[AMDGPU] gfx1250: MC support for 64-bit literals (#147861)

Author: rampitec

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -1130,6 +1130,12 @@ def FeaturePointSampleAccel : SubtargetFeature<"point-sample-accel",
   "Has point sample acceleration feature"
 >;
 
+def Feature64BitLiterals : SubtargetFeature<"64-bit-literals",
+  "Has64BitLiterals",
+  "true",
+  "Can use 64-bit literals with single DWORD instructions"
+>;
+
 def FeatureWaitXcnt : SubtargetFeature<"wait-xcnt",
   "HasWaitXcnt",
   "true",
@@ -1931,6 +1937,7 @@ def FeatureISAVersion12_50 : FeatureSet<
   [FeatureGFX12,
    FeatureGFX1250Insts,
    FeatureCuMode,
+   Feature64BitLiterals,
    FeatureLDSBankCount32,
    FeatureDLInsts,
    FeatureFmacF64Inst,
@@ -2678,6 +2685,9 @@ def HasPrngInst : Predicate<"Subtarget->hasPrngInst()">,
 def HasBVHDualAndBVH8Insts : Predicate<"Subtarget->hasBVHDualAndBVH8Insts()">,
   AssemblerPredicate<(all_of FeatureBVHDualAndBVH8Insts)>;
 
+def Has64BitLiterals : Predicate<"Subtarget->has64BitLiterals()">,
+  AssemblerPredicate<(all_of Feature64BitLiterals)>;
+
 def HasWaitXcnt : Predicate<"Subtarget->hasWaitXcnt()">,
   AssemblerPredicate<(all_of FeatureWaitXcnt)>;
 

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

@@ -81,6 +81,7 @@ class AMDGPUOperand : public MCParsedAsmOperand {
     bool Neg = false;
     bool Sext = false;
     bool Lit = false;
+    bool Lit64 = false;
 
     bool hasFPModifiers() const { return Abs || Neg; }
     bool hasIntModifiers() const { return Sext; }
@@ -480,7 +481,10 @@ class AMDGPUOperand : public MCParsedAsmOperand {
   bool isSSrc_b64() const {
     // TODO: Find out how SALU supports extension of 32-bit literals to 64 bits.
     // See isVSrc64().
-    return isSCSrc_b64() || isLiteralImm(MVT::i64);
+    return isSCSrc_b64() || isLiteralImm(MVT::i64) ||
+           (((const MCTargetAsmParser *)AsmParser)
+                ->getAvailableFeatures()[AMDGPU::Feature64BitLiterals] &&
+            isExpr());
   }
 
   bool isSSrc_f32() const {
@@ -1537,6 +1541,10 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
     return getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm];
   }
 
+  bool has64BitLiterals() const {
+    return getFeatureBits()[AMDGPU::Feature64BitLiterals];
+  }
+
   bool hasFlatOffsets() const {
     return getFeatureBits()[AMDGPU::FeatureFlatInstOffsets];
   }
@@ -1663,10 +1671,10 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
   bool isOpcodeModifierWithVal(const AsmToken &Token, const AsmToken &NextToken) const;
   bool parseSP3NegModifier();
   ParseStatus parseImm(OperandVector &Operands, bool HasSP3AbsModifier = false,
-                       bool HasLit = false);
+                       bool HasLit = false, bool HasLit64 = false);
   ParseStatus parseReg(OperandVector &Operands);
   ParseStatus parseRegOrImm(OperandVector &Operands, bool HasSP3AbsMod = false,
-                            bool HasLit = false);
+                            bool HasLit = false, bool HasLit64 = false);
   ParseStatus parseRegOrImmWithFPInputMods(OperandVector &Operands,
                                            bool AllowImm = true);
   ParseStatus parseRegOrImmWithIntInputMods(OperandVector &Operands,
@@ -2123,6 +2131,9 @@ bool AMDGPUOperand::isLiteralImm(MVT type) const {
     return false;
   }
 
+  bool Allow64Bit =
+      (type == MVT::i64 || type == MVT::f64) && AsmParser->has64BitLiterals();
+
   if (!Imm.IsFPImm) {
     // We got int literal token.
 
@@ -2134,8 +2145,11 @@ bool AMDGPUOperand::isLiteralImm(MVT type) const {
     }
 
     unsigned Size = type.getSizeInBits();
-    if (Size == 64)
+    if (Size == 64) {
+      if (Allow64Bit && !AMDGPU::isValid32BitLiteral(Imm.Val, false))
+        return true;
       Size = 32;
+    }
 
     // FIXME: 64-bit operands can zero extend, sign extend, or pad zeroes for FP
     // types.
@@ -2287,12 +2301,18 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
       }
 
       // Non-inlineable
-      if (AMDGPU::isSISrcFPOperand(InstDesc, OpNum)) { // Expected 64-bit fp operand
+      if (AMDGPU::isSISrcFPOperand(InstDesc,
+                                   OpNum)) { // Expected 64-bit fp operand
+        bool HasMandatoryLiteral =
+            AMDGPU::hasNamedOperand(Inst.getOpcode(), AMDGPU::OpName::imm);
         // For fp operands we check if low 32 bits are zeros
-        if (Literal.getLoBits(32) != 0) {
-          const_cast<AMDGPUAsmParser *>(AsmParser)->Warning(Inst.getLoc(),
-          "Can't encode literal as exact 64-bit floating-point operand. "
-          "Low 32-bits will be set to zero");
+        if (Literal.getLoBits(32) != 0 &&
+            (InstDesc.getSize() != 4 || !AsmParser->has64BitLiterals()) &&
+            !HasMandatoryLiteral) {
+          const_cast<AMDGPUAsmParser *>(AsmParser)->Warning(
+              Inst.getLoc(),
+              "Can't encode literal as exact 64-bit floating-point operand. "
+              "Low 32-bits will be set to zero");
           Val &= 0xffffffff00000000u;
         }
 
@@ -2392,8 +2412,25 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
     return;
 
   case AMDGPU::OPERAND_REG_IMM_INT64:
-  case AMDGPU::OPERAND_REG_IMM_FP64:
   case AMDGPU::OPERAND_REG_INLINE_C_INT64:
+    if (AMDGPU::isInlinableLiteral64(Val, AsmParser->hasInv2PiInlineImm())) {
+      Inst.addOperand(MCOperand::createImm(Val));
+      setImmKindConst();
+      return;
+    }
+
+    // When the 32 MSBs are not zero (effectively means it can't be safely
+    // truncated to uint32_t), if the target doesn't support 64-bit literals, or
+    // the lit modifier is explicitly used, we need to truncate it to the 32
+    // LSBs.
+    if (!AsmParser->has64BitLiterals() || getModifiers().Lit)
+      Val = Lo_32(Val);
+
+    Inst.addOperand(MCOperand::createImm(Val));
+    setImmKindLiteral();
+    return;
+
+  case AMDGPU::OPERAND_REG_IMM_FP64:
   case AMDGPU::OPERAND_REG_INLINE_C_FP64:
   case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
     if (AMDGPU::isInlinableLiteral64(Val, AsmParser->hasInv2PiInlineImm())) {
@@ -2402,8 +2439,20 @@ void AMDGPUOperand::addLiteralImmOperand(MCInst &Inst, int64_t Val, bool ApplyMo
       return;
     }
 
-    Val = AMDGPU::isSISrcFPOperand(InstDesc, OpNum) ? (uint64_t)Val << 32
-                                                    : Lo_32(Val);
+    // If the target doesn't support 64-bit literals, we need to use the
+    // constant as the high 32 MSBs of a double-precision floating point value.
+    if (!AsmParser->has64BitLiterals()) {
+      Val = static_cast<uint64_t>(Val) << 32;
+    } else {
+      // Now the target does support 64-bit literals, there are two cases
+      // where we still want to use src_literal encoding:
+      // 1) explicitly forced by using lit modifier;
+      // 2) the value is a valid 32-bit representation (signed or unsigned),
+      // meanwhile not forced by lit64 modifier.
+      if (getModifiers().Lit ||
+          (!getModifiers().Lit64 && (isInt<32>(Val) || isUInt<32>(Val))))
+        Val = static_cast<uint64_t>(Val) << 32;
+    }
 
     Inst.addOperand(MCOperand::createImm(Val));
     setImmKindLiteral();
@@ -3151,19 +3200,20 @@ AMDGPUAsmParser::parseRegister(bool RestoreOnFailure) {
 }
 
 ParseStatus AMDGPUAsmParser::parseImm(OperandVector &Operands,
-                                      bool HasSP3AbsModifier, bool HasLit) {
+                                      bool HasSP3AbsModifier, bool HasLit,
+                                      bool HasLit64) {
   // TODO: add syntactic sugar for 1/(2*PI)
 
-  if (isRegister())
+  if (isRegister() || isModifier())
     return ParseStatus::NoMatch;
-  assert(!isModifier());
 
-  if (!HasLit) {
-    HasLit = trySkipId("lit");
-    if (HasLit) {
+  if (!HasLit && !HasLit64) {
+    HasLit64 = trySkipId("lit64");
+    HasLit = !HasLit64 && trySkipId("lit");
+    if (HasLit || HasLit64) {
       if (!skipToken(AsmToken::LParen, "expected left paren after lit"))
         return ParseStatus::Failure;
-      ParseStatus S = parseImm(Operands, HasSP3AbsModifier, HasLit);
+      ParseStatus S = parseImm(Operands, HasSP3AbsModifier, HasLit, HasLit64);
       if (S.isSuccess() &&
           !skipToken(AsmToken::RParen, "expected closing parentheses"))
         return ParseStatus::Failure;
@@ -3185,6 +3235,7 @@ ParseStatus AMDGPUAsmParser::parseImm(OperandVector &Operands,
 
   AMDGPUOperand::Modifiers Mods;
   Mods.Lit = HasLit;
+  Mods.Lit64 = HasLit64;
 
   if (IsReal) {
     // Floating-point expressions are not supported.
@@ -3235,7 +3286,7 @@ ParseStatus AMDGPUAsmParser::parseImm(OperandVector &Operands,
       AMDGPUOperand &Op = static_cast<AMDGPUOperand &>(*Operands.back());
       Op.setModifiers(Mods);
     } else {
-      if (HasLit)
+      if (HasLit || HasLit64)
         return ParseStatus::NoMatch;
       Operands.push_back(AMDGPUOperand::CreateExpr(this, Expr, S));
     }
@@ -3259,13 +3310,14 @@ ParseStatus AMDGPUAsmParser::parseReg(OperandVector &Operands) {
 }
 
 ParseStatus AMDGPUAsmParser::parseRegOrImm(OperandVector &Operands,
-                                           bool HasSP3AbsMod, bool HasLit) {
+                                           bool HasSP3AbsMod, bool HasLit,
+                                           bool HasLit64) {
   ParseStatus Res = parseReg(Operands);
   if (!Res.isNoMatch())
     return Res;
   if (isModifier())
     return ParseStatus::NoMatch;
-  return parseImm(Operands, HasSP3AbsMod, HasLit);
+  return parseImm(Operands, HasSP3AbsMod, HasLit, HasLit64);
 }
 
 bool
@@ -3361,7 +3413,7 @@ AMDGPUAsmParser::parseRegOrImmWithFPInputMods(OperandVector &Operands,
                                               bool AllowImm) {
   bool Neg, SP3Neg;
   bool Abs, SP3Abs;
-  bool Lit;
+  bool Lit64, Lit;
   SMLoc Loc;
 
   // Disable ambiguous constructs like '--1' etc. Should use neg(-1) instead.
@@ -3381,7 +3433,15 @@ AMDGPUAsmParser::parseRegOrImmWithFPInputMods(OperandVector &Operands,
   if (Abs && !skipToken(AsmToken::LParen, "expected left paren after abs"))
     return ParseStatus::Failure;
 
-  Lit = trySkipId("lit");
+  Lit64 = trySkipId("lit64");
+  if (Lit64) {
+    if (!skipToken(AsmToken::LParen, "expected left paren after lit64"))
+      return ParseStatus::Failure;
+    if (!has64BitLiterals())
+      return Error(Loc, "lit64 is not supported on this GPU");
+  }
+
+  Lit = !Lit64 && trySkipId("lit");
   if (Lit && !skipToken(AsmToken::LParen, "expected left paren after lit"))
     return ParseStatus::Failure;
 
@@ -3392,14 +3452,16 @@ AMDGPUAsmParser::parseRegOrImmWithFPInputMods(OperandVector &Operands,
 
   ParseStatus Res;
   if (AllowImm) {
-    Res = parseRegOrImm(Operands, SP3Abs, Lit);
+    Res = parseRegOrImm(Operands, SP3Abs, Lit, Lit64);
   } else {
     Res = parseReg(Operands);
   }
   if (!Res.isSuccess())
-    return (SP3Neg || Neg || SP3Abs || Abs || Lit) ? ParseStatus::Failure : Res;
+    return (SP3Neg || Neg || SP3Abs || Abs || Lit || Lit64)
+               ? ParseStatus::Failure
+               : Res;
 
-  if (Lit && !Operands.back()->isImm())
+  if ((Lit || Lit64) && !Operands.back()->isImm())
     Error(Loc, "expected immediate with lit modifier");
 
   if (SP3Abs && !skipToken(AsmToken::Pipe, "expected vertical bar"))
@@ -3408,15 +3470,17 @@ AMDGPUAsmParser::parseRegOrImmWithFPInputMods(OperandVector &Operands,
     return ParseStatus::Failure;
   if (Neg && !skipToken(AsmToken::RParen, "expected closing parentheses"))
     return ParseStatus::Failure;
-  if (Lit && !skipToken(AsmToken::RParen, "expected closing parentheses"))
+  if ((Lit || Lit64) &&
+      !skipToken(AsmToken::RParen, "expected closing parentheses"))
     return ParseStatus::Failure;
 
   AMDGPUOperand::Modifiers Mods;
   Mods.Abs = Abs || SP3Abs;
   Mods.Neg = Neg || SP3Neg;
   Mods.Lit = Lit;
+  Mods.Lit64 = Lit64;
 
-  if (Mods.hasFPModifiers() || Lit) {
+  if (Mods.hasFPModifiers() || Lit || Lit64) {
     AMDGPUOperand &Op = static_cast<AMDGPUOperand &>(*Operands.back());
     if (Op.isExpr())
       return Error(Op.getStartLoc(), "expected an absolute expression");
@@ -4588,7 +4652,7 @@ bool AMDGPUAsmParser::validateSOPLiteral(const MCInst &Inst) const {
 
   unsigned NumExprs = 0;
   unsigned NumLiterals = 0;
-  uint32_t LiteralValue;
+  uint64_t LiteralValue;
 
   for (int OpIdx : OpIndices) {
     if (OpIdx == -1) break;
@@ -4597,7 +4661,7 @@ bool AMDGPUAsmParser::validateSOPLiteral(const MCInst &Inst) const {
     // Exclude special imm operands (like that used by s_set_gpr_idx_on)
     if (AMDGPU::isSISrcOperand(Desc, OpIdx)) {
       if (MO.isImm() && !isInlineConstant(Inst, OpIdx)) {
-        uint32_t Value = static_cast<uint32_t>(MO.getImm());
+        uint64_t Value = static_cast<uint64_t>(MO.getImm());
         if (NumLiterals == 0 || LiteralValue != Value) {
           LiteralValue = Value;
           ++NumLiterals;

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

@@ -1484,6 +1484,20 @@ MCOperand AMDGPUDisassembler::decodeLiteralConstant(bool ExtendFP64) const {
   return MCOperand::createImm(ExtendFP64 ? Literal64 : Literal);
 }
 
+MCOperand AMDGPUDisassembler::decodeLiteral64Constant() const {
+  assert(STI.hasFeature(AMDGPU::Feature64BitLiterals));
+
+  if (!HasLiteral) {
+    if (Bytes.size() < 8) {
+      return errOperand(0, "cannot read literal64, inst bytes left " +
+                               Twine(Bytes.size()));
+    }
+    HasLiteral = true;
+    Literal64 = eatBytes<uint64_t>(Bytes);
+  }
+  return MCOperand::createImm(Literal64);
+}
+
 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
   using namespace AMDGPU::EncValues;
 
@@ -1767,6 +1781,10 @@ MCOperand AMDGPUDisassembler::decodeNonVGPRSrcOp(unsigned Width,
       Val == LITERAL_CONST)
     return MCOperand::createImm(Val);
 
+  if (Val == LITERAL64_CONST && STI.hasFeature(AMDGPU::Feature64BitLiterals)) {
+    return decodeLiteral64Constant();
+  }
+
   switch (Width) {
   case 32:
   case 16:

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.h

@@ -179,6 +179,7 @@ class AMDGPUDisassembler : public MCDisassembler {
 
   MCOperand decodeMandatoryLiteralConstant(unsigned Imm) const;
   MCOperand decodeLiteralConstant(bool ExtendFP64) const;
+  MCOperand decodeLiteral64Constant() const;
 
   MCOperand decodeSrcOp(unsigned Width, unsigned Val) const;
 

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -231,6 +231,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   bool HasSALUFloatInsts = false;
   bool HasPseudoScalarTrans = false;
   bool HasRestrictedSOffset = false;
+  bool Has64BitLiterals = false;
   bool HasBitOp3Insts = false;
   bool HasTransposeLoadF4F6Insts = false;
   bool HasPrngInst = false;
@@ -1384,6 +1385,9 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   /// GFX1250.
   bool hasWaitXCnt() const { return HasWaitXcnt; }
 
+  // A single DWORD instructions can use a 64-bit literal.
+  bool has64BitLiterals() const { return Has64BitLiterals; }
+
   bool hasPointSampleAccel() const { return HasPointSampleAccel; }
 
   bool hasLdsBarrierArriveAtomic() const { return HasLdsBarrierArriveAtomic; }