Skip to content

LoongArch64: avoid SGEMM LA464 ncopy fallthrough from N2 to N1#5902

Merged
martin-frbg merged 1 commit into
OpenMathLib:developfrom
wxpppp:develop
Jul 11, 2026
Merged

LoongArch64: avoid SGEMM LA464 ncopy fallthrough from N2 to N1#5902
martin-frbg merged 1 commit into
OpenMathLib:developfrom
wxpppp:develop

Conversation

@wxpppp

@wxpppp wxpppp commented Jul 9, 2026

Copy link
Copy Markdown
Contributor

Problem

The .L_N2 tail path handles a remaining 2-column block. After that path completes, the original code could directly jump or fall through into .L_N1.

However, .L_N1 handles a remaining 1-column block, so entering .L_N1 should depend on N & 1.

The original code used M & 1:

.L_N2_M1:
    andi      I,     M,    0x01
    beq       I,     ZERO, .L_N1

M & 1 only indicates whether the current 2-column block has one remaining row. It does not indicate whether there is one remaining column.

This can cause the kernel to process an extra column after .L_N2, leading to an out-of-bounds read when the column tail contains 2 but not 1.

For example:

N % 8 == 2
N % 8 == 6

Fix

Add an .L_N2_END block and check N & 1 before falling through into .L_N1.

.L_N2_M1:
    andi      I,     M,    0x01
    beq       I,     ZERO, .L_N2_END
    ...

.L_N2_END:
    andi      J,     N,   0x01
    beq       ZERO,  J,   .L_N0

.align 5
.L_N1:

@martin-frbg martin-frbg added this to the 0.3.34 milestone Jul 11, 2026
@martin-frbg

Copy link
Copy Markdown
Collaborator

Thank you

@martin-frbg martin-frbg merged commit 992a536 into OpenMathLib:develop Jul 11, 2026
101 of 106 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants