[rom_ctrl,util] Rework memory collision check

[rivos-eblot]

As ROMs get bigger, scrambled memory collision checker in mem.py becomes slower.

On my local machine using this revised implementation, scrambling the second rom only takes 1.75 seconds vs. 7.68 seconds with the previous implementation (> 4x speed up).

I've tested collisions detections forcing several of them and checking the collision report remains the same:

diff --git a/hw/ip/rom_ctrl/util/scramble_image.py b/hw/ip/rom_ctrl/util/scramble_image.py
index c3e1cd0d63..6d70c2144e 100755
--- a/hw/ip/rom_ctrl/util/scramble_image.py
+++ b/hw/ip/rom_ctrl/util/scramble_image.py
@@ -493,6 +493,8 @@ def main() -> int:
         scr_mem.add_ecc32()
         assert scr_mem.width == 39

+    scr_mem.chunks[0].words[4] = scr_mem.chunks[0].words[78] = scr_mem.chunks[0].words[453]
+    scr_mem.chunks[0].words[12] = scr_mem.chunks[0].words[1000]
     # Check for collisions
     collisions = scr_mem.collisions()
     if collisions:

[rswarbrick]

This looks sensible to me (and I clearly did something a little silly with the original implementation). But I think we can be lazier: we don't actually care all that much about the collisions, only about whether they exist. So a better approach would probably be something like this: (completely untested!)

    def first_collision(self) -> Optional[Tuple[int, int]]:
        '''Return the address of the first pair of colliding addresses

        If there is no such pair (which is hopefully the case), return None.
        '''
        known = {}
        for idx, chunk in enumerate(self.chunks):
            for word_idx, word in enumerate(chunk.words):
                addr = chunk.base_addr + word_idx
                if word in known:
                    return (known[word], addr)
                else:
                    known[word] = addr
        return None

Then the call site would look something like this:

    collision = scr_mem.first_collision()
    if collision is not None:
        print(
            'ERROR: This combination of ROM contents and scrambling\n'
            '       key results in one or more collisions where\n'
            '       different addresses have the same data.\n'
            '\n'
            '       Looks like we\'ve been (very) unlucky with the\n'
            '       birthday problem. As a work-around, try again after\n'
            '       generating some different RndCnst* parameters.\n',
            file=sys.stderr)
        addr0, addr1 = collision
        print(f'First colliding addresses: {addr0:#010x}, {addr1:#010x}')
        return 1

[rivos-eblot]

Sure. I wrote the patch to get the exact same output as the original implementation. If only the first collision is of some interest, I can simplify this quite a bit. Pls let me know.

[rivos-eblot]

Hi @rswarbrick. Is this new version ok for you? It seems the CI fails with some unrelated issues to this change though.

[rswarbrick]

I'm really sorry for not replying before. Yes, this looks really good to me!

[rivos-eblot]

cc @sameo

[rivos-eblot]

Thanks