[PATCH] Optimize redundant insertvalue instructions

[Hal Finkel]

Michael,

+/// Here the second instruction inserts values at the same indices, as the

+/// first one, making it redundant.

You mean, "making the first one redundant."

+  Value *V = dyn_cast<Value>(&I);

There should be no cast necessary here (and it certainly shouldn't be a dyn_cast).

+    V = dyn_cast<Value>(UserInsInst);

Same here.

+  while (V->hasOneUse()) {

Should we have a cut-off here? If this function is called on every InsertValue instruction, then this algorithm will be O(N^2) in the size of the InsertValue chain.

 -Hal

----- Original Message -----
> From: "Michael Zolotukhin" <mzolotukhin at apple.com>
> To: "LLVM Commits" <llvm-commits at cs.uiuc.edu>
> Sent: Tuesday, May 6, 2014 9:42:12 AM
> Subject: [PATCH] Optimize redundant insertvalue instructions
> 
> 
> Hi,
> 
> 
> This patch implements a simple optimization of redundant insertvalue
> instructions.
> 
> 
> Here is a small example where it will help:
> 
> %0 = insertvalue { i8, i32 } undef, i8 %x, 0
> %1 = insertvalue { i8, i32 } %0, i8 %y, 0
> 
> 
> In this example, the second instruction inserts values at the same
> indices, as the first one, and consequently, the first instruction
> is redundant.
> This code will be transformed to:
> 
> %0 = insertvalue { i8, i32 } undef, i8 %y, 0
> 
> 
> This optimization catches nothing on specs and test-suite, but the
> example was inspired by a real world application.
> 
> 
> As an experiment, I tried a more sophisticated implementation, which
> was capable of dealing with several users and with partially
> overlapping indices, but that also caught nothing on specs (and for
> my application the simple version was enough). If there is an
> interest, I could share a patch with this more complicated
> implementation as well.
> 
> 
> Is it ok to commit this patch?
> 
> 
> 
> 
> 
> 
> 
> Thanks,
> Michael

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory