[llvm] [AArch64] Use GISel for optnone functions (PR #174746)

[Ryan Cowan via llvm-commits]

================
@@ -387,12 +387,19 @@ AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT,
     // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB)
     this->Options.TLSSize = 24;
 
-  // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is
-  // MachO/CodeModel::Large, which GlobalISel does not support.
-  if (static_cast<int>(getOptLevel()) <= EnableGlobalISelAtO &&
+  const bool TargetSupportsGISel =
       TT.getArch() != Triple::aarch64_32 &&
       TT.getEnvironment() != Triple::GNUILP32 &&
-      !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) {
+      !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO());
+
+  const bool GlobalISelFlag =
+      getCGPassBuilderOption().EnableGlobalISelOption.value_or(false);
+
+  // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is
+  // MachO/CodeModel::Large, which GlobalISel does not support.
+  if (TargetSupportsGISel &&
+      (static_cast<int>(getOptLevel()) <= EnableGlobalISelAtO ||
+       (!GlobalISelFlag && !Options.EnableGlobalISel))) {
----------------
HolyMolyCowMan wrote:

It was really quite tricky to get this logic to do what we wanted without exposing a point of control.

This comment roughly explains the idea behind the check.
```
  // GlobalISel is currently only enabled when the opt level is less than or
  // equal to EnableGlobalISelAt or it was explicitly enabled via the CLI. If we
  // encounter this check, we know GlobalISel was enabled. If not by these two,
  // it must have been used as part of the SDAG pipeline to use GlobalISel for
  // optnone.
 ```
 
I used the absence of the GlobalISel flag & the EnableGlobalISel option to determine whether GlobalISel had been enabled purposefully, in which case we can avoid the SDAG pipeline or as part of a fallback from the SDAG pipeline. I would have preferred to add some kind of control we could use to explicitly alter this.

I do think that part of the rationale behind this work was to disable the use of fastisel at O0 but I'd want to check that.

I've had a look at your potential fix, it looks like it would work but I'm wondering why you check the optlevel isn't 0?

https://github.com/llvm/llvm-project/pull/174746