From 450986db9607b1444cb50740fa4c102f20f92f6b Mon Sep 17 00:00:00 2001
From: Jakob Stoklund Olesen <stoklund@2pi.dk>
Date: Tue, 23 Feb 2010 22:44:02 +0000
Subject: [PATCH] Remember to handle sub-registers when moving imp-defs to a
 rematted instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@96995 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/SimpleRegisterCoalescing.cpp      | 19 ++++++-
 .../X86/2010-02-23-RematImplicitSubreg.ll     | 49 +++++++++++++++++++
 2 files changed, 66 insertions(+), 2 deletions(-)
 create mode 100644 test/CodeGen/X86/2010-02-23-RematImplicitSubreg.ll

diff --git a/lib/CodeGen/SimpleRegisterCoalescing.cpp b/lib/CodeGen/SimpleRegisterCoalescing.cpp
index dcbbb1a0045..73111750a1c 100644
--- a/lib/CodeGen/SimpleRegisterCoalescing.cpp
+++ b/lib/CodeGen/SimpleRegisterCoalescing.cpp
@@ -702,7 +702,8 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
     for (const unsigned* SR = tri_->getSubRegisters(DstReg); *SR; ++SR) {
       if (!li_->hasInterval(*SR))
         continue;
-      DLR = li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
+      const LiveRange *DLR =
+          li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
       if (DLR && DLR->valno->getCopy() == CopyMI)
         DLR->valno->setCopy(0);
     }
@@ -741,9 +742,21 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
       NewMI->addOperand(MO);
     if (MO.isDef() && li_->hasInterval(MO.getReg())) {
       unsigned Reg = MO.getReg();
-      DLR = li_->getInterval(Reg).getLiveRangeContaining(DefIdx);
+      const LiveRange *DLR =
+          li_->getInterval(Reg).getLiveRangeContaining(DefIdx);
       if (DLR && DLR->valno->getCopy() == CopyMI)
         DLR->valno->setCopy(0);
+      // Handle subregs as well
+      if (TargetRegisterInfo::isPhysicalRegister(Reg)) {
+        for (const unsigned* SR = tri_->getSubRegisters(Reg); *SR; ++SR) {
+          if (!li_->hasInterval(*SR))
+            continue;
+          const LiveRange *DLR =
+              li_->getInterval(*SR).getLiveRangeContaining(DefIdx);
+          if (DLR && DLR->valno->getCopy() == CopyMI)
+            DLR->valno->setCopy(0);
+        }
+      }
     }
   }
 
@@ -752,6 +765,7 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt,
   CopyMI->eraseFromParent();
   ReMatCopies.insert(CopyMI);
   ReMatDefs.insert(DefMI);
+  DEBUG(dbgs() << "Remat: " << *NewMI);
   ++NumReMats;
   return true;
 }
@@ -1705,6 +1719,7 @@ bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) {
         (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI) ||
          RemoveCopyByCommutingDef(SrcInt, DstInt, CopyMI))) {
       JoinedCopies.insert(CopyMI);
+      DEBUG(dbgs() << "Trivial!\n");
       return true;
     }
 
diff --git a/test/CodeGen/X86/2010-02-23-RematImplicitSubreg.ll b/test/CodeGen/X86/2010-02-23-RematImplicitSubreg.ll
new file mode 100644
index 00000000000..4a26ba088e5
--- /dev/null
+++ b/test/CodeGen/X86/2010-02-23-RematImplicitSubreg.ll
@@ -0,0 +1,49 @@
+; RUN: llc < %s
+; PR6372
+;
+; This test produces a move instruction with an implicitly defined super-register:
+;
+;   %DL<def> = MOV8rr %reg1038<kill>, %RDX<imp-def>
+;
+; When %DL is rematerialized, we must remember to update live intervals for
+; sub-registers %DX and %EDX.
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-apple-darwin10.0.0"
+
+define noalias i8* @foo() nounwind ssp {
+entry:
+  br i1 undef, label %for.end, label %for.body
+
+for.body:                                         ; preds = %if.end40, %entry
+  %tmp6 = load i8* undef, align 2                 ; <i8> [#uses=3]
+  %conv11 = sext i8 %tmp6 to i64                  ; <i64> [#uses=1]
+  %cmp15 = icmp slt i64 %conv11, undef            ; <i1> [#uses=1]
+  br i1 %cmp15, label %if.end, label %if.then
+
+if.then:                                          ; preds = %for.body
+  %conv18 = sext i8 %tmp6 to i32                  ; <i32> [#uses=1]
+  %call = tail call i32 (...)* @invalid(i32 0, i32 0, i32 %conv18) nounwind ; <i32> [#uses=0]
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %for.body
+  %index.0 = phi i8 [ 0, %if.then ], [ %tmp6, %for.body ] ; <i8> [#uses=1]
+  store i8 %index.0, i8* undef
+  %tmp24 = load i8* undef                         ; <i8> [#uses=2]
+  br i1 undef, label %if.end40, label %if.then36
+
+if.then36:                                        ; preds = %if.end
+  %conv38 = sext i8 %tmp24 to i32                 ; <i32> [#uses=1]
+  %call39 = tail call i32 (...)* @invalid(i32 0, i32 0, i32 %conv38) nounwind ; <i32> [#uses=0]
+  br label %if.end40
+
+if.end40:                                         ; preds = %if.then36, %if.end
+  %index.1 = phi i8 [ 0, %if.then36 ], [ %tmp24, %if.end ] ; <i8> [#uses=1]
+  store i8 %index.1, i8* undef
+  br i1 false, label %for.body, label %for.end
+
+for.end:                                          ; preds = %if.end40, %entry
+  ret i8* undef
+}
+
+declare i32 @invalid(...)