[llvm] [AMDGPU] Add option to prevent insns straddling half cache-line boundaries (PR #150239)

[Matt Arsenault via llvm-commits]

================
@@ -274,12 +274,72 @@ static void emitVGPRBlockComment(const MachineInstr *MI, const SIInstrInfo *TII,
   OS.emitRawComment(" transferring at most " + TransferredRegs);
 }
 
+extern cl::opt<bool> PreventHalfCacheLineStraddling;
+
+static unsigned getMCInstSizeInBytes(const MCInst &LoweredMCI,
+                                     const GCNSubtarget &STI,
+                                     MCContext &OutContext) {
+  SmallVector<MCFixup, 4> Fixups;
+  SmallVector<char, 16> CodeBytes;
+
+  std::unique_ptr<MCCodeEmitter> InstEmitter(
+      createAMDGPUMCCodeEmitter(*STI.getInstrInfo(), OutContext));
+  InstEmitter->encodeInstruction(LoweredMCI, CodeBytes, Fixups, STI);
+  return (CodeBytes.size());
+};
+
 void AMDGPUAsmPrinter::emitInstruction(const MachineInstr *MI) {
   // FIXME: Enable feature predicate checks once all the test pass.
   // AMDGPU_MC::verifyInstructionPredicates(MI->getOpcode(),
   //                                        getSubtargetInfo().getFeatureBits());
 
+  auto AvoidHalfCacheLineBoundary = [this](const MachineInstr *MI,
+                                           const MachineFunction *MF,
+                                           const MCInst &LoweredMCI) -> void {
+    const GCNSubtarget &STI = MF->getSubtarget<GCNSubtarget>();
+    SIMachineFunctionInfo *MFI = const_cast<SIMachineFunctionInfo *>(
+        MF->getInfo<SIMachineFunctionInfo>());
+
+    unsigned InstSizeInBytes = STI.getInstrInfo()->getInstSizeInBytes(*MI);
----------------
arsenm wrote:

MC should never use getInstSizeInBytes

https://github.com/llvm/llvm-project/pull/150239