[AArch64][SVE] Add SubtargetFeature to disable lowering unpredicated loads/stores as LDR/STR

PR llvm#127837 changed the lowering for unpredicated loads/stores to use LDR/STR instead of LD1/ST1.
However, on some CPUs, such as A64FX, there is a performance difference between LD1/ST1 and LDR/STR.
As a result, the lowering introduced in llvm#127837 can cause a performance regression on these targets.
This patch adds a SubtargetFeature to disable this lowering and prevent the regression.

Author: kinoshita-fj

llvm/lib/Target/AArch64/AArch64Features.td

@@ -915,6 +915,10 @@ def FeatureUseWzrToVecMove : SubtargetFeature<"use-wzr-to-vec-move",
                                               "UseWzrToVecMove", "true",
                                               "Move from WZR to insert 0 into vector registers">;
 
+def FeatureDisableUnpredicatedLdStLower : SubtargetFeature<
+    "disable-unpredicated-ld-st-lower", "DisableUnpredicatedLdStLower",
+    "true", "Disable lowering unpredicated loads/stores as LDR/STR">;
+
 //===----------------------------------------------------------------------===//
 // Architectures.
 //

llvm/lib/Target/AArch64/AArch64InstrInfo.td

@@ -443,6 +443,8 @@ def AllowMisalignedMemAccesses
 
 def UseWzrToVecMove : Predicate<"Subtarget->useWzrToVecMove()">;
 
+def AllowUnpredicatedLdStLower
+                    : Predicate<"!Subtarget->disableUnpredicatedLdStLower()">;
 
 //===----------------------------------------------------------------------===//
 // AArch64-specific DAG Nodes.

llvm/lib/Target/AArch64/AArch64Processors.td

@@ -306,7 +306,8 @@ def TuneA64FX : SubtargetFeature<"a64fx", "ARMProcFamily", "A64FX",
                                  FeatureAggressiveFMA,
                                  FeatureArithmeticBccFusion,
                                  FeatureStorePairSuppress,
-                                 FeaturePredictableSelectIsExpensive]>;
+                                 FeaturePredictableSelectIsExpensive,
+                                 FeatureDisableUnpredicatedLdStLower]>;
 
 def TuneMONAKA : SubtargetFeature<"fujitsu-monaka", "ARMProcFamily", "MONAKA",
                                  "Fujitsu FUJITSU-MONAKA processors", [

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

@@ -3164,7 +3164,7 @@ let Predicates = [HasSVE_or_SME] in {
   }
 
   // Allow using LDR/STR to avoid the predicate dependence.
-  let Predicates = [HasSVE_or_SME, IsLE, AllowMisalignedMemAccesses] in
+  let Predicates = [HasSVE_or_SME, IsLE, AllowMisalignedMemAccesses, AllowUnpredicatedLdStLower] in
     foreach Ty = [ nxv16i8, nxv8i16, nxv4i32, nxv2i64, nxv8f16, nxv4f32, nxv2f64, nxv8bf16 ] in {
       let AddedComplexity = 2 in {
         def : Pat<(Ty (load (am_sve_indexed_s9 GPR64sp:$base, simm9:$offset))),

llvm/test/CodeGen/AArch64/sve-disable-unpredicated-load-store-lower.ll

@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve,+disable-unpredicated-ld-st-lower < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve < %s | FileCheck --check-prefix CHECK-DEFAULT %s
+; RUN: llc -mcpu=a64fx < %s | FileCheck --check-prefix CHECK-A64FX %s
+
+define void @nxv2i64(ptr %ldptr, ptr %stptr) {
+; CHECK-LABEL: nxv2i64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-NEXT:    st1d { z0.d }, p0, [x1]
+; CHECK-NEXT:    ret
+;
+; CHECK-DEFAULT-LABEL: nxv2i64:
+; CHECK-DEFAULT:       // %bb.0:
+; CHECK-DEFAULT-NEXT:    ldr z0, [x0]
+; CHECK-DEFAULT-NEXT:    str z0, [x1]
+; CHECK-DEFAULT-NEXT:    ret
+;
+; CHECK-A64FX-LABEL: nxv2i64:
+; CHECK-A64FX:       // %bb.0:
+; CHECK-A64FX-NEXT:    ptrue p0.d
+; CHECK-A64FX-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-A64FX-NEXT:    st1d { z0.d }, p0, [x1]
+; CHECK-A64FX-NEXT:    ret
+  %l3 = load <vscale x 2 x i64>, ptr %ldptr, align 8
+  store <vscale x 2 x i64> %l3, ptr %stptr, align 8
+  ret void
+}