For AArch64, add missing register cost calculation for big value types like v4i64...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196456 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index f4b5af4c2c61da31e2474a97abf29a9c4acf1be4..4fdb667b95399a408594c977af9da9c81b105ec8 100644 (file)
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -66,7 +66,7 @@ AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
     addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
     addRegisterClass(MVT::v1f32, &AArch64::FPR32RegClass);
     addRegisterClass(MVT::v1f64, &AArch64::FPR64RegClass);
-    addRegisterClass(MVT::v8i8, &AArch64::FPR64RegClass);
+    addRegisterClass(MVT::v8i8,  &AArch64::FPR64RegClass);
     addRegisterClass(MVT::v4i16, &AArch64::FPR64RegClass);
     addRegisterClass(MVT::v2i32, &AArch64::FPR64RegClass);
     addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
@@ -409,6 +409,29 @@ static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
   StrOpc = StoreOps[Log2_32(Size)];
 }
 
+// FIXME: AArch64::DTripleRegClass and AArch64::QTripleRegClass don't really
+// have value type mapped, and they are both being defined as MVT::untyped.
+// Without knowing the MVT type, MachineLICM::getRegisterClassIDAndCost
+// would fail to figure out the register pressure correctly.
+std::pair<const TargetRegisterClass*, uint8_t>
+AArch64TargetLowering::findRepresentativeClass(MVT VT) const{
+  const TargetRegisterClass *RRC = 0;
+  uint8_t Cost = 1;
+  switch (VT.SimpleTy) {
+  default:
+    return TargetLowering::findRepresentativeClass(VT);
+  case MVT::v4i64:
+    RRC = &AArch64::QPairRegClass;
+    Cost = 2;
+    break;
+  case MVT::v8i64:
+    RRC = &AArch64::QQuadRegClass;
+    Cost = 4;
+    break;
+  }
+  return std::make_pair(RRC, Cost);
+}
+
 MachineBasicBlock *
 AArch64TargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
                                         unsigned Size,

diff --git a/lib/Target/AArch64/AArch64ISelLowering.h b/lib/Target/AArch64/AArch64ISelLowering.h
index 4cc21351435816905c6e415c9ec6781de30e456b..8ad5a79a33ee0a54bbcc051e0bd82f2a83688923 100644 (file)
--- a/lib/Target/AArch64/AArch64ISelLowering.h
+++ b/lib/Target/AArch64/AArch64ISelLowering.h
@@ -343,6 +343,10 @@ public:
   virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info, const CallInst &I,
                                   unsigned Intrinsic) const LLVM_OVERRIDE;
 
+protected:
+  std::pair<const TargetRegisterClass*, uint8_t>
+  findRepresentativeClass(MVT VT) const;
+
 private:
   const InstrItineraryData *Itins;
 

diff --git a/test/CodeGen/AArch64/neon-simd-ldst.ll b/test/CodeGen/AArch64/neon-simd-ldst.ll
new file mode 100644 (file)
index 0000000..afc0901
--- /dev/null
+++ b/test/CodeGen/AArch64/neon-simd-ldst.ll
@@ -0,0 +1,164 @@
+; RUN: llc < %s -O2 -verify-machineinstrs -mtriple=aarch64-none-linux-gnu -mattr=+neon | FileCheck %s
+
+define void @test_ldstq_4v(i8* noalias %io, i32 %count) {
+; CHECK-LABEL: test_ldstq_4v
+; CHECK: ld4     {v0.16b, v1.16b, v2.16b, v3.16b}, [x0]
+; CHECK: st4     {v0.16b, v1.16b, v2.16b, v3.16b}, [x0]
+entry:
+  %tobool62 = icmp eq i32 %count, 0
+  br i1 %tobool62, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %count.addr.063 = phi i32 [ %dec, %while.body ], [ %count, %entry ]
+  %dec = add i32 %count.addr.063, -1
+  %vld4 = tail call { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> } @llvm.arm.neon.vld4.v16i8(i8* %io, i32 1)
+  %vld4.fca.0.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> } %vld4, 0
+  %vld4.fca.1.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> } %vld4, 1
+  %vld4.fca.2.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> } %vld4, 2
+  %vld4.fca.3.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> } %vld4, 3
+  tail call void @llvm.arm.neon.vst4.v16i8(i8* %io, <16 x i8> %vld4.fca.0.extract, <16 x i8> %vld4.fca.1.extract, <16 x i8> %vld4.fca.2.extract, <16 x i8> %vld4.fca.3.extract, i32 1)
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare { <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8> } @llvm.arm.neon.vld4.v16i8(i8*, i32)
+
+declare void @llvm.arm.neon.vst4.v16i8(i8*, <16 x i8>, <16 x i8>, <16 x i8>, <16 x i8>, i32)
+
+define void @test_ldstq_3v(i8* noalias %io, i32 %count) {
+; CHECK-LABEL: test_ldstq_3v
+; CHECK: ld3     {v0.16b, v1.16b, v2.16b}, [x0]
+; CHECK: st3     {v0.16b, v1.16b, v2.16b}, [x0]
+entry:
+  %tobool47 = icmp eq i32 %count, 0
+  br i1 %tobool47, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %count.addr.048 = phi i32 [ %dec, %while.body ], [ %count, %entry ]
+  %dec = add i32 %count.addr.048, -1
+  %vld3 = tail call { <16 x i8>, <16 x i8>, <16 x i8> } @llvm.arm.neon.vld3.v16i8(i8* %io, i32 1)
+  %vld3.fca.0.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8> } %vld3, 0
+  %vld3.fca.1.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8> } %vld3, 1
+  %vld3.fca.2.extract = extractvalue { <16 x i8>, <16 x i8>, <16 x i8> } %vld3, 2
+  tail call void @llvm.arm.neon.vst3.v16i8(i8* %io, <16 x i8> %vld3.fca.0.extract, <16 x i8> %vld3.fca.1.extract, <16 x i8> %vld3.fca.2.extract, i32 1)
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare { <16 x i8>, <16 x i8>, <16 x i8> } @llvm.arm.neon.vld3.v16i8(i8*, i32)
+
+declare void @llvm.arm.neon.vst3.v16i8(i8*, <16 x i8>, <16 x i8>, <16 x i8>, i32)
+
+define void @test_ldstq_2v(i8* noalias %io, i32 %count) {
+; CHECK-LABEL: test_ldstq_2v
+; CHECK: ld2     {v0.16b, v1.16b}, [x0]
+; CHECK: st2     {v0.16b, v1.16b}, [x0]
+entry:
+  %tobool22 = icmp eq i32 %count, 0
+  br i1 %tobool22, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %count.addr.023 = phi i32 [ %dec, %while.body ], [ %count, %entry ]
+  %dec = add i32 %count.addr.023, -1
+  %vld2 = tail call { <16 x i8>, <16 x i8> } @llvm.arm.neon.vld2.v16i8(i8* %io, i32 1)
+  %vld2.fca.0.extract = extractvalue { <16 x i8>, <16 x i8> } %vld2, 0
+  %vld2.fca.1.extract = extractvalue { <16 x i8>, <16 x i8> } %vld2, 1
+  tail call void @llvm.arm.neon.vst2.v16i8(i8* %io, <16 x i8> %vld2.fca.0.extract, <16 x i8> %vld2.fca.1.extract, i32 1)
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare { <16 x i8>, <16 x i8> } @llvm.arm.neon.vld2.v16i8(i8*, i32)
+
+declare void @llvm.arm.neon.vst2.v16i8(i8*, <16 x i8>, <16 x i8>, i32)
+
+define void @test_ldst_4v(i8* noalias %io, i32 %count) {
+; CHECK-LABEL: test_ldst_4v
+; CHECK: ld4     {v0.8b, v1.8b, v2.8b, v3.8b}, [x0]
+; CHECK: st4     {v0.8b, v1.8b, v2.8b, v3.8b}, [x0]
+entry:
+  %tobool42 = icmp eq i32 %count, 0
+  br i1 %tobool42, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %count.addr.043 = phi i32 [ %dec, %while.body ], [ %count, %entry ]
+  %dec = add i32 %count.addr.043, -1
+  %vld4 = tail call { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> } @llvm.arm.neon.vld4.v8i8(i8* %io, i32 1)
+  %vld4.fca.0.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> } %vld4, 0
+  %vld4.fca.1.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> } %vld4, 1
+  %vld4.fca.2.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> } %vld4, 2
+  %vld4.fca.3.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> } %vld4, 3
+  tail call void @llvm.arm.neon.vst4.v8i8(i8* %io, <8 x i8> %vld4.fca.0.extract, <8 x i8> %vld4.fca.1.extract, <8 x i8> %vld4.fca.2.extract, <8 x i8> %vld4.fca.3.extract, i32 1)
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare { <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8> } @llvm.arm.neon.vld4.v8i8(i8*, i32)
+
+declare void @llvm.arm.neon.vst4.v8i8(i8*, <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8>, i32)
+
+define void @test_ldst_3v(i8* noalias %io, i32 %count) {
+; CHECK-LABEL: test_ldst_3v
+; CHECK: ld3     {v0.8b, v1.8b, v2.8b}, [x0]
+; CHECK: st3     {v0.8b, v1.8b, v2.8b}, [x0]
+entry:
+  %tobool32 = icmp eq i32 %count, 0
+  br i1 %tobool32, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %count.addr.033 = phi i32 [ %dec, %while.body ], [ %count, %entry ]
+  %dec = add i32 %count.addr.033, -1
+  %vld3 = tail call { <8 x i8>, <8 x i8>, <8 x i8> } @llvm.arm.neon.vld3.v8i8(i8* %io, i32 1)
+  %vld3.fca.0.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8> } %vld3, 0
+  %vld3.fca.1.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8> } %vld3, 1
+  %vld3.fca.2.extract = extractvalue { <8 x i8>, <8 x i8>, <8 x i8> } %vld3, 2
+  tail call void @llvm.arm.neon.vst3.v8i8(i8* %io, <8 x i8> %vld3.fca.0.extract, <8 x i8> %vld3.fca.1.extract, <8 x i8> %vld3.fca.2.extract, i32 1)
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare { <8 x i8>, <8 x i8>, <8 x i8> } @llvm.arm.neon.vld3.v8i8(i8*, i32)
+
+declare void @llvm.arm.neon.vst3.v8i8(i8*, <8 x i8>, <8 x i8>, <8 x i8>, i32)
+
+define void @test_ldst_2v(i8* noalias %io, i32 %count) {
+; CHECK-LABEL: test_ldst_2v
+; CHECK: ld2     {v0.8b, v1.8b}, [x0]
+; CHECK: st2     {v0.8b, v1.8b}, [x0]
+entry:
+  %tobool22 = icmp eq i32 %count, 0
+  br i1 %tobool22, label %while.end, label %while.body
+
+while.body:                                       ; preds = %entry, %while.body
+  %count.addr.023 = phi i32 [ %dec, %while.body ], [ %count, %entry ]
+  %dec = add i32 %count.addr.023, -1
+  %vld2 = tail call { <8 x i8>, <8 x i8> } @llvm.arm.neon.vld2.v8i8(i8* %io, i32 1)
+  %vld2.fca.0.extract = extractvalue { <8 x i8>, <8 x i8> } %vld2, 0
+  %vld2.fca.1.extract = extractvalue { <8 x i8>, <8 x i8> } %vld2, 1
+  tail call void @llvm.arm.neon.vst2.v8i8(i8* %io, <8 x i8> %vld2.fca.0.extract, <8 x i8> %vld2.fca.1.extract, i32 1)
+  %tobool = icmp eq i32 %dec, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.end:                                        ; preds = %while.body, %entry
+  ret void
+}
+
+declare { <8 x i8>, <8 x i8> } @llvm.arm.neon.vld2.v8i8(i8*, i32)
+
+declare void @llvm.arm.neon.vst2.v8i8(i8*, <8 x i8>, <8 x i8>, i32)
+