[AMDGPU] Add s_set_vgpr_msb gfx1250 instruction (#156524)

Author: rampitec

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -1210,6 +1210,12 @@ def Feature64BitLiterals : SubtargetFeature<"64-bit-literals",
   "Can use 64-bit literals with single DWORD instructions"
 >;
 
+def Feature1024AddressableVGPRs : SubtargetFeature<"1024-addressable-vgprs",
+  "Has1024AddressableVGPRs",
+  "true",
+  "Has 1024 addressable VGPRs"
+>;
+
 def FeatureWaitXcnt : SubtargetFeature<"wait-xcnt",
   "HasWaitXcnt",
   "true",
@@ -2033,6 +2039,7 @@ def FeatureISAVersion12_50 : FeatureSet<
    FeatureCUStores,
    FeatureAddressableLocalMemorySize327680,
    FeatureCuMode,
+   Feature1024AddressableVGPRs,
    Feature64BitLiterals,
    FeatureLDSBankCount32,
    FeatureDLInsts,
@@ -2841,6 +2848,9 @@ def HasBVHDualAndBVH8Insts : Predicate<"Subtarget->hasBVHDualAndBVH8Insts()">,
 def Has64BitLiterals : Predicate<"Subtarget->has64BitLiterals()">,
   AssemblerPredicate<(all_of Feature64BitLiterals)>;
 
+def Has1024AddressableVGPRs : Predicate<"Subtarget->has1024AddressableVGPRs()">,
+  AssemblerPredicate<(all_of Feature1024AddressableVGPRs)>;
+
 def HasWaitXcnt : Predicate<"Subtarget->hasWaitXcnt()">,
   AssemblerPredicate<(all_of FeatureWaitXcnt)>;
 

llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

@@ -1886,6 +1886,7 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
   bool validateTHAndScopeBits(const MCInst &Inst, const OperandVector &Operands,
                               const unsigned CPol);
   bool validateTFE(const MCInst &Inst, const OperandVector &Operands);
+  bool validateSetVgprMSB(const MCInst &Inst, const OperandVector &Operands);
   std::optional<StringRef> validateLdsDirect(const MCInst &Inst);
   bool validateWMMA(const MCInst &Inst, const OperandVector &Operands);
   unsigned getConstantBusLimit(unsigned Opcode) const;
@@ -5542,6 +5543,22 @@ bool AMDGPUAsmParser::validateTFE(const MCInst &Inst,
   return true;
 }
 
+bool AMDGPUAsmParser::validateSetVgprMSB(const MCInst &Inst,
+                                         const OperandVector &Operands) {
+  if (Inst.getOpcode() != AMDGPU::S_SET_VGPR_MSB_gfx12)
+    return true;
+
+  int Simm16Pos =
+      AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::simm16);
+  if ((unsigned)Inst.getOperand(Simm16Pos).getImm() > 255) {
+    SMLoc Loc = Operands[1]->getStartLoc();
+    Error(Loc, "s_set_vgpr_msb accepts values in range [0..255]");
+    return false;
+  }
+
+  return true;
+}
+
 bool AMDGPUAsmParser::validateWMMA(const MCInst &Inst,
                                    const OperandVector &Operands) {
   unsigned Opc = Inst.getOpcode();
@@ -5706,6 +5723,9 @@ bool AMDGPUAsmParser::validateInstruction(const MCInst &Inst,
   if (!validateTFE(Inst, Operands)) {
     return false;
   }
+  if (!validateSetVgprMSB(Inst, Operands)) {
+    return false;
+  }
   if (!validateWMMA(Inst, Operands)) {
     return false;
   }

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -236,6 +236,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   bool HasPseudoScalarTrans = false;
   bool HasRestrictedSOffset = false;
   bool Has64BitLiterals = false;
+  bool Has1024AddressableVGPRs = false;
   bool HasBitOp3Insts = false;
   bool HasTanhInsts = false;
   bool HasTensorCvtLutInsts = false;
@@ -1437,6 +1438,8 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
 
   bool hasAddPC64Inst() const { return GFX1250Insts; }
 
+  bool has1024AddressableVGPRs() const { return Has1024AddressableVGPRs; }
+
   bool hasMinimum3Maximum3PKF16() const {
     return HasMinimum3Maximum3PKF16;
   }

llvm/lib/Target/AMDGPU/SOPInstructions.td

@@ -1844,6 +1844,13 @@ let SubtargetPredicate = HasWaitXcnt, hasSideEffects = 1 in {
     SOPP_Pseudo<"s_wait_xcnt", (ins s16imm:$simm16), "$simm16">;
 } // End SubtargetPredicate = hasWaitXcnt, hasSideEffects = 1
 
+let SubtargetPredicate = Has1024AddressableVGPRs in {
+  def S_SET_VGPR_MSB : SOPP_Pseudo<"s_set_vgpr_msb" , (ins i16imm:$simm16), "$simm16"> {
+    let hasSideEffects = 1;
+    let Defs = [MODE];
+  }
+}
+
 //===----------------------------------------------------------------------===//
 // SOP1 Patterns
 //===----------------------------------------------------------------------===//
@@ -2691,6 +2698,7 @@ defm S_WAIT_STORECNT_DSCNT  : SOPP_Real_32_gfx12<0x049>;
 //===----------------------------------------------------------------------===//
 // SOPP - GFX1250 only.
 //===----------------------------------------------------------------------===//
+defm S_SET_VGPR_MSB   : SOPP_Real_32_gfx12<0x006>;
 defm S_SETPRIO_INC_WG : SOPP_Real_32_gfx12<0x03e>;
 defm S_WAIT_XCNT      : SOPP_Real_32_gfx12<0x045>;
 defm S_WAIT_ASYNCCNT  : SOPP_Real_32_gfx12<0x04a>;

llvm/test/MC/AMDGPU/gfx1250_asm_sopp.s

@@ -37,6 +37,14 @@ s_setprio_inc_wg 100
 // GFX1250: [0x64,0x00,0xbe,0xbf]
 // GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
 
+s_set_vgpr_msb 10
+// GFX1250: [0x0a,0x00,0x86,0xbf]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+s_set_vgpr_msb 255
+// GFX1250: [0xff,0x00,0x86,0xbf]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
 s_monitor_sleep 1
 // GFX1250: s_monitor_sleep 1                       ; encoding: [0x01,0x00,0x84,0xbf]
 // GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU

llvm/test/MC/AMDGPU/gfx1250_err.s

@@ -1,5 +1,15 @@
 // RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1250 -show-encoding %s 2>&1 | FileCheck --check-prefixes=GFX1250-ERR --implicit-check-not=error: -strict-whitespace %s
 
+s_set_vgpr_msb -1
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: s_set_vgpr_msb accepts values in range [0..255]
+// GFX1250-ERR: s_set_vgpr_msb -1
+// GFX1250-ERR:                ^
+
+s_set_vgpr_msb 256
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: s_set_vgpr_msb accepts values in range [0..255]
+// GFX1250-ERR: s_set_vgpr_msb 256
+// GFX1250-ERR:                ^
+
 s_load_b32 s4, s[2:3], 10 th:TH_LOAD_NT th:TH_LOAD_NT
 // GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
 // GFX1250-ERR: s_load_b32 s4, s[2:3], 10 th:TH_LOAD_NT th:TH_LOAD_NT

llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_sopp.txt

@@ -27,6 +27,12 @@
 # GFX1250: s_setprio_inc_wg 0x64 ; encoding: [0x64,0x00,0xbe,0xbf]
 0x64,0x00,0xbe,0xbf
 
+# GFX1250: s_set_vgpr_msb 10 ; encoding: [0x0a,0x00,0x86,0xbf]
+0x0a,0x00,0x86,0xbf
+
+# GFX1250: s_set_vgpr_msb 0xff ; encoding: [0xff,0x00,0x86,0xbf]
+0xff,0x00,0x86,0xbf
+
 # GFX1250: s_monitor_sleep 0                       ; encoding: [0x00,0x00,0x84,0xbf]
 0x00,0x00,0x84,0xbf