[xiaoyeli]

Summary

With -DXSDK_INDEX_SIZE=64, any call with lwork > 0 (L and U stored in a user-supplied work[] array) corrupted the factors. Reproducer:

cmake -B build -DXSDK_INDEX_SIZE=64 -Denable_internal_blaslib=ON
cmake --build build
cd build/TESTING && ./s_test -t "LA" -l 1000000 -n 9

This segfaulted in sPivotGrowth (called from sgssvx) on the first factorization; d_test aborted in free() instead. lwork == 0 and the 32-bit index build were unaffected.

Root cause

In xLUMemInit (SRC/{s,d,c,z}memory.c), the user-memory branch carves the column pointer arrays xlsub, xlusup and xusub out of work[] with (n+1) * iword bytes each, where iword = sizeof(int). Those arrays are int_t, so with 64-bit indices each gets half the bytes it needs. xusub overruns into the lusup/ucol block placed immediately after it, and the numeric values in turn overwrite the U column pointers. The heap path uses intMalloc(n+1), which is sized correctly, which is why only lwork > 0 failed.

Changes

In all four xmemory.c files:

• Allocate xlsub, xlusup and xusub with sizeof(int_t) (new local liword, matching the naming already used in xmemory_usage). xsup and supno remain int.
• Use sizeof(int_t) for the lsub/usub share in the matching xuser_free of the shrink loop.
• Make the lwork == -1 estimate count the two int arrays and the three int_t arrays separately, and the lsub/usub blocks with sizeof(int_t), so the reported requirement is sufficient on 64-bit builds.
• xStackCompress (currently not called anywhere) computed the byte offsets of the int_t lsub/usub blocks with sizeof(int) and the end of ucol with the integer size instead of the value size. Corrected for consistency.

Testing

Build	Before	After
64-bit index, s_test -t LA -l 1000000 -n 9	segfault	passes, 1158 tests
64-bit index, s/d/c/z testers, lwork 0 / 1000000 / 10000000, n 9 / 19, and the g20.rua SP case	segfault or abort for lwork > 0	all pass
64-bit index, ctest (24 tests)	–	100% pass
32-bit index, s/d/c/z testers and ctest (24 tests)	pass	pass

The corresponding change still needs to be made in SRC/xmemory.base of the base repository; I can open that PR separately.

---

Generated by Claude Code