[compiler-rt] [compiler-rt] Implement __extendxftf2 and __trunctfxf2 for x86_64 (PR #66918)

[Eli Friedman via llvm-commits]

================
@@ -81,19 +94,37 @@ static __inline dst_t __extendXfYf2__(src_t a) {
     // bit (if needed) and right-aligning the rest of the trailing NaN
     // payload field.
     absResult = (dst_rep_t)dstInfExp << dstSigBits;
-    absResult |= (dst_rep_t)(aAbs & srcQNaN) << (dstSigBits - srcSigBits);
-    absResult |= (dst_rep_t)(aAbs & srcNaNCode) << (dstSigBits - srcSigBits);
+    absResult |= (dst_rep_t)(aAbs & srcQNaN)
+                 << (dstSigFracBits - srcSigFracBits);
+    absResult |= (dst_rep_t)(aAbs & srcNaNCode)
+                 << (dstSigFracBits - srcSigFracBits);
   }
 
   else if (aAbs) {
     // a is denormal.
-    // renormalize the significand and clear the leading bit, then insert
-    // the correct adjusted exponent in the destination type.
-    const int scale = src_rep_t_clz(aAbs) - src_rep_t_clz(srcMinNormal);
-    absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits + scale);
-    absResult ^= dstMinNormal;
-    const int resultExponent = dstExpBias - srcExpBias - scale + 1;
-    absResult |= (dst_rep_t)resultExponent << dstSigBits;
+    if (srcExpBits == dstExpBits) {
----------------
efriedma-quic wrote:

If I understand correctly, the key points here are:

- The exponent fields are identical
- This is a denormal number, so all the non-significand bits are zero.

So we just need to shift the bits.  (And this code could be used with other floating-point formats, e.g. converting bfloat16->float.  The differences between fp80 and other formats aren't relevant.)

Please update the comment to indicate f80->f128 is just a special-case of this general issue.

https://github.com/llvm/llvm-project/pull/66918