[PATCH] D117247: [RISCV] Add inline expansion for vector fround.

[Craig Topper via Phabricator via llvm-commits]

craig.topper added inline comments.


================
Comment at: llvm/lib/Target/RISCV/RISCVISelLowering.cpp:1792
+// aren't compiling with trapping math, we can emulate this with
+// floor(X + copysign(nextafter(0.5, 0.0), X)).
+// FIXME: Could be shorter by changing rounding mode, but we don't have FRM
----------------
If I remember right, we have to use 0.49999999(mantissa of all ones) instead of 0.5 to deal with at least one corner case.

If the starting value is already 0.499999(mantissa all ones) then adding 0.5 would produce .9999999 with one extra mantissa bit than can be represented before the fadd rounds. The fadd would round that to 1.0. It would still be 1.0 after the conversion to int and back to fp. But the original 0.4999999(mantissa all ones) should produce 0.0.

By using 0.49999999(mantissa of all ones) adding that to 0.49999999 produces 0.49999999*2 which has the same number of ones in the mantissa just increments the exponent by 1. So the addition doesn't round to 1.0 it stays as 0.999999. The float->int->float conversion will turn that into 0.0.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D117247/new/

https://reviews.llvm.org/D117247