[X86] Add custom promotion of narrow fp_to_uint/fp_to_sint operations under -x86-experimental-vector-widening-legalization.

This tries to force the result type to vXi32 followed by a truncate. This can help avoid scalarization that would otherwise occur.

There's some annoying examples of an avx512 truncate instruction followed by a packus where we should really be able to just use one truncate. But overall this is still a net improvement.

llvm-svn: 347105

1 files changed, 48 insertions, 3 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5c07d7d906f..8f422ce524b 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -899,10 +899,18 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::FP_TO_SINT,         MVT::v4i32, Legal);
     setOperationAction(ISD::FP_TO_SINT,         MVT::v2i32, Custom);
     setOperationAction(ISD::FP_TO_SINT,         MVT::v2i16, Custom);
-    // Custom legalize these to avoid over promotion.
+
+    // Custom legalize these to avoid over promotion or custom promotion.
     setOperationAction(ISD::FP_TO_SINT,         MVT::v2i8,  Custom);
-    setOperationAction(ISD::FP_TO_UINT,         MVT::v2i16, Custom);
+    setOperationAction(ISD::FP_TO_SINT,         MVT::v4i8,  Custom);
+    setOperationAction(ISD::FP_TO_SINT,         MVT::v8i8,  Custom);
+    setOperationAction(ISD::FP_TO_SINT,         MVT::v2i16, Custom);
+    setOperationAction(ISD::FP_TO_SINT,         MVT::v4i16, Custom);
     setOperationAction(ISD::FP_TO_UINT,         MVT::v2i8,  Custom);
+    setOperationAction(ISD::FP_TO_UINT,         MVT::v4i8,  Custom);
+    setOperationAction(ISD::FP_TO_UINT,         MVT::v8i8,  Custom);
+    setOperationAction(ISD::FP_TO_UINT,         MVT::v2i16, Custom);
+    setOperationAction(ISD::FP_TO_UINT,         MVT::v4i16, Custom);
 
     setOperationAction(ISD::SINT_TO_FP,         MVT::v4i32, Legal);
     setOperationAction(ISD::SINT_TO_FP,         MVT::v2i32, Custom);
@@ -26287,7 +26295,7 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
     // Promote these manually to avoid over promotion to v2i64. Type
     // legalization will revisit the v2i32 operation for more cleanup.
     if ((VT == MVT::v2i8 || VT == MVT::v2i16) &&
-        getTypeAction(*DAG.getContext(), VT) != TypeWidenVector) {
+        getTypeAction(*DAG.getContext(), VT) == TypePromoteInteger) {
       // AVX512DQ provides instructions that produce a v2i64 result.
       if (Subtarget.hasDQI())
         return;
@@ -26302,6 +26310,43 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
       return;
     }
 
+    if (VT.isVector() && VT.getScalarSizeInBits() < 32) {
+      if (getTypeAction(*DAG.getContext(), VT) != TypeWidenVector)
+        return;
+
+      // Try to create a 128 bit vector, but don't exceed a 32 bit element.
+      unsigned NewEltWidth = std::min(128 / VT.getVectorNumElements(), 32U);
+      MVT PromoteVT = MVT::getVectorVT(MVT::getIntegerVT(NewEltWidth),
+                                       VT.getVectorNumElements());
+      unsigned Opc = N->getOpcode();
+      if (PromoteVT == MVT::v2i32 || PromoteVT == MVT::v4i32)
+        Opc = ISD::FP_TO_SINT;
+
+      SDValue Res = DAG.getNode(Opc, dl, PromoteVT, Src);
+
+      // Preserve what we know about the size of the original result. Except
+      // when the result is v2i32 since we can't widen the assert.
+      if (PromoteVT != MVT::v2i32)
+        Res = DAG.getNode(N->getOpcode() == ISD::FP_TO_UINT ? ISD::AssertZext
+                                                            : ISD::AssertSext,
+                          dl, PromoteVT, Res,
+                          DAG.getValueType(VT.getVectorElementType()));
+
+      // Truncate back to the original width.
+      Res = DAG.getNode(ISD::TRUNCATE, dl, VT, Res);
+
+      // Now widen to 128 bits.
+      unsigned NumConcats = 128 / VT.getSizeInBits();
+      MVT ConcatVT = MVT::getVectorVT(VT.getSimpleVT().getVectorElementType(),
+                                      VT.getVectorNumElements() * NumConcats);
+      SmallVector<SDValue, 8> ConcatOps(NumConcats, DAG.getUNDEF(VT));
+      ConcatOps[0] = Res;
+      Res = DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatVT, ConcatOps);
+      Results.push_back(Res);
+      return;
+    }
+
+
     if (VT == MVT::v2i32) {
       assert((IsSigned || Subtarget.hasAVX512()) &&
              "Can only handle signed conversion without AVX512");