Fix early-clobber handling in shrinkToUses.

I broke this in r144515, it affected most ARM testers.

<rdar://problem/10441389>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@144547 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/LiveIntervalAnalysis.cpp b/lib/CodeGen/LiveIntervalAnalysis.cpp
index 2ec2cbc5995da58bcec37b873f86273b26a60ecf..edcfebe866ef5158fb9069402620bd9055414b06 100644 (file)
--- a/lib/CodeGen/LiveIntervalAnalysis.cpp
+++ b/lib/CodeGen/LiveIntervalAnalysis.cpp
@@ -659,7 +659,9 @@ bool LiveIntervals::shrinkToUses(LiveInterval *li,
     if (UseMI->isDebugValue() || !UseMI->readsVirtualRegister(li->reg))
       continue;
     SlotIndex Idx = getInstructionIndex(UseMI).getRegSlot();
-    VNInfo *VNI = li->getVNInfoAt(Idx.getBaseIndex());
+    // Note: This intentionally picks up the wrong VNI in case of an EC redef.
+    // See below.
+    VNInfo *VNI = li->getVNInfoBefore(Idx);
     if (!VNI) {
       // This shouldn't happen: readsVirtualRegister returns true, but there is
       // no live value. It is likely caused by a target getting <undef> flags
@@ -669,10 +671,11 @@ bool LiveIntervals::shrinkToUses(LiveInterval *li,
                     << *li << '\n');
       continue;
     }
-    if (VNI->def == Idx.getRegSlot(true)) {
-      // Special case: An early-clobber tied operand reads and writes the
-      // register one slot early.
-      Idx = Idx.getRegSlot(true);
+    // Special case: An early-clobber tied operand reads and writes the
+    // register one slot early.  The getVNInfoBefore call above would have
+    // picked up the value defined by UseMI.  Adjust the kill slot and value.
+    if (SlotIndex::isSameInstr(VNI->def, Idx)) {
+      Idx = VNI->def;
       VNI = li->getVNInfoBefore(Idx);
       assert(VNI && "Early-clobber tied value not available");
     }
@@ -687,13 +690,6 @@ bool LiveIntervals::shrinkToUses(LiveInterval *li,
     if (VNI->isUnused())
       continue;
     NewLI.addRange(LiveRange(VNI->def, VNI->def.getDeadSlot(), VNI));
-
-    // A use tied to an early-clobber def ends at the load slot and isn't caught
-    // above. Catch it here instead. This probably only ever happens for inline
-    // assembly.
-    if (VNI->def.isEarlyClobber())
-      if (VNInfo *UVNI = li->getVNInfoBefore(VNI->def))
-        WorkList.push_back(std::make_pair(VNI->def, UVNI));
   }
 
   // Keep track of the PHIs that are in use.

diff --git a/test/CodeGen/ARM/2011-11-14-EarlyClobber.ll b/test/CodeGen/ARM/2011-11-14-EarlyClobber.ll
new file mode 100644 (file)
index 0000000..52aa0bf
--- /dev/null
+++ b/test/CodeGen/ARM/2011-11-14-EarlyClobber.ll
@@ -0,0 +1,62 @@
+; RUN: llc < %s -mcpu=cortex-a8 -verify-regalloc
+target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
+target triple = "thumbv7-apple-ios"
+
+; This test calls shrinkToUses with an early-clobber redefined live range during
+; spilling.
+;
+;   Shrink: %vreg47,1.158257e-02 = [384r,400e:0)[400e,420r:1)  0@384r 1@400e
+;
+; The early-clobber instruction is an str:
+;
+;   %vreg12<earlyclobber,def> = t2STR_PRE %vreg6, %vreg12, 32, pred:14, pred:%noreg
+;
+; This tests that shrinkToUses handles the EC redef correctly.
+
+%struct.Transform_Struct.0.11.12.17.43.46.56.58.60 = type { [4 x [4 x double]] }
+
+define void @Compute_Axis_Rotation_Transform(%struct.Transform_Struct.0.11.12.17.43.46.56.58.60* nocapture %transform, double* nocapture %V1, double %angle) nounwind {
+entry:
+  store double 1.000000e+00, double* null, align 4
+  %arrayidx5.1.i = getelementptr inbounds %struct.Transform_Struct.0.11.12.17.43.46.56.58.60* %transform, i32 0, i32 0, i32 0, i32 1
+  store double 0.000000e+00, double* %arrayidx5.1.i, align 4
+  %arrayidx5.2.i = getelementptr inbounds %struct.Transform_Struct.0.11.12.17.43.46.56.58.60* %transform, i32 0, i32 0, i32 0, i32 2
+  store double 0.000000e+00, double* %arrayidx5.2.i, align 4
+  %arrayidx5.114.i = getelementptr inbounds %struct.Transform_Struct.0.11.12.17.43.46.56.58.60* %transform, i32 0, i32 0, i32 1, i32 0
+  store double 0.000000e+00, double* %arrayidx5.114.i, align 4
+  %arrayidx5.1.1.i = getelementptr inbounds %struct.Transform_Struct.0.11.12.17.43.46.56.58.60* %transform, i32 0, i32 0, i32 1, i32 1
+  store double 1.000000e+00, double* %arrayidx5.1.1.i, align 4
+  store double 0.000000e+00, double* null, align 4
+  store double 1.000000e+00, double* null, align 4
+  store double 0.000000e+00, double* null, align 4
+  %call = tail call double @cos(double %angle) nounwind readnone
+  %call1 = tail call double @sin(double %angle) nounwind readnone
+  %0 = load double* %V1, align 4
+  %arrayidx2 = getelementptr inbounds double* %V1, i32 1
+  %1 = load double* %arrayidx2, align 4
+  %mul = fmul double %0, %1
+  %sub = fsub double 1.000000e+00, %call
+  %mul3 = fmul double %mul, %sub
+  %2 = load double* undef, align 4
+  %mul5 = fmul double %2, %call1
+  %add = fadd double %mul3, %mul5
+  store double %add, double* %arrayidx5.1.i, align 4
+  %3 = load double* %V1, align 4
+  %mul11 = fmul double %3, undef
+  %mul13 = fmul double %mul11, %sub
+  %4 = load double* %arrayidx2, align 4
+  %mul15 = fmul double %4, %call1
+  %sub16 = fsub double %mul13, %mul15
+  store double %sub16, double* %arrayidx5.2.i, align 4
+  %5 = load double* %V1, align 4
+  %6 = load double* %arrayidx2, align 4
+  %mul22 = fmul double %5, %6
+  %mul24 = fmul double %mul22, %sub
+  %sub27 = fsub double %mul24, undef
+  store double %sub27, double* %arrayidx5.114.i, align 4
+  ret void
+}
+
+declare double @cos(double) nounwind readnone
+
+declare double @sin(double) nounwind readnone