[IMPALA-5980] Upgrade to LLVM 5.0.1

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: Impala 2.13.0
• Component/s: Backend
• Labels: codegen
• Epic Color: ghx-label-4
• Assignee: Bikramjeet Vig
• Reporter: Bikramjeet Vig
• Created: 25/Sep/17 18:16
• Updated: 29/Mar/18 20:09
• Resolved: 29/Mar/18 20:09

Description

As a part of IMPALA-5907 it was identified that the GVN optimization pass is slow for large cases, specifically when huge expression trees are passed to it.
This slow GVN pass problem is a known issue (https://groups.google.com/forum/#!msg/llvm-dev/ozsAceDmOuA/2O_QC6EiDgAJ) that has motivated the llvm devs to develop a new GVN pass which is claimed to be faster and is available in LLVM 5.0.1
Therefore after upgrading to LLVM 5.0.1 we should try moving to the NewGVN pass and figure out if it is helpful by doing a perf analysis.

Issue Links

• is depended upon by: IMPALA-6259 Remove workaround for posix_memalign in SystemAllocator after LLVM 6.0 upgrade
• is depended upon by: IMPALA-6304 Consider re-enabling AVX512 codegen once we have LLVM 5 + test coverage
• is related to: IMPALA-5907 Increased codegen time on specific queries in Impala 2.6
• relates to: IMPALA-6177 Clean up handcrafted IRs if they encounter error during creation