[AMDGPU, AsmParser] Enable the mnemonic spell corrector.

[Quolyk]

Summary

Related: D40552, D33128

[arsenm]

LGTM

[sdardis]

test/MC/AMDGPU/invalid-instructions-spellcheck.s

1

You can drop -show-encoding from this as you're not testing the encoding.

[Quolyk]

Deleted -show-encoding from test

[Quolyk]

BTW, I don't have commit access.

[Quolyk]

@arsenm please commit

[arsenm]

There is a test failure with this. This seems to do the wrong thing when there are subtarget differences with the instruction.

Failing test test/MC/AMDGPU/vop1-gfx9-err.s

[arsenm]

I fixed the test, but the errors are still not great. They weren't great before, so it is only tangentially related.

[arsenm]

r321202

lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

public:
 
   SMLoc getStartLoc() const override {
     return StartLoc;
   }
 
   SMLoc getEndLoc() const override {
     return EndLoc;
   }
   
+  SMRange getLocRange() const { 
+    return SMRange(StartLoc, EndLoc); 
+  }
+
   Modifiers getModifiers() const {
     assert(isRegKind() || isImmTy(ImmTyNone));
     return isRegKind() ? Reg.Mods : Imm.Mods;
   }
 
   void setModifiers(Modifiers Mods) {
     assert(isRegKind() || isImmTy(ImmTyNone));
     if (isRegKind())
if (!validateIntClampSupported(Inst)) {
     Error(IDLoc,
       "integer clamping is not supported on this GPU");
     return false;
   }
 
   return true;
 }
 
+static std::string AMDGPUMnemonicSpellCheck(StringRef S, uint64_t FBS,
+                                            unsigned VariantID = 0);
+        
 bool AMDGPUAsmParser::MatchAndEmitInstruction(SMLoc IDLoc, unsigned &Opcode,
                                               OperandVector &Operands,
                                               MCStreamer &Out,
                                               uint64_t &ErrorInfo,
                                               bool MatchingInlineAsm) {
   MCInst Inst;
   unsigned Result = Match_Success;
   for (auto Variant : getMatchedVariants()) {
case Match_Success:
     }
     Inst.setLoc(IDLoc);
     Out.EmitInstruction(Inst, getSTI());
     return false;
 
   case Match_MissingFeature:
     return Error(IDLoc, "instruction not supported on this GPU");
 
-  case Match_MnemonicFail:
-    return Error(IDLoc, "unrecognized instruction mnemonic");
+  case Match_MnemonicFail: {
+    uint64_t FBS = ComputeAvailableFeatures(getSTI().getFeatureBits());
+    std::string Suggestion = AMDGPUMnemonicSpellCheck(
+        ((AMDGPUOperand &)*Operands[0]).getToken(), FBS);
+    return Error(IDLoc, "invalid instruction" + Suggestion,
+                 ((AMDGPUOperand &)*Operands[0]).getLocRange());
+  }
 
   case Match_InvalidOperand: {
     SMLoc ErrorLoc = IDLoc;
     if (ErrorInfo != ~0ULL) {
       if (ErrorInfo >= Operands.size()) {
         return Error(IDLoc, "too few operands for instruction");
       }
       ErrorLoc = ((AMDGPUOperand &)*Operands[ErrorInfo]).getStartLoc();

 /// Force static initialization.
 extern "C" void LLVMInitializeAMDGPUAsmParser() {
   RegisterMCAsmParser<AMDGPUAsmParser> A(getTheAMDGPUTarget());
   RegisterMCAsmParser<AMDGPUAsmParser> B(getTheGCNTarget());
 }
 
 #define GET_REGISTER_MATCHER
 #define GET_MATCHER_IMPLEMENTATION
+#define GET_MNEMONIC_SPELL_CHECKER
 #include "AMDGPUGenAsmMatcher.inc"
 
 // This fuction should be defined after auto-generated include so that we have
 // MatchClassKind enum defined
 unsigned AMDGPUAsmParser::validateTargetOperandClass(MCParsedAsmOperand &Op,
                                                      unsigned Kind) {
   // Tokens like "glc" would be parsed as immediate operands in ParseOperand().
   // But MatchInstructionImpl() expects to meet token and fails to validate

test/MC/AMDGPU/invalid-instructions-spellcheck.s

+# RUN: not llvm-mc -triple amdgcn -show-encoding < %s 2>&1 | FileCheck %s
+
+# This tests the mnemonic spell checker.
+
+# First check what happens when an instruction is omitted:
+
+v2, v4, v6
+
+# CHECK:      unknown token in expression
+# CHECK-NEXT: v2, v4, v6
+# CHECK-NEXT:   ^
+
+# CHECK:      error: not a valid operand.
+# CHECK-NEXT: v2, v4, v6
+# CHECK-NEXT:     ^
+
+# We don't want to see a suggestion here; the edit distance is too large to
+# give sensible suggestions:
+
+aaaaaaaaaaaaaaa v1, v2, v3
+
+# CHECK:      error: invalid instruction
+# CHECK-NEXT: aaaaaaaaaaaaaaa v1, v2, v3
+# CHECK-NEXT: ^
+
+# Check that we get one suggestion: 'dsc_write_src2_b64' is 1 edit away, i.e. an deletion.
+
+dsc_write_src2_b64 r1, r2, r3
+
+# CHECK:      error: invalid instruction, did you mean: ds_write_src2_b64?
+# CHECK-NEXT: dsc_write_src2_b64 r1, r2, r3
+# CHECK-NEXT: ^
+
+# Check edit distance 1 and 2, just insertions:
+
+s_mov_b v1, v2
+
+# CHECK:      error: invalid instruction, did you mean: s_mov_b32, s_mov_b64?
+# CHECK-NEXT: s_mov_b v1, v2
+# CHECK-NEXT: ^
+
+# Check an instruction that is 2 edits away, and also has a lot of candidates:
+
+s_load_dwordx v1, v2, v3
+
+# CHECK:      error: invalid instruction, did you mean: s_load_dword, s_load_dwordx16, s_load_dwordx2, s_load_dwordx4, s_load_dwordx8?
+# CHECK-NEXT: s_load_dwordx v1, v2, v3
+# CHECK-NEXT: ^