CBLAS: fix infinite loop in c/z gemv and gbmv (RowMajor, ConjTrans, M=0)

[EylonKrause]

Problem

The hand-written row-major shims in the C interface loop forever on a valid (if degenerate) call.

In cblas_cgemv, cblas_zgemv, cblas_cgbmv and cblas_zgbmv, the CblasRowMajor + CblasConjTrans path conjugates Y (and sets the loop stride i and the sentinel st) only inside:

if (M > 0) {
   ...
   if (N > 0) { i = ...; st = y + n; do { *y = -(*y); y += i; } while (y != st); ... }
}

but the post-call loop that un-conjugates Y is gated on if (N > 0) alone:

if (x != X) free(x);
if (N > 0) {           /* <-- runs even when M == 0 */
   do { *y = -(*y); y += i; } while (y != st);
}

When M == 0 and N > 0, the forward block is skipped, so i is still 0 and st is still NULL. The restore loop then advances y by 0 and compares against NULL, so it never terminates — an infinite loop that also keeps negating Y[0]. M = 0 is a documented, in-range argument for these routines (xerbla only rejects M < 0), so this is reachable from a conforming caller.

Fix

Gate the un-conjugation on the same condition under which Y was conjugated, if (M > 0 && N > 0), in all four files. One line each; a normal M > 0 ConjTrans call is byte-for-byte unchanged.

Reproducer / validation

BLIS carries a copy of these same netlib CBLAS wrappers, so I reproduced it there against a real build:

float A[256]={0}, X[256]={0}, Y[256]={0};
float alpha[2]={1,0}, beta[2]={1,0};
cblas_cgbmv(CblasRowMajor, CblasConjTrans, /*M*/0, /*N*/2, /*KL*/0, /*KU*/0,
            alpha, A, 1, X, 1, beta, Y, 1);   /* never returns */

Linked against the unpatched wrapper this is killed by a 5 s timeout; with the one-line gate it returns immediately and leaves Y unchanged. The gemv path behaves the same way.

Disclosure: this contribution was authored with an AI coding assistant (Claude) and reviewed and validated before submission.

[langou]

Thanks @EylonKrause ! Well spotted, great fix.

[codecov]

❌ 4 Tests Failed:

Tests completed	Failed	Passed	Skipped
236	4	232	0

View the full list of 4 ❄️ flaky test(s)

LAPACK.eig::ded.out (DOUBLE PRECISION Nonsymmetric Eigenvalue)

Flake rate in main: 40.97% (Passed 657 times, Failed 456 times)

Stack Traces | 0.31s run time

4 numerical error(s), 4 other error(s) (illegal: 0, info: 4), 14678 test(s) run
 DDRVES: DGEES1 returned INFO=     6.
 DDRVES: DGEES1 returned INFO=     6.
 DES:    2 out of  3810 tests failed to pass the threshold
 DGET24: DGEESX1 returned INFO=     6.
 DGET24: DGEESX1 returned INFO=     6.
 DSX:    2 out of  3494 tests failed to pass the threshold

LAPACK.eig::snep.out (REAL Nonsymmetric Eigenvalue Problem)

Flake rate in main: 28.48% (Passed 796 times, Failed 317 times)

Stack Traces | 0.03s run time

1 numerical error(s), 10080 test(s) run
 SHS:    1 out of  2016 tests failed to pass the threshold

LAPACK_64.eig::ded_64.out (DOUBLE PRECISION Nonsymmetric Eigenvalue)

Flake rate in main: 38.45% (Passed 730 times, Failed 456 times)

Stack Traces | 0.3s run time

4 numerical error(s), 4 other error(s) (illegal: 0, info: 4), 14678 test(s) run
 DDRVES_64: DGEES1 returned INFO=     6.
 DDRVES_64: DGEES1 returned INFO=     6.
 DES:    2 out of  3810 tests failed to pass the threshold
 DGET24_64: DGEESX1 returned INFO=     6.
 DGET24_64: DGEESX1 returned INFO=     6.
 DSX:    2 out of  3494 tests failed to pass the threshold

LAPACK_64.eig::snep_64.out (REAL Nonsymmetric Eigenvalue Problem)

Flake rate in main: 26.73% (Passed 869 times, Failed 317 times)

Stack Traces | 0.03s run time

1 numerical error(s), 10080 test(s) run
 SHS:    1 out of  2016 tests failed to pass the threshold

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}

[EylonKrause]

Closing this — I'll route this fix through a different channel rather than here. Sorry for the noise.