[LLVMbugs] [Bug 761] NEW: TargetData needs to be parameterized better in the .bc/.ll files

bugzilla-daemon at cs.uiuc.edu bugzilla-daemon at cs.uiuc.edu
Mon May 1 21:01:56 PDT 2006


http://llvm.org/bugs/show_bug.cgi?id=761

           Summary: TargetData needs to be parameterized better in the
                    .bc/.ll files
           Product: libraries
           Version: 1.0
          Platform: All
        OS/Version: All
            Status: NEW
          Severity: minor
          Priority: P2
         Component: Target Description Classes
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: sabre at nondot.org


Currently, when 'opt' is run, it configures and adds a TargetData to the pass manager based on the 
settings of the "target" flags in the Module.  These flags indicate whether the target is 32/64 bit, and 
whether it is big or little endian.

This has two problems:

1. If the module has no target info, we default to SparcV9 (this is Bug 760).
2. Our mapping from ptrsize/endianness to a full targetdata is a total hack.

#2 is true because a TargetData has many more fields than just ptrsize and endianness, thus the 
mapping from these two fields to a full TargetData is inherently a few-to-many mapping, which is 
impossible to be right in the general case (we have it carefully rigged to work for our current targets 
though).

Instead of encoding ptrsize/endianness into the module, it would be far better to encode a whole target 
data into it.  I propose that we do this defining a "default" target data, then encoding any diffs from this 
into the module as a string.  For example, if the default was a 64-bit, big endian, target with 8-byte 
aligned doubles, a target like X86 could be described as "p3eLd4".

As TargetData is extended in the future, more letters could be gracefully added without breaking 
old .bc files.

-Chris



------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.




More information about the llvm-bugs mailing list