#14528: LLVM's CallAnalyzer Breaks -------------------------------------+------------------------------------- Reporter: kavon | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.2.1 (LLVM) | Keywords: inline | Operating System: Unknown/Multiple Architecture: | Type of failure: Compile-time Unknown/Multiple | crash or panic Test Case: | Blocked By: Blocking: 11295 | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- I found this issue while working on #11295. It seems that after first running `-mem2reg -inline` on some nofib programs, such as `imaginary/integrate` and `real/lift` (using `llvm-link` to roll the IR files together), running another pass that depends on LLVM's CallAnalyzer will cause `opt` to overflow the stack. For example, here are some pass sequences that break `opt` when optimizing `integrate`: * `-mem2reg -inline -gvn` * `-mem2reg -inline -instcombine` * `-mem2reg -inline -early-cse` This happens in LLVM 5 and tip-of-tree (LLVM 6). It probably is present in LLVM 4 as well. I'm working on putting together a reduced test case to report a bug to LLVM developers. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14528 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler