Fix register coalescer failure to prune value

Register coalescer fails for the test in the patch with the assertion in
JoinVals::ConflictResolution `DefMI != nullptr'. It attempts to join
live intervals for two adjacent instructions and erase the copy:

    %2:vreg_256 = COPY %1
    %3:vreg_256 = COPY killed %1

The LI needs to be adjusted to kill subrange for the erased instruction
and extend the subrange of the original def. That was done for the main
interval only but not for the subrange. As a result subrange had a VNI
pointing to the erased slot resulting in the above failure.

Differential Revision: https://reviews.llvm.org/D62162

llvm-svn: 361293

2 files changed, 43 insertions, 2 deletions

diff --git a/llvm/lib/CodeGen/RegisterCoalescer.cpp b/llvm/lib/CodeGen/RegisterCoalescer.cpp
index ca53fe4f393..58e29d613f0 100644
--- a/llvm/lib/CodeGen/RegisterCoalescer.cpp
+++ b/llvm/lib/CodeGen/RegisterCoalescer.cpp
@@ -3032,7 +3032,9 @@ void JoinVals::pruneSubRegValues(LiveInterval &LI, LaneBitmask &ShrinkMask) {
       // If a subrange starts at the copy then an undefined value has been
       // copied and we must remove that subrange value as well.
       VNInfo *ValueOut = Q.valueOutOrDead();
-      if (ValueOut != nullptr && Q.valueIn() == nullptr) {
+      if (ValueOut != nullptr && (Q.valueIn() == nullptr ||
+                                  (V.Identical && V.Resolution == CR_Erase &&
+                                   ValueOut->def == Def))) {
         LLVM_DEBUG(dbgs() << "\t\tPrune sublane " << PrintLaneMask(S.LaneMask)
                           << " at " << Def << "\n");
         SmallVector<SlotIndex,8> EndPoints;
@@ -3041,7 +3043,7 @@ void JoinVals::pruneSubRegValues(LiveInterval &LI, LaneBitmask &ShrinkMask) {
         // Mark value number as unused.
         ValueOut->markUnused();
 
-        if (V.Identical && S.Query(OtherDef).valueOut()) {
+        if (V.Identical && S.Query(OtherDef).valueOutOrDead()) {
           // If V is identical to V.OtherVNI (and S was live at OtherDef),
           // then we can't simply prune V from S. V needs to be replaced
           // with V.OtherVNI.
diff --git a/llvm/test/CodeGen/AMDGPU/coalescer-subranges-prune-kill-copy.mir b/llvm/test/CodeGen/AMDGPU/coalescer-subranges-prune-kill-copy.mir
new file mode 100644
index 00000000000..919a7b10ec6
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/coalescer-subranges-prune-kill-copy.mir
@@ -0,0 +1,39 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass=simple-register-coalescing -o - %s | FileCheck -check-prefix=GCN %s
+
+# Test used to crash with message:
+# JoinVals::ConflictResolution (anonymous namespace)::JoinVals::analyzeValue(unsigned int, (anonymous namespace)::JoinVals &): Assertion `DefMI != nullptr' failed
+
+---
+name:            test
+tracksRegLiveness: true
+body:             |
+  ; GCN-LABEL: name: test
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.2(0x80000000)
+  ; GCN:   undef %1.sub0:vreg_128 = IMPLICIT_DEF
+  ; GCN:   %1.sub1:vreg_128 = IMPLICIT_DEF
+  ; GCN:   S_BRANCH %bb.2
+  ; GCN: bb.1:
+  ; GCN:   successors: %bb.2(0x80000000)
+  ; GCN:   [[DEF:%[0-9]+]]:vreg_128 = IMPLICIT_DEF
+  ; GCN: bb.2:
+  ; GCN:   [[DEF]].sub2:vreg_128 = COPY undef %3:sreg_32
+  ; GCN:   S_ENDPGM 0, implicit [[DEF]]
+  bb.0:
+    undef %0.sub0:vreg_128 = IMPLICIT_DEF
+    %0.sub1:vreg_128 = IMPLICIT_DEF
+    %1:vreg_128 = COPY %0
+    %2:vreg_128 = COPY killed %0
+    S_BRANCH %bb.2
+
+  bb.1:
+    %1:vreg_128 = COPY killed undef %1
+
+  bb.2:
+    %1.sub2:vreg_128 = COPY undef %3:sreg_32
+    %2:vreg_128 = COPY killed %1
+    %4:vreg_128 = COPY killed %2
+    S_ENDPGM 0, implicit killed %4
+
+...