Some parts of CodeGen rely on the IR order (e.g. ScheduleDAGRRList) and debug intrinsics previously affected the ordering. This problem was found using check_cfc.py. With this patch, the build failures for llvm's test-suite with check_cfc.py go from 113 down to 77 in x86_64.
Details
Details
Diff Detail
Diff Detail
Event Timeline
Comment Actions
Can't say that I fully understand the implication of this (I don't know what the node order is used for), but given your explanation this appears to be fine.
lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp | ||
---|---|---|
4975 ↗ | (On Diff #79757) | Extra whitespace :-) |
lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp | ||
---|---|---|
873 ↗ | (On Diff #79757) | Any way the purpose of this condition could be documented? |
lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp | ||
---|---|---|
873 ↗ | (On Diff #79757) | I hope my comment change makes is clearer. |
Comment Actions
After having a closer look at D25318 I think it solves the problem in a slightly better way.