Add more expectations in vector.ll and select.ll based on command-line option combinations.
Also, remove hard-coded shadow width references to enable fast8 transition.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
llvm/test/Instrumentation/DataFlowSanitizer/call.ll | ||
---|---|---|
2 | fast16 and legacy modes have the same code path tested by this file: it does not check ld/st much. | |
llvm/test/Instrumentation/DataFlowSanitizer/vector.ll | ||
11–12 | Does ARGS_ABI-LABEL @"dfs$pass_vector"(<4 x i4> %[[VEC:.*]], i[[#SBITS]] %[[LABEL:.*]]) also work? |
llvm/test/Instrumentation/DataFlowSanitizer/call.ll | ||
---|---|---|
2 | That's true but the legacy mode will be removed eventually. To enable this transition while retaining the same test coverage we will have to make all tests runnable under fast labels. I think having both options for a while, even though redundant (as you point out), is safer. When we are ready to deprecate the default legacy mode, removing a few RUN lines will be guaranteed to work no matter how much the code has diverged between the two by then. | |
llvm/test/Instrumentation/DataFlowSanitizer/vector.ll | ||
11–12 | No. The *-LABEL lines cannot include any variables. |
llvm/test/Instrumentation/DataFlowSanitizer/vector.ll | ||
---|---|---|
11–12 | Got it. Thanks. |
fast16 and legacy modes have the same code path tested by this file: it does not check ld/st much.
It is fine to ignore -dfsan-fast-16-labels.
We may add the new 8bit mode if its codegen introduces different code path.