So it's easier to test with it on by default.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | false is the default value so this appears to be NFC... |
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | I can more easily change the default value to true and test this way. |
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | Adding/removing cl::init(true) is very straightforward. Why is this NFC change needed? |
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | It's easier for me to replace "false" with "true" than to type the whole thing out. |
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | If adding a cmake variable is like moving from 0 to 1, then this NFC patch seems to me 0 to 0.1 I wonder why we didn't move directly to 1. This patch added unnecessary churn to the history. |
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | Ok, maybe the CMake argument isn't so strong, but this patch still does make my life just a little easier when testing NPM with check-llvm. I'm not sure where the threshold for a patch's usefulness is for it to be worth a commit. |
llvm/tools/opt/opt.cpp | ||
---|---|---|
75 | I don't think churning the version history is a concern here, and I don't see that the patch is doing anything weird here, just spelling out the default value and making it easier to flip with a one-liner. If this makes Arthur's work towards enabling the new pm easier, I believe that certainly puts it above the threshold. |
false is the default value so this appears to be NFC...