[LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

[Martin J. O'Riordan]

Regarding your proposed solution (2) - could this be further abstracted from just the volatile iteration variables so that all conditional dependencies between BBs could be tracked?  For VLIW architectures with predication support, it is often advantageous to either clone a BB predecessor (with appropriate predication) to each of the successor BBs, or to move a dependency from successor BBs into the predecessor BBs (with appropriate predication).  The approach you propose might help with this kind of ILP problem.  Currently LLVM does not seem to do a great job at tracking this kind of ILP data edge dependency across BBs.  I can achieve this now, but only by using a large amount of custom coding and I would prefer to use a more generalised abstraction.

 

            MartinO

 

From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu] On Behalf Of Hyojin Sung
Sent: 15 July 2015 20:52
To: llvmdev at cs.uiuc.edu
Subject: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

 

Hi all,

I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable.

*** Problem statement
Nested loops in LCALS Subset B ( <https://urldefense.proofpoint.com/v2/url?u=https-3A__codesign.llnl.gov_LCALS.php&d=AwMGaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=Mfk2qtn1LTDThVkh6-oGglNfMADXfJdty4_bhmuhMHA&m=aWKfvN4c8lvUSvVn8J0Z2ajTctlBJf0198Au28epBr0&s=4d9dt5ODcDWHHatSrwu5ZYT9ebgVzNEtpOlIR87izCM&e=> https://codesign.llnl.gov/LCALS.php) are not vectorized with LLVM -O3 because the LLVM loop vectorizer fails the test whether the loop latch and exiting block of a loop is the same. The loops are vectorizable, and get vectorized with LLVM -O2 and also with other commercial compilers (icc, xlc). 

*** Details
These loops ended up with different loop latch and exiting block after a series of optimizations including loop unswitching, jump threading, simplify-the-CFG, and loop simplify. The fundamental problem here is that the above optimizations cannot recognize a loop with a volatile iteration variable and do not preserve its canonical loop structure.

(1) Loop unswitching generates several empty placeholder BBs only with PHI nodes after separating out a shorter path with no inner loop execution from a standard path. 

(2) Jump threading and simplify-the-CFG passes independently calls TryToSimplifyUnconditionalBranchFromEmptyBlock() in Transforms/Utils/Local.cpp to get rid of almost empty BBs. 

(3) TryToSimplifyUnconditionalBranchFromEmtpyBlock() eliminates the placeholder BBs after loop unswitching and merges them into subsequent blocks including the header of the inner loop. Before eliminating the blocks, the function checks if the block is a loop header by looking at its PHI nodes so that it can be saved, but the test fails with the loops with a volatile iteration variable. The outer loop is now collapsed into the inner loop with multiple backedges.

(4) LoopSimplify checks if a loop with multiple backedges can be separated to nested loops by looking at PHI nodes in the loop header. The test fails with the loops with a volatile iteration variable. 

(5) LoopSimplify then creates a unique backedge block for the loop, and the loop now has a different loop latch (the unique backedge block created in (3)) and exiting block (a block where the volatile outer loop variable is incremented and tested).  

*** Proposed solutions
(1) Make LoopInfo available in Jump Threading and Simplify-the-CFG passes and use LoopInfo to test whether an almost empty BB is a loop header. If yes, do not eliminate the BB.
      Pros: Leverages existing analysis results, small code changes / Cons: Add pass dependencies 

(2) Instead of using LoopInfo, iterate through BB’s to identify backedges and loop headers in TryToSimplifyUnconditionalBranchFromEmptyBlock() and use the results to test if a BB is a loop header. If yes, do not eliminate the BB. Jump Threading has functions that do similar cheap loop analysis.
      Pros: No need to depend on external analysis results / Cons: More lines to be added

*** Summary
I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable. On top of the current algorithm relying on PHI nodes to recognize loops, actual loop analysis to test if a BB belongs to a loop and etc. can be used. 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150715/9b7e4398/attachment.html>