hexagon: find software divide calls using binary inspection tool

[trivikram-reddy1]

Overview

Add --swdiv mode to ggml-hexagon-inspect.py that lists functions calling the Hexagon software divide/modulo helpers. Hexagon has no hardware unit for division, so these calls are expensive, and this mode helps find hot paths that use them.

This PR also adds presets to build HTP skel with DWARF sections needed for hexagon-addr2line to map address to line numbers in src code.

Additional information

Details

$ ./scripts/snapdragon/ggml-hexagon-inspect.py --swdiv --lib pkg-adb/master-dbginfo/lib/libggml-htp-v81.so

Address    | Function                                    | In-Loop | Total | Call Sites
-----------|---------------------------------------------|---------|-------|----------------------------------
0x00093180 | pad_job_per_thread_hvx_circular_dma         |       0 |    21 | +0x260, pad-ops.c:396, pad-ops.c:397, pad-ops.c:414, pad-ops.c:458, pad-ops.c:459
0x00008d20 | op_matmul                                   |       0 |    13 | matmul-ops.c:3777, matmul-ops.c:3780
0x00046fc0 | op_binary                                   |       0 |    11 | binary-ops.c:1181
0x00086d20 | op_cpy                                      |       0 |    11 | cpy-ops.c:511
...
Total functions analyzed: 533 | Reported: 70 | Functions with sw divides: 70 | Total sw divide calls: 258 | In-loop: 0

$ ./scripts/snapdragon/ggml-hexagon-inspect.py --swdiv --inline --func "op_cpy" --lib pkg-adb/master-dbginfo/lib/libggml-htp-v81.so
--------------------------------------------------------------------------------------------------------------
Address    | Function                                 | In-Loop | Total | Call Sites
-----------|------------------------------------------|---------|-------|-------------------------------------
0x00086d20 | op_cpy                                   |       0 |    11 | cpy-ops.c:511
    +0x294   udivdi3  hex-fastdiv.h:22 <- cpy-ops.c:412 <- cpy-ops.c:511
    +0x2c8   udivdi3  hex-fastdiv.h:22 <- cpy-ops.c:413 <- cpy-ops.c:511
    +0x328   umodsi3  hex-common.h:83 <- cpy-ops.c:419 <- cpy-ops.c:511
    +0x338   udivsi3  cpy-ops.c:419 <- cpy-ops.c:511
    +0x484   udivdi3  hex-fastdiv.h:22 <- cpy-ops.c:468 <- cpy-ops.c:511
    ...

Because of LTO and inlining, most divides sit inside helpers several levels below the reported function. In the table above, all 11 divides in op_cpy show up as one line (cpy-ops.c:511). --inline prints one line per call with its full inline chain, innermost frame first and the op code that led to it (whether it is a 32-bit or a much slower 64-bit (udivdi3) divide).

Requirements

• I have read and agree with the contributing guidelines Yes
• AI usage disclosure: Yes to refactor code

[max-krasnyansky]

@lhez for second ack