Index: llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
===================================================================
--- llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -1428,7 +1428,7 @@
   RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
   if (RN && RN->getReg() == 0)
     Base = CurDAG->getRegister(0, MVT::i64);
-  else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(N)) {
+  else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(Base)) {
     // Base could already be %rip, particularly in the x32 ABI.
     Base = SDValue(CurDAG->getMachineNode(
                        TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
Index: llvm/trunk/lib/Target/X86/X86RegisterInfo.cpp
===================================================================
--- llvm/trunk/lib/Target/X86/X86RegisterInfo.cpp
+++ llvm/trunk/lib/Target/X86/X86RegisterInfo.cpp
@@ -489,6 +489,12 @@
   else
     BasePtr = (TFI->hasFP(MF) ? FramePtr : StackPtr);
 
+  // For LEA64_32r when BasePtr is 32-bits (X32) we can use full-size 64-bit
+  // register as source operand, semantic is the same and destination is
+  // 32-bits. It saves one byte per lea in code since 0x67 prefix is avoided.
+  if (Opc == X86::LEA64_32r && X86::GR32RegClass.contains(BasePtr))
+    BasePtr = getX86SubSuperRegister(BasePtr, MVT::i64, false);
+
   // This must be part of a four operand memory reference.  Replace the
   // FrameIndex with base register with EBP.  Add an offset to the offset.
   MI.getOperand(FIOperandNum).ChangeToRegister(BasePtr, false);
Index: llvm/trunk/test/CodeGen/X86/lea-2.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/lea-2.ll
+++ llvm/trunk/test/CodeGen/X86/lea-2.ll
@@ -1,4 +1,7 @@
-; RUN: llc < %s -march=x86 -x86-asm-syntax=intel | FileCheck %s
+; RUN: llc < %s -mtriple=i686-linux -x86-asm-syntax=intel | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux -x86-asm-syntax=intel | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 -x86-asm-syntax=intel | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-nacl -x86-asm-syntax=intel | FileCheck %s
 
 define i32 @test1(i32 %A, i32 %B) {
   %tmp1 = shl i32 %A, 2
Index: llvm/trunk/test/CodeGen/X86/lea-3.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/lea-3.ll
+++ llvm/trunk/test/CodeGen/X86/lea-3.ll
@@ -1,4 +1,6 @@
 ; RUN: llc < %s -mtriple=x86_64-linux | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-nacl | FileCheck %s
 ; RUN: llc < %s -mtriple=x86_64-win32 | FileCheck %s
 
 ; CHECK: leaq (,[[A0:%rdi|%rcx]],4), %rax
Index: llvm/trunk/test/CodeGen/X86/lea-4.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/lea-4.ll
+++ llvm/trunk/test/CodeGen/X86/lea-4.ll
@@ -1,4 +1,7 @@
-; RUN: llc < %s -march=x86-64 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-nacl | FileCheck %s
+
 
 define zeroext i16 @t1(i32 %on_off) nounwind {
 entry:
Index: llvm/trunk/test/CodeGen/X86/lea-5.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/lea-5.ll
+++ llvm/trunk/test/CodeGen/X86/lea-5.ll
@@ -0,0 +1,59 @@
+; test for more complicated forms of lea operands which can be generated
+; in loop optimized cases.
+; See also http://llvm.org/bugs/show_bug.cgi?id=20016
+
+; RUN: llc < %s -mtriple=x86_64-linux -O2        | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 -O2 | FileCheck %s -check-prefix=X32
+; RUN: llc < %s -mtriple=x86_64-nacl -O2 | FileCheck %s -check-prefix=X32
+
+; Function Attrs: nounwind readnone uwtable
+define void @foo(i32 %x, i32 %d) #0 {
+entry:
+  %a = alloca [8 x i32], align 16
+  br label %while.cond
+
+while.cond:                                       ; preds = %while.cond, %entry
+  %d.addr.0 = phi i32 [ %d, %entry ], [ %inc, %while.cond ]
+  %arrayidx = getelementptr inbounds [8 x i32]* %a, i32 0, i32 %d.addr.0
+
+; CHECK: leaq	-40(%rsp,%r{{[^,]*}},4), %rax
+; X32:   leal	-40(%rsp,%r{{[^,]*}},4), %eax
+  %0 = load i32* %arrayidx, align 4
+  %cmp1 = icmp eq i32 %0, 0
+  %inc = add nsw i32 %d.addr.0, 1
+
+; CHECK: leaq	4(%r{{[^,]*}}), %r{{[^,]*}}
+; X32:   leal	4(%r{{[^,]*}}), %e{{[^,]*}}
+  br i1 %cmp1, label %while.end, label %while.cond
+
+while.end:                                        ; preds = %while.cond
+  ret void
+}
+
+; The same test as above but with enforsed stack realignment (%a aligned by 64)
+; to check one more case of correct lea generation.
+
+; Function Attrs: nounwind readnone uwtable
+define void @bar(i32 %x, i32 %d) #0 {
+entry:
+  %a = alloca [8 x i32], align 64
+  br label %while.cond
+
+while.cond:                                       ; preds = %while.cond, %entry
+  %d.addr.0 = phi i32 [ %d, %entry ], [ %inc, %while.cond ]
+  %arrayidx = getelementptr inbounds [8 x i32]* %a, i32 0, i32 %d.addr.0
+
+; CHECK: leaq	(%rsp,%r{{[^,]*}},4), %rax
+; X32:   leal	(%rsp,%r{{[^,]*}},4), %eax
+  %0 = load i32* %arrayidx, align 4
+  %cmp1 = icmp eq i32 %0, 0
+  %inc = add nsw i32 %d.addr.0, 1
+
+; CHECK: leaq	4(%r{{[^,]*}}), %r{{[^,]*}}
+; X32:   leal	4(%r{{[^,]*}}), %e{{[^,]*}}
+  br i1 %cmp1, label %while.end, label %while.cond
+
+while.end:                                        ; preds = %while.cond
+  ret void
+}
+
Index: llvm/trunk/test/CodeGen/X86/lea.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/lea.ll
+++ llvm/trunk/test/CodeGen/X86/lea.ll
@@ -1,5 +1,7 @@
 ; RUN: llc < %s -mtriple=x86_64-linux | FileCheck %s
 ; RUN: llc < %s -mtriple=x86_64-win32 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 | FileCheck %s
+; RUN: llc < %s -mtriple=x86_64-nacl | FileCheck %s
 
 define i32 @test1(i32 %x) nounwind {
         %tmp1 = shl i32 %x, 3