[flang-commits] [PATCH] D131099: [flang] Handle subnormals while folding SCALE, SET_EXPONENT, & FRACTION

[Peter Klausler via Phabricator via flang-commits]

klausler created this revision.
klausler added a reviewer: vdonaldson.
klausler added a project: Flang.
Herald added a subscriber: jdoerfert.
Herald added a project: All.
klausler requested review of this revision.

The implementations of folding for the intrinsic functions SCALE and SET_EXPONENT
do not cope correctly with numbers in the subnormal range.  Fix SCALE,
then modify SET_EXPONENT to be a special case of SCALE.


https://reviews.llvm.org/D131099

Files:
  flang/include/flang/Evaluate/real.h
  flang/lib/Evaluate/real.cpp
  flang/test/Evaluate/folding07.f90


Index: flang/test/Evaluate/folding07.f90
===================================================================
--- flang/test/Evaluate/folding07.f90
+++ flang/test/Evaluate/folding07.f90
@@ -280,5 +280,6 @@
   logical, parameter :: test_set_exponent_0 = set_exponent(1., 0) == 0.5
   logical, parameter :: test_set_exponent_1 = set_exponent(1., 1) == 1.
   logical, parameter :: test_set_exponent_2 = set_exponent(1., 2) == 2.
+  logical, parameter :: test_set_exponent_min = set_exponent(1., -149) == 1.40129846432481707092372958328991613128026194187651577175706828388979108268586060148663818836212158203125e-45_4
 
 end module
Index: flang/lib/Evaluate/real.cpp
===================================================================
--- flang/lib/Evaluate/real.cpp
+++ flang/lib/Evaluate/real.cpp
@@ -765,9 +765,7 @@
   } else if (IsZero()) {
     return *this;
   } else {
-    Real result;
-    result.Normalize(IsNegative(), expo + exponentBias - 1, GetFraction());
-    return result;
+    return SCALE(Integer<32>(expo - UnbiasedExponent() - 1)).value;
   }
 }
 
Index: flang/include/flang/Evaluate/real.h
===================================================================
--- flang/include/flang/Evaluate/real.h
+++ flang/include/flang/Evaluate/real.h
@@ -177,17 +177,21 @@
   template <typename INT>
   ValueWithRealFlags<Real> SCALE(const INT &by,
       Rounding rounding = TargetCharacteristics::defaultRounding) const {
-    auto expo{exponentBias + by.ToInt64()};
+    // Normalize a fraction with just its LSB set and then multiply.
+    // (Set the LSB, not the MSB, in case the scale factor needs to
+    //  be subnormal.)
+    auto adjust{exponentBias + binaryPrecision - 1};
+    auto expo{adjust + by.ToInt64()};
     if (IsZero()) {
       expo = exponentBias; // ignore by, don't overflow
     } else if (by > INT{maxExponent}) {
-      expo = maxExponent;
-    } else if (by < INT{-exponentBias}) {
+      expo = maxExponent + binaryPrecision - 1;
+    } else if (by < INT{-adjust}) {
       expo = -1;
     }
     Real twoPow;
     RealFlags flags{
-        twoPow.Normalize(false, static_cast<int>(expo), Fraction::MASKL(1))};
+        twoPow.Normalize(false, static_cast<int>(expo), Fraction::MASKR(1))};
     ValueWithRealFlags<Real> result{Multiply(twoPow, rounding)};
     result.flags |= flags;
     return result;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D131099.449749.patch
Type: text/x-patch
Size: 2342 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/flang-commits/attachments/20220803/25777840/attachment.bin>