[LLVMdev] proposal to simplify isel/asmprinter interaction with globals

Chris Lattner clattner at apple.com
Sat Jun 20 19:30:16 PDT 2009

Hi All,

I'm working on various cleanups and simplifications to the  
asmprinters.  One thing that is driving me nuts is that the  
asmprinters currently "reverse engineer" a lot of information when  
printing an operand that isel had when it created it.

I'm specifically talking about all the suffixes generated by isel,  
like $non_lazy_ptr, @TLSGD, @NTPOFF, (%rip) etc.  These are all  
"magic" things that the assembler uses (for example, it generates a  
reference to a global's GOT entry instead of to the global itself).

The thing that is really frustrating to me is that the asmprinters  
need to reverse engineer what isel *meant* in order to output the  
right thing.  For example, to determine whether to emit $non_lazy_ptr,  
it uses a logic tree like this:

     if (shouldPrintStub(TM, Subtarget)) {
       // Link-once, declaration, or Weakly-linked global variables need
       // non-lazily-resolved stubs
       if (GV->isDeclaration() || GV->isWeakForLinker()) {
         // Dynamically-resolved functions need a stub for the function.
         if (GV->hasHiddenVisibility()) {
           if (!GV->isDeclaration() && !GV->hasCommonLinkage())
           else {
             printSuffixedName(Name, "$non_lazy_ptr");
         } else {
           printSuffixedName(Name, "$non_lazy_ptr");

  where shouldPrintStub is:

static inline bool shouldPrintStub(TargetMachine &TM, const  
X86Subtarget* ST) {
   return ST->isPICStyleStub() && TM.getRelocationModel() !=  

This is really redundant with isel, because isel also needs to know  
exactly which GV references go through a stub so that it isels the  
access correctly.

My proposed fix for this is to add an 'unsigned char' slot to  
MachineOperand that holds a target-specific enum value.  The code in  
asmprinter would be reduced to:

switch (theoperandenum) {
case X86::MO_Flag_non_lazy_ptr:
   O << "$non_lazy_ptr";
case X86::MO_Flag_TLSGD:
   O << "@TLSGD";
case X86::MO_Flag_NTPOFF:
   O << "@NTPOFF";

etc.  The possible set of suffixes and modifiers are all target- 
specific, so the main code generator would just pass them through (as  
it does now).

Does anyone have any objections to this?


More information about the llvm-dev mailing list