apps/fft benchmark fails for GPU targets at 12x12 or larger

[steven-johnson]

Run apps/fft/bench_fft at 12x12 or larger, using a GPU target (e.g., host-metal), and fail; on OSX, the error was

bench_fft(9110,0x1048f9600) malloc: *** error for object 0xffffffffffffffff: pointer being freed was not allocated
bench_fft(9110,0x1048f9600) malloc: *** set a breakpoint in malloc_error_break to debug

[steven-johnson]

Debugging further, the metal case is crashing inside LLVM:

    frame #3: 0x000000010f7e6b86 libHalide.dylib`llvm::slpvectorizer::BoUpSLP::getEntryCost(llvm::slpvectorizer::BoUpSLP::TreeEntry const*, llvm::ArrayRef<llvm::Value*>) + 20118
    frame #4: 0x000000010f7ebff5 libHalide.dylib`llvm::slpvectorizer::BoUpSLP::getTreeCost(llvm::ArrayRef<llvm::Value*>) + 1013
    frame #5: 0x000000010f806b1e libHalide.dylib`llvm::SLPVectorizerPass::tryToVectorizeList(llvm::ArrayRef<llvm::Value*>, llvm::slpvectorizer::BoUpSLP&, bool) + 3118
    frame #6: 0x000000010f80b51c libHalide.dylib`llvm::SLPVectorizerPass::vectorizeInsertElementInst(llvm::InsertElementInst*, llvm::BasicBlock*, llvm::slpvectorizer::BoUpSLP&) + 556
    frame #7: 0x000000010f80b73c libHalide.dylib`llvm::SLPVectorizerPass::vectorizeSimpleInstructions(llvm::SmallSetVector<llvm::Instruction*, 8u>&, llvm::BasicBlock*, llvm::slpvectorizer::BoUpSLP&, bool) + 508
    frame #8: 0x000000010f802ac4 libHalide.dylib`llvm::SLPVectorizerPass::vectorizeChainsInBlock(llvm::BasicBlock*, llvm::slpvectorizer::BoUpSLP&) + 3620
    frame #9: 0x000000010f80106c libHalide.dylib`llvm::SLPVectorizerPass::runImpl(llvm::Function&, llvm::ScalarEvolution*, llvm::TargetTransformInfo*, llvm::TargetLibraryInfo*, llvm::AAResults*, llvm::LoopInfo*, llvm::DominatorTree*, llvm::AssumptionCache*, llvm::DemandedBits*, llvm::OptimizationRemarkEmitter*) + 2012
    frame #10: 0x000000010f800444 libHalide.dylib`llvm::SLPVectorizerPass::run(llvm::Function&, llvm::AnalysisManager<llvm::Function>&) + 996
    frame #11: 0x000000010f3ecbf2 libHalide.dylib`llvm::detail::PassModel<llvm::Function, llvm::SLPVectorizerPass, llvm::PreservedAnalyses, llvm::AnalysisManager<llvm::Function> >::run(llvm::Function&, llvm::AnalysisManager<llvm::Function>&) + 18
    frame #12: 0x00000001105f9c3b libHalide.dylib`llvm::PassManager<llvm::Function, llvm::AnalysisManager<llvm::Function> >::run(llvm::Function&, llvm::AnalysisManager<llvm::Function>&) + 347
    frame #13: 0x000000010e401152 libHalide.dylib`llvm::detail::PassModel<llvm::Function, llvm::PassManager<llvm::Function, llvm::AnalysisManager<llvm::Function> >, llvm::PreservedAnalyses, llvm::AnalysisManager<llvm::Function> >::run(llvm::Function&, llvm::AnalysisManager<llvm::Function>&) + 18
    frame #14: 0x00000001105fe8dc libHalide.dylib`llvm::ModuleToFunctionPassAdaptor::run(llvm::Module&, llvm::AnalysisManager<llvm::Module>&) + 492
    frame #15: 0x000000010e400f32 libHalide.dylib`llvm::detail::PassModel<llvm::Module, llvm::ModuleToFunctionPassAdaptor, llvm::PreservedAnalyses, llvm::AnalysisManager<llvm::Module> >::run(llvm::Module&, llvm::AnalysisManager<llvm::Module>&) + 18
    frame #16: 0x00000001105f8e7b libHalide.dylib`llvm::PassManager<llvm::Module, llvm::AnalysisManager<llvm::Module> >::run(llvm::Module&, llvm::AnalysisManager<llvm::Module>&) + 347
    frame #17: 0x000000010c8de09a libHalide.dylib`Halide::Internal::CodeGen_LLVM::optimize_module(this=0x0000000100f2de50) at CodeGen_LLVM.cpp:1249:9

[steven-johnson]

Issues doesn't occur in LLVM15. Going to try to bisect in LLVM16 to see if it's an injection from upstream. (Could have been there for a while, as our buildbots had been mistakenly not testing this for, uh, a while)