127961974489d
+Don't clobber f16-f18.
+
+diff -r 0771124b9f13 -r 394bdf8fdaee mpn/ia64/divrem_2.asm
+--- a/mpn/ia64/divrem_2.asm Mon May 20 16:56:05 2013 +0200
++++ b/mpn/ia64/divrem_2.asm Wed May 22 12:02:34 2013 +0200
+@@ -1,6 +1,6 @@
+ dnl IA-64 mpn_divrem_2 -- Divide an mpn number by a normalized 2-limb number.
+
+-dnl Copyright 2010 Free Software Foundation, Inc.
++dnl Copyright 2010, 2013 Free Software Foundation, Inc.
+
+ dnl This file is part of the GNU MP Library.
+
+@@ -98,17 +98,17 @@
+ br.call.sptk.many b0 = mpn_invert_limb
+ ;;
+ setf.sig f11 = r8 // di (non-final)
+- setf.sig f18 = r39 // d1
+- setf.sig f17 = r36 // d0
++ setf.sig f34 = r39 // d1
++ setf.sig f33 = r36 // d0
+ mov r1 = r43
+ ;;
+ mov r17 = 1
+ setf.sig f9 = r38 // n2
+- xma.l f6 = f11, f18, f0 // t0 = LO(di * d1)
++ xma.l f6 = f11, f34, f0 // t0 = LO(di * d1)
+ ;;
+ setf.sig f10 = r37 // n1
+ setf.sig f15 = r17 // 1
+- xma.hu f8 = f11, f17, f0 // s0 = HI(di * d0)
++ xma.hu f8 = f11, f33, f0 // s0 = HI(di * d0)
+ ;;
+ getf.sig r17 = f6
+ getf.sig r16 = f8
+@@ -178,7 +178,7 @@
+ (p9) br.cond.dptk .L52
+ .L46:
+ ')
+- setf.sig f16 = r8 // di
++ setf.sig f32 = r8 // di
+ shladd r32 = r35, 3, r32
+ ;;
+
+@@ -189,8 +189,8 @@
+ ;;
+ (p8) mov r37 = r0
+ (p9) ld8 r37 = [r34], -8
+- xma.hu f8 = f9, f16, f10 // 0,29
+- xma.l f12 = f9, f16, f10 // 0
++ xma.hu f8 = f9, f32, f10 // 0,29
++ xma.l f12 = f9, f32, f10 // 0
+ ;;
+ getf.sig r20 = f12 // q0 4
+ xma.l f13 = f15, f8, f9 // q += n2 4
+@@ -198,8 +198,8 @@
+ ;;
+ getf.sig r18 = f13 // 8
+ xma.l f7 = f14, f13, f10 // 8
+- xma.l f6 = f17, f13, f17 // t0 = LO(d0*q+d0) 8
+- xma.hu f9 = f17, f13, f17 // t1 = HI(d0*q+d0) 9
++ xma.l f6 = f33, f13, f33 // t0 = LO(d0*q+d0) 8
++ xma.hu f9 = f33, f13, f33 // t1 = HI(d0*q+d0) 9
+ ;;
+ getf.sig r38 = f7 // n1 12
+ getf.sig r16 = f6 // 13
+
generated by cgit v1.2.3 (git 2.39.1) at 2025-05-21 20:36:35 +0000