On Mon, Mar 11, 2013 at 12:13 AM, Steven Bosscher <stevenb....@gmail.com> wrote: > On Sun, Mar 10, 2013 at 22:54, Matt Davis wrote: >> I have a particular instance of a >> function call within a function that I am analyzing (and >> transforming). I want the address of that function call, for which I >> do have a gimple_call instance of. I want to somehow create a symbol >> identifying that particular address/location for that instance of the >> function call. I need this address at runtime. I was inserting labels >> just after each function call, but the placement of my label was not >> always immediately after the function call, even if I set the label as >> being addressable and/or volatile. > > You have the GIMPLE_CALL gimple statement, and you can use > gsi_for_stmt to get an insertion point for the labels. But at this > point in the compilation chain, you can't expect this label to be the > exact call site, e.g. because arguments need to be set up, registers > saved/restored around the call, etc. Also, a label will count as a > basic block split point, so your call and your label may end up being > separated by a jump when basic blocks are re-ordered.
Yep! That was what I was seeing. The return value results/registers being adjusted immediately after the function call, resulting in my label getting moved. > If you want to get the address of the actual call site at runtime, > you're going to have to put the label before the call insn in the > "final" pass. Perhaps you could do that by inserting > NOTE_INSN_DELETED_LABEL notes before the call insn, or by modifying > the assembly output routines for calls to put a label just before the > call. But even this may not work reliable if a call output pattern > emits multiple assembly instructions for a single call RTL > instruction. Thanks for this suggestion! -Matt
