[AMDGPU][MC] Fix for Bug 30829 + LIT tests

Added code to check constant bus restrictions for VOP formats (only one SGPR value or literal-constant may be used by the instruction).
Note that the same checks are performed by SIInstrInfo::verifyInstruction (used by lowering code).
Added LIT tests.

llvm-svn: 296873

7 files changed, 163 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 77bf0a618bb..ea76043eb5f 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -980,6 +980,12 @@ private:
   void errorExpTgt();
   OperandMatchResultTy parseExpTgtImpl(StringRef Str, uint8_t &Val);
 
+  bool validateOperandLimitations(const MCInst &Inst);
+  bool usesConstantBus(const MCInst &Inst, unsigned OpIdx);
+  bool isInlineConstant(const MCInst &Inst, unsigned OpIdx) const;
+  unsigned findImplicitSGPRReadInVOP(const MCInst &Inst) const;
+  bool isSGPR(unsigned Reg);
+
 public:
   OperandMatchResultTy parseOptionalOperand(OperandVector &Operands);
 
@@ -1886,6 +1892,128 @@ ArrayRef<unsigned> AMDGPUAsmParser::getMatchedVariants() const {
   return makeArrayRef(Variants);
 }
 
+unsigned AMDGPUAsmParser::findImplicitSGPRReadInVOP(const MCInst &Inst) const {
+  const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
+  const unsigned Num = Desc.getNumImplicitUses();
+  for (unsigned i = 0; i < Num; ++i) {
+    unsigned Reg = Desc.ImplicitUses[i];
+    switch (Reg) {
+    case AMDGPU::FLAT_SCR:
+    case AMDGPU::VCC:
+    case AMDGPU::M0:
+      return Reg;
+    default:
+      break;
+    }
+  }
+  return AMDGPU::NoRegister;
+}
+
+bool AMDGPUAsmParser::isSGPR(unsigned Reg) {
+  const MCRegisterInfo *TRI = getContext().getRegisterInfo();
+  const MCRegisterClass SGPRClass = TRI->getRegClass(AMDGPU::SReg_32RegClassID);
+  const unsigned FirstSubReg = TRI->getSubReg(Reg, 1);
+  return SGPRClass.contains(FirstSubReg != 0 ? FirstSubReg : Reg) ||
+         Reg == AMDGPU::SCC;
+}
+
+// NB: This code is correct only when used to check constant
+// bus limitations because GFX7 support no f16 inline constants.
+// Note that there are no cases when a GFX7 opcode violates
+// constant bus limitations due to the use of an f16 constant.
+bool AMDGPUAsmParser::isInlineConstant(const MCInst &Inst,
+                                       unsigned OpIdx) const {
+  const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
+
+  if (!AMDGPU::isSISrcOperand(Desc, OpIdx)) {
+    return false;
+  }
+
+  const MCOperand &MO = Inst.getOperand(OpIdx);
+
+  int64_t Val = MO.getImm();
+  auto OpSize = AMDGPU::getOperandSize(Desc, OpIdx);
+
+  switch (OpSize) { // expected operand size
+  case 8:
+    return AMDGPU::isInlinableLiteral64(Val, hasInv2PiInlineImm());
+  case 4:
+    return AMDGPU::isInlinableLiteral32(Val, hasInv2PiInlineImm());
+  case 2: {
+    const unsigned OperandType = Desc.OpInfo[OpIdx].OperandType;
+    if (OperandType == AMDGPU::OPERAND_REG_INLINE_C_V2INT16 ||
+        OperandType == AMDGPU::OPERAND_REG_INLINE_C_V2FP16) {
+      return AMDGPU::isInlinableLiteralV216(Val, hasInv2PiInlineImm());
+    } else {
+      return AMDGPU::isInlinableLiteral16(Val, hasInv2PiInlineImm());
+    }
+  }
+  default:
+    llvm_unreachable("invalid operand size");
+  }
+}
+
+bool AMDGPUAsmParser::usesConstantBus(const MCInst &Inst, unsigned OpIdx) {
+  const MCOperand &MO = Inst.getOperand(OpIdx);
+  if (MO.isImm()) {
+    return !isInlineConstant(Inst, OpIdx);
+  }
+  return !MO.isReg() || isSGPR(mc2PseudoReg(MO.getReg()));
+}
+
+bool AMDGPUAsmParser::validateOperandLimitations(const MCInst &Inst) {
+  const unsigned Opcode = Inst.getOpcode();
+  const MCInstrDesc &Desc = MII.get(Opcode);
+  unsigned ConstantBusUseCount = 0;
+
+  if (Desc.TSFlags &
+      (SIInstrFlags::VOPC |
+       SIInstrFlags::VOP1 | SIInstrFlags::VOP2 |
+       SIInstrFlags::VOP3 | SIInstrFlags::VOP3P)) {
+
+    // Check special imm operands (used by madmk, etc)
+    if (AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::imm) != -1) {
+      ++ConstantBusUseCount;
+    }
+
+    unsigned SGPRUsed = findImplicitSGPRReadInVOP(Inst);
+    if (SGPRUsed != AMDGPU::NoRegister) {
+      ++ConstantBusUseCount;
+    }
+
+    const int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
+    const int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
+    const int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
+
+    const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx };
+
+    for (int OpIdx : OpIndices) {
+      if (OpIdx == -1) break;
+
+      const MCOperand &MO = Inst.getOperand(OpIdx);
+      if (usesConstantBus(Inst, OpIdx)) {
+        if (MO.isReg()) {
+          const unsigned Reg = mc2PseudoReg(MO.getReg());
+          // Pairs of registers with a partial intersections like these
+          //   s0, s[0:1]
+          //   flat_scratch_lo, flat_scratch
+          //   flat_scratch_lo, flat_scratch_hi
+          // are theoretically valid but they are disabled anyway.
+          // Note that this code mimics SIInstrInfo::verifyInstruction
+          if (Reg != SGPRUsed) {
+            ++ConstantBusUseCount;
+          }
+          SGPRUsed = Reg;
+        } else { // Expression or a literal
+          ++ConstantBusUseCount;
+        }
+      }
+    }
+  }
+
+  return ConstantBusUseCount <= 1;
+}
+
 bool AMDGPUAsmParser::MatchAndEmitInstruction(SMLoc IDLoc, unsigned &Opcode,
                                               OperandVector &Operands,
                                               MCStreamer &Out,
@@ -1918,6 +2046,10 @@ bool AMDGPUAsmParser::MatchAndEmitInstruction(SMLoc IDLoc, unsigned &Opcode,
   switch (Result) {
   default: break;
   case Match_Success:
+    if (!validateOperandLimitations(Inst)) {
+      return Error(IDLoc,
+                   "invalid operand (violates constant bus restrictions)");
+    }
     Inst.setLoc(IDLoc);
     Out.EmitInstruction(Inst, getSTI());
     return false;
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index cdcc0f08177..de0fda4be6f 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -547,6 +547,25 @@ unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI) {
   return Reg;
 }
 
+unsigned mc2PseudoReg(unsigned Reg) {
+  switch (Reg) {
+  case AMDGPU::FLAT_SCR_ci:
+  case AMDGPU::FLAT_SCR_vi:
+    return FLAT_SCR;
+
+  case AMDGPU::FLAT_SCR_LO_ci:
+  case AMDGPU::FLAT_SCR_LO_vi:
+    return AMDGPU::FLAT_SCR_LO;
+
+  case AMDGPU::FLAT_SCR_HI_ci:
+  case AMDGPU::FLAT_SCR_HI_vi:
+    return AMDGPU::FLAT_SCR_HI;
+
+  default:
+    return Reg;
+  }
+}
+
 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo) {
   assert(OpNo < Desc.NumOperands);
   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index c20a2093755..96171562ebe 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -263,6 +263,10 @@ bool isVI(const MCSubtargetInfo &STI);
 /// \p STI otherwise return \p Reg.
 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI);
 
+/// \brief Convert hardware register \p Reg to a pseudo register
+LLVM_READNONE
+unsigned mc2PseudoReg(unsigned Reg);
+
 /// \brief Can this operand also contain immediate values?
 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo);
 
diff --git a/llvm/lib/Target/AMDGPU/VOP1Instructions.td b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
index ff2f9338407..dcef11ee7bf 100644
--- a/llvm/lib/Target/AMDGPU/VOP1Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
@@ -75,6 +75,8 @@ class VOP1_Real <VOP1_Pseudo ps, int EncodingFamily> :
   let Constraints        = ps.Constraints;
   let DisableEncoding    = ps.DisableEncoding;
   let TSFlags            = ps.TSFlags;
+  let UseNamedOperandTable = ps.UseNamedOperandTable;
+  let Uses                 = ps.Uses;
 }
 
 class VOP1_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index bfd73d44b42..4476fd56763 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -93,6 +93,8 @@ class VOP2_Real <VOP2_Pseudo ps, int EncodingFamily> :
   let Constraints        = ps.Constraints;
   let DisableEncoding    = ps.DisableEncoding;
   let TSFlags            = ps.TSFlags;
+  let UseNamedOperandTable = ps.UseNamedOperandTable;
+  let Uses                 = ps.Uses;
 }
 
 class VOP2_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
diff --git a/llvm/lib/Target/AMDGPU/VOPCInstructions.td b/llvm/lib/Target/AMDGPU/VOPCInstructions.td
index bb05fb7bae7..c64a18aaee0 100644
--- a/llvm/lib/Target/AMDGPU/VOPCInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOPCInstructions.td
@@ -93,6 +93,8 @@ class VOPC_Real <VOPC_Pseudo ps, int EncodingFamily> :
   let Constraints        = ps.Constraints;
   let DisableEncoding    = ps.DisableEncoding;
   let TSFlags            = ps.TSFlags;
+  let UseNamedOperandTable = ps.UseNamedOperandTable;
+  let Uses                 = ps.Uses;
 }
 
 class VOPC_SDWA_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
diff --git a/llvm/lib/Target/AMDGPU/VOPInstructions.td b/llvm/lib/Target/AMDGPU/VOPInstructions.td
index acb5b066896..cc4ad934465 100644
--- a/llvm/lib/Target/AMDGPU/VOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOPInstructions.td
@@ -136,6 +136,8 @@ class VOP3_Real <VOP3_Pseudo ps, int EncodingFamily> :
   let Constraints        = ps.Constraints;
   let DisableEncoding    = ps.DisableEncoding;
   let TSFlags            = ps.TSFlags;
+  let UseNamedOperandTable = ps.UseNamedOperandTable;
+  let Uses                 = ps.Uses;
 }
 
 // XXX - Is there any reason to distingusih this from regular VOP3