In PR https://github.com/compiler-explorer/compiler-explorer/pull/7976,
I forgot to handle the case where same instructions have different
descriptions. This PR fixes the script and removes the duplicates from
the generated file.
The SASS docs are extracted from multiple tables from
https://docs.nvidia.com/cuda/cuda-binary-utilities/index.html#instruction-set-reference.
For example,
- Table 4. Maxwell and Pascal Instruction Set
- ...
- Table 8. Hopper Instruction Set
- Table 8. Blackwell Instruction Set (wait...what? both tables are table
8? whatever...)
Previously, only the duplications for the (opcode, descriptions) pair
were removed. In this PR, we only check the duplications based on the
opcode and only keep the last one (based on the assumption that the
newer ones are better).
I manually went through the removed duplicates, and found that
- No instruction actually changed the meaning across generations (I
would be very surprised if they did so).
- Most of the changes are minor fixes like changing capitalization
(e.g., `generic Memory` -> `Generic Memory`) or fixing typo (e.g.,
`Local within Shared Memory Window` -> `Load within Shared Memory
Window`)
- A couple of the descriptions are for whatever reason simplified (e.g.,
`LOP3`: `3-input Logic Operation` -> `Logic Operation`), which is a bit
unfortunate.
9c72429e08