Index: lib/CodeGen/RegisterCoalescer.cpp
===================================================================
--- lib/CodeGen/RegisterCoalescer.cpp
+++ lib/CodeGen/RegisterCoalescer.cpp
@@ -1496,6 +1496,10 @@
   unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
   isMoveInstr(*TRI, CopyMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx);
 
+  // This optimization is valid only for a copy that writes the whole register.
+  if (DstSubIdx)
+    return nullptr;
+
   SlotIndex Idx = LIS->getInstructionIndex(*CopyMI);
   const LiveInterval &SrcLI = LIS->getInterval(SrcReg);
   // CopyMI is undef iff SrcReg is not live before the instruction.
Index: test/CodeGen/AMDGPU/coalescing-subreg-removed-undef-copy.mir
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/coalescing-subreg-removed-undef-copy.mir
@@ -0,0 +1,165 @@
+# RUN: llc -march=amdgcn -mcpu=gfx803 -run-pass simple-register-coalescing -verify-machineinstrs -o - %s | FileCheck --check-prefix=GCN %s
+#
+# This is another example of a test giving "Couldn't join subrange!"
+#
+# This shows a case where a subreg was defined by a copy from an undef register,
+# and that copy got removed and the subreg became undef at a later whole reg use.
+#
+# GCN: {{^body}}
+
+---
+name:            _amdgpu_ps_main
+alignment:       0
+exposesReturnsTwice: false
+legalized:       false
+regBankSelected: false
+selected:        false
+failedISel:      false
+tracksRegLiveness: true
+liveins:         
+frameInfo:       
+  isFrameAddressTaken: false
+  isReturnAddressTaken: false
+  hasStackMap:     false
+  hasPatchPoint:   false
+  stackSize:       0
+  offsetAdjustment: 0
+  maxAlignment:    0
+  adjustsStack:    false
+  hasCalls:        false
+  stackProtector:  ''
+  maxCallFrameSize: 4294967295
+  hasOpaqueSPAdjustment: false
+  hasVAStart:      false
+  hasMustTailInVarArgFunc: false
+  localFrameSize:  0
+  savePoint:       ''
+  restorePoint:    ''
+fixedStack:      
+stack:           
+constants:       
+body:             |
+  bb.0:
+    successors: %bb.2(0x40000000), %bb.1(0x40000000)
+  
+    %23:sreg_32_xm0 = S_MOV_B32 0
+    undef %24.sub0:sreg_128 = COPY %23
+    %24.sub1:sreg_128 = COPY %23
+    %24.sub3:sreg_128 = COPY %23
+    %110:sreg_128 = COPY killed %24
+    S_CBRANCH_SCC1 %bb.2, implicit undef $scc
+    S_BRANCH %bb.1
+  
+  bb.1:
+    successors: %bb.2(0x80000000)
+  
+    %110:sreg_128 = IMPLICIT_DEF
+  
+  bb.2:
+    successors: %bb.3(0x40000000), %bb.4(0x40000000)
+  
+    %0:sreg_128 = COPY killed %110
+    S_CBRANCH_SCC1 %bb.4, implicit undef $scc
+    S_BRANCH %bb.3
+  
+  bb.3:
+    successors: %bb.5(0x80000000)
+  
+    %88:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    undef %87.sub2:vreg_128 = COPY killed %88
+    %43:sreg_32_xm0 = S_MOV_B32 0
+    undef %44.sub0:sreg_256 = COPY %43
+    %44.sub1:sreg_256 = COPY %43
+    %44.sub2:sreg_256 = COPY %43
+    %44.sub3:sreg_256 = COPY %43
+    %44.sub4:sreg_256 = COPY %43
+    %44.sub5:sreg_256 = COPY %43
+    %44.sub6:sreg_256 = COPY %43
+    %44.sub7:sreg_256 = COPY killed %43
+    %45:vgpr_32 = IMAGE_LOAD_MIP_V1_V4 killed %87, killed %44, 1, -1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from constant-pool, addrspace 4)
+    dead %89:vgpr_32 = V_BFE_U32 killed %45, 7, 1, implicit $exec
+    %49:sreg_32_xm0 = S_MOV_B32 1
+    %50:sreg_128 = COPY killed %0
+    %50.sub2:sreg_128 = COPY killed %49
+    %91:vreg_128 = COPY killed %50
+    %90:vreg_128 = COPY killed %91
+    %90.sub3:vreg_128 = COPY undef %89
+    %111:vreg_128 = COPY killed %90
+    S_BRANCH %bb.5
+  
+  bb.4:
+    successors: %bb.5(0x80000000)
+  
+    %34:sreg_32_xm0 = S_MOV_B32 0
+    %4:sreg_128 = COPY killed %0
+    %4.sub2:sreg_128 = COPY killed %34
+    %93:vreg_128 = COPY killed %4
+    %111:vreg_128 = COPY killed %93
+  
+  bb.5:
+    successors: %bb.8(0x40000000), %bb.6(0x40000000)
+  
+    %92:vreg_128 = COPY killed %111
+    %108:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    undef %107.sub2:vreg_128 = COPY killed %108
+    %60:sgpr_32 = S_MOV_B32 0
+    undef %61.sub0:sreg_256 = COPY %60
+    %61.sub1:sreg_256 = COPY %60
+    %61.sub2:sreg_256 = COPY %60
+    %61.sub3:sreg_256 = COPY %60
+    %61.sub4:sreg_256 = COPY %60
+    %61.sub5:sreg_256 = COPY %60
+    %61.sub6:sreg_256 = COPY %60
+    %61.sub7:sreg_256 = COPY killed %60
+    %62:vgpr_32 = IMAGE_LOAD_MIP_V1_V4 killed %107, killed %61, 1, -1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from constant-pool, addrspace 4)
+    %109:vgpr_32 = V_AND_B32_e32 24, killed %62, implicit $exec
+    %67:sreg_64_xexec = V_CMP_EQ_U32_e64 0, killed %109, implicit $exec
+    dead %69:vgpr_32 = V_CNDMASK_B32_e64 0, -1, %67, implicit $exec
+    %112:vreg_128 = IMPLICIT_DEF
+    %114:sreg_64 = COPY $exec, implicit-def $exec
+    %115:sreg_64 = S_AND_B64 %114, %67, implicit-def dead $scc
+    %8:sreg_64 = S_XOR_B64 %115, %114, implicit-def dead $scc
+    $exec = S_MOV_B64_term killed %115
+    SI_MASK_BRANCH %bb.6, implicit $exec
+    S_BRANCH %bb.8
+  
+  bb.6:
+    successors: %bb.7(0x40000000), %bb.9(0x40000000)
+  
+    %116:sreg_64 = COPY killed %8
+    %10:sreg_64 = S_OR_SAVEEXEC_B64 %116, implicit-def $exec, implicit-def $scc, implicit $exec
+    %105:vreg_128 = COPY killed %112
+    %113:vreg_128 = COPY killed %105
+    $exec = S_XOR_B64_term $exec, %10, implicit-def $scc
+    SI_MASK_BRANCH %bb.9, implicit $exec
+    S_BRANCH %bb.7
+  
+  bb.7:
+    successors: %bb.9(0x80000000)
+  
+    dead %97:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    undef %95.sub1:vreg_128 = COPY killed %92.sub1
+    %113:vreg_128 = COPY killed %95
+    S_BRANCH %bb.9
+  
+  bb.8:
+    successors: %bb.6(0x80000000)
+  
+    %103:vreg_128 = COPY %92
+    %103.sub2:vreg_128 = COPY undef %69
+    %104:vreg_128 = COPY killed %103
+    %104.sub3:vreg_128 = COPY undef %69
+    %112:vreg_128 = COPY killed %104
+    S_BRANCH %bb.6
+  
+  bb.9:
+    $exec = S_OR_B64 $exec, killed %10, implicit-def $scc
+    %99:vreg_128 = COPY killed %113
+    %77:sreg_64_xexec = V_CMP_EQ_U32_e64 target-flags(amdgpu-gotprel32-lo) 0, killed %99.sub1, implicit $exec
+    %79:vgpr_32 = V_CNDMASK_B32_e64 0, 1065353216, killed %77, implicit $exec
+    %81:vgpr_32 = V_MUL_F32_e32 0, killed %79, implicit $exec
+    %82:vgpr_32 = V_CVT_PKRTZ_F16_F32_e64 0, undef %83:vgpr_32, 0, killed %81, 0, 0, implicit $exec
+    EXP_DONE 0, undef %84:vgpr_32, killed %82, undef %85:vgpr_32, undef %86:vgpr_32, -1, -1, 15, implicit $exec
+    S_ENDPGM
+
+...