[RISCV] Optimize floating-point "dominant value" BUILD_VECTORs

[frasercrmck]

Summary

This patch aims to improve the performance of BUILD_VECTORs which are
identified as containing a dominant element. Given that most
floating-point constants themselves require a load from the constant
pool, it was possible for the optimization to actually increase the
number of individual loads on small vectors. The exception is the zero
constant -- +0.0 -- which can be materialized efficiently.

While this optimization could do with a proper cost model to weigh the
benfits of a single vector load vs. the manipulation of individual
elements -- even for integer vectors which often require several
instructions to materialize -- without a concrete RVV implementation to
work with any heuristic is likely to be both more obtuse and inaccurate.

Until then, this patch fixes at least one known obvious deficiency.

[craig.topper]

LGTM

[frasercrmck]

Cheers! Do you or the person who reported this need the fix in LLVM 13?

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
   // "insert" the upper element, and an insert of the lower element at position
   // 0, which improves codegen.
   SDValue DominantValue;
   unsigned MostCommonCount = 0;
   DenseMap<SDValue, unsigned> ValueCounts;
   unsigned NumUndefElts =
       count_if(Op->op_values(), [](const SDValue &V) { return V.isUndef(); });
 
+  // Track the number of scalar loads we know we'd be inserting, estimated as
+  // any non-zero floating-point constant. Other kinds of element are either
+  // already in registers or are materialized on demand. The threshold at which
+  // a vector load is more desirable than several scalar materializion and
+  // vector-insertion instructions is not known.
+  unsigned NumScalarLoads = 0;
+
   for (SDValue V : Op->op_values()) {
     if (V.isUndef())
       continue;
 
     ValueCounts.insert(std::make_pair(V, 0));
     unsigned &Count = ValueCounts[V];
 
+    if (auto *CFP = dyn_cast<ConstantFPSDNode>(V))
+      NumScalarLoads += !CFP->isExactlyValue(+0.0);
+
     // Is this value dominant? In case of a tie, prefer the highest element as
     // it's cheaper to insert near the beginning of a vector than it is at the
     // end.
     if (++Count >= MostCommonCount) {
       DominantValue = V;
       MostCommonCount = Count;
     }
   }
 
   assert(DominantValue && "Not expecting an all-undef BUILD_VECTOR");
   unsigned NumDefElts = NumElts - NumUndefElts;
   unsigned DominantValueCountThreshold = NumDefElts <= 2 ? 0 : NumDefElts - 2;
 
   // Don't perform this optimization when optimizing for size, since
   // materializing elements and inserting them tends to cause code bloat.
-  if (!DAG.shouldOptForSize() &&
+  if (!DAG.shouldOptForSize() && NumScalarLoads < NumElts &&
       ((MostCommonCount > DominantValueCountThreshold) ||
        (ValueCounts.size() <= Log2_32(NumDefElts)))) {
     // Start by splatting the most common element.
     SDValue Vec = DAG.getSplatBuildVector(VT, DL, DominantValue);
 
     DenseSet<SDValue> Processed{DominantValue};
     MVT SelMaskTy = VT.changeVectorElementType(MVT::i1);
     for (const auto &OpIdx : enumerate(Op->ops())) {

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-fp-buildvec.ll

 ; CHECK-NEXT:    vfmv.s.f v25, ft0
 ; CHECK-NEXT:    vsetvli zero, zero, e32, mf2, ta, mu
 ; CHECK-NEXT:    vse32.v v25, (a0)
 ; CHECK-NEXT:    ret
   store <2 x float> <float 0.0, float 1.0>, <2 x float>* %x
   ret void
 }
 
-; FIXME: We "optimize" this one 2-element load from the constant pool to two
-; loads from the constant pool.
+; We don't want to lower this to the insertion of two scalar elements as above,
+; as each would require their own load from the constant pool.
 
 define void @buildvec_dominant1_v2f32(<2 x float>* %x) {
 ; CHECK-LABEL: buildvec_dominant1_v2f32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    lui a1, %hi(.LCPI3_0)
 ; CHECK-NEXT:    addi a1, a1, %lo(.LCPI3_0)
 ; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, mu
-; CHECK-NEXT:    vlse32.v v25, (a1), zero
-; CHECK-NEXT:    lui a1, %hi(.LCPI3_1)
-; CHECK-NEXT:    flw ft0, %lo(.LCPI3_1)(a1)
-; CHECK-NEXT:    vsetvli zero, zero, e32, mf2, tu, mu
-; CHECK-NEXT:    vfmv.s.f v25, ft0
-; CHECK-NEXT:    vsetvli zero, zero, e32, mf2, ta, mu
+; CHECK-NEXT:    vle32.v v25, (a1)
 ; CHECK-NEXT:    vse32.v v25, (a0)
 ; CHECK-NEXT:    ret
   store <2 x float> <float 1.0, float 2.0>, <2 x float>* %x
   ret void
 }
 
 define void @buildvec_dominant0_v4f32(<4 x float>* %x) {
 ; CHECK-LABEL: buildvec_dominant0_v4f32: