[llvm] [AArch64][SelectionDAG] Avoid cross-bank copy for NEON vcvtfp2fx result (PR #210275)

[via llvm-commits]

================
@@ -9121,6 +9121,49 @@ defm FCVTZS : SIMDFPScalarRShift<0, 0b11111, "fcvtzs">;
 defm FCVTZU : SIMDFPScalarRShift<1, 0b11111, "fcvtzu">;
 defm SCVTF  : SIMDFPScalarRShift<0, 0b11100, "scvtf">;
 defm UCVTF  : SIMDFPScalarRShift<1, 0b11100, "ucvtf">;
+
+// Transformation for SIMD shift imm to fixed point imm for FPR-to-GPR result
+def fixedpoint_scalar_xform : SDNodeXForm<timm, [{
+  (void)N;
+  return V;
+}]>;
+
+multiclass FPToFixedScalarPats<SDPatternOperator OpN, string INST > {
+// Allow integer result to remain in GPR registers
+// SelectionDAG-only as GIsel doesn't import fixedpoint_scalar_xform
+// Give priority over generic FPR fallback
+let AddedComplexity = 1 in {
+  def : Pat<(i32 (OpN FPR32:$Rn, vecshiftR32:$imm)),
+          (!cast<Instruction>(INST # "SWSri") FPR32:$Rn, (fixedpoint_scalar_xform vecshiftR32:$imm))>;
+  def : Pat<(i64 (OpN (f64 FPR64:$Rn), vecshiftR64:$imm)),
+          (!cast<Instruction>(INST # "SXDri") FPR64:$Rn, (fixedpoint_scalar_xform vecshiftR64:$imm))>;
+  
+  def : Pat<(i32 (OpN (f16 FPR16:$Rn), vecshiftR32:$imm)),
+          (!cast<Instruction>(INST # "SWHri") FPR16:$Rn, (fixedpoint_scalar_xform vecshiftR32:$imm))>;
+  def : Pat<(i64 (OpN (f16 FPR16:$Rn), vecshiftR64:$imm)),
+          (!cast<Instruction>(INST # "SXHri") FPR16:$Rn, (fixedpoint_scalar_xform vecshiftR64:$imm))>;
+  }
+
+  // Bitcast results kept in FP/SIMD registers.
+  def : Pat<(f32 (bitconvert(i32 (OpN FPR32:$Rn, vecshiftR32:$imm)))),
+          (!cast<Instruction>(INST # "s") FPR32:$Rn, vecshiftR32:$imm)>;
+  def : Pat<(f64 (bitconvert(i64 (OpN (f64 FPR64:$Rn), vecshiftR64:$imm)))),
+          (!cast<Instruction>(INST # "d") FPR64:$Rn, vecshiftR64:$imm)>;
+
+  def : Pat<(f32(bitconvert (i32 (OpN (f16 FPR16:$Rn), vecshiftR32:$imm)))),
----------------
Lukacma wrote:

These patterns shouldn't be added as they are incorrect based on newly proposed spec (see https://github.com/llvm/llvm-project/pull/201425/changes#r3603735005 ). Also this PR and #201425 are adding the same patterns for FCVT conversions. @CarolineConcatto could coordinate with @kieroxide  on the merging strategy of these 2 patches ?

https://github.com/llvm/llvm-project/pull/210275