[PATCH] D84345: [AMDGPU] Set the default globals address space to 1

[Alexander Richardson via Phabricator via llvm-commits]

arichardson added inline comments.


================
Comment at: llvm/lib/IR/AutoUpgrade.cpp:4297
+  // address space of 1.
+  if (T.isAMDGPU() && !DL.contains("-G") && !DL.startswith("G")) {
+    return DL.empty() ? std::string("G1") : (DL + "-G1").str();
----------------
dylanmckay wrote:
> dylanmckay wrote:
> > arichardson wrote:
> > > akhuang wrote:
> > > > arichardson wrote:
> > > > > arsenm wrote:
> > > > > > I would expect datalayout upgrades to work by parsing the old string, and checking the field values inside. I guess directly checking the string isn't a new problem here
> > > > > I agree that would be less error prone. I wonder if there are cases where the old string may fail to parse so you have to do the textual upgrade first. I'm happy to make this change.
> > > > > 
> > > > > @akhuang is there a reason you used string parsing in D67631? Any objections to changing the code to parse the datalayout and add missing attributes?
> > > > I don't think so; parsing the datalayout sounds better to me too. 
> > > I just looked into parsing the DataLayout instead. Unfortunately the resulting code is more complicated since there are no setters in DataLayout and no way to create a normalized representation.
> > > There's also no way to differentiate between no `-G ` passed and `-G0` so  something like `e-p:64:64-G0` will be converted to `e-p:64:64-G0-G1`
> > > 
> > I suspect it would be possible to use the existing `DataLayout(StringRef)` constructor on the string, then call `getDefaultGlobalsAddressSpace()` on it, explicitly ignoring modifying the datalayout for the special case of an explicit `-G0`.
> > 
> > For example,
> > 
> > ```cpp
> >   DataLayout ParsedDL = DataLayout(DL);
> >   if (T.isAMDGPU() && !DL.contains("-G0") &&ParsedDL.getDefaultGlobalsAddressSpace() != 1) {
> >     return DL.empty() ? std::string("G1") : (DL + "-G1").str();
> >   }
> > ```
> > 
> > As I understand it, this would cover the fact that we cannot distinguish between an explicit default globals space of zero, and a datalayout without a default globals space (also `DL::getDefaultGlobalsAddressSpace() == 0`) by explicitly excluding the special case `-G0`
> To be completely correct it should not assume that the global address space is not at the very start of the data layout as my initial snippet did. I've removed the `-` prefix from the `contains` check
> 
> ```
>   DataLayout ParsedDL = DataLayout(DL);
>   if (T.isAMDGPU() && !DL.contains("G0") &&ParsedDL.getDefaultGlobalsAddressSpace() != 1) {
>     return DL.empty() ? std::string("G1") : (DL + "-G1").str();
>   }
> ```
We then end up overwriting explicitly specified `G<N>` flags. Maybe some tests would like to check that overriding the globals address space works? Also I believe this can result in an invalid datalayout: `"...-G2"` will be converted to `"...-G2-G1"`.

Not sure what the correct approach is @arsenm ?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D84345/new/

https://reviews.llvm.org/D84345