[ConstantFPRange] Add support for cast operations

llvm/include/llvm/IR/ConstantFPRange.h

@@ -216,6 +216,12 @@ class [[nodiscard]] ConstantFPRange {
   /// Get the range without infinities. It is useful when we apply ninf flag to
   /// range of operands/results.
   LLVM_ABI ConstantFPRange getWithoutInf() const;
+
+  /// Return a new range in the specified format with the specified rounding
+  /// mode.
+  LLVM_ABI ConstantFPRange
+  cast(const fltSemantics &DstSem,
+       APFloat::roundingMode RM = APFloat::rmNearestTiesToEven) const;
 };
 
 inline raw_ostream &operator<<(raw_ostream &OS, const ConstantFPRange &CR) {

llvm/lib/IR/ConstantFPRange.cpp

@@ -326,6 +326,8 @@ std::optional<bool> ConstantFPRange::getSignBit() const {
 }
 
 bool ConstantFPRange::operator==(const ConstantFPRange &CR) const {
+  assert(&getSemantics() == &CR.getSemantics() &&
+         "Should only use the same semantics");
   if (MayBeSNaN != CR.MayBeSNaN || MayBeQNaN != CR.MayBeQNaN)
     return false;
   return Lower.bitwiseIsEqual(CR.Lower) && Upper.bitwiseIsEqual(CR.Upper);
@@ -425,3 +427,20 @@ ConstantFPRange ConstantFPRange::getWithoutInf() const {
   return ConstantFPRange(std::move(NewLower), std::move(NewUpper), MayBeQNaN,
                          MayBeSNaN);
 }
+
+ConstantFPRange ConstantFPRange::cast(const fltSemantics &DstSem,
+                                      APFloat::roundingMode RM) const {
+  bool LosesInfo;
+  APFloat NewLower = Lower;
+  APFloat NewUpper = Upper;
+  // For conservative, return full range if conversion is invalid.
+  if (NewLower.convert(DstSem, RM, &LosesInfo) == APFloat::opInvalidOp ||
+      NewLower.isNaN())
+    return getFull(DstSem);
+  if (NewUpper.convert(DstSem, RM, &LosesInfo) == APFloat::opInvalidOp ||
+      NewUpper.isNaN())
+    return getFull(DstSem);
+  return ConstantFPRange(std::move(NewLower), std::move(NewUpper),
+                         /*MayBeQNaNVal=*/MayBeQNaN || MayBeSNaN,
+                         /*MayBeSNaNVal=*/false);
+}

llvm/unittests/IR/ConstantFPRangeTest.cpp

@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/IR/ConstantFPRange.h"
+#include "llvm/ADT/APFloat.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Operator.h"
 #include "gtest/gtest.h"
@@ -818,4 +819,110 @@ TEST_F(ConstantFPRangeTest, getWithout) {
                 APFloat::getLargest(Sem, /*Negative=*/true), APFloat(3.0)));
 }
 
+TEST_F(ConstantFPRangeTest, cast) {
+  const fltSemantics &F16Sem = APFloat::IEEEhalf();
+  const fltSemantics &BF16Sem = APFloat::BFloat();
+  const fltSemantics &F32Sem = APFloat::IEEEsingle();
+  const fltSemantics &F8NanOnlySem = APFloat::Float8E4M3FN();
+  // normal -> normal (exact)
+  EXPECT_EQ(ConstantFPRange::getNonNaN(APFloat(1.0), APFloat(2.0)).cast(F32Sem),
+            ConstantFPRange::getNonNaN(APFloat(1.0f), APFloat(2.0f)));
+  EXPECT_EQ(
+      ConstantFPRange::getNonNaN(APFloat(-2.0f), APFloat(-1.0f)).cast(Sem),
+      ConstantFPRange::getNonNaN(APFloat(-2.0), APFloat(-1.0)));
+  // normal -> normal (inexact)
+  EXPECT_EQ(
+      ConstantFPRange::getNonNaN(APFloat(3.141592653589793),
+                                 APFloat(6.283185307179586))
+          .cast(F32Sem),
+      ConstantFPRange::getNonNaN(APFloat(3.14159274f), APFloat(6.28318548f)));
+  // normal -> subnormal
+  EXPECT_EQ(ConstantFPRange::getNonNaN(APFloat(-5e-8), APFloat(5e-8))
+                .cast(F16Sem)
+                .classify(),
+            fcSubnormal | fcZero);
+  // normal -> zero
+  EXPECT_EQ(ConstantFPRange::getNonNaN(
+                APFloat::getSmallestNormalized(Sem, /*Negative=*/true),
+                APFloat::getSmallestNormalized(Sem, /*Negative=*/false))
+                .cast(F32Sem)
+                .classify(),
+            fcZero);
+  // normal -> inf
+  EXPECT_EQ(ConstantFPRange::getNonNaN(APFloat(-65536.0), APFloat(65536.0))
+                .cast(F16Sem),
+            ConstantFPRange::getNonNaN(F16Sem));
+  // nan -> qnan
+  EXPECT_EQ(
+      ConstantFPRange::getNaNOnly(Sem, /*MayBeQNaN=*/true, /*MayBeSNaN=*/false)
+          .cast(F32Sem),
+      ConstantFPRange::getNaNOnly(F32Sem, /*MayBeQNaN=*/true,
+                                  /*MayBeSNaN=*/false));
+  EXPECT_EQ(
+      ConstantFPRange::getNaNOnly(Sem, /*MayBeQNaN=*/false, /*MayBeSNaN=*/true)
+          .cast(F32Sem),
+      ConstantFPRange::getNaNOnly(F32Sem, /*MayBeQNaN=*/true,
+                                  /*MayBeSNaN=*/false));
+  EXPECT_EQ(
+      ConstantFPRange::getNaNOnly(Sem, /*MayBeQNaN=*/true, /*MayBeSNaN=*/true)
+          .cast(F32Sem),
+      ConstantFPRange::getNaNOnly(F32Sem, /*MayBeQNaN=*/true,
+                                  /*MayBeSNaN=*/false));
+  // For BF16 -> F32, signaling bit is still lost.
+  EXPECT_EQ(ConstantFPRange::getNaNOnly(BF16Sem, /*MayBeQNaN=*/true,
+                                        /*MayBeSNaN=*/true)
+                .cast(F32Sem),
+            ConstantFPRange::getNaNOnly(F32Sem, /*MayBeQNaN=*/true,
+                                        /*MayBeSNaN=*/false));
+  // inf -> nan only (return full set for now)
+  EXPECT_EQ(ConstantFPRange::getNonNaN(APFloat::getInf(Sem, /*Negative=*/true),
+                                       APFloat::getInf(Sem, /*Negative=*/false))
+                .cast(F8NanOnlySem),
+            ConstantFPRange::getFull(F8NanOnlySem));
+  // other rounding modes
+  EXPECT_EQ(
+      ConstantFPRange::getNonNaN(APFloat::getSmallest(Sem, /*Negative=*/true),
+                                 APFloat::getSmallest(Sem, /*Negative=*/false))
+          .cast(F32Sem, APFloat::rmTowardNegative),
+      ConstantFPRange::getNonNaN(
+          APFloat::getSmallest(F32Sem, /*Negative=*/true),
+          APFloat::getZero(F32Sem, /*Negative=*/false)));
+  EXPECT_EQ(
+      ConstantFPRange::getNonNaN(APFloat::getSmallest(Sem, /*Negative=*/true),
+                                 APFloat::getSmallest(Sem, /*Negative=*/false))
+          .cast(F32Sem, APFloat::rmTowardPositive),
+      ConstantFPRange::getNonNaN(
+          APFloat::getZero(F32Sem, /*Negative=*/true),
+          APFloat::getSmallest(F32Sem, /*Negative=*/false)));
+  EXPECT_EQ(
+      ConstantFPRange::getNonNaN(
+          APFloat::getSmallestNormalized(Sem, /*Negative=*/true),
+          APFloat::getSmallestNormalized(Sem, /*Negative=*/false))
+          .cast(F32Sem, APFloat::rmTowardZero),
+      ConstantFPRange::getNonNaN(APFloat::getZero(F32Sem, /*Negative=*/true),
+                                 APFloat::getZero(F32Sem, /*Negative=*/false)));
+
+  EnumerateValuesInConstantFPRange(
+      ConstantFPRange::getFull(APFloat::Float8E4M3()),
+      [&](const APFloat &V) {
+        bool LosesInfo = false;
+
+        APFloat DoubleV = V;
+        DoubleV.convert(Sem, APFloat::rmNearestTiesToEven, &LosesInfo);
+        ConstantFPRange DoubleCR = ConstantFPRange(V).cast(Sem);
+        EXPECT_TRUE(DoubleCR.contains(DoubleV))
+            << "Casting " << V << " to double failed. " << DoubleCR
+            << " doesn't contain " << DoubleV;
+
+        auto &FP4Sem = APFloat::Float4E2M1FN();
+        APFloat FP4V = V;
+        FP4V.convert(FP4Sem, APFloat::rmNearestTiesToEven, &LosesInfo);
+        ConstantFPRange FP4CR = ConstantFPRange(V).cast(FP4Sem);
+        EXPECT_TRUE(FP4CR.contains(FP4V))
+            << "Casting " << V << " to FP4E2M1FN failed. " << FP4CR
+            << " doesn't contain " << FP4V;
+      },
+      /*IgnoreNaNPayload=*/true);
+}
+
 } // anonymous namespace