[med-svn] [Git][med-team/mrbayes][master] 2 commits: Fix i386 rate normalization infinite loop
Andreas Tille (@tille)
gitlab at salsa.debian.org
Mon Nov 9 07:22:12 GMT 2026
Andreas Tille pushed to branch master at Debian Med / mrbayes
Commits:
e5d4efe7 by Andreas Tille at 2026-11-09T08:11:50+01:00
Fix i386 rate normalization infinite loop
On i386, x87 excess precision may evaluate the float RATE_MIN constant in wider precision than the double rate values. A rate that has just been clamped to RATE_MIN then still compares as strictly smaller than RATE_MIN, so the normalization loops in Move_Revmat_Dir and Move_Revmat_DirMix clamp it again on every iteration and never terminate, making the build-time tests hang.
Use a MrBFlt copy of RATE_MIN for the comparisons so that the threshold and the rate values are always evaluated at the same precision.
Assisted-By: DeepSeek V4.1 Flash Free
- - - - -
bf4d314b by Andreas Tille at 2026-11-09T08:21:45+01:00
Upload to unstable
- - - - -
3 changed files:
- debian/changelog
- + debian/patches/fix_i386_rate_min_loop.patch
- debian/patches/series
Changes:
=====================================
debian/changelog
=====================================
@@ -1,11 +1,11 @@
-mrbayes (3.2.9-1) UNRELEASED; urgency=medium
+mrbayes (3.2.9-1) unstable; urgency=medium
- TODO: Fix build time test for i386
* New upstream version
* Drop 'Rules-Requires-Root: no' from d/control (routine-update)
* Use dh_missing --fail-missing default
+ * Fix build-time test hang on i386: use a double-precision copy of RATE_MIN
- -- Andreas Tille <tille at debian.org> Thu, 05 Nov 2026 15:58:23 +0100
+ -- Andreas Tille <tille at debian.org> Mon, 09 Nov 2026 08:12:12 +0100
mrbayes (3.2.7a-8) unstable; urgency=medium
=====================================
debian/patches/fix_i386_rate_min_loop.patch
=====================================
@@ -0,0 +1,116 @@
+Description: Fix infinite rate-normalization loop on i386
+ The rate normalization loops in Move_Revmat_Dir and Move_Revmat_DirMix
+ compare double-precision rate values against the float constant RATE_MIN.
+ On i386, x87 excess precision may evaluate the constant in wider precision
+ than the stored double value, so that a rate which has just been clamped to
+ RATE_MIN still compares as strictly smaller than RATE_MIN. The loop then
+ clamps it again on every iteration and never terminates, which makes the
+ build-time test suite hang on i386.
+ .
+ Use a MrBFlt (double) copy of RATE_MIN for the comparisons so that the
+ threshold and the rate values always have the same precision.
+Author: Andreas Tille <tille at debian.org>
+Last-Update: 2026-11-09
+Assisted-By: DeepSeek V4.1 Flash Free
+
+diff --git a/src/proposal.c b/src/proposal.c
+index 4e1f7b7..4691eb6 100644
+--- a/src/proposal.c
++++ b/src/proposal.c
+@@ -14873,7 +14873,7 @@ int Move_Revmat_Dir (Param *param, int chain, RandLong *seed, MrBFlt *lnPriorRat
+ /* change revMat using Dirichlet proposal */
+
+ int i, nRates,isValid;
+- MrBFlt oldRate[200], newRate[200], dirParm[200], *value, sum, x, y, rate_pot, *alphaDir, alphaPi;
++ MrBFlt oldRate[200], newRate[200], dirParm[200], *value, sum, x, y, rate_pot, rateMin, *alphaDir, alphaPi;
+ ModelParams *mp;
+ ModelInfo *m;
+
+@@ -14906,6 +14906,12 @@ int Move_Revmat_Dir (Param *param, int chain, RandLong *seed, MrBFlt *lnPriorRat
+ DirichletRandomVariable (dirParm, newRate, nRates, seed);
+
+ /* check new values. we rely on newRate be already normalized */
++ /* Use a MrBFlt copy of RATE_MIN for the comparisons below. RATE_MIN
++ is a float constant; on i386 (x87 excess precision) it may be
++ evaluated in wider precision than the double values it is compared
++ against, so that a value stored as RATE_MIN still tests as smaller
++ than RATE_MIN. The normalization loop would then never terminate. */
++ rateMin = RATE_MIN;
+ while (1)
+ {
+ sum = 0.0;
+@@ -14913,14 +14919,14 @@ int Move_Revmat_Dir (Param *param, int chain, RandLong *seed, MrBFlt *lnPriorRat
+ isValid=1;
+ for (i=0; i<nRates; i++)
+ {
+- if (newRate[i] <= RATE_MIN)
++ if (newRate[i] <= rateMin)
+ {
+- if (newRate[i] < RATE_MIN)
++ if (newRate[i] < rateMin)
+ {
+- newRate[i] = RATE_MIN;
++ newRate[i] = rateMin;
+ isValid=0;
+ }
+- rate_pot -= RATE_MIN;
++ rate_pot -= rateMin;
+ }
+ else
+ sum += newRate[i];
+@@ -14928,7 +14934,7 @@ int Move_Revmat_Dir (Param *param, int chain, RandLong *seed, MrBFlt *lnPriorRat
+ if (isValid==1) break;
+ for (i=0; i<nRates; i++)
+ {
+- if (newRate[i]!=RATE_MIN)
++ if (newRate[i]!=rateMin)
+ newRate[i] = rate_pot * newRate[i] / sum;
+ }
+ }
+@@ -14988,7 +14994,7 @@ int Move_Revmat_Dir (Param *param, int chain, RandLong *seed, MrBFlt *lnPriorRat
+ int Move_Revmat_DirMix (Param *param, int chain, RandLong *seed, MrBFlt *lnPriorRatio, MrBFlt *lnProposalRatio, MrBFlt *mvp)
+ {
+ int i, j, k, isValid, *growthFxn, nRates, groupSize[6];
+- MrBFlt *value, dirParm[6], newRate[6], oldRate[6], alphaPi, symDir, sum, rate_pot, x, y;
++ MrBFlt *value, dirParm[6], newRate[6], oldRate[6], alphaPi, symDir, sum, rate_pot, rateMin, x, y;
+ ModelParams *mp;
+
+ /* get model params and settings */
+@@ -15026,6 +15032,9 @@ int Move_Revmat_DirMix (Param *param, int chain, RandLong *seed, MrBFlt *lnPrior
+ DirichletRandomVariable (dirParm, newRate, nRates, seed);
+
+ /* check new values. we rely on newRate be already normalized */
++ /* Use a MrBFlt copy of RATE_MIN so that the comparisons are done
++ consistently in double precision (see Move_Revmat_Dir). */
++ rateMin = RATE_MIN;
+ while (1)
+ {
+ sum = 0.0;
+@@ -15033,14 +15042,14 @@ int Move_Revmat_DirMix (Param *param, int chain, RandLong *seed, MrBFlt *lnPrior
+ isValid=1;
+ for (i=0; i<nRates; i++)
+ {
+- if (newRate[i] <= RATE_MIN)
++ if (newRate[i] <= rateMin)
+ {
+- if (newRate[i] < RATE_MIN)
++ if (newRate[i] < rateMin)
+ {
+- newRate[i] = RATE_MIN;
++ newRate[i] = rateMin;
+ isValid=0;
+ }
+- rate_pot -= RATE_MIN;
++ rate_pot -= rateMin;
+ }
+ else
+ sum += newRate[i];
+@@ -15048,7 +15057,7 @@ int Move_Revmat_DirMix (Param *param, int chain, RandLong *seed, MrBFlt *lnPrior
+ if (isValid==1) break;
+ for (i=0; i<nRates; i++)
+ {
+- if (newRate[i] != RATE_MIN)
++ if (newRate[i] != rateMin)
+ newRate[i] = rate_pot * newRate[i] / sum;
+ }
+ }
=====================================
debian/patches/series
=====================================
@@ -1,3 +1,4 @@
no_rpath.patch
autoreconf.patch
no_baseline_violation.patch
+fix_i386_rate_min_loop.patch
View it on GitLab: https://salsa.debian.org/med-team/mrbayes/-/compare/ee1f23fe273b1657b2baf7e817a3a22dec11d6eb...bf4d314b79d075f5b6cf5abc1332a8c014e10631
--
View it on GitLab: https://salsa.debian.org/med-team/mrbayes/-/compare/ee1f23fe273b1657b2baf7e817a3a22dec11d6eb...bf4d314b79d075f5b6cf5abc1332a8c014e10631
You're receiving this email because of your account on salsa.debian.org. Manage all notifications: https://salsa.debian.org/-/profile/notifications | Help: https://salsa.debian.org/help
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://alioth-lists.debian.net/pipermail/debian-med-commit/attachments/20261109/7999bcbb/attachment-0001.htm>
More information about the debian-med-commit
mailing list