fuzz: Add libfuzzer compatible bounds fuzzer #7549
Conversation
|
See #4088 |
test/fuzz/bounds.cpp
Outdated
| @@ -96,108 +97,98 @@ Expr random_condition(Type T, int depth, bool maybe_scalar) { | |||
| }; | |||
| const int op_count = sizeof(make_bin_op) / sizeof(make_bin_op[0]); | |||
There was a problem hiding this comment.
This is unused, it must be removed
|
I took the liberty of harvesting fixes from #7551 and #7548 and applying them here (since our buildbot doesn't test this stuff, I pulled this into Google to test)... unfortunately it's still broken, as running it with default args ends up trying to construct a binary operation with vector elements with mismatched lanes. Gonna debug some more. Specifically: ...it seems puzzling that we would be generating random expressions that return vector-expressions with a lane count different from the input types; I wouldn't think that would ever work. |
|
hrm, this may have found a real bug: it seems we are generating two uint32x6: But if we attempt to cast either to uint1x6 (ie, bool), we fail inside cast at IROperator.cpp:1946 because we don't expect the type to be a combination of Broadcast-and-Ramp. Pondering the right way to handle this. |
|
See #7556 for possible fix |
|
LGTM. There are still issues to be resolved with the fuzzers but this change seems OK. |
|
ready to land? |
|
All good on my end! |
* fuzz: Add libfuzzer compatible bounds fuzzer * Remove unused constant * Style fix * Fix handling of binary ops * Handle casting to vector-of-bool properly * fuzz: Alphabetically sort targets in CMake --------- Co-authored-by: Steven Johnson <srj@google.com>
No description provided.