[ARM][MVE] Add target flag for narrowing insts

Add a flag, 'RetainsPreviousHalfElement', for operations that operate
on top/bottom halves of their input and only write to half of their
destination, leaving the other half to retain its previous value.

Differential Revision: https://reviews.llvm.org/D76608

Author: sparker-arm

llvm/lib/Target/ARM/ARMInstrFormats.td

@@ -408,6 +408,7 @@ class InstTemplate<AddrMode am, int sz, IndexMode im,
   bit thumbArithFlagSetting = 0;
 
   bit validForTailPredication = 0;
+  bit retainsPreviousHalfElement = 0;
 
   // If this is a pseudo instruction, mark it isCodeGenOnly.
   let isCodeGenOnly = !eq(!cast<string>(f), "Pseudo");
@@ -421,6 +422,7 @@ class InstTemplate<AddrMode am, int sz, IndexMode im,
   let TSFlags{18-15} = D.Value;
   let TSFlags{19}    = thumbArithFlagSetting;
   let TSFlags{20}    = validForTailPredication;
+  let TSFlags{21}    = retainsPreviousHalfElement;
 
   let Constraints = cstr;
   let Itinerary = itin;

llvm/lib/Target/ARM/ARMInstrMVE.td

@@ -2688,6 +2688,7 @@ class MVE_VxSHRN<string iname, string suffix, bit bit_12, bit bit_28,
   let Inst{4} = 0b0;
   let Inst{0} = 0b1;
   let validForTailPredication = 1;
+  let retainsPreviousHalfElement = 1;
 }
 
 def MVE_VRSHRNi16bh : MVE_VxSHRN<"vrshrnb", "i16", 0b0, 0b1, shr_imm8> {
@@ -2730,6 +2731,7 @@ class MVE_VxQRSHRUN<string iname, string suffix, bit bit_28, bit bit_12,
   let Inst{4} = 0b0;
   let Inst{0} = 0b0;
   let validForTailPredication = 1;
+  let retainsPreviousHalfElement = 1;
 }
 
 def MVE_VQRSHRUNs16bh : MVE_VxQRSHRUN<
@@ -2779,6 +2781,7 @@ class MVE_VxQRSHRN<string iname, string suffix, bit bit_0, bit bit_12,
   let Inst{4} = 0b0;
   let Inst{0} = bit_0;
   let validForTailPredication = 1;
+  let retainsPreviousHalfElement = 1;
 }
 
 multiclass MVE_VxQRSHRN_types<string iname, bit bit_0, bit bit_12> {
@@ -4492,6 +4495,7 @@ class MVE_VxMOVxN<string iname, string suffix, bit bit_28, bit bit_17,
   let Inst{7} = !if(!eq(bit_17, 0), 1, 0);
   let Inst{0} = 0b1;
   let validForTailPredication = 1;
+  let retainsPreviousHalfElement = 1;
 }
 
 multiclass MVE_VxMOVxN_halves<string iname, string suffix,
@@ -4589,6 +4593,7 @@ class MVE_VCVT_ff<string iname, string suffix, bit op, bit T,
   let Inst{0} = 0b1;
 
   let Predicates = [HasMVEFloat];
+  let retainsPreviousHalfElement = 1;
 }
 
 multiclass MVE_VCVT_f2h_m<string iname, int half> {

llvm/lib/Target/ARM/MCTargetDesc/ARMBaseInfo.h

@@ -396,6 +396,10 @@ namespace ARMII {
     // Whether an instruction can be included in an MVE tail-predicated loop.
     ValidForTailPredication = 1 << 20,
 
+    // Whether an instruction writes to the top/bottom half of a vector element
+    // and leaves the other half untouched.
+    RetainsPreviousHalfElement = 1 << 21,
+
     //===------------------------------------------------------------------===//
     // Code ___domain.
     DomainShift   = 15,

llvm/unittests/Target/ARM/MachineInstrTest.cpp

@@ -10,11 +10,110 @@
 
 using namespace llvm;
 
+TEST(MachineInstructionRetainsPreviousHalfElement, IsCorrect) {
+  using namespace ARM;
+
+  auto RetainsPreviousHalfElement = [](unsigned Opcode) {
+    switch (Opcode) {
+    default:
+      break;
+    case MVE_VMOVNi16bh:
+    case MVE_VMOVNi16th:
+    case MVE_VMOVNi32bh:
+    case MVE_VMOVNi32th:
+    case MVE_VQMOVNs16bh:
+    case MVE_VQMOVNs16th:
+    case MVE_VQMOVNs32bh:
+    case MVE_VQMOVNs32th:
+    case MVE_VQMOVNu16bh:
+    case MVE_VQMOVNu16th:
+    case MVE_VQMOVNu32bh:
+    case MVE_VQMOVNu32th:
+    case MVE_VQMOVUNs16bh:
+    case MVE_VQMOVUNs16th:
+    case MVE_VQMOVUNs32bh:
+    case MVE_VQMOVUNs32th:
+    case MVE_VQRSHRNbhs16:
+    case MVE_VQRSHRNbhs32:
+    case MVE_VQRSHRNbhu16:
+    case MVE_VQRSHRNbhu32:
+    case MVE_VQRSHRNths16:
+    case MVE_VQRSHRNths32:
+    case MVE_VQRSHRNthu16:
+    case MVE_VQRSHRNthu32:
+    case MVE_VQRSHRUNs16bh:
+    case MVE_VQRSHRUNs16th:
+    case MVE_VQRSHRUNs32bh:
+    case MVE_VQRSHRUNs32th:
+    case MVE_VQSHRNbhs16:
+    case MVE_VQSHRNbhs32:
+    case MVE_VQSHRNbhu16:
+    case MVE_VQSHRNbhu32:
+    case MVE_VQSHRNths16:
+    case MVE_VQSHRNths32:
+    case MVE_VQSHRNthu16:
+    case MVE_VQSHRNthu32:
+    case MVE_VQSHRUNs16bh:
+    case MVE_VQSHRUNs16th:
+    case MVE_VQSHRUNs32bh:
+    case MVE_VQSHRUNs32th:
+    case MVE_VRSHRNi16bh:
+    case MVE_VRSHRNi16th:
+    case MVE_VRSHRNi32bh:
+    case MVE_VRSHRNi32th:
+    case MVE_VSHRNi16bh:
+    case MVE_VSHRNi16th:
+    case MVE_VSHRNi32bh:
+    case MVE_VSHRNi32th:
+    case MVE_VCVTf16f32bh:
+    case MVE_VCVTf16f32th:
+    case MVE_VCVTf32f16bh:
+    case MVE_VCVTf32f16th:
+      return true;
+    }
+    return false;
+  };
+
+  LLVMInitializeARMTargetInfo();
+  LLVMInitializeARMTarget();
+  LLVMInitializeARMTargetMC();
+
+  auto TT(Triple::normalize("thumbv8.1m.main-arm-none-eabi"));
+  std::string Error;
+  const Target *T = TargetRegistry::lookupTarget(TT, Error);
+  if (!T) {
+    dbgs() << Error;
+    return;
+  }
+
+  TargetOptions Options;
+  auto TM = std::unique_ptr<LLVMTargetMachine>(
+    static_cast<LLVMTargetMachine*>(
+      T->createTargetMachine(TT, "generic", "", Options, None, None,
+                             CodeGenOpt::Default)));
+  ARMSubtarget ST(TM->getTargetTriple(), std::string(TM->getTargetCPU()),
+                  std::string(TM->getTargetFeatureString()),
+                  *static_cast<const ARMBaseTargetMachine *>(TM.get()), false);
+  const ARMBaseInstrInfo *TII = ST.getInstrInfo();
+  auto MII = TM->getMCInstrInfo();
+
+  for (unsigned i = 0; i < ARM::INSTRUCTION_LIST_END; ++i) {
+    const MCInstrDesc &Desc = TII->get(i);
+
+    uint64_t Flags = Desc.TSFlags;
+    if ((Flags & ARMII::DomainMask) != ARMII::DomainMVE)
+      continue;
+
+    bool Valid = (Flags & ARMII::RetainsPreviousHalfElement) != 0;
+    ASSERT_EQ(RetainsPreviousHalfElement(i), Valid)
+              << MII->getName(i)
+              << ": mismatched expectation for tail-predicated safety\n";
+  }
+}
 // Test for instructions that aren't immediately obviously valid within a
 // tail-predicated loop. This should be marked up in their tablegen
 // descriptions. Currently we, conservatively, disallow:
 // - cross beat carries.
-// - narrowing of results.
 // - complex operations.
 // - horizontal operations.
 // - byte swapping.