[llvm] r225503 - RegisterCoalescer: Fix removeCopyByCommutingDef with subreg liveness

[Matthias Braun]

Author: matze
Date: Thu Jan  8 21:01:31 2015
New Revision: 225503

URL: http://llvm.org/viewvc/llvm-project?rev=225503&view=rev
Log:
RegisterCoalescer: Fix removeCopyByCommutingDef with subreg liveness

The code that eliminated additional coalescable copies in
removeCopyByCommutingDef() used MergeValueNumberInto() which internally
may merge A into B or B into A. In this case A and B had different Def
points, so we have to reset ValNo.Def to the intended one after merging.

Added:
    llvm/trunk/test/CodeGen/X86/coalesce_commute_subreg.ll
Modified:
    llvm/trunk/lib/CodeGen/RegisterCoalescer.cpp

Modified: llvm/trunk/lib/CodeGen/RegisterCoalescer.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/RegisterCoalescer.cpp?rev=225503&r1=225502&r2=225503&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/RegisterCoalescer.cpp (original)
+++ llvm/trunk/lib/CodeGen/RegisterCoalescer.cpp Thu Jan  8 21:01:31 2015
@@ -751,7 +751,9 @@ bool RegisterCoalescer::removeCopyByComm
       if (!SubDVNI)
         continue;
       VNInfo *SubBValNo = S.getVNInfoAt(CopyIdx);
-      S.MergeValueNumberInto(SubBValNo, SubDVNI);
+      assert(SubBValNo->def == CopyIdx);
+      VNInfo *Merged = S.MergeValueNumberInto(SubBValNo, SubDVNI);
+      Merged->def = CopyIdx;
     }
 
     ErasedInstrs.insert(UseMI);

Added: llvm/trunk/test/CodeGen/X86/coalesce_commute_subreg.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/coalesce_commute_subreg.ll?rev=225503&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/coalesce_commute_subreg.ll (added)
+++ llvm/trunk/test/CodeGen/X86/coalesce_commute_subreg.ll Thu Jan  8 21:01:31 2015
@@ -0,0 +1,51 @@
+; RUN: llc -mtriple="x86_64-apple-darwin" -o - -verify-machineinstrs %s
+
+define void @make_wanted() #0 {
+entry:
+  br i1 undef, label %for.end20, label %for.cond1.preheader.lr.ph
+
+for.cond1.preheader.lr.ph:
+  br label %for.body3
+
+for.body3:
+  %cmp20.i = icmp eq i32 undef, 0
+  %.col.057 = select i1 %cmp20.i, i32 0, i32 undef
+  br i1 undef, label %while.cond.i, label %for.body5.lr.ph.i
+
+for.body5.lr.ph.i:
+  %0 = sext i32 %.col.057 to i64
+  %1 = sub i32 0, %.col.057
+  %2 = zext i32 %1 to i64
+  %3 = add nuw nsw i64 %2, 1
+  %n.vec110 = and i64 %3, 8589934588
+  %end.idx.rnd.down111 = add nsw i64 %n.vec110, %0
+  br i1 undef, label %middle.block105, label %vector.ph103
+
+vector.ph103:
+  br i1 undef, label %middle.block105, label %vector.body104
+
+vector.body104:
+  %4 = icmp eq i64 undef, %end.idx.rnd.down111
+  br i1 %4, label %middle.block105, label %vector.body104
+
+middle.block105:
+  %resume.val114 = phi i64 [ %0, %for.body5.lr.ph.i ], [ %end.idx.rnd.down111, %vector.body104 ], [ %end.idx.rnd.down111, %vector.ph103 ]
+  %cmp.n116 = icmp eq i64 undef, %resume.val114
+  br i1 %cmp.n116, label %while.cond.i, label %for.body5.i.preheader
+
+for.body5.i.preheader:
+  %lcmp.or182 = or i1 undef, undef
+  br i1 %lcmp.or182, label %for.body5.i.prol, label %while.cond.i
+
+for.body5.i.prol:
+  br i1 undef, label %for.body5.i.prol, label %while.cond.i
+
+while.cond.i:
+  br i1 undef, label %while.cond.i, label %if.then
+
+if.then:
+  br label %for.body3
+
+for.end20:
+  ret void
+}