Add option to force emission of const return function (#57824)

This is mostly for GPUCompiler

(cherry picked from commit e78cf6c)

Author: gbaraldi

base/reflection.jl

@@ -149,23 +149,22 @@ struct CodegenParams
     use_jlplt::Cint
 
     """
-    If enabled, only provably reachable code (from functions marked with `entrypoint`) is included
-    in the output system image. Errors or warnings can be given for call sites too dynamic to handle.
-    The option is disabled by default. (0=>disabled, 1=>safe (static errors), 2=>unsafe, 3=>unsafe plus warnings)
+        If enabled emit LLVM IR for all functions even if wouldn't be compiled
+        for some reason (i.e functions that return a constant value).
     """
-    trim::Cint
+    force_emit_all::Cint
 
     function CodegenParams(; track_allocations::Bool=true, code_coverage::Bool=true,
                    prefer_specsig::Bool=false,
                    gnu_pubnames::Bool=true, debug_info_kind::Cint = default_debug_info_kind(),
                    debug_info_level::Cint = Cint(JLOptions().debug_level), safepoint_on_entry::Bool=true,
-                   gcstack_arg::Bool=true, use_jlplt::Bool=true, trim::Cint=Cint(0))
+                   gcstack_arg::Bool=true, use_jlplt::Bool=true, force_emit_all::Bool=false)
         return new(
             Cint(track_allocations), Cint(code_coverage),
             Cint(prefer_specsig),
             Cint(gnu_pubnames), debug_info_kind,
             debug_info_level, Cint(safepoint_on_entry),
-            Cint(gcstack_arg), Cint(use_jlplt), Cint(trim))
+            Cint(gcstack_arg), Cint(use_jlplt), Cint(force_emit_all))
     end
 end
 

src/aotcompile.cpp

@@ -775,7 +775,7 @@ void *jl_emit_native_impl(jl_array_t *codeinfos, LLVMOrcThreadSafeModuleRef llvm
                     params.tsctx, clone.getModuleUnlocked()->getDataLayout(),
                     Triple(clone.getModuleUnlocked()->getTargetTriple()));
             jl_llvm_functions_t decls;
-            if (jl_atomic_load_relaxed(&codeinst->invoke) == jl_fptr_const_return_addr)
+            if (!(params.params->force_emit_all) && jl_atomic_load_relaxed(&codeinst->invoke) == jl_fptr_const_return_addr)
                 decls.functionObject = "jl_fptr_const_return";
             else
                 decls = jl_emit_codeinst(result_m, codeinst, src, params);

src/cgutils.cpp

@@ -4410,7 +4410,8 @@ static int compare_cgparams(const jl_cgparams_t *a, const jl_cgparams_t *b)
            (a->debug_info_kind == b->debug_info_kind) &&
            (a->safepoint_on_entry == b->safepoint_on_entry) &&
            (a->gcstack_arg == b->gcstack_arg) &&
-           (a->use_jlplt == b->use_jlplt);
+           (a->use_jlplt == b->use_jlplt) &&
+           (a->force_emit_all == b->force_emit_all);
 }
 #endif
 

src/init.c

@@ -739,7 +739,8 @@ JL_DLLEXPORT jl_cgparams_t jl_default_cgparams = {
         /* debug_info_level */ 0, // later jl_options.debug_level,
         /* safepoint_on_entry */ 1,
         /* gcstack_arg */ 1,
-        /* use_jlplt*/ 1 };
+        /* use_jlplt*/ 1 ,
+        /*force_emit_all=*/ 0};
 
 static void init_global_mutexes(void) {
     JL_MUTEX_INIT(&jl_modules_mutex, "jl_modules_mutex");

src/julia.h

@@ -2727,6 +2727,7 @@ typedef struct {
     int gcstack_arg; // Pass the ptls value as an argument with swiftself
 
     int use_jlplt; // Whether to use the Julia PLT mechanism or emit symbols directly
+    int force_emit_all; // Force emission of code for const return functions
 } jl_cgparams_t;
 extern JL_DLLEXPORT int jl_default_debug_info_kind;
 extern JL_DLLEXPORT jl_cgparams_t jl_default_cgparams;