Can we retire the yk-no-fallthrough ykllvm hack?

[vext01]

This hack is slowing AOT code down:

+----+-------------------------------------------------------------------------------+----------------+---------+
| #  | flags turned off (O2 passes are set)                                          | execution time | 95% CI  |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 1  | baseline (all flags are turned on)                                            | 89.9ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 2  | —yk-no-fallthrough                                                            | 76.1ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 3  | —yk-block-disambiguate                                                        | 78.7ms         | 0.00189 |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 4  | —yk-patch-control-point                                                       | 84.4ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 5  | —yk-insert-stackmaps                                                          | 85.8ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 6  | —yk-split-blocks-after-calls                                                  | 82ms           | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 7  | —yk-no-fallthrough + —yk-block-disambiguate                                   | 68.4ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 8  | —yk-no-fallthrough + —yk-block-disambiguate + —yk-patch-control-point         | 63.9ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 9  | —yk-no-fallthrough + —yk-block-disambiguate + —yk-insert-stackmaps            | 58.4ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 10 | —yk-no-fallthrough + —yk-block-disambiguate + —yk-split-blocks-after-calls    | 66.8ms         | 0.00189 |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 11 | —yk-no-fallthrough + —yk-block-disambiguate + —yk-patch-control-point +       |                |         |
|    | —yk-insert-stackmaps                                                          | 55.1ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 12 | —yk-no-fallthrough + —yk-block-disambiguate + —yk-patch-control-point +       |                |         |
|    | —yk-split-blocks-after-calls                                                  | 64.2ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 13 | —yk-no-fallthrough + —yk-block-disambiguate + —yk-insert-stackmaps +          |                |         |
|    | —yk-split-blocks-after-calls                                                  | 58.2ms         | 0.002   |
+----+-------------------------------------------------------------------------------+----------------+---------+
| 14 | —yk-no-fallthrough + —yk-block-disambiguate + —yk-patch-control-point +       |                |         |
|    | —yk-insert-stackmaps + —yk-split-blocks-after-calls                           | 54.5ms         | 0.003   |
+----+-------------------------------------------------------------------------------+----------------+

(Data collected by @nmdis1999)

The pass also seems to compound with some of our other passes making even worse AOT performance.

As far as I can see, this hack stops consecutive blocks from falling through (by keeping jump terminators around). This seems to be because PT would see fallthrough as one large block instead of two (or small) smaller blocks and that might break the mapper.

Questions:

• Does it break the mapper though?
• If it does, why should it? Blockmap queries have the ability to return more than one block. Combined with the corr_bbs stuff, in my head this shouldn't be an issue...

[ltratt]

@nmdis1999 Has there been any movement on this?