Regression in dotnet9.0 builds caused by 23da169

[tstellar]

After 23da169 the dotnet9.0 builds in Fedora started failing with this error:

Consolidate compiler generated dependencies of target utilcodestaticnohost
make[3]: Leaving directory '/builddir/build/BUILD/dotnet9.0-9.0.102-build/dotnet-9.0.1/src/runtime/artifacts/obj/coreclr/linux.ppc64le.Release'
make  -f utilcode/CMakeFiles/utilcodestaticnohost.dir/build.make utilcode/CMakeFiles/utilcodestaticnohost.dir/build
/builddir/build/BUILD/dotnet9.0-9.0.102-build/dotnet-9.0.1/src/runtime/src/coreclr/pal/src/arch/ppc64le/context2.S:140:15: error: invalid register name
              ld %r0, 0 +8(%R3)
                           ^

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: Tom Stellard (tstellar)

After 23da169 the dotnet9.0 builds in Fedora started failing with this error:

Consolidate compiler generated dependencies of target utilcodestaticnohost
make[3]: Leaving directory '/builddir/build/BUILD/dotnet9.0-9.0.102-build/dotnet-9.0.1/src/runtime/artifacts/obj/coreclr/linux.ppc64le.Release'
make  -f utilcode/CMakeFiles/utilcodestaticnohost.dir/build.make utilcode/CMakeFiles/utilcodestaticnohost.dir/build
/builddir/build/BUILD/dotnet9.0-9.0.102-build/dotnet-9.0.1/src/runtime/src/coreclr/pal/src/arch/ppc64le/context2.S:140:15: error: invalid register name
              ld %r0, 0 +8(%R3)
                           ^

[tstellar]

I might be missing something, can you please provide more info as to why you think this patch is the issue?

@lei137 I used git bisect to find this commit as the one that introduces the problem. I can try to get a reduced test case if this helps.

[tstellar]

@lei137 Just from inspection, I think the problem is here. The TableGen'd MatchRegisterName() function only parses lowercase register names, so it returns NoRegister when it sees R3 and then matchRegisterName() returns early.

  MCRegister RegNo = MatchRegisterName(Name);
  if (!RegNo)
    return RegNo;

[tstellar]

Closing this. Discussion is on going at #111553, and don't want to have two threads on it.

[lei137]

@tstellar Can you please see if you can get a reduce testcase for this?

[nikic]

@lei137 Isn't the reproducer just this?

# RUN: llvm-mc --triple=powerpc64le-unknown-linx-gnu %s
test:
    ld %r0, 0(%R3)
    blr

This prints now:

test.S:3:15: error: invalid register name

It used to be accepted and produced:

	.text
test:
	ld 0, 0(3)
	blr

[nikic]

I checked that powerpc64-linux-gnu-as also accepts this assembly.