[llvm] [APFloat] Add APFloat support for E8M0 type (PR #107127)

[Matt Arsenault via llvm-commits]

================
@@ -3819,6 +3882,40 @@ void IEEEFloat::initFromPPCDoubleDoubleAPInt(const APInt &api) {
   }
 }
 
+// The E8M0 format has the following characteristics:
+// It is an 8-bit unsigned format with only exponents (no actual significand)
+// No encodings for {zero, infinities or denorms}
+// NaN is represented by all 1's
+// Bias is 127
+void IEEEFloat::initFromFloat8E8M0FNAPInt(const APInt &api) {
+  const uint64_t exponent_mask = 0xff;
+  uint64_t val = api.getRawData()[0];
+  uint64_t myexponent = (val & exponent_mask);
+
+  initialize(&semFloat8E8M0FN);
+  assert(partCount() == 1);
+
+  // This format has unsigned representation only
+  sign = 0;
+
+  // Set the significand
+  // This format does not have any significand but the 'Pth' precision bit is
+  // always set to 1 for consistency in APFloat's internal representation.
+  uint64_t mysignificand = 1;
+  significandParts()[0] = mysignificand;
+
+  // This format can either have a NaN or fcNormal
+  // All 1's i.e. 255 is a NaN
+  if (val == exponent_mask) {
+    category = fcNaN;
+    exponent = exponentNaN();
+    return;
+  }
+  // Handle fcNormal...
+  category = fcNormal;
+  exponent = myexponent - 127; // 127 is bias
----------------
arsenm wrote:

There should be some helper or semantics field for the bias?

https://github.com/llvm/llvm-project/pull/107127