AMDGPU: Add instruction definitions for VGPR indexing

VI added a second method of indexing into VGPRs
besides using v_movrel*

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@284027 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index 14c79d2c8168244bcdc23e05d9556bf92cec7a3d..e0bde330449ef42c52cde9c769339180a25fe02a 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/lib/Target/AMDGPU/AMDGPU.td
@@ -168,6 +168,18 @@ def Feature16BitInsts : SubtargetFeature<"16-bit-insts",
   "Has i16/f16 instructions"
 >;
 
+def FeatureMovrel : SubtargetFeature<"movrel",
+  "HasMovrel",
+  "true",
+  "Has v_movrel*_b32 instructions"
+>;
+
+def FeatureVGPRIndexMode : SubtargetFeature<"vgpr-index-mode",
+  "HasVGPRIndexMode",
+  "true",
+  "Has VGPR mode register indexing"
+>;
+
 //===------------------------------------------------------------===//
 // Subtarget Features (options and debugging)
 //===------------------------------------------------------------===//
@@ -295,20 +307,20 @@ def FeatureNorthernIslands : SubtargetFeatureGeneration<"NORTHERN_ISLANDS",
 def FeatureSouthernIslands : SubtargetFeatureGeneration<"SOUTHERN_ISLANDS",
   [FeatureFP64, FeatureLocalMemorySize32768,
   FeatureWavefrontSize64, FeatureGCN, FeatureGCN1Encoding,
-  FeatureLDSBankCount32]
+  FeatureLDSBankCount32, FeatureMovrel]
 >;
 
 def FeatureSeaIslands : SubtargetFeatureGeneration<"SEA_ISLANDS",
   [FeatureFP64, FeatureLocalMemorySize65536,
   FeatureWavefrontSize64, FeatureGCN, FeatureFlatAddressSpace,
-  FeatureGCN1Encoding, FeatureCIInsts]
+  FeatureGCN1Encoding, FeatureCIInsts, FeatureMovrel]
 >;
 
 def FeatureVolcanicIslands : SubtargetFeatureGeneration<"VOLCANIC_ISLANDS",
   [FeatureFP64, FeatureLocalMemorySize65536,
    FeatureWavefrontSize64, FeatureFlatAddressSpace, FeatureGCN,
    FeatureGCN3Encoding, FeatureCIInsts, Feature16BitInsts,
-   FeatureSMemRealTime
+   FeatureSMemRealTime, FeatureVGPRIndexMode, FeatureMovrel
   ]
 >;
 

diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index 586d0478a5d13ae8e0aa27abc4dfe02ab26a45d2..aba12635453d419bc1a6fbfe49982ee2cba7c56e 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -107,6 +107,8 @@ AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     SGPRInitBug(false),
     HasSMemRealTime(false),
     Has16BitInsts(false),
+    HasMovrel(false),
+    HasVGPRIndexMode(false),
     FlatAddressSpace(false),
 
     R600ALUInst(false),

diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.h b/lib/Target/AMDGPU/AMDGPUSubtarget.h
index c0ebab06f768564b9bfff3465e3832b3908232e4..1dd6506520159374f5cab625b16b1170c1adf661 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -99,6 +99,8 @@ protected:
   bool SGPRInitBug;
   bool HasSMemRealTime;
   bool Has16BitInsts;
+  bool HasMovrel;
+  bool HasVGPRIndexMode;
   bool FlatAddressSpace;
   bool R600ALUInst;
   bool CaymanISA;
@@ -501,6 +503,14 @@ public:
     return Has16BitInsts;
   }
 
+  bool hasMovrel() const {
+    return HasMovrel;
+  }
+
+  bool hasVGPRIndexMode() const {
+    return HasVGPRIndexMode;
+  }
+
   bool hasScalarCompareEq64() const {
     return getGeneration() >= VOLCANIC_ISLANDS;
   }

diff --git a/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 6d8a7a5d4f6645396e6b51fe2b65aa4cc7472be5..957233caa9f8433ef50be6664812492c8bd53df2 100644 (file)
--- a/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -344,6 +344,7 @@ public:
   bool isSMRDOffset() const;
   bool isSMRDLiteralOffset() const;
   bool isDPPCtrl() const;
+  bool isGPRIdxMode() const;
 
   StringRef getExpressionAsToken() const {
     assert(isExpr());
@@ -2744,6 +2745,10 @@ bool AMDGPUOperand::isDPPCtrl() const {
   return false;
 }
 
+bool AMDGPUOperand::isGPRIdxMode() const {
+  return isImm() && isUInt<4>(getImm());
+}
+
 AMDGPUAsmParser::OperandMatchResultTy
 AMDGPUAsmParser::parseDPPCtrl(OperandVector &Operands) {
   SMLoc S = Parser.getTok().getLoc();

diff --git a/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp b/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp
index 56a6a0413c8a782b773ccf5ae50d15621fe7cc4b..cb10ec11671389383e93c45fd74df4a2a6eb76b5 100644 (file)
--- a/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp
+++ b/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.cpp
@@ -34,6 +34,7 @@ void AMDGPUInstPrinter::printInst(const MCInst *MI, raw_ostream &OS,
 }
 
 void AMDGPUInstPrinter::printU4ImmOperand(const MCInst *MI, unsigned OpNo,
+                                          const MCSubtargetInfo &STI,
                                           raw_ostream &O) {
   O << formatHex(MI->getOperand(OpNo).getImm() & 0xf);
 }
@@ -510,14 +511,14 @@ void AMDGPUInstPrinter::printRowMask(const MCInst *MI, unsigned OpNo,
                                      const MCSubtargetInfo &STI,
                                      raw_ostream &O) {
   O << " row_mask:";
-  printU4ImmOperand(MI, OpNo, O);
+  printU4ImmOperand(MI, OpNo, STI, O);
 }
 
 void AMDGPUInstPrinter::printBankMask(const MCInst *MI, unsigned OpNo,
                                       const MCSubtargetInfo &STI,
                                       raw_ostream &O) {
   O << " bank_mask:";
-  printU4ImmOperand(MI, OpNo, O);
+  printU4ImmOperand(MI, OpNo, STI, O);
 }
 
 void AMDGPUInstPrinter::printBoundCtrl(const MCInst *MI, unsigned OpNo,
@@ -598,6 +599,28 @@ void AMDGPUInstPrinter::printInterpSlot(const MCInst *MI, unsigned OpNo,
   }
 }
 
+void AMDGPUInstPrinter::printVGPRIndexMode(const MCInst *MI, unsigned OpNo,
+                                           const MCSubtargetInfo &STI,
+                                           raw_ostream &O) {
+  unsigned Val = MI->getOperand(OpNo).getImm();
+  if (Val == 0) {
+    O << " 0";
+    return;
+  }
+
+  if (Val & VGPRIndexMode::DST_ENABLE)
+    O << " dst";
+
+  if (Val & VGPRIndexMode::SRC0_ENABLE)
+    O << " src0";
+
+  if (Val & VGPRIndexMode::SRC1_ENABLE)
+    O << " src1";
+
+  if (Val & VGPRIndexMode::SRC2_ENABLE)
+    O << " src2";
+}
+
 void AMDGPUInstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
                                         const MCSubtargetInfo &STI,
                                         raw_ostream &O) {

diff --git a/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.h b/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.h
index 53b1966dfe07dd199f15648888331273dac55dd4..2c54e4832716a236a383b6922b9757bc9739f119 100644 (file)
--- a/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.h
+++ b/lib/Target/AMDGPU/InstPrinter/AMDGPUInstPrinter.h
@@ -34,7 +34,8 @@ public:
                               const MCRegisterInfo &MRI);
 
 private:
-  void printU4ImmOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printU4ImmOperand(const MCInst *MI, unsigned OpNo,
+                         const MCSubtargetInfo &STI, raw_ostream &O);
   void printU8ImmOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O);
   void printU16ImmOperand(const MCInst *MI, unsigned OpNo,
                           const MCSubtargetInfo &STI, raw_ostream &O);
@@ -107,6 +108,8 @@ private:
                           const MCSubtargetInfo &STI, raw_ostream &O);
   void printInterpSlot(const MCInst *MI, unsigned OpNo,
                        const MCSubtargetInfo &STI, raw_ostream &O);
+  void printVGPRIndexMode(const MCInst *MI, unsigned OpNo,
+                          const MCSubtargetInfo &STI, raw_ostream &O);
   void printMemOperand(const MCInst *MI, unsigned OpNo,
                        const MCSubtargetInfo &STI, raw_ostream &O);
   static void printIfSet(const MCInst *MI, unsigned OpNo, raw_ostream &O,

diff --git a/lib/Target/AMDGPU/SIDefines.h b/lib/Target/AMDGPU/SIDefines.h
index 2e4d0ecc8ce50bbcfcb8498cfd01593a71a57fdc..ad1bec25801d9ebe1f71c6efdd5362989333afc8 100644 (file)
--- a/lib/Target/AMDGPU/SIDefines.h
+++ b/lib/Target/AMDGPU/SIDefines.h
@@ -110,6 +110,15 @@ namespace SIOutMods {
   };
 }
 
+namespace VGPRIndexMode {
+  enum {
+    SRC0_ENABLE = 1 << 0,
+    SRC1_ENABLE = 1 << 1,
+    SRC2_ENABLE = 1 << 2,
+    DST_ENABLE = 1 << 3
+  };
+}
+
 namespace AMDGPUAsmVariants {
   enum {
     DEFAULT = 0,

diff --git a/lib/Target/AMDGPU/SIInstructions.td b/lib/Target/AMDGPU/SIInstructions.td
index cbff609684bc16fbf33e2a456422662a794dd817..38374000a9587ef47734ecbc7a07f01e13cc0077 100644 (file)
--- a/lib/Target/AMDGPU/SIInstructions.td
+++ b/lib/Target/AMDGPU/SIInstructions.td
@@ -20,6 +20,10 @@ def isSI : Predicate<"Subtarget->getGeneration() "
 
 def has16BankLDS : Predicate<"Subtarget->getLDSBankCount() == 16">;
 def has32BankLDS : Predicate<"Subtarget->getLDSBankCount() == 32">;
+def HasVGPRIndexMode : Predicate<"Subtarget->hasVGPRIndexMode()">,
+                      AssemblerPredicate<"FeatureVGPRIndexMode">;
+def HasMovrel : Predicate<"Subtarget->hasMovrel()">,
+                AssemblerPredicate<"FeatureMovrel">;
 
 include "VOPInstructions.td"
 include "SOPInstructions.td"

diff --git a/lib/Target/AMDGPU/SOPInstructions.td b/lib/Target/AMDGPU/SOPInstructions.td
index 404ee4260aaac6b6def3847737be4c00b7a525a1..7056b41c6956bb9f1b6a701235637eee0419e9cf 100644 (file)
--- a/lib/Target/AMDGPU/SOPInstructions.td
+++ b/lib/Target/AMDGPU/SOPInstructions.td
@@ -7,6 +7,18 @@
 //
 //===----------------------------------------------------------------------===//
 
+def GPRIdxModeMatchClass : AsmOperandClass {
+  let Name = "GPRIdxMode";
+  let PredicateMethod = "isGPRIdxMode";
+  let RenderMethod = "addImmOperands";
+}
+
+def GPRIdxMode : Operand<i32> {
+  let PrintMethod = "printVGPRIndexMode";
+  let ParserMatchClass = GPRIdxModeMatchClass;
+  let OperandType = "OPERAND_IMMEDIATE";
+}
+
 //===----------------------------------------------------------------------===//
 // SOP1 Instructions
 //===----------------------------------------------------------------------===//
@@ -63,6 +75,13 @@ class SOP1_32 <string opName, list<dag> pattern=[]> : SOP1_Pseudo <
   "$sdst, $src0", pattern
 >;
 
+// 32-bit input, no output.
+class SOP1_0_32 <string opName, list<dag> pattern = []> : SOP1_Pseudo <
+  opName, (outs), (ins SSrc_b32:$src0),
+  "$src0", pattern> {
+  let has_sdst = 0;
+}
+
 class SOP1_64 <string opName, list<dag> pattern=[]> : SOP1_Pseudo <
   opName, (outs SReg_64:$sdst), (ins SSrc_b64:$src0),
   "$sdst, $src0", pattern
@@ -198,6 +217,12 @@ def S_ABS_I32 : SOP1_32 <"s_abs_i32">;
 } // End Defs = [SCC]
 def S_MOV_FED_B32 : SOP1_32 <"s_mov_fed_b32">;
 
+let SubtargetPredicate = HasVGPRIndexMode in {
+def S_SET_GPR_IDX_IDX : SOP1_0_32<"s_set_gpr_idx_idx"> {
+  let Uses = [M0];
+  let Defs = [M0];
+}
+}
 
 //===----------------------------------------------------------------------===//
 // SOP2 Instructions
@@ -597,7 +622,8 @@ class SOPCe <bits<7> op> : Enc32 {
   let Inst{31-23} = 0x17e;
 }
 
-class SOPC <bits<7> op, dag outs, dag ins, string asm, list<dag> pattern> :
+class SOPC <bits<7> op, dag outs, dag ins, string asm,
+            list<dag> pattern = []> :
   InstSI<outs, ins, asm, pattern>, SOPCe <op> {
   let mayLoad = 0;
   let mayStore = 0;
@@ -670,6 +696,17 @@ def S_CMP_EQ_U64 : SOPC_CMP_64 <0x12, "s_cmp_eq_u64", COND_EQ>;
 def S_CMP_LG_U64 : SOPC_CMP_64 <0x13, "s_cmp_lg_u64", COND_NE>;
 }
 
+let SubtargetPredicate = HasVGPRIndexMode in {
+def S_SET_GPR_IDX_ON : SOPC <0x11,
+  (outs),
+  (ins SSrc_b32:$src0, GPRIdxMode:$src1),
+  "s_set_gpr_idx_on $src0,$src1"> {
+  let Defs = [M0]; // No scc def
+  let Uses = [M0]; // Other bits of m0 unmodified.
+  let hasSideEffects = 1; // Sets mode.gpr_idx_en
+}
+}
+
 //===----------------------------------------------------------------------===//
 // SOPP Instructions
 //===----------------------------------------------------------------------===//
@@ -809,8 +846,20 @@ def S_DECPERFLEVEL : SOPP <0x00000015, (ins i32imm:$simm16), "s_decperflevel $si
 def S_TTRACEDATA : SOPP <0x00000016, (ins), "s_ttracedata"> {
   let simm16 = 0;
 }
+
+let SubtargetPredicate = HasVGPRIndexMode in {
+def S_SET_GPR_IDX_OFF : SOPP<0x1c, (ins), "s_set_gpr_idx_off"> {
+  let simm16 = 0;
+}
+}
 } // End hasSideEffects
 
+let SubtargetPredicate = HasVGPRIndexMode in {
+def S_SET_GPR_IDX_MODE : SOPP<0x1d, (ins GPRIdxMode:$simm16),
+  "s_set_gpr_idx_mode$simm16"> {
+  let Defs = [M0];
+}
+}
 
 let Predicates = [isGCN] in {
 
@@ -1071,6 +1120,7 @@ def S_CBRANCH_JOIN_vi      : SOP1_Real_vi <0x2e, S_CBRANCH_JOIN>;
 def S_MOV_REGRD_B32_vi     : SOP1_Real_vi <0x2f, S_MOV_REGRD_B32>;
 def S_ABS_I32_vi           : SOP1_Real_vi <0x30, S_ABS_I32>;
 def S_MOV_FED_B32_vi       : SOP1_Real_vi <0x31, S_MOV_FED_B32>;
+def S_SET_GPR_IDX_IDX_vi   : SOP1_Real_vi <0x32, S_SET_GPR_IDX_IDX>;
 
 def S_ADD_U32_vi           : SOP2_Real_vi <0x00, S_ADD_U32>;
 def S_ADD_I32_vi           : SOP2_Real_vi <0x02, S_ADD_I32>;

diff --git a/lib/Target/AMDGPU/VOP1Instructions.td b/lib/Target/AMDGPU/VOP1Instructions.td
index f108922b213f16f44d5723f59656f7379ce922c1..77b39d7172705129353e31872c562721c1adf1ec 100644 (file)
--- a/lib/Target/AMDGPU/VOP1Instructions.td
+++ b/lib/Target/AMDGPU/VOP1Instructions.td
@@ -229,7 +229,7 @@ def VOP_MOVRELD : VOPProfile<[untyped, i32, untyped, untyped]> {
   let EmitDst = 1; // force vdst emission
 }
 
-let Uses = [M0, EXEC] in {
+let SubtargetPredicate = HasMovrel, Uses = [M0, EXEC] in {
 // v_movreld_b32 is a special case because the destination output
  // register is really a source. It isn't actually read (but may be
  // written), and is only to provide the base register to start

diff --git a/test/MC/AMDGPU/sop1.s b/test/MC/AMDGPU/sop1.s
index c0985f144c9764837235117358882f680f86cf88..f611b022fb4b8f4ad35d8a393043be95c2b39d35 100644 (file)
--- a/test/MC/AMDGPU/sop1.s
+++ b/test/MC/AMDGPU/sop1.s
@@ -1,5 +1,5 @@
-// RUN: llvm-mc -arch=amdgcn -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SICI %s
-// RUN: llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SICI %s
+// RUN: not llvm-mc -arch=amdgcn -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SICI %s
+// RUN: not llvm-mc -arch=amdgcn -show-encoding %s 2>&1 | FileCheck --check-prefix=NOSICI %s
 // RUN: not llvm-mc -arch=amdgcn -mcpu=fiji -show-encoding %s 2>&1 | FileCheck --check-prefix=GCN --check-prefix=VI %s
 // RUN: not llvm-mc -arch=amdgcn -mcpu=fiji -show-encoding %s 2>&1 | FileCheck --check-prefix=NOVI %s
 
@@ -242,3 +242,7 @@ s_abs_i32 s1, s2
 
 s_mov_fed_b32 s1, s2
 // SICI: s_mov_fed_b32 s1, s2 ; encoding: [0x02,0x35,0x81,0xbe]
+
+s_set_gpr_idx_idx s0
+// VI: s_set_gpr_idx_idx s0 ; encoding: [0x00,0x32,0x80,0xbe]
+// NOSICI: error: instruction not supported on this GPU

diff --git a/test/MC/AMDGPU/sopc-err.s b/test/MC/AMDGPU/sopc-err.s
new file mode 100644 (file)
index 0000000..b6d9542
--- /dev/null
+++ b/test/MC/AMDGPU/sopc-err.s
@@ -0,0 +1,10 @@
+// RUN: not llvm-mc -arch=amdgcn -mcpu=tonga %s 2>&1 | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+s_set_gpr_idx_on s0, s1
+// GCN: error: invalid operand for instruction
+
+s_set_gpr_idx_on s0, 16
+// GCN: error: invalid operand for instruction
+
+s_set_gpr_idx_on s0, -1
+// GCN: error: invalid operand for instruction

diff --git a/test/MC/AMDGPU/sopc.s b/test/MC/AMDGPU/sopc.s
index ef4351d06599d028398793294788712feb069fe4..dc922b65a6e9f373c47ece614b04ef30cb1286f4 100644 (file)
--- a/test/MC/AMDGPU/sopc.s
+++ b/test/MC/AMDGPU/sopc.s
@@ -64,3 +64,19 @@ s_cmp_eq_u64 s[0:1], s[2:3]
 s_cmp_lg_u64 s[0:1], s[2:3]
 // VI: s_cmp_lg_u64 s[0:1], s[2:3] ; encoding: [0x00,0x02,0x13,0xbf]
 // NOSICI: error: instruction not supported on this GPU
+
+s_set_gpr_idx_on s0, 0
+// VI: s_set_gpr_idx_on s0, 0 ; encoding: [0x00,0x00,0x11,0xbf]
+// NOSICI: error: instruction not supported on this GPU
+
+s_set_gpr_idx_on s0, 1
+// VI: s_set_gpr_idx_on s0, src0 ; encoding: [0x00,0x01,0x11,0xbf]
+// NOSICI: error: instruction not supported on this GPU
+
+s_set_gpr_idx_on s0, 3
+// VI: s_set_gpr_idx_on s0, src0 src1 ; encoding: [0x00,0x03,0x11,0xbf]
+// NOSICI: error: instruction not supported on this GPU
+
+s_set_gpr_idx_on s0, 15
+// VI: s_set_gpr_idx_on s0, dst src0 src1 src2 ; encoding: [0x00,0x0f,0x11,0xbf]
+// NOSICI: error: instruction not supported on this GPU

diff --git a/test/MC/AMDGPU/sopp.s b/test/MC/AMDGPU/sopp.s
index 59de24533e0e37d574f2cb5443a636eacba617b3..b073c8dfc635e95dbfbc0817910bda23f225d578 100644 (file)
--- a/test/MC/AMDGPU/sopp.s
+++ b/test/MC/AMDGPU/sopp.s
@@ -1,5 +1,5 @@
-// RUN: llvm-mc -arch=amdgcn -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SICI %s
-// RUN: llvm-mc -arch=amdgcn -mcpu=SI -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SICI %s
+// RUN: not llvm-mc -arch=amdgcn -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=SICI %s
+// RUN: not llvm-mc -arch=amdgcn -show-encoding %s 2>&1 | FileCheck %s --check-prefix=NOSICI
 // RUN: llvm-mc -arch=amdgcn -mcpu=fiji -show-encoding %s | FileCheck --check-prefix=GCN --check-prefix=VI %s
 
 //===----------------------------------------------------------------------===//
@@ -176,3 +176,15 @@ s_decperflevel 6
 
 s_ttracedata
 // GCN: s_ttracedata ; encoding: [0x00,0x00,0x96,0xbf]
+
+s_set_gpr_idx_off
+// VI:         s_set_gpr_idx_off ; encoding: [0x00,0x00,0x9c,0xbf]
+// NOSICI: error: instruction not supported on this GPU
+
+s_set_gpr_idx_mode 0
+// VI: s_set_gpr_idx_mode 0 ; encoding: [0x00,0x00,0x9d,0xbf]
+// NOSICI: error: instruction not supported on this GPU
+
+s_set_gpr_idx_mode 15
+// VI: s_set_gpr_idx_mode dst src0 src1 src2 ; encoding: [0x0f,0x00,0x9d,0xbf]
+// NOSICI: error: instruction not supported on this GPU