Allow MachineCSE to coalesce trivial subregister copies the same way

that it coalesces normal copies.

Without this, MachineCSE is powerless to handle redundant operations
with truncated source operands.

This required fixing the 2-addr pass to handle tied subregisters. It
isn't clear what combinations of subregisters can legally be tied, but
the simple case of truncated source operands is now safely handled:

     %vreg11<def> = COPY %vreg1:sub_32bit; GR32:%vreg11 GR64:%vreg1
     %vreg12<def> = COPY %vreg2:sub_32bit; GR32:%vreg12 GR64:%vreg2
     %vreg13<def,tied1> = ADD32rr %vreg11<tied0>, %vreg12<kill>, %EFLAGS<imp-def>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@197414 91177308-0d34-0410-b5e6-96231b3b80d8

test/CodeGen/X86/cse-add-with-overflow.ll

@@ -0,0 +1,42 @@
+; RUN: llc < %s -mtriple=x86_64-darwin -mcpu=generic | FileCheck %s
+; rdar:15661073 simple example of redundant adds
+;
+; MachineCSE should coalesce trivial subregister copies.
+;
+; The extra movl+addl should be removed during MachineCSE.
+; CHECK-LABEL: redundantadd
+; CHECK: cmpq
+; CHECK: movq
+; CHECK-NOT: movl
+; CHECK: addl
+; CHECK-NOT: addl
+; CHECK: ret
+
+define i64 @redundantadd(i64* %a0, i64* %a1) {
+entry:
+  %tmp8 = load i64* %a0, align 8
+  %tmp12 = load i64* %a1, align 8
+  %tmp13 = icmp ult i64 %tmp12, -281474976710656
+  br i1 %tmp13, label %exit1, label %body
+
+exit1:
+  unreachable
+
+body:
+  %tmp14 = trunc i64 %tmp8 to i32
+  %tmp15 = trunc i64 %tmp12 to i32
+  %tmp16 = tail call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %tmp14, i32 %tmp15)
+  %tmp17 = extractvalue { i32, i1 } %tmp16, 1
+  br i1 %tmp17, label %exit2, label %return
+
+exit2:
+  unreachable
+
+return:
+  %tmp18 = add i64 %tmp12, %tmp8
+  %tmp19 = and i64 %tmp18, 4294967295
+  %tmp20 = or i64 %tmp19, -281474976710656
+  ret i64 %tmp20
+}
+
+declare { i32, i1 } @llvm.sadd.with.overflow.i32(i32, i32)