[SystemZ] Add STOC and STOCG

These instructions are allowed to trap even if the condition is false,
so for now they are only used for "*ptr = (cond ? x : *ptr)"-style
constructs.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@187111 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/SystemZ/SystemZISelLowering.cpp

@@ -1773,22 +1773,36 @@ SystemZTargetLowering::emitSelect(MachineInstr *MI,
 
 // Implement EmitInstrWithCustomInserter for pseudo CondStore* instruction MI.
 // StoreOpcode is the store to use and Invert says whether the store should
-// happen when the condition is false rather than true.
+// happen when the condition is false rather than true.  If a STORE ON
+// CONDITION is available, STOCOpcode is its opcode, otherwise it is 0.
 MachineBasicBlock *
 SystemZTargetLowering::emitCondStore(MachineInstr *MI,
                                      MachineBasicBlock *MBB,
-                                     unsigned StoreOpcode, bool Invert) const {
+                                     unsigned StoreOpcode, unsigned STOCOpcode,
+                                     bool Invert) const {
   const SystemZInstrInfo *TII = TM.getInstrInfo();
 
-  MachineOperand Base = MI->getOperand(0);
+  unsigned SrcReg     = MI->getOperand(0).getReg();
-  int64_t Disp        = MI->getOperand(1).getImm();
+  MachineOperand Base = MI->getOperand(1);
-  unsigned IndexReg   = MI->getOperand(2).getReg();
+  int64_t Disp        = MI->getOperand(2).getImm();
-  unsigned SrcReg     = MI->getOperand(3).getReg();
+  unsigned IndexReg   = MI->getOperand(3).getReg();
   unsigned CCMask     = MI->getOperand(4).getImm();
   DebugLoc DL         = MI->getDebugLoc();
 
   StoreOpcode = TII->getOpcodeForOffset(StoreOpcode, Disp);
 
+  // Use STOCOpcode if possible.  We could use different store patterns in
+  // order to avoid matching the index register, but the performance trade-offs
+  // might be more complicated in that case.
+  if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) {
+    if (Invert)
+      CCMask = CCMask ^ SystemZ::CCMASK_ANY;
+    BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
+      .addReg(SrcReg).addOperand(Base).addImm(Disp).addImm(CCMask);
+    MI->eraseFromParent();
+    return MBB;
+  }
+
   // Get the condition needed to branch around the store.
   if (!Invert)
     CCMask = CCMask ^ SystemZ::CCMASK_ANY;
@@ -2249,41 +2263,41 @@ EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const {
     return emitSelect(MI, MBB);
 
   case SystemZ::CondStore8_32:
-    return emitCondStore(MI, MBB, SystemZ::STC32, false);
+    return emitCondStore(MI, MBB, SystemZ::STC32, 0, false);
   case SystemZ::CondStore8_32Inv:
-    return emitCondStore(MI, MBB, SystemZ::STC32, true);
+    return emitCondStore(MI, MBB, SystemZ::STC32, 0, true);
   case SystemZ::CondStore16_32:
-    return emitCondStore(MI, MBB, SystemZ::STH32, false);
+    return emitCondStore(MI, MBB, SystemZ::STH32, 0, false);
   case SystemZ::CondStore16_32Inv:
-    return emitCondStore(MI, MBB, SystemZ::STH32, true);
+    return emitCondStore(MI, MBB, SystemZ::STH32, 0, true);
   case SystemZ::CondStore32_32:
-    return emitCondStore(MI, MBB, SystemZ::ST32, false);
+    return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, false);
   case SystemZ::CondStore32_32Inv:
-    return emitCondStore(MI, MBB, SystemZ::ST32, true);
+    return emitCondStore(MI, MBB, SystemZ::ST32, SystemZ::STOC32, true);
   case SystemZ::CondStore8:
-    return emitCondStore(MI, MBB, SystemZ::STC, false);
+    return emitCondStore(MI, MBB, SystemZ::STC, 0, false);
   case SystemZ::CondStore8Inv:
-    return emitCondStore(MI, MBB, SystemZ::STC, true);
+    return emitCondStore(MI, MBB, SystemZ::STC, 0, true);
   case SystemZ::CondStore16:
-    return emitCondStore(MI, MBB, SystemZ::STH, false);
+    return emitCondStore(MI, MBB, SystemZ::STH, 0, false);
   case SystemZ::CondStore16Inv:
-    return emitCondStore(MI, MBB, SystemZ::STH, true);
+    return emitCondStore(MI, MBB, SystemZ::STH, 0, true);
   case SystemZ::CondStore32:
-    return emitCondStore(MI, MBB, SystemZ::ST, false);
+    return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, false);
   case SystemZ::CondStore32Inv:
-    return emitCondStore(MI, MBB, SystemZ::ST, true);
+    return emitCondStore(MI, MBB, SystemZ::ST, SystemZ::STOC, true);
   case SystemZ::CondStore64:
-    return emitCondStore(MI, MBB, SystemZ::STG, false);
+    return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, false);
   case SystemZ::CondStore64Inv:
-    return emitCondStore(MI, MBB, SystemZ::STG, true);
+    return emitCondStore(MI, MBB, SystemZ::STG, SystemZ::STOCG, true);
   case SystemZ::CondStoreF32:
-    return emitCondStore(MI, MBB, SystemZ::STE, false);
+    return emitCondStore(MI, MBB, SystemZ::STE, 0, false);
   case SystemZ::CondStoreF32Inv:
-    return emitCondStore(MI, MBB, SystemZ::STE, true);
+    return emitCondStore(MI, MBB, SystemZ::STE, 0, true);
   case SystemZ::CondStoreF64:
-    return emitCondStore(MI, MBB, SystemZ::STD, false);
+    return emitCondStore(MI, MBB, SystemZ::STD, 0, false);
   case SystemZ::CondStoreF64Inv:
-    return emitCondStore(MI, MBB, SystemZ::STD, true);
+    return emitCondStore(MI, MBB, SystemZ::STD, 0, true);
 
   case SystemZ::AEXT128_64:
     return emitExt128(MI, MBB, false, SystemZ::subreg_low);

lib/Target/SystemZ/SystemZISelLowering.h

@@ -211,7 +211,8 @@ private:
                                 MachineBasicBlock *BB) const;
   MachineBasicBlock *emitCondStore(MachineInstr *MI,
                                    MachineBasicBlock *BB,
-                                   unsigned StoreOpcode, bool Invert) const;
+                                   unsigned StoreOpcode, unsigned STOCOpcode,
+                                   bool Invert) const;
   MachineBasicBlock *emitExt128(MachineInstr *MI,
                                 MachineBasicBlock *MBB,
                                 bool ClearEven, unsigned SubReg) const;

lib/Target/SystemZ/SystemZInstrFormats.td

@@ -451,9 +451,11 @@ class InstSS<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern>
 // Instruction definitions with semantics
 //===----------------------------------------------------------------------===//
 //
-// These classes have the form <Category><Format>, where <Format> is one
+// These classes have the form [Cond]<Category><Format>, where <Format> is one
 // of the formats defined above and where <Category> describes the inputs
-// and outputs.  <Category> can be one of:
+// and outputs.  "Cond" is used if the instruction is conditional,
+// in which case the 4-bit condition-code mask is added as a final operand.
+// <Category> can be one of:
 //
 //   Inherent:
 //     One register output operand and no input operands.
@@ -618,6 +620,40 @@ multiclass StoreSIPair<string mnemonic, bits<8> siOpcode, bits<16> siyOpcode,
   }
 }
 
+class CondStoreRSY<string mnemonic, bits<16> opcode,
+                   RegisterOperand cls, bits<5> bytes,
+                   AddressingMode mode = bdaddr20only>
+  : InstRSY<opcode, (outs), (ins cls:$R1, mode:$BD2, cond4:$R3),
+            mnemonic#"$R3\t$R1, $BD2", []>,
+    Requires<[FeatureLoadStoreOnCond]> {
+  let mayStore = 1;
+  let AccessBytes = bytes;
+}
+
+// Like CondStoreRSY, but used for the raw assembly form.  The condition-code
+// mask is the third operand rather than being part of the mnemonic.
+class AsmCondStoreRSY<string mnemonic, bits<16> opcode,
+                      RegisterOperand cls, bits<5> bytes,
+                      AddressingMode mode = bdaddr20only>
+  : InstRSY<opcode, (outs), (ins cls:$R1, mode:$BD2, uimm8zx4:$R3),
+            mnemonic#"\t$R1, $BD2, $R3", []>,
+    Requires<[FeatureLoadStoreOnCond]> {
+  let mayStore = 1;
+  let AccessBytes = bytes;
+}
+
+// Like CondStoreRSY, but with a fixed CC mask.
+class FixedCondStoreRSY<string mnemonic, bits<16> opcode,
+                        RegisterOperand cls, bits<4> ccmask, bits<5> bytes,
+                        AddressingMode mode = bdaddr20only>
+  : InstRSY<opcode, (outs), (ins cls:$R1, mode:$BD2),
+            mnemonic#"\t$R1, $BD2", []>,
+    Requires<[FeatureLoadStoreOnCond]> {
+  let mayStore = 1;
+  let AccessBytes = bytes;
+  let R3 = ccmask;
+}
+
 class UnaryRR<string mnemonic, bits<8> opcode, SDPatternOperator operator,
               RegisterOperand cls1, RegisterOperand cls2>
   : InstRR<opcode, (outs cls1:$R1), (ins cls2:$R2),
@@ -1151,12 +1187,12 @@ class SelectWrapper<RegisterOperand cls>
 multiclass CondStores<RegisterOperand cls, SDPatternOperator store,
                       SDPatternOperator load, AddressingMode mode> {
   let Defs = [CC], Uses = [CC], usesCustomInserter = 1 in {
-    def "" : Pseudo<(outs), (ins mode:$addr, cls:$new, i8imm:$cc),
+    def "" : Pseudo<(outs), (ins cls:$new, mode:$addr, uimm8zx4:$cc),
                     [(store (z_select_ccmask cls:$new, (load mode:$addr),
-                                             imm:$cc), mode:$addr)]>;
+                                             uimm8zx4:$cc), mode:$addr)]>;
-    def Inv : Pseudo<(outs), (ins mode:$addr, cls:$new, i8imm:$cc),
+    def Inv : Pseudo<(outs), (ins cls:$new, mode:$addr, uimm8zx4:$cc),
                      [(store (z_select_ccmask (load mode:$addr), cls:$new,
-                                              imm:$cc), mode:$addr)]>;
+                                              uimm8zx4:$cc), mode:$addr)]>;
   }
 }
 

lib/Target/SystemZ/SystemZInstrInfo.td

@@ -107,26 +107,28 @@ defm AsmC : CompareBranches<uimm8zx4, "", "$M3, ">;
 // (integer or floating-point)
 multiclass CondExtendedMnemonic<bits<4> ccmask, string name> {
   let R1 = ccmask in {
-    def "" : InstRI<0xA74, (outs), (ins brtarget16:$I2),
+    def J : InstRI<0xA74, (outs), (ins brtarget16:$I2),
                    "j"##name##"\t$I2", []>;
-    def L  : InstRIL<0xC04, (outs), (ins brtarget32:$I2),
+    def JG : InstRIL<0xC04, (outs), (ins brtarget32:$I2),
                      "jg"##name##"\t$I2", []>;
   }
+  def STOC  : FixedCondStoreRSY<"stoc"##name,  0xEBF3, GR32, ccmask, 4>;
+  def STOCG : FixedCondStoreRSY<"stocg"##name, 0xEBE3, GR64, ccmask, 8>;
 }
-defm AsmJO   : CondExtendedMnemonic<1,  "o">;
+defm AsmO   : CondExtendedMnemonic<1,  "o">;
-defm AsmJH   : CondExtendedMnemonic<2,  "h">;
+defm AsmH   : CondExtendedMnemonic<2,  "h">;
-defm AsmJNLE : CondExtendedMnemonic<3,  "nle">;
+defm AsmNLE : CondExtendedMnemonic<3,  "nle">;
-defm AsmJL   : CondExtendedMnemonic<4,  "l">;
+defm AsmL   : CondExtendedMnemonic<4,  "l">;
-defm AsmJNHE : CondExtendedMnemonic<5,  "nhe">;
+defm AsmNHE : CondExtendedMnemonic<5,  "nhe">;
-defm AsmJLH  : CondExtendedMnemonic<6,  "lh">;
+defm AsmLH  : CondExtendedMnemonic<6,  "lh">;
-defm AsmJNE  : CondExtendedMnemonic<7,  "ne">;
+defm AsmNE  : CondExtendedMnemonic<7,  "ne">;
-defm AsmJE   : CondExtendedMnemonic<8,  "e">;
+defm AsmE   : CondExtendedMnemonic<8,  "e">;
-defm AsmJNLH : CondExtendedMnemonic<9,  "nlh">;
+defm AsmNLH : CondExtendedMnemonic<9,  "nlh">;
-defm AsmJHE  : CondExtendedMnemonic<10, "he">;
+defm AsmHE  : CondExtendedMnemonic<10, "he">;
-defm AsmJNL  : CondExtendedMnemonic<11, "nl">;
+defm AsmNL  : CondExtendedMnemonic<11, "nl">;
-defm AsmJLE  : CondExtendedMnemonic<12, "le">;
+defm AsmLE  : CondExtendedMnemonic<12, "le">;
-defm AsmJNH  : CondExtendedMnemonic<13, "nh">;
+defm AsmNH  : CondExtendedMnemonic<13, "nh">;
-defm AsmJNO  : CondExtendedMnemonic<14, "no">;
+defm AsmNO  : CondExtendedMnemonic<14, "no">;
 
 // Define AsmParser mnemonics for each integer condition-code mask.
 // This is like the list above, except that condition 3 is not possible
@@ -274,6 +276,17 @@ let isCodeGenOnly = 1 in
   def STRL32 : StoreRILPC<"strl", 0xC4F, aligned_store, GR32>;
 def STGRL : StoreRILPC<"stgrl", 0xC4B, aligned_store, GR64>;
 
+// Store on condition.
+let isCodeGenOnly = 1, Uses = [CC] in {
+  def STOC32 : CondStoreRSY<"stoc",  0xEBF3, GR32, 4>;
+  def STOC   : CondStoreRSY<"stoc",  0xEBF3, GR64, 4>;
+  def STOCG  : CondStoreRSY<"stocg", 0xEBE3, GR64, 8>;
+}
+let Uses = [CC] in {
+  def AsmSTOC  : AsmCondStoreRSY<"stoc",  0xEBF3, GR32, 4>;
+  def AsmSTOCG : AsmCondStoreRSY<"stocg", 0xEBE3, GR64, 8>;
+}
+
 // 8-bit immediate stores to 8-bit fields.
 defm MVI : StoreSIPair<"mvi", 0x92, 0xEB52, truncstorei8, imm32zx8trunc>;
 

lib/Target/SystemZ/SystemZProcessors.td

@@ -21,6 +21,13 @@ def FeatureDistinctOps : SystemZFeature<
   "Assume that the distinct-operands facility is installed"
 >;
 
+def FeatureLoadStoreOnCond : SystemZFeature<
+  "load-store-on-cond", "LoadStoreOnCond",
+  "Assume that the load/store-on-condition facility is installed"
+>;
+
 def : Processor<"z10",   NoItineraries, []>;
-def : Processor<"z196",  NoItineraries, [FeatureDistinctOps]>;
+def : Processor<"z196",  NoItineraries,
-def : Processor<"zEC12", NoItineraries, [FeatureDistinctOps]>;
+                [FeatureDistinctOps, FeatureLoadStoreOnCond]>;
+def : Processor<"zEC12", NoItineraries,
+                [FeatureDistinctOps, FeatureLoadStoreOnCond]>;

lib/Target/SystemZ/SystemZSubtarget.cpp

@@ -21,7 +21,7 @@ SystemZSubtarget::SystemZSubtarget(const std::string &TT,
                                    const std::string &CPU,
                                    const std::string &FS)
   : SystemZGenSubtargetInfo(TT, CPU, FS), HasDistinctOps(false),
-    TargetTriple(TT) {
+    HasLoadStoreOnCond(false), TargetTriple(TT) {
   std::string CPUName = CPU;
   if (CPUName.empty())
     CPUName = "z10";

lib/Target/SystemZ/SystemZSubtarget.h

@@ -28,6 +28,7 @@ class StringRef;
 class SystemZSubtarget : public SystemZGenSubtargetInfo {
 protected:
   bool HasDistinctOps;
+  bool HasLoadStoreOnCond;
 
 private:
   Triple TargetTriple;
@@ -42,6 +43,9 @@ public:
   // Return true if the target has the distinct-operands facility.
   bool hasDistinctOps() const { return HasDistinctOps; }
 
+  // Return true if the target has the load/store-on-condition facility.
+  bool hasLoadStoreOnCond() const { return HasLoadStoreOnCond; }
+
   // Return true if GV can be accessed using LARL for reloc model RM
   // and code model CM.
   bool isPC32DBLSymbol(const GlobalValue *GV, Reloc::Model RM,

test/CodeGen/SystemZ/cond-store-03.ll

@@ -1,6 +1,6 @@
 ; Test 32-bit conditional stores that are presented as selects.
 ;
-; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | FileCheck %s
 
 declare void @foo(i32 *)
 

test/CodeGen/SystemZ/cond-store-04.ll

@@ -1,6 +1,6 @@
 ; Test 64-bit conditional stores that are presented as selects.
 ;
-; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z10 | FileCheck %s
 
 declare void @foo(i64 *)
 

test/CodeGen/SystemZ/cond-store-07.ll

@@ -0,0 +1,186 @@
+; Test STOCs that are presented as selects.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z196 | FileCheck %s
+
+declare void @foo(i32 *)
+
+; Test the simple case, with the loaded value first.
+define void @f1(i32 *%ptr, i32 %alt, i32 %limit) {
+; CHECK-LABEL: f1:
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %orig, i32 %alt
+  store i32 %res, i32 *%ptr
+  ret void
+}
+
+; ...and with the loaded value second
+define void @f2(i32 *%ptr, i32 %alt, i32 %limit) {
+; CHECK-LABEL: f2:
+; CHECK: clfi %r4, 42
+; CHECK: stocl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %alt, i32 %orig
+  store i32 %res, i32 *%ptr
+  ret void
+}
+
+; Test cases where the value is explicitly sign-extended to 64 bits, with the
+; loaded value first.
+define void @f3(i32 *%ptr, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f3:
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %ext = sext i32 %orig to i64
+  %res = select i1 %cond, i64 %ext, i64 %alt
+  %trunc = trunc i64 %res to i32
+  store i32 %trunc, i32 *%ptr
+  ret void
+}
+
+; ...and with the loaded value second
+define void @f4(i32 *%ptr, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f4:
+; CHECK: clfi %r4, 42
+; CHECK: stocl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %ext = sext i32 %orig to i64
+  %res = select i1 %cond, i64 %alt, i64 %ext
+  %trunc = trunc i64 %res to i32
+  store i32 %trunc, i32 *%ptr
+  ret void
+}
+
+; Test cases where the value is explicitly zero-extended to 32 bits, with the
+; loaded value first.
+define void @f5(i32 *%ptr, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f5:
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %ext = zext i32 %orig to i64
+  %res = select i1 %cond, i64 %ext, i64 %alt
+  %trunc = trunc i64 %res to i32
+  store i32 %trunc, i32 *%ptr
+  ret void
+}
+
+; ...and with the loaded value second
+define void @f6(i32 *%ptr, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f6:
+; CHECK: clfi %r4, 42
+; CHECK: stocl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %ext = zext i32 %orig to i64
+  %res = select i1 %cond, i64 %alt, i64 %ext
+  %trunc = trunc i64 %res to i32
+  store i32 %trunc, i32 *%ptr
+  ret void
+}
+
+; Check the high end of the aligned STOC range.
+define void @f7(i32 *%base, i32 %alt, i32 %limit) {
+; CHECK-LABEL: f7:
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, 524284(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i32 *%base, i64 131071
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %orig, i32 %alt
+  store i32 %res, i32 *%ptr
+  ret void
+}
+
+; Check the next word up.  Other sequences besides this one would be OK.
+define void @f8(i32 *%base, i32 %alt, i32 %limit) {
+; CHECK-LABEL: f8:
+; CHECK: agfi %r2, 524288
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, 0(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i32 *%base, i64 131072
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %orig, i32 %alt
+  store i32 %res, i32 *%ptr
+  ret void
+}
+
+; Check the low end of the STOC range.
+define void @f9(i32 *%base, i32 %alt, i32 %limit) {
+; CHECK-LABEL: f9:
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, -524288(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i32 *%base, i64 -131072
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %orig, i32 %alt
+  store i32 %res, i32 *%ptr
+  ret void
+}
+
+; Check the next word down, with the same comments as f8.
+define void @f10(i32 *%base, i32 %alt, i32 %limit) {
+; CHECK-LABEL: f10:
+; CHECK: agfi %r2, -524292
+; CHECK: clfi %r4, 42
+; CHECK: stocnl %r3, 0(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i32 *%base, i64 -131073
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %orig, i32 %alt
+  store i32 %res, i32 *%ptr
+  ret void
+}
+
+; Try a frame index base.
+define void @f11(i32 %alt, i32 %limit) {
+; CHECK-LABEL: f11:
+; CHECK: brasl %r14, foo@PLT
+; CHECK: stocnl {{%r[0-9]+}}, {{[0-9]+}}(%r15)
+; CHECK: brasl %r14, foo@PLT
+; CHECK: br %r14
+  %ptr = alloca i32
+  call void @foo(i32 *%ptr)
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i32 *%ptr
+  %res = select i1 %cond, i32 %orig, i32 %alt
+  store i32 %res, i32 *%ptr
+  call void @foo(i32 *%ptr)
+  ret void
+}
+
+; Test that conditionally-executed stores do not use STOC, since STOC
+; is allowed to trap even when the condition is false.
+define void @f12(i32 %a, i32 %b, i32 *%dest) {
+; CHECK-LABEL: f12:
+; CHECK-NOT: stoc
+; CHECK: br %r14
+entry:
+  %cmp = icmp ule i32 %a, %b
+  br i1 %cmp, label %store, label %exit
+
+store:
+  store i32 %b, i32 *%dest
+  br label %exit
+
+exit:
+  ret void
+}

test/CodeGen/SystemZ/cond-store-08.ll

@@ -0,0 +1,124 @@
+; Test STOCGs that are presented as selects.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z196 | FileCheck %s
+
+declare void @foo(i64 *)
+
+; Test with the loaded value first.
+define void @f1(i64 *%ptr, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f1:
+; CHECK: clfi %r4, 42
+; CHECK: stocgnl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %orig, i64 %alt
+  store i64 %res, i64 *%ptr
+  ret void
+}
+
+; ...and with the loaded value second
+define void @f2(i64 *%ptr, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f2:
+; CHECK: clfi %r4, 42
+; CHECK: stocgl %r3, 0(%r2)
+; CHECK: br %r14
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %alt, i64 %orig
+  store i64 %res, i64 *%ptr
+  ret void
+}
+
+; Check the high end of the aligned STOCG range.
+define void @f3(i64 *%base, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f3:
+; CHECK: clfi %r4, 42
+; CHECK: stocgnl %r3, 524280(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i64 *%base, i64 65535
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %orig, i64 %alt
+  store i64 %res, i64 *%ptr
+  ret void
+}
+
+; Check the next doubleword up.  Other sequences besides this one would be OK.
+define void @f4(i64 *%base, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f4:
+; CHECK: agfi %r2, 524288
+; CHECK: clfi %r4, 42
+; CHECK: stocgnl %r3, 0(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i64 *%base, i64 65536
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %orig, i64 %alt
+  store i64 %res, i64 *%ptr
+  ret void
+}
+
+; Check the low end of the STOCG range.
+define void @f5(i64 *%base, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f5:
+; CHECK: clfi %r4, 42
+; CHECK: stocgnl %r3, -524288(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i64 *%base, i64 -65536
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %orig, i64 %alt
+  store i64 %res, i64 *%ptr
+  ret void
+}
+
+; Check the next doubleword down, with the same comments as f4.
+define void @f6(i64 *%base, i64 %alt, i32 %limit) {
+; CHECK-LABEL: f6:
+; CHECK: agfi %r2, -524296
+; CHECK: clfi %r4, 42
+; CHECK: stocgnl %r3, 0(%r2)
+; CHECK: br %r14
+  %ptr = getelementptr i64 *%base, i64 -65537
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %orig, i64 %alt
+  store i64 %res, i64 *%ptr
+  ret void
+}
+
+; Try a frame index base.
+define void @f7(i64 %alt, i32 %limit) {
+; CHECK-LABEL: f7:
+; CHECK: brasl %r14, foo@PLT
+; CHECK: stocgnl {{%r[0-9]+}}, {{[0-9]+}}(%r15)
+; CHECK: brasl %r14, foo@PLT
+; CHECK: br %r14
+  %ptr = alloca i64
+  call void @foo(i64 *%ptr)
+  %cond = icmp ult i32 %limit, 42
+  %orig = load i64 *%ptr
+  %res = select i1 %cond, i64 %orig, i64 %alt
+  store i64 %res, i64 *%ptr
+  call void @foo(i64 *%ptr)
+  ret void
+}
+
+; Test that conditionally-executed stores do not use STOC, since STOC
+; is allowed to trap even when the condition is false.
+define void @f8(i64 %a, i64 %b, i64 *%dest) {
+; CHECK-LABEL: f8:
+; CHECK-NOT: stocg %r3, 0(%r4)
+; CHECK: br %r14
+entry:
+  %cmp = icmp ule i64 %a, %b
+  br i1 %cmp, label %store, label %exit
+
+store:
+  store i64 %b, i64 *%dest
+  br label %exit
+
+exit:
+  ret void
+}

test/MC/Disassembler/SystemZ/insns.txt

@@ -6106,6 +6106,102 @@
 # CHECK: st %r15, 0
 0x50 0xf0 0x00 0x00
 
+# CHECK: stoc %r1, 2(%r3), 0
+0xeb 0x10 0x30 0x02 0x00 0xf3
+
+# CHECK: stoco %r1, 2(%r3)
+0xeb 0x11 0x30 0x02 0x00 0xf3
+
+# CHECK: stoch %r1, 2(%r3)
+0xeb 0x12 0x30 0x02 0x00 0xf3
+
+# CHECK: stocnle %r1, 2(%r3)
+0xeb 0x13 0x30 0x02 0x00 0xf3
+
+# CHECK: stocl %r1, 2(%r3)
+0xeb 0x14 0x30 0x02 0x00 0xf3
+
+# CHECK: stocnhe %r1, 2(%r3)
+0xeb 0x15 0x30 0x02 0x00 0xf3
+
+# CHECK: stoclh %r1, 2(%r3)
+0xeb 0x16 0x30 0x02 0x00 0xf3
+
+# CHECK: stocne %r1, 2(%r3)
+0xeb 0x17 0x30 0x02 0x00 0xf3
+
+# CHECK: stoce %r1, 2(%r3)
+0xeb 0x18 0x30 0x02 0x00 0xf3
+
+# CHECK: stocnlh %r1, 2(%r3)
+0xeb 0x19 0x30 0x02 0x00 0xf3
+
+# CHECK: stoche %r1, 2(%r3)
+0xeb 0x1a 0x30 0x02 0x00 0xf3
+
+# CHECK: stocnl %r1, 2(%r3)
+0xeb 0x1b 0x30 0x02 0x00 0xf3
+
+# CHECK: stocle %r1, 2(%r3)
+0xeb 0x1c 0x30 0x02 0x00 0xf3
+
+# CHECK: stocnh %r1, 2(%r3)
+0xeb 0x1d 0x30 0x02 0x00 0xf3
+
+# CHECK: stocno %r1, 2(%r3)
+0xeb 0x1e 0x30 0x02 0x00 0xf3
+
+# CHECK: stoc %r1, 2(%r3), 15
+0xeb 0x1f 0x30 0x02 0x00 0xf3
+
+# CHECK: stocg %r1, 2(%r3), 0
+0xeb 0x10 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgo %r1, 2(%r3)
+0xeb 0x11 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgh %r1, 2(%r3)
+0xeb 0x12 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgnle %r1, 2(%r3)
+0xeb 0x13 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgl %r1, 2(%r3)
+0xeb 0x14 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgnhe %r1, 2(%r3)
+0xeb 0x15 0x30 0x02 0x00 0xe3
+
+# CHECK: stocglh %r1, 2(%r3)
+0xeb 0x16 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgne %r1, 2(%r3)
+0xeb 0x17 0x30 0x02 0x00 0xe3
+
+# CHECK: stocge %r1, 2(%r3)
+0xeb 0x18 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgnlh %r1, 2(%r3)
+0xeb 0x19 0x30 0x02 0x00 0xe3
+
+# CHECK: stocghe %r1, 2(%r3)
+0xeb 0x1a 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgnl %r1, 2(%r3)
+0xeb 0x1b 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgle %r1, 2(%r3)
+0xeb 0x1c 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgnh %r1, 2(%r3)
+0xeb 0x1d 0x30 0x02 0x00 0xe3
+
+# CHECK: stocgno %r1, 2(%r3)
+0xeb 0x1e 0x30 0x02 0x00 0xe3
+
+# CHECK: stocg %r1, 2(%r3), 15
+0xeb 0x1f 0x30 0x02 0x00 0xe3
+
 # CHECK: s %r0, 0
 0x5b 0x00 0x00 0x00
 

test/MC/SystemZ/insn-bad-z196.s

@@ -65,3 +65,37 @@
 	srlk	%r0,%r0,524288
 	srlk	%r0,%r0,0(%r0)
 	srlk	%r0,%r0,0(%r1,%r2)
+
+#CHECK: error: invalid operand
+#CHECK: stoc	%r0,0,-1
+#CHECK: error: invalid operand
+#CHECK: stoc	%r0,0,16
+#CHECK: error: invalid operand
+#CHECK: stoc	%r0,-524289,1
+#CHECK: error: invalid operand
+#CHECK: stoc	%r0,524288,1
+#CHECK: error: invalid use of indexed addressing
+#CHECK: stoc	%r0,0(%r1,%r2),1
+
+	stoc	%r0,0,-1
+	stoc	%r0,0,16
+	stoc	%r0,-524289,1
+	stoc	%r0,524288,1
+	stoc	%r0,0(%r1,%r2),1
+
+#CHECK: error: invalid operand
+#CHECK: stocg	%r0,0,-1
+#CHECK: error: invalid operand
+#CHECK: stocg	%r0,0,16
+#CHECK: error: invalid operand
+#CHECK: stocg	%r0,-524289,1
+#CHECK: error: invalid operand
+#CHECK: stocg	%r0,524288,1
+#CHECK: error: invalid use of indexed addressing
+#CHECK: stocg	%r0,0(%r1,%r2),1
+
+	stocg	%r0,0,-1
+	stocg	%r0,0,16
+	stocg	%r0,-524289,1
+	stocg	%r0,524288,1
+	stocg	%r0,0(%r1,%r2),1

test/MC/SystemZ/insn-good-z196.s

@@ -295,6 +295,102 @@
 	srlk	%r0,%r0,524287(%r1)
 	srlk	%r0,%r0,524287(%r15)
 
+#CHECK: stoc	%r0, 0, 0               # encoding: [0xeb,0x00,0x00,0x00,0x00,0xf3]
+#CHECK: stoc	%r0, 0, 15              # encoding: [0xeb,0x0f,0x00,0x00,0x00,0xf3]
+#CHECK: stoc	%r0, -524288, 0         # encoding: [0xeb,0x00,0x00,0x00,0x80,0xf3]
+#CHECK: stoc	%r0, 524287, 0          # encoding: [0xeb,0x00,0x0f,0xff,0x7f,0xf3]
+#CHECK: stoc	%r0, 0(%r1), 0          # encoding: [0xeb,0x00,0x10,0x00,0x00,0xf3]
+#CHECK: stoc	%r0, 0(%r15), 0         # encoding: [0xeb,0x00,0xf0,0x00,0x00,0xf3]
+#CHECK: stoc	%r15, 0, 0              # encoding: [0xeb,0xf0,0x00,0x00,0x00,0xf3]
+#CHECK: stoc	%r1, 4095(%r2), 3       # encoding: [0xeb,0x13,0x2f,0xff,0x00,0xf3]
+
+	stoc	%r0,0,0
+	stoc	%r0,0,15
+	stoc	%r0,-524288,0
+	stoc	%r0,524287,0
+	stoc	%r0,0(%r1),0
+	stoc	%r0,0(%r15),0
+	stoc	%r15,0,0
+	stoc	%r1,4095(%r2),3
+
+#CHECK: stoco   %r1, 2(%r3)             # encoding: [0xeb,0x11,0x30,0x02,0x00,0xf3]
+#CHECK: stoch   %r1, 2(%r3)             # encoding: [0xeb,0x12,0x30,0x02,0x00,0xf3]
+#CHECK: stocnle %r1, 2(%r3)             # encoding: [0xeb,0x13,0x30,0x02,0x00,0xf3]
+#CHECK: stocl   %r1, 2(%r3)             # encoding: [0xeb,0x14,0x30,0x02,0x00,0xf3]
+#CHECK: stocnhe %r1, 2(%r3)             # encoding: [0xeb,0x15,0x30,0x02,0x00,0xf3]
+#CHECK: stoclh  %r1, 2(%r3)             # encoding: [0xeb,0x16,0x30,0x02,0x00,0xf3]
+#CHECK: stocne  %r1, 2(%r3)             # encoding: [0xeb,0x17,0x30,0x02,0x00,0xf3]
+#CHECK: stoce   %r1, 2(%r3)             # encoding: [0xeb,0x18,0x30,0x02,0x00,0xf3]
+#CHECK: stocnlh %r1, 2(%r3)             # encoding: [0xeb,0x19,0x30,0x02,0x00,0xf3]
+#CHECK: stoche  %r1, 2(%r3)             # encoding: [0xeb,0x1a,0x30,0x02,0x00,0xf3]
+#CHECK: stocnl  %r1, 2(%r3)             # encoding: [0xeb,0x1b,0x30,0x02,0x00,0xf3]
+#CHECK: stocle  %r1, 2(%r3)             # encoding: [0xeb,0x1c,0x30,0x02,0x00,0xf3]
+#CHECK: stocnh  %r1, 2(%r3)             # encoding: [0xeb,0x1d,0x30,0x02,0x00,0xf3]
+#CHECK: stocno  %r1, 2(%r3)             # encoding: [0xeb,0x1e,0x30,0x02,0x00,0xf3]
+
+	stoco   %r1,2(%r3)
+	stoch   %r1,2(%r3)
+	stocnle %r1,2(%r3)
+	stocl   %r1,2(%r3)
+	stocnhe %r1,2(%r3)
+	stoclh  %r1,2(%r3)
+	stocne  %r1,2(%r3)
+	stoce   %r1,2(%r3)
+	stocnlh %r1,2(%r3)
+	stoche  %r1,2(%r3)
+	stocnl  %r1,2(%r3)
+	stocle  %r1,2(%r3)
+	stocnh  %r1,2(%r3)
+	stocno  %r1,2(%r3)
+
+#CHECK: stocg	%r0, 0, 0               # encoding: [0xeb,0x00,0x00,0x00,0x00,0xe3]
+#CHECK: stocg	%r0, 0, 15              # encoding: [0xeb,0x0f,0x00,0x00,0x00,0xe3]
+#CHECK: stocg	%r0, -524288, 0         # encoding: [0xeb,0x00,0x00,0x00,0x80,0xe3]
+#CHECK: stocg	%r0, 524287, 0          # encoding: [0xeb,0x00,0x0f,0xff,0x7f,0xe3]
+#CHECK: stocg	%r0, 0(%r1), 0          # encoding: [0xeb,0x00,0x10,0x00,0x00,0xe3]
+#CHECK: stocg	%r0, 0(%r15), 0         # encoding: [0xeb,0x00,0xf0,0x00,0x00,0xe3]
+#CHECK: stocg	%r15, 0, 0              # encoding: [0xeb,0xf0,0x00,0x00,0x00,0xe3]
+#CHECK: stocg	%r1, 4095(%r2), 3       # encoding: [0xeb,0x13,0x2f,0xff,0x00,0xe3]
+
+	stocg	%r0,0,0
+	stocg	%r0,0,15
+	stocg	%r0,-524288,0
+	stocg	%r0,524287,0
+	stocg	%r0,0(%r1),0
+	stocg	%r0,0(%r15),0
+	stocg	%r15,0,0
+	stocg	%r1,4095(%r2),3
+
+#CHECK: stocgo   %r1, 2(%r3)            # encoding: [0xeb,0x11,0x30,0x02,0x00,0xe3]
+#CHECK: stocgh   %r1, 2(%r3)            # encoding: [0xeb,0x12,0x30,0x02,0x00,0xe3]
+#CHECK: stocgnle %r1, 2(%r3)            # encoding: [0xeb,0x13,0x30,0x02,0x00,0xe3]
+#CHECK: stocgl   %r1, 2(%r3)            # encoding: [0xeb,0x14,0x30,0x02,0x00,0xe3]
+#CHECK: stocgnhe %r1, 2(%r3)            # encoding: [0xeb,0x15,0x30,0x02,0x00,0xe3]
+#CHECK: stocglh  %r1, 2(%r3)            # encoding: [0xeb,0x16,0x30,0x02,0x00,0xe3]
+#CHECK: stocgne  %r1, 2(%r3)            # encoding: [0xeb,0x17,0x30,0x02,0x00,0xe3]
+#CHECK: stocge   %r1, 2(%r3)            # encoding: [0xeb,0x18,0x30,0x02,0x00,0xe3]
+#CHECK: stocgnlh %r1, 2(%r3)            # encoding: [0xeb,0x19,0x30,0x02,0x00,0xe3]
+#CHECK: stocghe  %r1, 2(%r3)            # encoding: [0xeb,0x1a,0x30,0x02,0x00,0xe3]
+#CHECK: stocgnl  %r1, 2(%r3)            # encoding: [0xeb,0x1b,0x30,0x02,0x00,0xe3]
+#CHECK: stocgle  %r1, 2(%r3)            # encoding: [0xeb,0x1c,0x30,0x02,0x00,0xe3]
+#CHECK: stocgnh  %r1, 2(%r3)            # encoding: [0xeb,0x1d,0x30,0x02,0x00,0xe3]
+#CHECK: stocgno  %r1, 2(%r3)            # encoding: [0xeb,0x1e,0x30,0x02,0x00,0xe3]
+
+	stocgo   %r1,2(%r3)
+	stocgh   %r1,2(%r3)
+	stocgnle %r1,2(%r3)
+	stocgl   %r1,2(%r3)
+	stocgnhe %r1,2(%r3)
+	stocglh  %r1,2(%r3)
+	stocgne  %r1,2(%r3)
+	stocge   %r1,2(%r3)
+	stocgnlh %r1,2(%r3)
+	stocghe  %r1,2(%r3)
+	stocgnl  %r1,2(%r3)
+	stocgle  %r1,2(%r3)
+	stocgnh  %r1,2(%r3)
+	stocgno  %r1,2(%r3)
+
 #CHECK: xgrk	%r0, %r0, %r0           # encoding: [0xb9,0xe7,0x00,0x00]
 #CHECK: xgrk	%r0, %r0, %r15          # encoding: [0xb9,0xe7,0xf0,0x00]
 #CHECK: xgrk	%r0, %r15, %r0          # encoding: [0xb9,0xe7,0x00,0x0f]