Subzero: Aggressive LEA

src/IceTargetLoweringX86BaseImpl.h

 //===- subzero/src/IceTargetLoweringX86BaseImpl.h - x86 lowering -*- C++ -*-==//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
@@ skipping 2166 matching lines @@
       break;
     }
     return;
   }
   Variable *T_edx = nullptr;
   Variable *T = nullptr;
   switch (Instr->getOp()) {
   case InstArithmetic::_num:
     llvm_unreachable("Unknown arithmetic operator");
     break;
-  case InstArithmetic::Add:
-    _mov(T, Src0);
-    _add(T, Src1);
-    _mov(Dest, T);
-    break;
+  case InstArithmetic::Add: {

[John, 2016/07/11 21:28:43]

LEA are pretty powerful. You can use it as a general purpose, 4-address
instructions in x86. E.g.,

x = add i32 y, z     -> lea x, (y_reg, z_reg, 0)
x = add i32 y, const -> lea x, const(y_reg,,0)
x = add i32 const, z -> lea x, const(z_reg,,0)
x = sub i32 y, const -> lea x, -const(y_reg,,0)

If I am not mistaken, LEA can also generate 16 bit addresses with 32-bit address
modes.

Finally, you can't really optimize for i64 unles Traits::Is64Bit

+    Constant *Const = llvm::dyn_cast<Constant>(Instr->getSrc(1));
+    if (getFlags().getAggressiveLea() && Func->getOptLevel() == Opt_2 &&
+        Const != nullptr && (llvm::isa<ConstantInteger32>(Const) ||
+                             llvm::isa<ConstantRelocatable>(Const)) &&
+        (Ty == IceType_i32 || Ty == IceType_i64)) {
+      auto Var = legalize(Src0, Legal_Reg);
+      auto Mem = Traits::X86OperandMem::create(
+          Func, IceType_void, llvm::cast<Variable>(Var), Const);
+      T = makeReg(Ty);
+      _lea(T, _sandbox_mem_reference(Mem));
+      _mov(Dest, T);
+    } else {
+      _mov(T, Src0);
+      _add(T, Src1);
+      _mov(Dest, T);
+    }
+  } break;
   case InstArithmetic::And:
     _mov(T, Src0);
     _and(T, Src1);
     _mov(Dest, T);
     break;
   case InstArithmetic::Or:
     _mov(T, Src0);
     _or(T, Src1);
     _mov(Dest, T);
     break;
@@ skipping 5830 matching lines @@
         emitGlobal(*Var, SectionSuffix);
       }
     }
   } break;
   }
 }
 } // end of namespace X86NAMESPACE
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICETARGETLOWERINGX86BASEIMPL_H