[PATCH] D90217: [AMDGPU][GlobalISel] Fold a chain of two shift instructions with constant operands

[Jay Foad via Phabricator via llvm-commits]

foad added inline comments.


================
Comment at: llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp:1596-1597
+    }
+    // Arithmetic shift larger than scalar size has no effect.
+    Imm = Ty.getScalarSizeInBits() - 1;
+  }
----------------
foad wrote:
> You could also handle G_USHLSAT and G_SSHLSAT the same way as G_ASHR.
Sorry, I was wrong about G_USHLSAT. You can't handle it quite the same way as G_ASHR. With 8-bit ints:
1 ushlsat 5 ushlsat 5 = 0xff
1 ushlsat 7 = 0x80
So they are not the same.

If the combined shift amount for ushlsat is >= 8 then you could replace it with "x == 0 ? 0 : 0xff" but it's probably not worth doing this.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D90217/new/

https://reviews.llvm.org/D90217