From e6e2d8cd90ceb5190aa646dc06584027f7d492d0 Mon Sep 17 00:00:00 2001
From: Jakob Stoklund Olesen <stoklund@2pi.dk>
Date: Thu, 20 Sep 2012 23:08:42 +0000
Subject: [PATCH] Ignore PHI-defs for -new-coalescer interference checks.

A PHI can't create interference on its own. If two live ranges interfere
at a PHI, they must also interfere when leaving one of the PHI
predecessors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@164330 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/RegisterCoalescer.cpp    | 12 ++++---
 test/CodeGen/ARM/coalesce-subregs.ll | 48 ++++++++++++++++++++++++++++
 2 files changed, 56 insertions(+), 4 deletions(-)

diff --git a/lib/CodeGen/RegisterCoalescer.cpp b/lib/CodeGen/RegisterCoalescer.cpp
index e9bfac964e2..dd0f5488675 100644
--- a/lib/CodeGen/RegisterCoalescer.cpp
+++ b/lib/CodeGen/RegisterCoalescer.cpp
@@ -1402,7 +1402,6 @@ JoinVals::analyzeValue(unsigned ValNo, JoinVals &Other) {
   // values should be merged into one, but not into any preceding value.
   // The first value defined or visited gets CR_Keep, the other gets CR_Merge.
   if (VNInfo *OtherVNI = OtherLRQ.valueDefined()) {
-    DEBUG(dbgs() << "\t\tDouble def: " << VNI->def << '\n');
     assert(SlotIndex::isSameInstr(VNI->def, OtherVNI->def) && "Broken LRQ");
 
     // One value stays, the other is merged. Keep the earlier one, or the first
@@ -1420,7 +1419,11 @@ JoinVals::analyzeValue(unsigned ValNo, JoinVals &Other) {
     // Keep this value, check for conflicts when analyzing OtherVNI.
     if (!OtherV.isAnalyzed())
       return CR_Keep;
-    // Both sides have been analyzed now. Do they conflict?
+    // Both sides have been analyzed now.
+    // Allow overlapping PHI values. Any real interference would show up in a
+    // predecessor, the PHI itself can't introduce any conflicts.
+    if (VNI->isPHIDef())
+      return CR_Merge;
     if (V.ValidLanes & OtherV.ValidLanes)
       // Overlapping lanes can't be resolved.
       return CR_Impossible;
@@ -1441,9 +1444,10 @@ JoinVals::analyzeValue(unsigned ValNo, JoinVals &Other) {
   Other.computeAssignment(V.OtherVNI->id, *this);
   const Val &OtherV = Other.Vals[V.OtherVNI->id];
 
-  // Don't attempt resolving PHI values for now.
+  // Allow overlapping PHI values. Any real interference would show up in a
+  // predecessor, the PHI itself can't introduce any conflicts.
   if (VNI->isPHIDef())
-    return CR_Impossible;
+    return CR_Replace;
 
   // Check for simple erasable conflicts.
   if (DefMI->isImplicitDef())
diff --git a/test/CodeGen/ARM/coalesce-subregs.ll b/test/CodeGen/ARM/coalesce-subregs.ll
index e5a88be64d2..8d7ded5be06 100644
--- a/test/CodeGen/ARM/coalesce-subregs.ll
+++ b/test/CodeGen/ARM/coalesce-subregs.ll
@@ -141,3 +141,51 @@ if.end:                                           ; preds = %entry, %if.then
   tail call void @llvm.arm.neon.vst1.v2f32(i8* %0, <2 x float> %x.0, i32 4)
   ret void
 }
+
+; CHECK: f5
+; Coalesce vector lanes through phis.
+; CHECK: vmov.f32 {{.*}}, #1.0
+; CHECK-NOT: vmov
+; CHECK-NOT: vorr
+; CHECK: %if.end
+; We may leave the last insertelement in the if.end block.
+; It is inserting the %add value into a dead lane, but %add causes interference
+; in the entry block, and we don't do dead lane checks across basic blocks.
+define void @f5(float* %p, float* %q) nounwind ssp {
+entry:
+  %0 = bitcast float* %p to i8*
+  %vld1 = tail call <4 x float> @llvm.arm.neon.vld1.v4f32(i8* %0, i32 4)
+  %vecext = extractelement <4 x float> %vld1, i32 0
+  %vecext1 = extractelement <4 x float> %vld1, i32 1
+  %vecext2 = extractelement <4 x float> %vld1, i32 2
+  %vecext3 = extractelement <4 x float> %vld1, i32 3
+  %add = fadd float %vecext3, 1.000000e+00
+  %tobool = icmp eq float* %q, null
+  br i1 %tobool, label %if.end, label %if.then
+
+if.then:                                          ; preds = %entry
+  %arrayidx = getelementptr inbounds float* %q, i32 1
+  %1 = load float* %arrayidx, align 4
+  %add4 = fadd float %vecext, %1
+  %2 = load float* %q, align 4
+  %add6 = fadd float %vecext1, %2
+  %arrayidx7 = getelementptr inbounds float* %q, i32 2
+  %3 = load float* %arrayidx7, align 4
+  %add8 = fadd float %vecext2, %3
+  br label %if.end
+
+if.end:                                           ; preds = %entry, %if.then
+  %a.0 = phi float [ %add4, %if.then ], [ %vecext, %entry ]
+  %b.0 = phi float [ %add6, %if.then ], [ %vecext1, %entry ]
+  %c.0 = phi float [ %add8, %if.then ], [ %vecext2, %entry ]
+  %vecinit = insertelement <4 x float> undef, float %a.0, i32 0
+  %vecinit9 = insertelement <4 x float> %vecinit, float %b.0, i32 1
+  %vecinit10 = insertelement <4 x float> %vecinit9, float %c.0, i32 2
+  %vecinit11 = insertelement <4 x float> %vecinit10, float %add, i32 3
+  tail call void @llvm.arm.neon.vst1.v4f32(i8* %0, <4 x float> %vecinit11, i32 4)
+  ret void
+}
+
+declare <4 x float> @llvm.arm.neon.vld1.v4f32(i8*, i32) nounwind readonly
+
+declare void @llvm.arm.neon.vst1.v4f32(i8*, <4 x float>, i32) nounwind