Index: lib/Target/X86/X86DomainReassignment.cpp
===================================================================
--- lib/Target/X86/X86DomainReassignment.cpp
+++ lib/Target/X86/X86DomainReassignment.cpp
@@ -422,6 +422,7 @@
   if (Domain != RD)
     return;
 
+  EnclosedEdges.insert(Reg);
   Worklist.push_back(Reg);
 }
 
Index: test/CodeGen/X86/gpr-to-mask.ll
===================================================================
--- test/CodeGen/X86/gpr-to-mask.ll
+++ test/CodeGen/X86/gpr-to-mask.ll
@@ -556,3 +556,98 @@
   store <8 x float> %selected, <8 x float>* %fptrvec
   ret void
 }
+
+define void @test_sel(i32 %in, double %d1, double %d2, i32* %ptr) {
+; X86-64-LABEL: test_sel:
+; X86-64:       # %bb.0: # %entry
+; X86-64-NEXT:    vxorpd %xmm2, %xmm2, %xmm2
+; X86-64-NEXT:    movl $2, %eax
+; X86-64-NEXT:    vcmpeqsd %xmm2, %xmm0, %k0
+; X86-64-NEXT:    testl %edi, %edi
+; X86-64-NEXT:    je .LBB10_3
+; X86-64-NEXT:    .p2align 4, 0x90
+; X86-64-NEXT:  .LBB10_2: # %if.then.i
+; X86-64-NEXT:    # =>This Inner Loop Header: Depth=1
+; X86-64-NEXT:    kmovd %k0, %ecx
+; X86-64-NEXT:    movzbl %cl, %ecx
+; X86-64-NEXT:    leal 1(%rcx,%rcx,2), %ecx
+; X86-64-NEXT:    movl %ecx, (%rsi)
+; X86-64-NEXT:    testl %edi, %edi
+; X86-64-NEXT:    jne .LBB10_2
+; X86-64-NEXT:  .LBB10_3: # %if.else6.i
+; X86-64-NEXT:    # =>This Inner Loop Header: Depth=1
+; X86-64-NEXT:    vucomisd %xmm2, %xmm1
+; X86-64-NEXT:    setp %cl
+; X86-64-NEXT:    setne %dl
+; X86-64-NEXT:    orb %cl, %dl
+; X86-64-NEXT:    movl $8, %ecx
+; X86-64-NEXT:    cmovnel %eax, %ecx
+; X86-64-NEXT:    movl %ecx, (%rsi)
+; X86-64-NEXT:    testl %edi, %edi
+; X86-64-NEXT:    jne .LBB10_2
+; X86-64-NEXT:    jmp .LBB10_3
+;
+; X86-32-LABEL: test_sel:
+; X86-32:       # %bb.0: # %entry
+; X86-32-NEXT:    pushl %ebx
+; X86-32-NEXT:    .cfi_def_cfa_offset 8
+; X86-32-NEXT:    pushl %edi
+; X86-32-NEXT:    .cfi_def_cfa_offset 12
+; X86-32-NEXT:    pushl %esi
+; X86-32-NEXT:    .cfi_def_cfa_offset 16
+; X86-32-NEXT:    .cfi_offset %esi, -16
+; X86-32-NEXT:    .cfi_offset %edi, -12
+; X86-32-NEXT:    .cfi_offset %ebx, -8
+; X86-32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; X86-32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-32-NEXT:    vxorpd %xmm1, %xmm1, %xmm1
+; X86-32-NEXT:    movl $2, %edx
+; X86-32-NEXT:    vcmpeqsd {{[0-9]+}}(%esp), %xmm1, %k0
+; X86-32-NEXT:    kmovb %k0, %k0
+; X86-32-NEXT:    leal 1(%k0,%k0,2), %esi
+; X86-32-NEXT:    testl %ecx, %ecx
+; X86-32-NEXT:    je .LBB10_3
+; X86-32-NEXT:    .p2align 4, 0x90
+; X86-32-NEXT:  .LBB10_2: # %if.then.i
+; X86-32-NEXT:    # =>This Inner Loop Header: Depth=1
+; X86-32-NEXT:    movl %esi, (%eax)
+; X86-32-NEXT:    testl %ecx, %ecx
+; X86-32-NEXT:    jne .LBB10_2
+; X86-32-NEXT:  .LBB10_3: # %if.else6.i
+; X86-32-NEXT:    # =>This Inner Loop Header: Depth=1
+; X86-32-NEXT:    vucomisd %xmm1, %xmm0
+; X86-32-NEXT:    setp %bl
+; X86-32-NEXT:    setne %bh
+; X86-32-NEXT:    orb %bl, %bh
+; X86-32-NEXT:    movl $8, %edi
+; X86-32-NEXT:    cmovnel %edx, %edi
+; X86-32-NEXT:    movl %edi, (%eax)
+; X86-32-NEXT:    testl %ecx, %ecx
+; X86-32-NEXT:    jne .LBB10_2
+; X86-32-NEXT:    jmp .LBB10_3
+entry:
+  %0 = icmp eq i32 %in, 0
+  br label %for.cond
+
+for.cond:
+  %1 = phi i32 [0, %entry], [%7, %exit]
+  br i1 %0, label %if.else6.i, label %if.then.i
+
+if.then.i:
+  %2 = fcmp une double %d1, 0.000000e+00
+  %3 = select i1 %2, i32 1, i32 4
+  br label %exit
+
+if.else6.i:
+  %4 = fcmp une double %d2, 0.000000e+00
+  %5 = select i1 %4, i32 2, i32 8
+  br label %exit
+
+exit:
+  %6 = phi i32 [%3, %if.then.i], [%5, %if.else6.i]
+  store i32 %6, i32* %ptr
+  %7 = add i32 %1, 1
+  br label %for.cond
+  ret void
+}