Merge branch 'master' into test_compiler_codegen_ci

Compiler/extras/CompilerDevTools/src/CompilerDevTools.jl

@@ -9,12 +9,13 @@ struct SplitCacheInterp <: Compiler.AbstractInterpreter
     inf_params::Compiler.InferenceParams
     opt_params::Compiler.OptimizationParams
     inf_cache::Vector{Compiler.InferenceResult}
+    codegen_cache::IdDict{CodeInstance,CodeInfo}
     function SplitCacheInterp(;
         world::UInt = Base.get_world_counter(),
         inf_params::Compiler.InferenceParams = Compiler.InferenceParams(),
         opt_params::Compiler.OptimizationParams = Compiler.OptimizationParams(),
         inf_cache::Vector{Compiler.InferenceResult} = Compiler.InferenceResult[])
-        new(world, inf_params, opt_params, inf_cache)
+        new(world, inf_params, opt_params, inf_cache, IdDict{CodeInstance,CodeInfo}())
     end
 end
 
@@ -23,10 +24,11 @@ Compiler.OptimizationParams(interp::SplitCacheInterp) = interp.opt_params
 Compiler.get_inference_world(interp::SplitCacheInterp) = interp.world
 Compiler.get_inference_cache(interp::SplitCacheInterp) = interp.inf_cache
 Compiler.cache_owner(::SplitCacheInterp) = SplitCacheOwner()
+Compiler.codegen_cache(interp::SplitCacheInterp) = interp.codegen_cache
 
 import Core.OptimizedGenerics.CompilerPlugins: typeinf, typeinf_edge
 @eval @noinline typeinf(::SplitCacheOwner, mi::MethodInstance, source_mode::UInt8) =
-    Base.invoke_in_world(which(typeinf, Tuple{SplitCacheOwner, MethodInstance, UInt8}).primary_world, Compiler.typeinf_ext, SplitCacheInterp(; world=Base.tls_world_age()), mi, source_mode)
+    Base.invoke_in_world(which(typeinf, Tuple{SplitCacheOwner, MethodInstance, UInt8}).primary_world, Compiler.typeinf_ext_toplevel, SplitCacheInterp(; world=Base.tls_world_age()), mi, source_mode)
 
 @eval @noinline function typeinf_edge(::SplitCacheOwner, mi::MethodInstance, parent_frame::Compiler.InferenceState, world::UInt, source_mode::UInt8)
     # TODO: This isn't quite right, we're just sketching things for now

Compiler/src/typeinfer.jl

@@ -143,9 +143,10 @@ function finish!(interp::AbstractInterpreter, caller::InferenceState, validation
             ci, inferred_result, const_flag, first(result.valid_worlds), last(result.valid_worlds), encode_effects(result.ipo_effects),
             result.analysis_results, di, edges)
         engine_reject(interp, ci)
-        if !discard_src && isdefined(interp, :codegen) && uncompressed isa CodeInfo
+        codegen = codegen_cache(interp)
+        if !discard_src && codegen !== nothing && uncompressed isa CodeInfo
             # record that the caller could use this result to generate code when required, if desired, to avoid repeating n^2 work
-            interp.codegen[ci] = uncompressed
+            codegen[ci] = uncompressed
             if bootstrapping_compiler && inferred_result == nothing
                 # This is necessary to get decent bootstrapping performance
                 # when compiling the compiler to inject everything eagerly
@@ -185,8 +186,9 @@ function finish!(interp::AbstractInterpreter, mi::MethodInstance, ci::CodeInstan
     ccall(:jl_update_codeinst, Cvoid, (Any, Any, Int32, UInt, UInt, UInt32, Any, Any, Any),
         ci, nothing, const_flag, min_world, max_world, ipo_effects, nothing, di, edges)
     code_cache(interp)[mi] = ci
-    if isdefined(interp, :codegen)
-        interp.codegen[ci] = src
+    codegen = codegen_cache(interp)
+    if codegen !== nothing
+        codegen[ci] = src
     end
     engine_reject(interp, ci)
     return nothing
@@ -1168,7 +1170,10 @@ function typeinf_ext(interp::AbstractInterpreter, mi::MethodInstance, source_mod
 
     ci = result.ci # reload from result in case it changed
     @assert frame.cache_mode != CACHE_MODE_NULL
-    @assert is_result_constabi_eligible(result) || (!isdefined(interp, :codegen) || haskey(interp.codegen, ci))
+    @assert is_result_constabi_eligible(result) || begin
+            codegen = codegen_cache(interp)
+            codegen === nothing || haskey(codegen, ci)
+        end
     @assert is_result_constabi_eligible(result) == use_const_api(ci)
     @assert isdefined(ci, :inferred) "interpreter did not fulfill our expectations"
     if !is_cached(frame) && source_mode == SOURCE_MODE_ABI
@@ -1234,44 +1239,55 @@ function collectinvokes!(wq::Vector{CodeInstance}, ci::CodeInfo)
     end
 end
 
-# This is a bridge for the C code calling `jl_typeinf_func()` on a single Method match
-function typeinf_ext_toplevel(mi::MethodInstance, world::UInt, source_mode::UInt8)
-    interp = NativeInterpreter(world)
-    ci = typeinf_ext(interp, mi, source_mode)
-    if source_mode == SOURCE_MODE_ABI && ci isa CodeInstance && !ci_has_invoke(ci)
-        inspected = IdSet{CodeInstance}()
-        tocompile = Vector{CodeInstance}()
-        push!(tocompile, ci)
-        while !isempty(tocompile)
-            # ci_has_real_invoke(ci) && return ci # optimization: cease looping if ci happens to get compiled (not just jl_fptr_wait_for_compiled, but fully jl_is_compiled_codeinst)
-            callee = pop!(tocompile)
-            ci_has_invoke(callee) && continue
-            callee in inspected && continue
-            src = get(interp.codegen, callee, nothing)
+function add_codeinsts_to_jit!(interp::AbstractInterpreter, ci, source_mode::UInt8)
+    source_mode == SOURCE_MODE_ABI || return ci
+    ci isa CodeInstance && !ci_has_invoke(ci) || return ci
+    codegen = codegen_cache(interp)
+    codegen === nothing && return ci
+    inspected = IdSet{CodeInstance}()
+    tocompile = Vector{CodeInstance}()
+    push!(tocompile, ci)
+    while !isempty(tocompile)
+        # ci_has_real_invoke(ci) && return ci # optimization: cease looping if ci happens to get compiled (not just jl_fptr_wait_for_compiled, but fully jl_is_compiled_codeinst)
+        callee = pop!(tocompile)
+        ci_has_invoke(callee) && continue
+        callee in inspected && continue
+        src = get(codegen, callee, nothing)
+        if !isa(src, CodeInfo)
+            src = @atomic :monotonic callee.inferred
+            if isa(src, String)
+                src = _uncompressed_ir(callee, src)
+            end
             if !isa(src, CodeInfo)
-                src = @atomic :monotonic callee.inferred
-                if isa(src, String)
-                    src = _uncompressed_ir(callee, src)
-                end
-                if !isa(src, CodeInfo)
-                    newcallee = typeinf_ext(interp, callee.def, source_mode)
-                    if newcallee isa CodeInstance
-                        callee === ci && (ci = newcallee) # ci stopped meeting the requirements after typeinf_ext last checked, try again with newcallee
-                        push!(tocompile, newcallee)
-                    #else
-                    #    println("warning: could not get source code for ", callee.def)
-                    end
-                    continue
+                newcallee = typeinf_ext(interp, callee.def, source_mode)
+                if newcallee isa CodeInstance
+                    callee === ci && (ci = newcallee) # ci stopped meeting the requirements after typeinf_ext last checked, try again with newcallee
+                    push!(tocompile, newcallee)
+                #else
+                #    println("warning: could not get source code for ", callee.def)
                 end
+                continue
             end
-            push!(inspected, callee)
-            collectinvokes!(tocompile, src)
-            ccall(:jl_add_codeinst_to_jit, Cvoid, (Any, Any), callee, src)
         end
+        push!(inspected, callee)
+        collectinvokes!(tocompile, src)
+        ccall(:jl_add_codeinst_to_jit, Cvoid, (Any, Any), callee, src)
     end
     return ci
 end
 
+function typeinf_ext_toplevel(interp::AbstractInterpreter, mi::MethodInstance, source_mode::UInt8)
+    ci = typeinf_ext(interp, mi, source_mode)
+    ci = add_codeinsts_to_jit!(interp, ci, source_mode)
+    return ci
+end
+
+# This is a bridge for the C code calling `jl_typeinf_func()` on a single Method match
+function typeinf_ext_toplevel(mi::MethodInstance, world::UInt, source_mode::UInt8)
+    interp = NativeInterpreter(world)
+    return typeinf_ext_toplevel(interp, mi, source_mode)
+end
+
 # This is a bridge for the C code calling `jl_typeinf_func()` on set of Method matches
 function typeinf_ext_toplevel(methods::Vector{Any}, worlds::Vector{UInt}, trim::Bool)
     inspected = IdSet{CodeInstance}()

Compiler/src/types.jl

@@ -23,6 +23,10 @@ the following methods to satisfy the `AbstractInterpreter` API requirement:
 - `get_inference_world(interp::NewInterpreter)` - return the world age for this interpreter
 - `get_inference_cache(interp::NewInterpreter)` - return the local inference cache
 - `cache_owner(interp::NewInterpreter)` - return the owner of any new cache entries
+
+If `CodeInstance`s compiled using `interp::NewInterpreter` are meant to be executed with `invoke`,
+a method `codegen_cache(interp::NewInterpreter) -> IdDict{CodeInstance, CodeInfo}` must be defined,
+and inference must be triggered via `typeinf_ext_toplevel` with source mode `SOURCE_MODE_ABI`.
 """
 abstract type AbstractInterpreter end
 
@@ -430,6 +434,19 @@ to incorporate customized dispatches for the overridden methods.
 method_table(interp::AbstractInterpreter) = InternalMethodTable(get_inference_world(interp))
 method_table(interp::NativeInterpreter) = interp.method_table
 
+"""
+    codegen_cache(interp::AbstractInterpreter) -> Union{Nothing, IdDict{CodeInstance, CodeInfo}}
+
+Optionally return a cache associating a `CodeInfo` to a `CodeInstance` that should be added to the JIT
+for future execution via `invoke(f, ::CodeInstance, args...)`. This cache is used during `typeinf_ext_toplevel`,
+and may be safely discarded between calls to this function.
+
+By default, a value of `nothing` is returned indicating that `CodeInstance`s should not be added to the JIT.
+Attempting to execute them via `invoke` will result in an error.
+"""
+codegen_cache(interp::AbstractInterpreter) = nothing
+codegen_cache(interp::NativeInterpreter) = interp.codegen
+
 """
 By default `AbstractInterpreter` implements the following inference bail out logic:
 - `bail_out_toplevel_call(::AbstractInterpreter, sig, ::InferenceState)`: bail out from

Compiler/test/AbstractInterpreter.jl

@@ -534,3 +534,17 @@ let interp = DebugInterp()
     end
     @test found
 end
+
+@newinterp InvokeInterp
+struct InvokeOwner end
+codegen = IdDict{CodeInstance, CodeInfo}()
+Compiler.cache_owner(::InvokeInterp) = InvokeOwner()
+Compiler.codegen_cache(::InvokeInterp) = codegen
+let interp = InvokeInterp()
+    source_mode = Compiler.SOURCE_MODE_ABI
+    f = (+)
+    args = (1, 1)
+    mi = @ccall jl_method_lookup(Any[f, args...]::Ptr{Any}, (1+length(args))::Csize_t, Base.tls_world_age()::Csize_t)::Ref{Core.MethodInstance}
+    ci = Compiler.typeinf_ext_toplevel(interp, mi, source_mode)
+    @test invoke(f, ci, args...) == 2
+end

Compiler/test/codegen.jl

@@ -889,57 +889,6 @@ ex54166 = Union{Missing, Int64}[missing -2; missing -2];
 dims54166 = (1,2)
 @test (minimum(ex54166; dims=dims54166)[1] === missing)
 
-# #54109 - Excessive LLVM time for egal
-struct DefaultOr54109{T}
-    x::T
-    default::Bool
-end
-
-@eval struct Torture1_54109
-    $((Expr(:(::), Symbol("x$i"), DefaultOr54109{Float64}) for i = 1:897)...)
-end
-Torture1_54109() = Torture1_54109((DefaultOr54109(1.0, false) for i = 1:897)...)
-
-@eval struct Torture2_54109
-    $((Expr(:(::), Symbol("x$i"), DefaultOr54109{Float64}) for i = 1:400)...)
-    $((Expr(:(::), Symbol("x$(i+400)"), DefaultOr54109{Int16}) for i = 1:400)...)
-end
-Torture2_54109() = Torture2_54109((DefaultOr54109(1.0, false) for i = 1:400)..., (DefaultOr54109(Int16(1), false) for i = 1:400)...)
-
-@noinline egal_any54109(x, @nospecialize(y::Any)) = x === Base.compilerbarrier(:type, y)
-
-let ir1 = get_llvm(egal_any54109, Tuple{Torture1_54109, Any}),
-    ir2 = get_llvm(egal_any54109, Tuple{Torture2_54109, Any})
-
-    # We can't really do timing on CI, so instead, let's look at the length of
-    # the optimized IR. The original version had tens of thousands of lines and
-    # was slower, so just check here that we only have < 500 lines. If somebody,
-    # implements a better comparison that's larger than that, just re-benchmark
-    # this and adjust the threshold.
-
-    @test count(==('\n'), ir1) < 500
-    @test count(==('\n'), ir2) < 500
-end
-
-## Regression test for egal of a struct of this size without padding, but with
-## non-bitsegal, to make sure that it doesn't accidentally go down the accelerated
-## path.
-@eval struct BigStructAnyInt
-    $((Expr(:(::), Symbol("x$i"), Pair{Any, Int}) for i = 1:33)...)
-end
-BigStructAnyInt() = BigStructAnyInt((Union{Base.inferencebarrier(Float64), Int}=>i for i = 1:33)...)
-@test egal_any54109(BigStructAnyInt(), BigStructAnyInt())
-
-## For completeness, also test correctness, since we don't have a lot of
-## large-struct tests.
-
-# The two allocations of the same struct will likely have different padding,
-# we want to make sure we find them egal anyway - a naive memcmp would
-# accidentally look at it.
-@test egal_any54109(Torture1_54109(), Torture1_54109())
-@test egal_any54109(Torture2_54109(), Torture2_54109())
-@test !egal_any54109(Torture1_54109(), Torture1_54109((DefaultOr54109(2.0, false) for i = 1:897)...))
-
 bar54599() = Base.inferencebarrier(true) ? (Base.PkgId(Main),1) : nothing
 
 function foo54599()

base/loading.jl

@@ -3288,7 +3288,7 @@ function compilecache(pkg::PkgId, path::String, internal_stderr::IO = stderr, in
     if p.exitcode == 125
         return PrecompilableError()
     else
-        error("Failed to precompile $(repr("text/plain", pkg)) to $(repr(tmppath)).")
+        error("Failed to precompile $(repr("text/plain", pkg)) to $(repr(tmppath)) ($(Base.process_status(p))).")
     end
 end
 

src/cgutils.cpp

@@ -4710,10 +4710,9 @@ static jl_cgval_t emit_memoryref(jl_codectx_t &ctx, const jl_cgval_t &ref, jl_cg
             setName(ctx.emission_context, ovflw, "memoryref_ovflw");
         }
 #endif
-        boffset = ctx.builder.CreateMul(offset, elsz);
-        setName(ctx.emission_context, boffset, "memoryref_byteoffset");
-        newdata = ctx.builder.CreateGEP(getInt8Ty(ctx.builder.getContext()), data, boffset);
-        setName(ctx.emission_context, newdata, "memoryref_data_byteoffset");
+        Type *elty = isboxed ? ctx.types().T_prjlvalue : julia_type_to_llvm(ctx, jl_tparam1(ref.typ));
+        newdata = ctx.builder.CreateGEP(elty, data, offset);
+        setName(ctx.emission_context, newdata, "memoryref_data_offset");
         (void)boffset; // LLVM is very bad at handling GEP with types different from the load
         if (bc) {
             BasicBlock *failBB, *endBB;