This patch fixes failures in the SingleSource/Regression/C/uint64_to_float
test case on PowerPC caused by rounding errors when converting from a 64-bit
integer to a single-precision floating point. The reason for this are
double-rounding effects, since on PowerPC we have to convert to an
intermediate double-precision value first, which gets rounded to the
final single-precision result.
The patch fixes the problem by preparing the 64-bit integer so that the
first conversion step to double-precision will always be exact, and the
final rounding step will result in the correctly-rounded single-precision
result. The generated code sequence is equivalent to what GCC would generate.
When -enable-unsafe-fp-math is in effect, that extra effort is omitted
and we accept possible rounding errors (just like GCC does as well).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@166178 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index c18250a..36db4b5 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -4224,7 +4224,52 @@
return SDValue();
if (Op.getOperand(0).getValueType() == MVT::i64) {
- SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
+ SDValue SINT = Op.getOperand(0);
+ // When converting to single-precision, we actually need to convert
+ // to double-precision first and then round to single-precision.
+ // To avoid double-rounding effects during that operation, we have
+ // to prepare the input operand. Bits that might be truncated when
+ // converting to double-precision are replaced by a bit that won't
+ // be lost at this stage, but is below the single-precision rounding
+ // position.
+ //
+ // However, if -enable-unsafe-fp-math is in effect, accept double
+ // rounding to avoid the extra overhead.
+ if (Op.getValueType() == MVT::f32 &&
+ !DAG.getTarget().Options.UnsafeFPMath) {
+
+ // Twiddle input to make sure the low 11 bits are zero. (If this
+ // is the case, we are guaranteed the value will fit into the 53 bit
+ // mantissa of an IEEE double-precision value without rounding.)
+ // If any of those low 11 bits were not zero originally, make sure
+ // bit 12 (value 2048) is set instead, so that the final rounding
+ // to single-precision gets the correct result.
+ SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
+ SINT, DAG.getConstant(2047, MVT::i64));
+ Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
+ Round, DAG.getConstant(2047, MVT::i64));
+ Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
+ Round = DAG.getNode(ISD::AND, dl, MVT::i64,
+ Round, DAG.getConstant(-2048, MVT::i64));
+
+ // However, we cannot use that value unconditionally: if the magnitude
+ // of the input value is small, the bit-twiddling we did above might
+ // end up visibly changing the output. Fortunately, in that case, we
+ // don't need to twiddle bits since the original input will convert
+ // exactly to double-precision floating-point already. Therefore,
+ // construct a conditional to use the original value if the top 11
+ // bits are all sign-bit copies, and use the rounded value computed
+ // above otherwise.
+ SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
+ SINT, DAG.getConstant(53, MVT::i32));
+ Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
+ Cond, DAG.getConstant(1, MVT::i64));
+ Cond = DAG.getSetCC(dl, MVT::i32,
+ Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
+
+ SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
+ }
+ SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
if (Op.getValueType() == MVT::f32)
FP = DAG.getNode(ISD::FP_ROUND, dl,