Index: lib/Target/X86/X86ISelLowering.h
===================================================================
--- lib/Target/X86/X86ISelLowering.h
+++ lib/Target/X86/X86ISelLowering.h
@@ -835,6 +835,9 @@
     EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
                            EVT VT) const override;
 
+    bool targetShrinkDemandedConstant(SDValue Op, const APInt &Demanded,
+                                      TargetLoweringOpt &TLO) const override;
+
     /// Determine which of the bits specified in Mask are known to be either
     /// zero or one and return them in the KnownZero/KnownOne bitsets.
     void computeKnownBitsForTargetNode(const SDValue Op,
Index: lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -27810,6 +27810,65 @@
 //                           X86 Optimization Hooks
 //===----------------------------------------------------------------------===//
 
+bool
+X86TargetLowering::targetShrinkDemandedConstant(SDValue Op,
+                                                const APInt &Demanded,
+                                                TargetLoweringOpt &TLO) const {
+  // Only optimize Ands to prevent shrinking a constant that could be
+  // matched by movzx.
+  if (Op.getOpcode() != ISD::AND)
+    return false;
+
+  EVT VT = Op.getValueType();
+
+  // Ignore vectors.
+  if (VT.isVector())
+    return false;
+
+  unsigned Size = VT.getSizeInBits();
+
+  // Make sure the RHS really is a constant.
+  ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+  if (!C)
+    return false;
+
+  const APInt &Mask = C->getAPIntValue();
+
+  // Clear all non-demanded bits initially.
+  APInt ShrunkMask = Mask & Demanded;
+
+  // Find the width of the shrunk mask.
+  unsigned Width = ShrunkMask.getActiveBits();
+
+  // If the mask is all 0s there's nothing to do here.
+  if (Width == 0)
+    return false;
+
+  // Find the next power of 2 width, rounding up to a byte.
+  Width = PowerOf2Ceil(std::max(Width, 8U));
+  // Truncate the width to size to handle illegal types.
+  Width = std::min(Width, Size);
+
+  // Calculate a possible zero extend mask for this constant.
+  APInt ZeroExtendMask = APInt::getLowBitsSet(Size, Width);
+
+  // If we aren't changing the mask, just return true to keep it and prevent
+  // the caller from optimizing.
+  if (ZeroExtendMask == Mask)
+    return true;
+
+  // Make sure the bits in the ZeroExtendMask are also set in the original mask.
+  // TODO: We should be able to set bits that aren't demanded too.
+  if (!ZeroExtendMask.isSubsetOf(Mask))
+    return false;
+
+  // Replace the constant with the zero extend mask.
+  SDLoc DL(Op);
+  SDValue NewC = TLO.DAG.getConstant(ZeroExtendMask, DL, VT);
+  SDValue NewOp = TLO.DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), NewC);
+  return TLO.CombineTo(Op, NewOp);
+}
+
 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
                                                       KnownBits &Known,
                                                       const APInt &DemandedElts,
Index: lib/Target/X86/X86InstrCompiler.td
===================================================================
--- lib/Target/X86/X86InstrCompiler.td
+++ lib/Target/X86/X86InstrCompiler.td
@@ -1514,6 +1514,10 @@
           (EXTRACT_SUBREG GR16:$src, sub_8bit)>,
       Requires<[In64BitMode]>;
 
+def immff00_ffff  : ImmLeaf<i32, [{
+  return Imm >= 0xff00 && Imm <= 0xffff;
+}]>;
+
 // h-register tricks
 def : Pat<(i8 (trunc (srl_su GR16:$src, (i8 8)))),
           (EXTRACT_SUBREG GR16:$src, sub_8bit_hi)>,
@@ -1534,7 +1538,7 @@
           (MOVZX32_NOREXrr8 (EXTRACT_SUBREG GR16:$src, sub_8bit_hi))>;
 def : Pat<(and (srl_su GR32:$src, (i8 8)), (i32 255)),
           (MOVZX32_NOREXrr8 (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
-def : Pat<(srl (and_su GR32:$src, 0xff00), (i8 8)),
+def : Pat<(srl (and_su GR32:$src, immff00_ffff), (i8 8)),
           (MOVZX32_NOREXrr8 (EXTRACT_SUBREG GR32:$src, sub_8bit_hi))>;
 
 // h-register tricks.
Index: test/CodeGen/X86/3addr-or.ll
===================================================================
--- test/CodeGen/X86/3addr-or.ll
+++ test/CodeGen/X86/3addr-or.ll
@@ -10,12 +10,14 @@
   ret i32 %1
 }
 
+; This test no long requires or to be converted to 3 addr form because we are
+; are able to use a zero extend instead of an 'and' which gives the register
+; allocator freedom.
 define i64 @test2(i8 %A, i8 %B) nounwind {
 ; CHECK-LABEL: test2:
+; CHECK: movzbl
 ; CHECK: shrq $4
-; CHECK-NOT: movq
-; CHECK-NOT: orq
-; CHECK: leaq
+; CHECK: orq
 ; CHECK: ret
   %C = zext i8 %A to i64                          ; <i64> [#uses=1]
   %D = shl i64 %C, 4                              ; <i64> [#uses=1]
Index: test/CodeGen/X86/popcnt.ll
===================================================================
--- test/CodeGen/X86/popcnt.ll
+++ test/CodeGen/X86/popcnt.ll
@@ -70,8 +70,8 @@
 ; X32-NEXT:    shrl $2, %eax
 ; X32-NEXT:    andl $13107, %eax # imm = 0x3333
 ; X32-NEXT:    addl %ecx, %eax
+; X32-NEXT:    movzwl %ax, %eax
 ; X32-NEXT:    movl %eax, %ecx
-; X32-NEXT:    andl $32752, %ecx # imm = 0x7FF0
 ; X32-NEXT:    shrl $4, %ecx
 ; X32-NEXT:    addl %eax, %ecx
 ; X32-NEXT:    andl $3855, %ecx # imm = 0xF0F
@@ -93,15 +93,15 @@
 ; X64-NEXT:    shrl $2, %edi
 ; X64-NEXT:    andl $13107, %edi # imm = 0x3333
 ; X64-NEXT:    addl %eax, %edi
-; X64-NEXT:    movl %edi, %eax
-; X64-NEXT:    andl $32752, %eax # imm = 0x7FF0
-; X64-NEXT:    shrl $4, %eax
-; X64-NEXT:    addl %edi, %eax
-; X64-NEXT:    andl $3855, %eax # imm = 0xF0F
+; X64-NEXT:    movzwl %di, %eax
 ; X64-NEXT:    movl %eax, %ecx
-; X64-NEXT:    shll $8, %ecx
+; X64-NEXT:    shrl $4, %ecx
 ; X64-NEXT:    addl %eax, %ecx
-; X64-NEXT:    movzbl %ch, %eax # NOREX
+; X64-NEXT:    andl $3855, %ecx # imm = 0xF0F
+; X64-NEXT:    movl %ecx, %eax
+; X64-NEXT:    shll $8, %eax
+; X64-NEXT:    addl %ecx, %eax
+; X64-NEXT:    movzbl %ah, %eax # NOREX
 ; X64-NEXT:    # kill: def %ax killed %ax killed %eax
 ; X64-NEXT:    retq
 ;
Index: test/CodeGen/X86/pr21792.ll
===================================================================
--- test/CodeGen/X86/pr21792.ll
+++ test/CodeGen/X86/pr21792.ll
@@ -12,19 +12,18 @@
 ; CHECK-NEXT:    pushq %rax
 ; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    pand {{.*}}(%rip), %xmm0
-; CHECK-NEXT:    pextrq $1, %xmm0, %rdx
-; CHECK-NEXT:    movq %rdx, %rcx
-; CHECK-NEXT:    shrq $32, %rcx
-; CHECK-NEXT:    movq %xmm0, %rax
-; CHECK-NEXT:    movq %rax, %r9
+; CHECK-NEXT:    pextrq $1, %xmm0, %rax
+; CHECK-NEXT:    movzwl %ax, %ecx
+; CHECK-NEXT:    shrq $32, %rax
+; CHECK-NEXT:    movq %xmm0, %rdx
+; CHECK-NEXT:    movzwl %dx, %r8d
+; CHECK-NEXT:    movq %rdx, %r9
 ; CHECK-NEXT:    shrq $32, %r9
-; CHECK-NEXT:    andl $2032, %eax # imm = 0x7F0
-; CHECK-NEXT:    leaq stuff(%rax), %rdi
+; CHECK-NEXT:    leaq stuff(%r8), %rdi
 ; CHECK-NEXT:    leaq stuff(%r9), %rsi
-; CHECK-NEXT:    andl $2032, %edx # imm = 0x7F0
-; CHECK-NEXT:    leaq stuff(%rdx), %rdx
-; CHECK-NEXT:    leaq stuff(%rcx), %rcx
-; CHECK-NEXT:    leaq stuff+8(%rax), %r8
+; CHECK-NEXT:    leaq stuff(%rcx), %rdx
+; CHECK-NEXT:    leaq stuff(%rax), %rcx
+; CHECK-NEXT:    leaq stuff+8(%r8), %r8
 ; CHECK-NEXT:    leaq stuff+8(%r9), %r9
 ; CHECK-NEXT:    callq toto
 ; CHECK-NEXT:    popq %rax