[GlobalISel][TableGen] Fully qualify calls to llvm::format that have ::std:: types as paramters

[barcharcraz]

Summary

Hello, I work on the Microsoft standard standard library and we encountered a build failure upon enabling std::format in /std:c++latest mode.

This showed up in commit 4c263ed, it's "fixed" in main, but the underlying issue festers and will cause a new failure if headers are rearranged such that GIMatchDiag.cpp sees the content of <format>.

On C++ standard versions that have std::format these calls become ambiguous with std::format due to ADL triggered by the unique_ptr* parameters.

On another note it's a little odd to format &N as a pointer when N is a unique_ptr, that's for someone else to consider though, I don't know enough about the code in question.

This is my first patch for llvm, so do let me know if I've screwed up the submission process.

[Paul-C-Anagnostopoulos]

Did you run all the TableGen tests?

[barcharcraz]

In D119213#3304640, @Paul-C-Anagnostopoulos wrote:

Did you run all the TableGen tests?

Are they not included in the CI test suite? I'm not sure what's going on with the debian failures.

[Paul-C-Anagnostopoulos]

Sorry, I don't know what the CI test suite is. I don't know much about the testing facilities.

[barcharcraz]

It appears to run ninja check-llvm which seems to have passed on windows at least.

[Paul-C-Anagnostopoulos]

Looks good to me.

[nhaehnle]

Out of curiosity (not knowing all the intricacies of ADL), is the leading :: really needed or would llvm::format without that prefix suffice? Personally, I find the prefix a bit grating visually.

[barcharcraz]

llvm::format should also suffice, although it's slightly less specific (possibly matching llvm::llvm::format or similar). It won't trigger adl though.

[barcharcraz]

I've removed the global qualification on ::llvm::format

[barcharcraz]

@Paul-C-Anagnostopoulos Since you are the code-owner for TableGen... Is there anything else I need to do in order for this to be merged? It looks like header reorganization has made things no-longer break, but this fix will still make things more robust.

[Paul-C-Anagnostopoulos]

You are good to go.

llvm/utils/TableGen/GlobalISel/GIMatchDag.cpp

for (const auto &Assignment : N->user_assigned_operand_names())
       ToPrint.emplace_back(Assignment.first, Assignment.second);
     llvm::sort(ToPrint);
     StringRef Separator = "";
     for (const auto &Assignment : ToPrint) {
       OS << Separator << "$" << Assignment.second << "=getOperand("
          << Assignment.first << ")";
       Separator = ", ";
     }
-    OS << format("|%p|", &N);
+    OS << llvm::format("|%p|", &N);
     writePorts("d", N->getOperandInfo());
     OS << "}\"";
     if (N->isMatchRoot())
       OS << ",color=red";
     OS << "]\n";
   }
 
   for (const auto &E : Edges) {
for (const auto &E : Edges) {
     OS << "]\n";
   }
 
   for (const auto &N : PredicateNodes) {
     OS << "  " << format("Pred%p", &*N) << " [shape=record,label=\"{";
     writePorts("s", N->getOperandInfo());
     OS << "|" << N->getName() << "|";
     N->printDescription(OS);
-    OS << format("|%p|", &N);
+    OS << llvm::format("|%p|", &N);
     writePorts("d", N->getOperandInfo());
     OS << "}\",style=dotted]\n";
   }
 
   for (const auto &E : PredicateDependencies) {
     const char *FromMIFmt = "Node%p:e";
     const char *FromMOFmt = "Node%p:s%d:n";
     const char *ToFmt = "Pred%p:d%d:s";