[PATCH] D57445: [ARM] Fix TTI IntImmCost

[Eli Friedman via Phabricator via llvm-commits]

efriedma added inline comments.


================
Comment at: lib/Target/ARM/ARMTargetTransformInfo.cpp:99
+  if (BitWidth - LeadingZeros - TrailingZeros <= 8)
+    return 0;
+
----------------
Why are you throwing away the existing code using getT2SOImmVal/getSOImmVal and writing out the same logic inline?


================
Comment at: lib/Target/ARM/ARMTargetTransformInfo.cpp:108
+    // Use mov and movt
     return 2;
+  }
----------------
Should this be guarded by useMovt?


================
Comment at: lib/Target/ARM/ARMTargetTransformInfo.cpp:158
 
+  if (Opcode == Instruction::GetElementPtr) {
+    // FIXME Simplify by assuming 32-bit access, which uses an 8-bit offset
----------------
Not sure I follow this logic; what does "Imm" actually mean in this case?  Are we expecting that the result of the GEP will be folded into a load?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D57445/new/

https://reviews.llvm.org/D57445