[X86] canonicalize a v2f64 blendi of 2 registers

[spatel]

Summary

This canonicalization step saves us 3 pattern matching possibilities * 4 math ops for scalar math that uses xmm regs.

The tests in llvm/test/CodeGen/X86/sse-scalar-fp-arith.ll cover this scenario already, so I don't think we need to add any more tests.

Is it possible to add an assert to make sure we're not bypassing this canonicalization? Where would it go?

Also, I'm assuming that we don't turn this into a 'movsd' because of its partial register update, but should there be a 'FIXME' comment here when optimizing for size because 'movsd' is 2 bytes shorter?

[chandlerc]

Yea, there is a better way to do this: Add a target-specific combine (or even post-legalize transform in X86ISelDAGToDAG which does this canonicalization. That will run right before pattern matching through the tables, and we won't need to think about it in any other place we might form BLENDI.

[spatel]

I don't know how I blanked on the target-specific combine option. Sorry about that. Thanks for the pointers!

Patch updated to include a PerformBLENDICombine() that handles the canonicalization. I also added a comment about the size optimization potential of 'movsd'.

[chandlerc]

lib/Target/X86/X86ISelLowering.cpp

26211

Should also mention that the backend knows how to commute this instruction again as needed to match the register allocation.

lib/Target/X86/X86ISelLowering.cpp

static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
   // Create this as a scalar to vector to match the instruction pattern.
   SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
   // countS bits are ignored when loading from memory on insertps, which
   // means we don't need to explicitly set them to 0.
   return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
                      LoadScalarToVector, N->getOperand(2));
 }
 
+static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
+  SDValue V0 = N->getOperand(0);
+  SDValue V1 = N->getOperand(1);
+  SDLoc DL(N);
+  EVT VT = N->getValueType(0);
+
+  // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
+  // operands and changing the mask to 1. This saves us a bunch of
+  // pattern-matching possibilities related to scalar math ops in SSE/AVX.
+  
+  // TODO: If optimizing for size or a processor that doesn't suffer from
+  // partial register update stalls, this should be transformed into a MOVSD
+  // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
+
+  if (VT == MVT::v2f64)
+    if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
+      if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
+        SDValue NewMask = DAG.getConstant(1, MVT::i8);
+        return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
+      }
+
+  return SDValue();
+}
+
 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
 // as "sbb reg,reg", since it can be extended without zext and produces
 // an all-ones bit which is more useful than 0/1 in some cases.
 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
                                MVT VT) {
   if (VT == MVT::i8)
     return DAG.getNode(ISD::AND, DL, VT,
                        DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::FMA:            return PerformFMACombine(N, DAG, Subtarget);
   case ISD::INTRINSIC_WO_CHAIN:
     return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
   case X86ISD::INSERTPS: {
     if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
       return PerformINSERTPSCombine(N, DAG, Subtarget);
     break;
   }
+  case X86ISD::BLENDI:    return PerformBLENDICombine(N, DAG);
   case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
   }
 
   return SDValue();
 }
 
 /// isTypeDesirableForOp - Return true if the target has native support for
 /// the specified value type and it is 'desirable' to use the type for the

lib/Target/X86/X86InstrSSE.td

 // Previously we generated:
 //   addps %xmm0, %xmm1
 //   movss %xmm1, %xmm0
 //
 // We now generate:
 //   addss %xmm1, %xmm0
 
 // TODO: Some canonicalization in lowering would simplify the number of
-// patterns we have to try to match. In particular, the reversed order blends
-// seem unnecessary.
+// patterns we have to try to match.
 multiclass scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
   let Predicates = [UseSSE1] in {
     // extracted scalar math op with insert via movss
     def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
           (Op (f32 (vector_extract (v4f32 VR128:$dst), (iPTR 0))),
           FR32:$src))))),
       (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
           (COPY_TO_REGCLASS FR32:$src, VR128))>;
def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
           FR64:$src))), (i8 1))),
       (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
           (COPY_TO_REGCLASS FR64:$src, VR128))>;
           
     // vector math op with insert via blend
     def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
           (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
       (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
-
-    // vector math op with insert via blend (reversed order)
-    def : Pat<(v2f64 (X86Blendi
-          (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)),
-          (v2f64 VR128:$dst), (i8 2))),
-      (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
   }
 
-  // Repeat everything for AVX and add one more pattern
-  // (the scalar + blend reversed order) for good measure.
+  // Repeat everything for AVX.
   let Predicates = [HasAVX] in {
     // extracted scalar math op with insert via movsd
     def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
           (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
           FR64:$src))))),
       (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
           (COPY_TO_REGCLASS FR64:$src, VR128))>;
 
     // extracted scalar math op with insert via blend
     def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
           (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
           FR64:$src))), (i8 1))),
       (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
           (COPY_TO_REGCLASS FR64:$src, VR128))>;
 
-    // extracted scalar math op with insert via blend (reversed order)
-    def : Pat<(v2f64 (X86Blendi (v2f64 (scalar_to_vector
-          (Op (f64 (vector_extract (v2f64 VR128:$dst), (iPTR 0))),
-          FR64:$src))), (v2f64 VR128:$dst), (i8 2))),
-      (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
-          (COPY_TO_REGCLASS FR64:$src, VR128))>;
-
     // vector math op with insert via movsd
     def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
           (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
       (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
 
     // vector math op with insert via blend
     def : Pat<(v2f64 (X86Blendi (v2f64 VR128:$dst),
           (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)), (i8 1))),
       (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
-
-    // vector math op with insert via blend (reversed order)
-    def : Pat<(v2f64 (X86Blendi
-          (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)),
-          (v2f64 VR128:$dst), (i8 2))),
-      (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
   }
 }
 
 defm : scalar_math_f64_patterns<fadd, "ADD">;
 defm : scalar_math_f64_patterns<fsub, "SUB">;
 defm : scalar_math_f64_patterns<fmul, "MUL">;
 defm : scalar_math_f64_patterns<fdiv, "DIV">;