[AArch64] Sink splat to fmlal intrinsics

Similar to other neon index instructions, it is beneficial to sink the splat to
the instruction for fmlal in order for it to create the index.

Author: davemgreen

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -14273,6 +14273,16 @@ bool AArch64TargetLowering::shouldSinkOperands(
       if (isSplatShuffle(II->getOperand(1)))
         Ops.push_back(&II->getOperandUse(1));
       return !Ops.empty();
+    case Intrinsic::aarch64_neon_fmlal:
+    case Intrinsic::aarch64_neon_fmlal2:
+    case Intrinsic::aarch64_neon_fmlsl:
+    case Intrinsic::aarch64_neon_fmlsl2:
+      // Sink splats for index lane variants
+      if (isSplatShuffle(II->getOperand(1)))
+        Ops.push_back(&II->getOperandUse(1));
+      if (isSplatShuffle(II->getOperand(2)))
+        Ops.push_back(&II->getOperandUse(2));
+      return !Ops.empty();
     case Intrinsic::aarch64_sve_ptest_first:
     case Intrinsic::aarch64_sve_ptest_last:
       if (auto *IIOp = dyn_cast<IntrinsicInst>(II->getOperand(0)))

llvm/test/CodeGen/AArch64/fmlal-loreg.ll

@@ -86,13 +86,12 @@ define void @sink(ptr %out_tile, ptr %lhs_panel, ptr %rhs_panel, i32 noundef %K,
 ; CHECK-NEXT:    movi v1.2d, #0000000000000000
 ; CHECK-NEXT:    mov w8, w3
 ; CHECK-NEXT:    movi v2.2d, #0000000000000000
-; CHECK-NEXT:    dup v0.8h, v0.h[0]
 ; CHECK-NEXT:  .LBB2_1: // %for.body
 ; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    ldr q3, [x2], #2
 ; CHECK-NEXT:    subs x8, x8, #1
-; CHECK-NEXT:    fmlal v1.4s, v3.4h, v0.4h
-; CHECK-NEXT:    fmlal2 v2.4s, v3.4h, v0.4h
+; CHECK-NEXT:    fmlal v1.4s, v3.4h, v0.h[0]
+; CHECK-NEXT:    fmlal2 v2.4s, v3.4h, v0.h[0]
 ; CHECK-NEXT:    b.ne .LBB2_1
 ; CHECK-NEXT:  // %bb.2: // %for.cond.cleanup
 ; CHECK-NEXT:    stp q1, q2, [x0]