Add half precision floating point support (float16) to APFloat,

patch by Peter Johnson! (PR5195)


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@84239 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/ADT/APFloat.h

@@ -125,6 +125,7 @@ namespace llvm {
   public:
 
     /* We support the following floating point semantics.  */
+    static const fltSemantics IEEEhalf;
     static const fltSemantics IEEEsingle;
     static const fltSemantics IEEEdouble;
     static const fltSemantics IEEEquad;
@@ -321,12 +322,14 @@ namespace llvm {
     opStatus roundSignificandWithExponent(const integerPart *, unsigned int,
                                           int, roundingMode);
 
+    APInt convertHalfAPFloatToAPInt() const;
     APInt convertFloatAPFloatToAPInt() const;
     APInt convertDoubleAPFloatToAPInt() const;
     APInt convertQuadrupleAPFloatToAPInt() const;
     APInt convertF80LongDoubleAPFloatToAPInt() const;
     APInt convertPPCDoubleDoubleAPFloatToAPInt() const;
     void initFromAPInt(const APInt& api, bool isIEEE = false);
+    void initFromHalfAPInt(const APInt& api);
     void initFromFloatAPInt(const APInt& api);
     void initFromDoubleAPInt(const APInt& api);
     void initFromQuadrupleAPInt(const APInt &api);

lib/Support/APFloat.cpp

@@ -48,6 +48,7 @@ namespace llvm {
     unsigned int arithmeticOK;
   };
 
+  const fltSemantics APFloat::IEEEhalf = { 15, -14, 11, true };
   const fltSemantics APFloat::IEEEsingle = { 127, -126, 24, true };
   const fltSemantics APFloat::IEEEdouble = { 1023, -1022, 53, true };
   const fltSemantics APFloat::IEEEquad = { 16383, -16382, 113, true };
@@ -2812,6 +2813,35 @@ APFloat::convertFloatAPFloatToAPInt() const
                     (mysignificand & 0x7fffff)));
 }
 
+APInt
+APFloat::convertHalfAPFloatToAPInt() const
+{
+  assert(semantics == (const llvm::fltSemantics*)&IEEEhalf);
+  assert (partCount()==1);
+
+  uint32_t myexponent, mysignificand;
+
+  if (category==fcNormal) {
+    myexponent = exponent+15; //bias
+    mysignificand = (uint32_t)*significandParts();
+    if (myexponent == 1 && !(mysignificand & 0x400))
+      myexponent = 0;   // denormal
+  } else if (category==fcZero) {
+    myexponent = 0;
+    mysignificand = 0;
+  } else if (category==fcInfinity) {
+    myexponent = 0xff;
+    mysignificand = 0;
+  } else {
+    assert(category == fcNaN && "Unknown category!");
+    myexponent = 0xff;
+    mysignificand = (uint32_t)*significandParts();
+  }
+
+  return APInt(16, (((sign&1) << 15) | ((myexponent&0x1f) << 10) |
+                    (mysignificand & 0x3ff)));
+}
+
 // This function creates an APInt that is just a bit map of the floating
 // point constant as it would appear in memory.  It is not a conversion,
 // and treating the result as a normal integer is unlikely to be useful.
@@ -2819,6 +2849,9 @@ APFloat::convertFloatAPFloatToAPInt() const
 APInt
 APFloat::bitcastToAPInt() const
 {
+  if (semantics == (const llvm::fltSemantics*)&IEEEhalf)
+    return convertHalfAPFloatToAPInt();
+
   if (semantics == (const llvm::fltSemantics*)&IEEEsingle)
     return convertFloatAPFloatToAPInt();
 
@@ -3051,6 +3084,39 @@ APFloat::initFromFloatAPInt(const APInt & api)
   }
 }
 
+void
+APFloat::initFromHalfAPInt(const APInt & api)
+{
+  assert(api.getBitWidth()==16);
+  uint32_t i = (uint32_t)*api.getRawData();
+  uint32_t myexponent = (i >> 15) & 0x1f;
+  uint32_t mysignificand = i & 0x3ff;
+
+  initialize(&APFloat::IEEEhalf);
+  assert(partCount()==1);
+
+  sign = i >> 15;
+  if (myexponent==0 && mysignificand==0) {
+    // exponent, significand meaningless
+    category = fcZero;
+  } else if (myexponent==0x1f && mysignificand==0) {
+    // exponent, significand meaningless
+    category = fcInfinity;
+  } else if (myexponent==0x1f && mysignificand!=0) {
+    // sign, exponent, significand meaningless
+    category = fcNaN;
+    *significandParts() = mysignificand;
+  } else {
+    category = fcNormal;
+    exponent = myexponent - 15;  //bias
+    *significandParts() = mysignificand;
+    if (myexponent==0)    // denormal
+      exponent = -14;
+    else
+      *significandParts() |= 0x400; // integer bit
+  }
+}
+
 /// Treat api as containing the bits of a floating point number.  Currently
 /// we infer the floating point type from the size of the APInt.  The
 /// isIEEE argument distinguishes between PPC128 and IEEE128 (not meaningful
@@ -3058,7 +3124,9 @@ APFloat::initFromFloatAPInt(const APInt & api)
 void
 APFloat::initFromAPInt(const APInt& api, bool isIEEE)
 {
-  if (api.getBitWidth() == 32)
+  if (api.getBitWidth() == 16)
+    return initFromHalfAPInt(api);
+  else if (api.getBitWidth() == 32)
     return initFromFloatAPInt(api);
   else if (api.getBitWidth()==64)
     return initFromDoubleAPInt(api);