[PATCH] D74315: AMDGPU/GlobalISel: Start handling _L to _LZ optimization

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp:3560-3564
+    // FIXME: This isn't the cleanest way to handle this, but it's the easiest
+    // option the current infrastructure gives. We really should be changing the
+    // base intrinsic opcode, but the current searchable tables only gives us
+    // the final MI opcode. Eliminate the register here, and track with an
+    // immediate 0 so the final selection will know to do the opcode change.
----------------
arsenm wrote:
> nhaehnle wrote:
> > This no longer applies, does it? MIMGLZMappingTable contains L and LZ fields of type MIMGBaseOpcode now.
> This is still in terms of the machine opcode, not the intrinsic ID
How about doing a reverse lookup with a second index like
```
def getImageDimInstrinsicByBaseOpcode : SearchIndex {
  let Table = ImageDimIntrinsicTable;
  let Key = ["BaseOpcode", "Dim"];
}
```
which should auto-generate a C++ function with signature:
```
const ImageDimIntrinsicInfo *getImageDimIntrinsicByBaseOpcode(unsigned BaseOpcode, MIMGDim Dim);
```



CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D74315/new/

https://reviews.llvm.org/D74315