Remove temporary per-compile IR dump that taxed all downstream CI (#861)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: maleadt

Project.toml

@@ -1,6 +1,6 @@
 name = "GPUCompiler"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
-version = "1.22.5"
+version = "1.22.6"
 authors = ["Tim Besard <tim.besard@gmail.com>"]
 
 [workspace]

src/optim.jl

@@ -49,94 +49,33 @@ function aggressiveinstcombine_pass(@nospecialize(job::CompilerJob))
     end
 end
 
-# A crash in an optimization pass is a compiler bug, and the only reliable way to debug one
-# that reproduces only in CI is to get the offending module off the machine. On a failure we
-# write out the IR *entering* `optimize!` -- a faithful, re-optimizable reproducer (parse it
-# and call `optimize!` again), unlike a module recovered later via `optimize=false`, which has
-# already been through back-end IR finalization -- and, on a CI agent, publish it:
-#  - Buildkite: `buildkite-agent artifact upload`, plus an error annotation pointing to it.
-#  - GitHub Actions: drop it in `$RUNNER_TEMP` for an `actions/upload-artifact` step and
-#    surface it as a workflow notice.
-# Snapshotting costs a `string(mod)` per compile, so it is gated to CI (auto-detected) or an
-# explicit `JULIA_GPUCOMPILER_DUMP_IR`; everything else pays nothing.
-function dump_failing_ir_enabled()
-    if haskey(ENV, "JULIA_GPUCOMPILER_DUMP_IR")
-        return parse(Bool, ENV["JULIA_GPUCOMPILER_DUMP_IR"])
-    end
-    return haskey(ENV, "BUILDKITE") || get(ENV, "GITHUB_ACTIONS", "false") == "true"
-end
-
-function dump_failing_ir(ir::AbstractString, @nospecialize(err))
-    path = tempname() * ".ll"
-    try
-        write(path, ir)
-    catch e
-        @error "GPUCompiler: could not write failing IR" exception=(e, catch_backtrace())
-        return nothing
-    end
-    name = basename(path)
-
-    if parse(Bool, get(ENV, "BUILDKITE", "false"))
-        try
-            run(`buildkite-agent artifact upload $path`)
-            annotation = "GPUCompiler optimization failed: `$(sprint(showerror, err))`. " *
-                "Un-optimized IR uploaded as artifact `$name` (download with " *
-                "`buildkite-agent artifact download $name .` and re-run the optimizer on it " *
-                "to reproduce)."
-            run(pipeline(`buildkite-agent annotate --style error --context gpucompiler-optimization`;
-                         stdin=IOBuffer(annotation)))
-        catch e
-            @error "GPUCompiler: failed to upload failing IR via buildkite-agent" exception=(e, catch_backtrace())
-        end
-    elseif get(ENV, "GITHUB_ACTIONS", "false") == "true"
-        try
-            dir = mkpath(joinpath(get(ENV, "RUNNER_TEMP", tempdir()), "gpucompiler-dumps"))
-            cp(path, joinpath(dir, name); force=true)
-            println("::notice title=GPUCompiler optimization failure::wrote $name to $dir")
-        catch e
-            @error "GPUCompiler: failed to stage failing IR for upload" exception=(e, catch_backtrace())
-        end
-    end
-
-    @info "GPUCompiler: wrote failing module to $path"
-    return path
-end
-
 function optimize!(@nospecialize(job::CompilerJob), mod::LLVM.Module; opt_level=2)
     tm = llvm_machine(job.config.target)
     tti = llvm_targetinfo(job.config.target)
 
     global current_job
     current_job = job
 
-    # snapshot the (re-optimizable) input so a pass crash can surface a reproducer
-    input_ir = dump_failing_ir_enabled() ? string(mod) : nothing
-
-    try
-        @dispose pb=NewPMPassBuilder() begin
-            tti === nothing || LLVM.target_transform_info!(pb, tti)
-
-            register!(pb, GPULowerCPUFeaturesPass())
-            register!(pb, GPULowerPTLSPass())
-            register!(pb, GPULowerGCFramePass())
-            register!(pb, GPULinkRuntimePass())
-            register!(pb, GPULinkLibrariesPass())
-            register!(pb, GPUFinishRuntimeIntrinsicsPass())
-            register!(pb, AddKernelStatePass())
-            register!(pb, LowerKernelStatePass())
-            register!(pb, CleanupKernelStatePass())
-
-            add!(pb, NewPMModulePassManager()) do mpm
-                buildNewPMPipeline!(mpm, job, opt_level)
-            end
-            run!(pb, mod, tm)
+    @dispose pb=NewPMPassBuilder() begin
+        tti === nothing || LLVM.target_transform_info!(pb, tti)
+
+        register!(pb, GPULowerCPUFeaturesPass())
+        register!(pb, GPULowerPTLSPass())
+        register!(pb, GPULowerGCFramePass())
+        register!(pb, GPULinkRuntimePass())
+        register!(pb, GPULinkLibrariesPass())
+        register!(pb, GPUFinishRuntimeIntrinsicsPass())
+        register!(pb, AddKernelStatePass())
+        register!(pb, LowerKernelStatePass())
+        register!(pb, CleanupKernelStatePass())
+
+        add!(pb, NewPMModulePassManager()) do mpm
+            buildNewPMPipeline!(mpm, job, opt_level)
         end
-        optimize_module!(job, mod)
-        run!(DeadArgumentEliminationPass(), mod, tm)
-    catch err
-        input_ir === nothing || dump_failing_ir(input_ir, err)
-        rethrow()
+        run!(pb, mod, tm)
     end
+    optimize_module!(job, mod)
+    run!(DeadArgumentEliminationPass(), mod, tm)
     return
 end