[PATCH] D26855: New unsafe-fp-math implementation for X86 target

[Simon Pilgrim via llvm-commits]

RKSimon added inline comments.


================
Comment at: lib/Target/X86/X86InstrInfo.cpp:8726
+  return Rcp;
+}
+
----------------
This needs refactoring to support scalar and packed versions of SSE and AVX opcodes if possible.


================
Comment at: lib/Target/X86/X86InstrInfo.cpp:8774
+  case X86::DIVSSrr: // f32
+  case X86::DIVSDrr: // f64
+    assert(Root.getOperand(1).isReg() && Root.getOperand(2).isReg() &&
----------------
We don't do reciprocal estimates for double types - these all need removing.


================
Comment at: lib/Target/X86/X86InstrInfo.cpp:8816
+  //  if (getFSQRTPatterns(Root, Patterns))
+  //    return true;
+
----------------
Replace this with a basic TODO comment for future FSQRT pattern support - don't comment out code (especially when it doesn't exist).


https://reviews.llvm.org/D26855