Bug in solve

[jgdumas]

cat /tmp/f3.sms
3 3 M
1 1 1
2 2 1
2 3 1
3 2 -1
3 3 2
0 0 0
./rank /tmp/f3.sms
matrix is 3 by 3 (0.000211s (0.000207 cpu) [1])
Rank is 3 (0.00138617s (2e-06 cpu) [1] )

---

But with

cat /tmp/b3.txt
-379491943 1054657936 583190604
./solve /tmp/f3.sms /tmp/b3.txt
Computation is done over Q
A is 3 by 3
A:=[[1, 0, 0 ], [0, 1, 1 ], [0, -1, 2 ]];
B is [-379491943 1054657936 583190604 ]
ERROR in reconstruction ? (3)

try to reconstruct :
approximation: 1230831354298319
modulus: 3692492536769689
numbound: 4294967296
denbound: 2
(DenseElimination) Solution is [-379491943 1526125268 0 ] / 0

---

while with a smaller vector solution is less obviously wrong :

cat /tmp/b2.txt
-3491943 157936 590604
./solve /tmp/f3.sms /tmp/b2.txt
(DenseElimination) Solution is [-6983886 21398639 -21082767 ] / 2
(SparseElimination) Solution is [-10475829 -274732 748540 ] / 3

---

with even smaller, seems correct

./solve /tmp/f3.sms
Computation is done over Q
A is 3 by 3
A:=[[1, 0, 0 ], [0, 1, 1 ], [0, -1, 2 ]];
Creating a random {-1,1} vector U, B is AU
B is [-1 0 -3 ]
(DenseElimination) Solution is [-1 1 -1 ] / 1
(SparseElimination) Solution is [-1 1 -1 ] / 1

[jgdumas]

Problem is :
PRESENT on Xeon E5345 with blis
PRESENT on Xeon E5-4620 with openblas
PRESENT on Xeon E5-4620 with blis
ABSENT on Xeon E5-4620 with --stable=yes
PRESENT on i7-6700 with openblas
ABSENT on Gold 6126 with blis
ABSENT on breush machine
PRESENT on cpernet machine i7-8650U
PRESENT on i7-6600U, where an old executable from Nov. 30, 2018 does not show it

[jgdumas]

Seems related to test-smith-valence looping forever (bad solve of inv mat and non-zero vector gives a zero solution ...)

[jgdumas]

fixed in commit c593941