X86 cost model: Model cost for uitofp and sitofp on SSE2

The costs are overfitted so that I can still use the legalization factor.

For example the following kernel has about half the throughput vectorized than
unvectorized when compiled with SSE2. Before this patch we would vectorize it.

unsigned short A[1024];
double B[1024];
void f() {
  int i;
  for (i = 0; i < 1024; ++i) {
    B[i] = (double) A[i];
  }
}

radar://13599001

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179033 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 34 insertions, 3 deletions

diff --git a/lib/Target/X86/X86TargetTransformInfo.cpp b/lib/Target/X86/X86TargetTransformInfo.cpp
index a98c699119..488c2a42b2 100644
--- a/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -334,12 +334,43 @@ unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
   int ISD = TLI->InstructionOpcodeToISD(Opcode);
   assert(ISD && "Invalid opcode");
 
+  std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
+  std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
+
+  static const TypeConversionCostTblEntry<MVT> SSE2ConvTbl[] = {
+    // These are somewhat magic numbers justified by looking at the output of
+    // Intel's IACA, running some kernels and making sure when we take
+    // legalization into account the throughput will be overestimated.
+    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
+    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
+    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
+    { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
+    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
+    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
+    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
+    { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
+    // There are faster sequences for float conversions.
+    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
+    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
+    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
+    { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
+    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
+    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
+    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
+    { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
+  };
+
+  if (ST->hasSSE2() && !ST->hasAVX()) {
+    int Idx = ConvertCostTableLookup<MVT>(SSE2ConvTbl,
+                                          array_lengthof(SSE2ConvTbl),
+                                          ISD, LTDest.second, LTSrc.second);
+    if (Idx != -1)
+      return LTSrc.first * SSE2ConvTbl[Idx].Cost;
+  }
+
   EVT SrcTy = TLI->getValueType(Src);
   EVT DstTy = TLI->getValueType(Dst);
 
-  if (!SrcTy.isSimple() || !DstTy.isSimple())
-    return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
-
   static const TypeConversionCostTblEntry<MVT> AVXConversionTbl[] = {
     { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
     { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },