[APFloat] Fix incorrect fptrunc rouning when from semantics exp lower than minimal exp of target semantics

[kpolushin]

Summary

Incorrect calculation of the exponentChange in the case when the exponent
of the original semantics is lower than the minimum exponent of the target semantics,
led to an incorrect shift and an incorrect value of the lost fraction.

For example:

fptrunc double 0x10000001 to float

Round to the minimum positive float 0x36A0000000000000 (denorm),
although in the case of rounding to the nearest, this should have reduced to 0.
Since the lost fraction was calculated incorrectly, the function IEEEFloat::roundAwayFromZero,
called from IEEEFloat::normalize returned true (ignoring differences in changed exponent semantics),
which resulted in rounding up.

[uweigand]

I'm not sure what specific problem you're trying to solve here, but this change in itself looks incorrect.

The computation done here follows the same logic that would occur later in normalize(). exponentChange is (correctly) set to toSemantics.minExponent - exponent here so that the subsequent exponent += exponentChange will then set exponent to toSemantics.minExponent.

llvm/lib/Support/APFloat.cpp

IEEEFloat::opStatus IEEEFloat::convert(const fltSemantics &toSemantics,
   // If this is a truncation of a denormal number, and the target semantics
   // has larger exponent range than the source semantics (this can happen
   // when truncating from PowerPC double-double to double format), the
   // right shift could lose result mantissa bits.  Adjust exponent instead
   // of performing excessive shift.
   if (shift < 0 && isFiniteNonZero()) {
     int exponentChange = significandMSB() + 1 - fromSemantics.precision;
     if (exponent + exponentChange < toSemantics.minExponent)
-      exponentChange = toSemantics.minExponent - exponent;
+      exponentChange = exponent - toSemantics.minExponent;
     if (exponentChange < shift)
       exponentChange = shift;
     if (exponentChange < 0) {
       shift -= exponentChange;
       exponent += exponentChange;
     }
   }
 

llvm/test/Transforms/InstCombine/fpcast.ll

 ; CHECK-NEXT:    [[T2:%.*]] = fptrunc float [[T1]] to half
 ; CHECK-NEXT:    ret half [[T2]]
 ;
   %t1 = fptrunc double %V to float
   %t2 = fptrunc float %t1 to half
   ret half %t2
 }
 
+define float @fptrunc_double_to_float_rounding_to_zero0() {
+; CHECK-LABEL: @fptrunc_double_to_float_rounding_to_zero0(
+; CHECK-NEXT:    ret float 0.000000e+00
+;
+  %trunc = fptrunc double 0x10000001 to float
+  ret float %trunc
+}
+
+define float @fptrunc_double_to_float_rounding_to_zero1() {
+; CHECK-LABEL: @fptrunc_double_to_float_rounding_to_zero1(
+; CHECK-NEXT:    ret float 0.000000e+00
+;
+  %trunc = fptrunc double 0x11111111 to float
+  ret float %trunc
+}
+
+define float @fptrunc_double_to_float_rounding_to_zero2() {
+; CHECK-LABEL: @fptrunc_double_to_float_rounding_to_zero2(
+; CHECK-NEXT:    ret float 0.000000e+00
+;
+  %trunc = fptrunc double 0x10000000 to float
+  ret float %trunc
+}
+
+define float @fptrunc_double_to_float_rounding_lower() {
+; CHECK-LABEL: @fptrunc_double_to_float_rounding_lower(
+; CHECK-NEXT:    ret float 0.000000e+00
+;
+  %den_f = bitcast i32 1 to float
+  %den_d = fpext float %den_f to double
+  %div = fdiv double %den_d, 2.0
+  %bits = bitcast double %div to i64
+  %add = add i64 %bits, 0
+  %cast = bitcast i64 %add to double
+  %trunc = fptrunc double %cast to float
+  ret float %trunc
+}
+
+define float @fptrunc_double_to_float_rounding_upper() {
+; CHECK-LABEL: @fptrunc_double_to_float_rounding_upper(
+; CHECK-NEXT:    ret float 0x36A0000000000000
+;
+  %den_f = bitcast i32 1 to float
+  %den_d = fpext float %den_f to double
+  %div = fdiv double %den_d, 2.0
+  %bits = bitcast double %div to i64
+  %add = add i64 %bits, 1
+  %cast = bitcast i64 %add to double
+  %trunc = fptrunc double %cast to float
+  ret float %trunc
+}
+