<html>
  <head>
    <meta content="text/html; charset=UTF-8" http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    <div class="moz-cite-prefix">On 1/31/14 5:23 PM, Nick Lewycky wrote:<br>
    </div>
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">On 30 January 2014 09:55, Philip Reames <span
          dir="ltr"><<a moz-do-not-send="true"
            href="mailto:listmail@philipreames.com" target="_blank">listmail@philipreames.com</a>></span>
        wrote:<br>
        <div class="gmail_extra">
          <div class="gmail_quote">
            <blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">
              <div class="im">On 1/29/14 3:40 PM, Nick Lewycky wrote:<br>
                <blockquote class="gmail_quote" style="margin:0px 0px
                  0px
0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">The
                  LLVM Module has an optional target triple and target
                  datalayout. Without them, an llvm::DataLayout can't be
                  constructed with meaningful data. The benefit to
                  making them optional is to permit optimization that
                  would work across all possible DataLayouts, then allow
                  us to commit to a particular one at a later point in
                  time, thereby performing more optimization in advance.<br>
                  <br>
                  This feature is not being used. Instead, every user of
                  LLVM IR in a portability system defines one or more
                  standardized datalayouts for their platform, and shims
                  to place calls with the outside world. The primary
                  reason for this is that independence from DataLayout
                  is not sufficient to achieve portability because it
                  doesn't also represent ABI lowering constraints. If
                  you have a system that attempts to use LLVM IR in a
                  portable fashion and does it without standardizing on
                  a datalayout, please share your experience.<br>
                </blockquote>
              </div>
              Nick, I don't have a current system in place, but I do
              want to put forward an alternate perspective.<br>
              <br>
              We've been looking at doing late insertion of safepoints
              for garbage collection.  One of the properties that we end
              up needing to preserve through all the optimizations which
              precede our custom rewriting phase is that the optimizer
              has not chosen to "hide" pointers from us by using
              ptrtoint and integer math tricks. Currently, we're simply
              running a verification pass before our rewrite, but I'm
              very interested long term in constructing ways to ensure a
              "gc safe" set of optimization passes.<br>
            </blockquote>
            <div><br>
            </div>
            <div>
              <div>As a general rule passes need to support the whole of
                what the IR can support. Trying to operate on a subset
                of IR seems like a losing battle, unless you can show a
                mapping from one to the other (ie., using code
                duplication to remove all unnatural loops from IR, or
                collapsing a function to having a single exit node).</div>
            </div>
            <div><br>
            </div>
            <div>What language were you planning to do this for? Does
              the language permit the user to convert pointers to
              integers and vice versa? If so, what do you do if the user
              program writes a pointer out to a file, reads it back in
              later, and uses it?</div>
          </div>
        </div>
      </div>
    </blockquote>
    Java - which does not permit arbitrary pointer manipulation.  (Well,
    without resorting to mechanism like JNI and sun.misc.Unsafe.  Doing
    so would be explicitly undefined behavior though.)  We also use raw
    pointer manipulations in our implementation (which is eventually
    inlined), but this happens after the safepoint insertion rewrite.<br>
    <br>
    We strictly control the input IR.  As a result, I can insure that
    the initial IR meets our subset requirements.  In practice, all of
    the opto passes appear to preserve these invariants (i.e. not
    introducing inttoptr), but we'd like to justify that a bit more.  <br>
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div><br>
            </div>
            <blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">One
              of the ways I've been thinking about - but haven't
              actually implemented yet - is to deny the optimization
              passes information about pointer sizing.</blockquote>
            <div><br>
            </div>
            <div>Right, pointer size (address space size) will become
              known to all parts of the compiler. It's not even going to
              be just the optimizations, ConstantExpr::get is going to
              grow smarter because of this, as
              lib/Analysis/ConstantFolding.cpp merges into
              lib/IR/ConstantFold.cpp. That is one of the major benefits
              that's driving this. (All parts of the compiler will also
              know endian-ness, which means we can constant fold loads,
              too.)</div>
          </div>
        </div>
      </div>
    </blockquote>
    I would argue that all of the pieces you mentioned are performing
    optimizations.  :)  However, the exact semantics are unimportant for
    the overall discussion.  <br>
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div><br>
            </div>
            <blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">Under
              the assumption that an opto pass can't insert an ptrtoint
              cast without knowing a safe integer size to use, this
              seems like it would outlaw a class of optimizations we'd
              be broken by.<br>
            </blockquote>
            <div><br>
            </div>
            <div>Optimization passes generally prefer converting
              ptrtoint and inttoptr to GEPs whenever possible. </div>
          </div>
        </div>
      </div>
    </blockquote>
    This is good to hear and helps us.<br>
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div>I expect that we'll end up with *fewer* ptr<->int
              conversions with this change, because we'll know enough
              about the target to convert them into GEPs.</div>
          </div>
        </div>
      </div>
    </blockquote>
    Er, I'm confused by this.  Why would not knowing the size of a
    pointer case a GEP to be converted to a ptr <-> int
    conversion?  <br>
    <br>
    Or do you mean that after the change conversions in the original
    input IR are more likely to be recognized?<br>
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div><br>
            </div>
            <blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">My
              understanding is that the only current way to do this
              would be to not specify a DataLayout.  (And hack a few
              places with built in assumptions.  Let's ignore that for
              the moment.)  With your proposed change, would there be a
              clean way to express something like this?<br>
            </blockquote>
            <div><br>
            </div>
            <div>I think your GC placement algorithm needs to handle
              inttoptr and ptrtoint, whichever way this discussion goes.
              Sorry. I'd be happy to hear others chime in -- I know I'm
              not an expert in this area or about GCs -- but I don't
              find this rationale compelling.</div>
          </div>
        </div>
      </div>
    </blockquote>
    The key assumption I didn't initially explain is that the initial IR
    couldn't contain conversions.  With that added, do you still see
    concerns?  I'm fairly sure I don't need to handle general ptr
    <-> int conversions.  If I'm wrong, I'd really like to know
    it. 
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div><br>
            </div>
            <blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">p.s.
              From reading the mailing list a while back, I suspect that
              the SPIR folks might have similar needs.  (i.e. hiding
              pointer sizes, etc..)  Pure speculation on my part though.<br>
            </blockquote>
            <div><br>
            </div>
            <div>The SPIR spec specifies two target datalayouts, one for
              32 bits and one for 64 bits.</div>
          </div>
        </div>
      </div>
    </blockquote>
    Good to know.  Thanks.<br>
    <blockquote
cite="mid:CADbEz-jg7-UdWK77G5+u5n4ninz-H1wFpZGyeqv-GWr1hK-DUg@mail.gmail.com"
      type="cite">
      <div dir="ltr">
        <div class="gmail_extra">
          <div class="gmail_quote">
            <div><br>
            </div>
            <div>Nick</div>
            <div><br>
            </div>
          </div>
        </div>
      </div>
    </blockquote>
    Philip<br>
  </body>
</html>