[cfe-dev] Possible problem with debug metadata generated for empty 'return' statements

[Eli Bendersky]

Hello,

Consider this C code:

    void func(char c, char* d)
    {
        *d = c + 1;
        return;
    }

The LLVM IR generated with "clang -g -S -O0 -emit-llvm" is:

    define void @func(i8 signext %c, i8* %d) nounwind {
    entry:
      %c.addr = alloca i8, align 1
      %d.addr = alloca i8*, align 4
      store i8 %c, i8* %c.addr, align 1
      call void @llvm.dbg.declare(metadata !{i8* %c.addr}, metadata !11),
!dbg !13
      store i8* %d, i8** %d.addr, align 4
      call void @llvm.dbg.declare(metadata !{i8** %d.addr}, metadata !14),
!dbg !16
      %tmp = load i8* %c.addr, align 1, !dbg !17
      %conv = sext i8 %tmp to i32, !dbg !17
      %add = add nsw i32 %conv, 1, !dbg !17
      %conv1 = trunc i32 %add to i8, !dbg !17
      %tmp2 = load i8** %d.addr, align 4, !dbg !17
      store i8 %conv1, i8* %tmp2, !dbg !17
      ret void, !dbg !19
    }

Relevant metadata nodes:

    !17 = metadata !{i32 3, i32 5, metadata !18, null}
    !19 = metadata !{i32 5, i32 1, metadata !18, null}

As you can see, the metadata node attached to the 'ret void' instruction
reports line 5, although the 'return' is on line 4. This is an error, which
will cause a breakpoint placed on the 'return' not to stop there.
Some preliminary investigation in the source code of Clang reveals that in
all likeness Clang optimizes away the empty "return" at the end of the
function, using its own return inserted by
CodeGenFunction::EmitFunctionEpilog

Examination of emitted AST shows that the return statement appears there, so
the optimization is done after that.

Eli
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20110901/03a96464/attachment.html>