Index: lib/Target/AMDGPU/SIDefines.h
===================================================================
--- lib/Target/AMDGPU/SIDefines.h
+++ lib/Target/AMDGPU/SIDefines.h
@@ -40,7 +40,8 @@
   FLAT = 1 << 21,
   WQM = 1 << 22,
   VGPRSpill = 1 << 23,
-  VOPAsmPrefer32Bit = 1 << 24
+  VOPAsmPrefer32Bit = 1 << 24,
+  Gather4 = 1 << 25
 };
 }
 
Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -3133,7 +3133,8 @@
   const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
   unsigned Opcode = Node->getMachineOpcode();
 
-  if (TII->isMIMG(Opcode) && !TII->get(Opcode).mayStore())
+  if (TII->isMIMG(Opcode) && !TII->get(Opcode).mayStore() &&
+      !TII->isGather4(Opcode))
     adjustWritemask(Node, DAG);
 
   if (Opcode == AMDGPU::INSERT_SUBREG ||
Index: lib/Target/AMDGPU/SIInstrFormats.td
===================================================================
--- lib/Target/AMDGPU/SIInstrFormats.td
+++ lib/Target/AMDGPU/SIInstrFormats.td
@@ -48,6 +48,8 @@
   // is unable to infer the encoding from the operands.
   field bits<1> VOPAsmPrefer32Bit = 0;
 
+  field bits<1> Gather4 = 0;
+
   // These need to be kept in sync with the enum in SIInstrFlags.
   let TSFlags{0} = VM_CNT;
   let TSFlags{1} = EXP_CNT;
@@ -78,6 +80,7 @@
   let TSFlags{22} = WQM;
   let TSFlags{23} = VGPRSpill;
   let TSFlags{24} = VOPAsmPrefer32Bit;
+  let TSFlags{25} = Gather4;
 
   let SchedRW = [Write32Bit];
 
Index: lib/Target/AMDGPU/SIInstrInfo.h
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.h
+++ lib/Target/AMDGPU/SIInstrInfo.h
@@ -310,6 +310,14 @@
     return get(Opcode).TSFlags & SIInstrFlags::MIMG;
   }
 
+  static bool isGather4(const MachineInstr &MI) {
+    return MI.getDesc().TSFlags & SIInstrFlags::Gather4;
+  }
+
+  bool isGather4(uint16_t Opcode) const {
+    return get(Opcode).TSFlags & SIInstrFlags::Gather4;
+  }
+
   static bool isFLAT(const MachineInstr &MI) {
     return MI.getDesc().TSFlags & SIInstrFlags::FLAT;
   }
Index: lib/Target/AMDGPU/SIInstrInfo.td
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.td
+++ lib/Target/AMDGPU/SIInstrInfo.td
@@ -3557,8 +3557,8 @@
   // 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
   // (red,red,red,red) etc.) The ISA document doesn't mention
   // this.
-  // Therefore, disable all code which updates DMASK by setting these two:
-  let MIMG = 0;
+  // Therefore, disable all code which updates DMASK by setting this:
+  let Gather4 = 1;
   let hasPostISelHook = 0;
   let WQM = wqm;
 
Index: test/CodeGen/AMDGPU/llvm.SI.gather4.ll
===================================================================
--- test/CodeGen/AMDGPU/llvm.SI.gather4.ll
+++ test/CodeGen/AMDGPU/llvm.SI.gather4.ll
@@ -462,7 +462,27 @@
   ret void
 }
 
-
+;CHECK-LABEL: {{^}}gather4_sgpr_bug:
+;
+; This crashed at some point due to a bug in FixSGPRCopies. Derived from the
+; report in https://bugs.freedesktop.org/show_bug.cgi?id=96877
+;
+;TODO: the readfirstlanes are unnecessary, see http://reviews.llvm.org/D22217
+;
+;CHECK: v_readfirstlane_b32 s[[LO:[0-9]+]], v{{[0-9]+}}
+;CHECK: v_readfirstlane_b32
+;CHECK: v_readfirstlane_b32
+;CHECK: v_readfirstlane_b32 s[[HI:[0-9]+]], v{{[0-9]+}}
+;CHECK: image_gather4_lz {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, s{{\[}}[[LO]]:[[HI]]] dmask:0x8
+define amdgpu_ps float @gather4_sgpr_bug() {
+main_body:
+  %tmp = load <4 x i32>, <4 x i32> addrspace(2)* undef, align 16
+  %tmp1 = insertelement <4 x i32> %tmp, i32 0, i32 0
+  %tmp2 = call <4 x float> @llvm.SI.gather4.lz.v2i32(<2 x i32> undef, <8 x i32> undef, <4 x i32> %tmp1, i32 8, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0)
+  %tmp4 = extractelement <4 x float> %tmp2, i32 1
+  %tmp9 = fadd float undef, %tmp4
+  ret float %tmp9
+}
 
 declare <4 x float> @llvm.SI.gather4.v2i32(<2 x i32>, <8 x i32>, <4 x i32>, i32, i32, i32, i32, i32, i32, i32, i32) #0
 declare <4 x float> @llvm.SI.gather4.v4i32(<4 x i32>, <8 x i32>, <4 x i32>, i32, i32, i32, i32, i32, i32, i32, i32) #0