[PATCH] D130513: [Flang] Add -fconvert option to swap endianness for unformatted files

Tue Sep 20 03:20:14 PDT 2022

awarzynski added inline comments.

================
Comment at: flang/include/flang/Optimizer/Builder/Runtime/EnvironmentDefaults.h:7
+//
+//===----------------------------------------------------------------------===//
+
----------------
Could you document what these are? And what are they used for?

================
Comment at: flang/lib/Frontend/CompilerInvocation.cpp:403-413
+  if (const auto *arg =
+          args.getLastArg(clang::driver::options::OPT_fconvert_EQ)) {
+    auto parseConvertArg = [](const char *s) {
+      return llvm::StringSwitch<std::optional<const char *>>(s)
+          .Case("unknown", "UNKNOWN")
+          .Case("native", "NATIVE")
+          .Case("little-endian", "LITTLE_ENDIAN")
----------------
I'm OK with a lambda here, just pointing our that in other cases we added small hooks, e.g. `getOptimizationLevel`. I personally prefer hooks as this means that methods like `parseFrontendArgs` can be a bit shorter. But this is a very weak preference!

================
Comment at: flang/test/Driver/convert.f90:1
+! Ensure argument -fconvert=<value> works as expected.
+
----------------
Could you be more specific? IIUC, this is more about making sure that the option parser works correctly and reports invalid usage of `-fconvert` as an error, right?

================
Comment at: flang/test/Driver/emit-mlir.f90:16
 ! CHECK-NEXT: }
+! CHECK-NEXT: fir.global @_QQEnvironmentDefaults constant : !fir.ref<tuple<i[[int_size:.*]], !fir.ref<!fir.array<0xtuple<!fir.ref<i8>, !fir.ref<i8>>>>>> {
+! CHECK-NEXT:  %[[VAL_0:.*]] = fir.zero_bits !fir.ref<tuple<i[[int_size]], !fir.ref<!fir.array<0xtuple<!fir.ref<i8>, !fir.ref<i8>>>>>>
----------------
peixin wrote:
> jpenix-quic wrote:
> > peixin wrote:
> > > jpenix-quic wrote:
> > > > peixin wrote:
> > > > > jpenix-quic wrote:
> > > > > > peixin wrote:
> > > > > > > Is it possible not to generated this global variable if `fconvert=` is not specified?
> > > > > > I'm not entirely sure--the issue I was running into was how to handle this in Fortran_main.c in a way which worked for all of GCC/Clang/Visual Studio (and maybe others?). I was originally thinking of doing this by using a weak definition of _QQEnvironmentDefaults set to nullptr so fconvert, etc. could override this definition without explicitly generating the fallback case. For GCC/clang, I think I could use __attribute__((weak)), but I wasn't sure how to handle this if someone tried to build with Visual Studio (or maybe another toolchain). I saw a few workarounds (ex: https://devblogs.microsoft.com/oldnewthing/20200731-00/?p=104024) but I shied away from this since it seems to be an undocumented feature (and presumably only helps with Visual Studio). 
> > > > > > 
> > > > > > Do you know of a better or more general way I could do this? (Or, is there non-weak symbol approach that might be better that I'm missing?)
> > > > > How about generate one runtime function with the argument of `EnvironmentDefaultList`? This will avoid this and using one extern variable?
> > > > > 
> > > > > If users use one variable with bind C name `_QQEnvironmentDefaults` in fortran or one variable with name `_QQEnvironmentDefaults` in C, it is risky. Would using the runtime function and static variable with the type `EnvironmentDefaultList` in runtime be safer?
> > > > Agreed that there are potential risks with the current approach (although, are the `_Q*` names considered reserved?). Unfortunately, I think generating a call to set the environment defaults requires somewhat significant changes to the runtime. The runtime reads environment variables during initialization in `ExecutionEnvironment::Configure` which is ultimately called from the "hardcoded" `Fortran_main.c` and I need to set the defaults before this happens. So, I believe I'd either have to move the initialization to `_QQmain`  or make it so that `main` isn't hardcoded so that I could insert the appropriate runtime function.
> > > > 
> > > > @klausler I think I asked you about this when I was first trying to figure out how to implement the environment defaults and you suggested I try the extern approach--please let me know if you have thoughts/suggestions around this!
> > > This is what @klausler suggested:
> > > ```
> > > Instead of adding new custom APIs that let command-line options control behavior in a way that is redundant with the runtime environment, I suggest that you try a more general runtime library API by which the main program can specify a default environment variable setting, or a set of them. Then turn the command-line options into the equivalent environment settings and pass them as default settings that could be overridden by the actual environment.
> > > ```
> > > If I understand correctly, what I am suggesting match his comments. The "main program" he means should be fortran main program, not the `RTNAME(ProgramStart`. In your initial patch, you add the runtime specified for "convert option". I think @klausler suggest you making the runtime argument more general used for a set of runtime environment variable settings, not restricted to "convert option". And that is what you already added -- `EnvironmentDefaultList`. So, combining this patch and your initial patch will be the solution. Hope I understand it correctly.
> > The issue I hit with the suggested approach is that in order to use the pre-existing runtime environment variable handling to set the internal state I need to set the environment variable defaults before the environment variables are read by the runtime.
> > 
> > I might be misunderstanding/missing something, but given that the environment variables are read as part of `RTNAME(ProgramStart)` in `main` and the earliest I can place the call if I am generating it is `_QQmain`, I think that leaves three options: 1. don't hardcode `main` so that I can place the call early enough 2. delay or rerun the code [[ https://github.com/llvm/llvm-project/blob/c619d4f840dcba54751ff8c5aaafce0f173a4ad5/flang/runtime/environment.cpp#L50-L90 | here ]] that is responsible for initializing the runtime state so that it is called as part of `_QQmain` so I can insert my runtime function before it or 3. hardcode something like the `_QQEnvironmentDefaults` into Fortran_main.c so that the environment defaults are available early enough. Option 2 seems less than ideal to me, option 1 seems ideal but requires generating `main`, so option 3 is what I ended up going with. If options 1 or 2 would be preferable to what is currently implemented (or if there is another possibility I'm missing!) I'd be happy to switch and try to implement them. 
> > 
> What do other reviewers think? @klausler @awarzynski @kiranchandramohan @clementval @jeanPerier 
> The current approach has two drawbacks:
> 1. Add one extern variable `_QQEnvironmentDefaults` in runtime.
> 2. Generate the variable during lowering even if there is no `-fconvert` option.
> 
> Can we accept this?
Based on the analysis by @jpenix-quic, this is the path of least resistance and SGTM. It can always be refactored/improved in the future if need be. It would be helpful if this was documented  somewhere. But I'm not sure whether there's a good landing space for this ATM.

================
Comment at: flang/test/Lower/environment-defaults.f90:1
+! RUN: bbc -emit-fir -o - %s | FileCheck %s
+
----------------
Can you test with `flang-new` as well?

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130513/new/

https://reviews.llvm.org/D130513