[LLVMdev] [LV] possible `vector.memcheck` regression when using `llvm.loop` and `llvm.mem.parallel_loop_access`

Adam Nemet anemet at apple.com
Thu Mar 19 11:56:35 PDT 2015

> On Mar 19, 2015, at 10:18 AM, Josh Klontz <josh.klontz at gmail.com> wrote:
> Adam,
> Please find the attached test case (run with ToT opt -O3). As you can see, `y_body` successfully is vectorized, though %33 and %46 are deemed MayAlias despite their exclusive use in loads ands stores marked with `llvm.mem.parallel_loop_access`.

Looks like no bug here.  Your metadata is off.  As I understand the operand of llvm.mem.parallel_loop_access should reference a loop.  Your accesses use !1 but the loop is identified as !2.  Adjusting the loop like this removes the memchecks for me:

--- /tmp/test_case.ll   2015-03-19 11:52:52.000000000 -0700
+++ /tmp/test_case-2.ll 2015-03-19 11:53:00.000000000 -0700
@@ -84,7 +84,7 @@
   %x_increment = add nuw nsw i64 0, 1
   %y_increment = add nuw nsw i64 %y, 1
   %y_postcondition = icmp ne i64 %y_increment, %32
-  br i1 %y_postcondition, label %y_body, label %y_exit, !llvm.loop !2
+  br i1 %y_postcondition, label %y_body, label %y_exit, !llvm.loop !1

 y_exit:                                           ; preds = %y_body
   ret %i16SXY* %18


> Many Thanks,
> Josh
> On Thu, Mar 19, 2015 at 12:55 PM, Adam Nemet <anemet at apple.com <mailto:anemet at apple.com>> wrote:
> > On Mar 19, 2015, at 9:43 AM, Josh Klontz <josh.klontz at gmail.com <mailto:josh.klontz at gmail.com>> wrote:
> >
> > It seems that at some point in the not-so-distant-past that the loop vectorizer gained the ability to vectorize loops without explicit `llvm.loop` & `llvm.mem.parallel_loop_access` metadata. While that's awesome, there seems to be a regression in that `llvm.mem.parallel_loop_access` metadata doesn't make it into the alias analysis, and therefore a `vector.memcheck` basic block is inserted, where as before it was not.
> There has been active development in this are to generalize LV’s dependence analysis and memcheck infrastructure.  The changes should not have affected functionality minus bugs.  If you have a testcase I can look at this.
> Adam
> > It's unclear if this is a regression, as I assume that if I upgrade my frontend to use the new alias metadata instead of the loop metadata then I would expect this problem to disappear. Please advise, happy to provide exemplar code if helpful.
> >
> > v/r,
> > Josh
> > _______________________________________________
> > LLVM Developers mailing list
> > LLVMdev at cs.uiuc.edu <mailto:LLVMdev at cs.uiuc.edu>         http://llvm.cs.uiuc.edu <http://llvm.cs.uiuc.edu/>
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev <http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev>
> <test_case.ll>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150319/e7c53f6a/attachment.html>

More information about the llvm-dev mailing list