[NVPTX] Disable parens for identifiers starting with '$'

[asavonic]

Summary

ptxas fails to parse such syntax:

mov.u64 %rd1, ($str);
fatal   : Parsing error near '$str': syntax error

A new MCAsmInfo option was added because InParens parameter of
MCExpr::print is not sufficient to disable parens
completely. MCExpr::print resets it to false for a recursive call in
case of unary or binary expressions.

Targets that require parens around identifiers that start with '$'
should always pass MCAsmInfo to MCExpr::print.
Therefore 'operator<<(raw_ostream &, MCExpr&)' should be avoided
because it calls MCExpr::print with nullptr MAI.

[asavonic]

This is the same issue as in D119669:

In D119669#3320716, @tra wrote:

Looks like it broke CUDA:
https://lab.llvm.org/buildbot/#/builders/46/builds/23372/steps/3/logs/stdio

ptxas /tmp/complex-78abf1/complex-sm_75.s, line 250; fatal   : Parsing error near '$L__BB6_2': syntax error

[asavonic]

• Rebased and added patch context.

[tra]

LGTM.

llvm/lib/MC/MCExpr.cpp

79–82

Nit: LLVM code style typically omits braces around single statement bodies.
It may be even worth folding it all into:

bool UseParens = MAI && MAI->useParensForDollarSignNames() && !(InParens || Sym.getName().empty() || Sym.getName()[0] != '$');

llvm/test/CodeGen/NVPTX/no-extra-parens.ll

12

For the record, right now LLVM generates ($str) and that's what ptxas complains about. https://cuda.godbolt.org/z/s7ahYn5oM

[tra]

What happened this time? AFAICT CUDA bots were fine with the change.

[asavonic]

In D123702#3452464, @tra wrote:

What happened this time? AFAICT CUDA bots were fine with the change.

The patch failed on Mips32:
https://lab.llvm.org/buildbot#builders/109/builds/36628

# CHECK: # fixup A - offset: 0, value: ($tmp0), kind: fixup_Mips_26
<stdin>:580:2: note: possible intended match here
# fixup A - offset: 0, value: $tmp0, kind: fixup_Mips_26

It is probably because MAI is missing for some reason..

[asavonic]

• Fixed issues found in Mips LIT tests: MCAsmStreamer had one call to 'operator<<' function which doesn't pass MAI to MCExpr::print.

[tra]

llvm/lib/MC/MCAsmStreamer.cpp

2256

This is awkward.
Perhaps we should consider adding some sort of withMAI(F.getValue(), MAI) -> std::pair<MCExpr*, MCAsmInfo*> wrapper and an overloaded operator<< to catch it and pass MAI along to MCExpr::print().

[asavonic]

llvm/lib/MC/MCAsmStreamer.cpp

2256

Well, this is how MCExpr is printed for assembly output - there are more than 50 occurrences in this file (also in MCTargetDesc). I personally find these two forms equally readable, so it's probably not worth it to refactor everything...

[tra]

llvm/lib/MC/MCAsmStreamer.cpp

2256

Up to you. I would find having to break OS << A << B .... chain for some values, but it's indeed a cosmetic change. It could be done as a clean-up in a separate patch.

llvm/include/llvm/MC/MCAsmInfo.h

protected:
   /// True if dwarf register numbers are printed instead of symbolic register
   /// names in .cfi_* directives.  Defaults to false.
   bool DwarfRegNumForCFI = false;
 
   /// True if target uses parens to indicate the symbol variant instead of @.
   /// For example, foo(plt) instead of foo@plt.  Defaults to false.
   bool UseParensForSymbolVariant = false;
 
+  /// True if the target uses parens for symbol names starting with
+  /// '$' character to distinguish them from absolute names.
+  bool UseParensForDollarSignNames = true;
+
   /// True if the target supports flags in ".loc" directive, false if only
   /// location is allowed.
   bool SupportsExtendedDwarfLocDirective = true;
 
   //===--- Prologue State ----------------------------------------------===//
 
   std::vector<MCCFIInstruction> InitialFrameState;
 
public:
 
   bool doesDwarfUseRelocationsAcrossSections() const {
     return DwarfUsesRelocationsAcrossSections;
   }
 
   bool doDwarfFDESymbolsUseAbsDiff() const { return DwarfFDESymbolsUseAbsDiff; }
   bool useDwarfRegNumForCFI() const { return DwarfRegNumForCFI; }
   bool useParensForSymbolVariant() const { return UseParensForSymbolVariant; }
+  bool useParensForDollarSignNames() const {
+    return UseParensForDollarSignNames;
+  }
   bool supportsExtendedDwarfLocDirective() const {
     return SupportsExtendedDwarfLocDirective;
   }
 
   bool usesDwarfFileAndLocDirectives() const {
     return UsesDwarfFileAndLocDirectives;
   }
 

llvm/lib/MC/MCAsmStreamer.cpp

for (unsigned i = 0, e = Code.size(); i != e; ++i) {
     }
   }
   OS << "]\n";
 
   for (unsigned i = 0, e = Fixups.size(); i != e; ++i) {
     MCFixup &F = Fixups[i];
     const MCFixupKindInfo &Info =
         getAssembler().getBackend().getFixupKindInfo(F.getKind());
-    OS << "  fixup " << char('A' + i) << " - " << "offset: " << F.getOffset()
-       << ", value: " << *F.getValue() << ", kind: " << Info.Name << "\n";
+    OS << "  fixup " << char('A' + i) << " - "
+       << "offset: " << F.getOffset() << ", value: ";
+    F.getValue()->print(OS, MAI);
+    OS << ", kind: " << Info.Name << "\n";
   }
 }
 
 void MCAsmStreamer::emitInstruction(const MCInst &Inst,
                                     const MCSubtargetInfo &STI) {
   assert(getCurrentSectionOnly() &&
          "Cannot emit contents before setting section!");
 

llvm/lib/MC/MCExpr.cpp

else
       OS << Value;
     return;
   }
   case MCExpr::SymbolRef: {
     const MCSymbolRefExpr &SRE = cast<MCSymbolRefExpr>(*this);
     const MCSymbol &Sym = SRE.getSymbol();
     // Parenthesize names that start with $ so that they don't look like
     // absolute names.
-    bool UseParens =
-        !InParens && !Sym.getName().empty() && Sym.getName()[0] == '$';
+    bool UseParens = MAI && MAI->useParensForDollarSignNames() && !InParens &&
+                     !Sym.getName().empty() && Sym.getName()[0] == '$';
+
     if (UseParens) {
       OS << '(';
       Sym.print(OS, MAI);
       OS << ')';
     } else
       Sym.print(OS, MAI);
 
     const MCSymbolRefExpr::VariantKind Kind = SRE.getKind();
     if (Kind != MCSymbolRefExpr::VK_None) {
       if (MAI && MAI->useParensForSymbolVariant()) // ARM

llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXMCAsmInfo.cpp

NVPTXMCAsmInfo::NVPTXMCAsmInfo(const Triple &TheTriple,
   SupportsExtendedDwarfLocDirective = false;
   SupportsSignedData = false;
 
   // @TODO: Can we just disable this?
   WeakDirective = "\t// .weak\t";
   GlobalDirective = "\t// .globl\t";
 
   UseIntegratedAssembler = false;
+
+  // Avoid using parens for identifiers starting with $ - ptxas does
+  // not expect them.
+  UseParensForDollarSignNames = false;
 }

llvm/test/CodeGen/NVPTX/no-extra-parens.ll

+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s
+
+; ptxas has no special meaning for '$' character, so it should be used
+; without parens.
+
+@"$str" = private addrspace(1) constant [4 x i8] c"str\00"
+
+declare void @str2(i8* %str)
+define void @str1() {
+entry:
+;; CHECK: mov.u64 %rd{{[0-9]+}}, $str;
+  tail call void @str2(i8* getelementptr ([4 x i8], [4 x i8]* addrspacecast ([4 x i8] addrspace(1)* @"$str" to [4 x i8]*), i64 0, i64 0))
+  ret void
+}