From 2623a10c18646e054cdeb47b4a0b138654097f07 Mon Sep 17 00:00:00 2001
From: Tim Besard <tim.besard@gmail.com>
Date: Sat, 19 Nov 2022 12:16:04 +0100
Subject: [PATCH] Improve performance of global code by emitting fewer atomic
 barriers.

---
 src/codegen.cpp | 37 +++++++++++++++++++++----------------
 1 file changed, 21 insertions(+), 16 deletions(-)

diff --git a/src/codegen.cpp b/src/codegen.cpp
index cdc5e00a26281..f000eb9d57516 100644
--- a/src/codegen.cpp
+++ b/src/codegen.cpp
@@ -7791,13 +7791,6 @@ static jl_llvm_functions_t
             ctx.builder.SetInsertPoint(tryblk);
         }
         else {
-            if (!jl_is_method(ctx.linfo->def.method) && !ctx.is_opaque_closure) {
-                // TODO: inference is invalid if this has any effect (which it often does)
-                LoadInst *world = ctx.builder.CreateAlignedLoad(getSizeTy(ctx.builder.getContext()),
-                    prepare_global_in(jl_Module, jlgetworld_global), Align(sizeof(size_t)));
-                world->setOrdering(AtomicOrdering::Acquire);
-                ctx.builder.CreateAlignedStore(world, world_age_field, Align(sizeof(size_t)));
-            }
             emit_stmtpos(ctx, stmt, cursor);
             mallocVisitStmt(debuginfoloc, nullptr);
         }
@@ -8023,12 +8016,12 @@ static jl_llvm_functions_t
     }
 
     // step 12. Perform any delayed instantiations
-    if (ctx.debug_enabled) {
-        bool in_prologue = true;
-        for (auto &BB : *ctx.f) {
-            for (auto &I : BB) {
-                CallBase *call = dyn_cast<CallBase>(&I);
-                if (call && !I.getDebugLoc()) {
+    bool in_prologue = true;
+    for (auto &BB : *ctx.f) {
+        for (auto &I : BB) {
+            CallBase *call = dyn_cast<CallBase>(&I);
+            if (call) {
+                if (ctx.debug_enabled && !I.getDebugLoc()) {
                     // LLVM Verifier: inlinable function call in a function with debug info must have a !dbg location
                     // make sure that anything we attempt to call has some inlining info, just in case optimization messed up
                     // (except if we know that it is an intrinsic used in our prologue, which should never have its own debug subprogram)
@@ -8037,12 +8030,24 @@ static jl_llvm_functions_t
                         I.setDebugLoc(topdebugloc);
                     }
                 }
-                if (&I == &prologue_end)
-                    in_prologue = false;
+                if (toplevel && !ctx.is_opaque_closure && !in_prologue) {
+                    // we're at toplevel; insert an atomic barrier between every instruction
+                    // TODO: inference is invalid if this has any effect (which it often does)
+                    LoadInst *load_world = new LoadInst(getSizeTy(ctx.builder.getContext()),
+                        prepare_global_in(jl_Module, jlgetworld_global), Twine(),
+                        /*isVolatile*/false, Align(sizeof(size_t)), /*insertBefore*/&I);
+                    load_world->setOrdering(AtomicOrdering::Monotonic);
+                    StoreInst *store_world = new StoreInst(load_world, world_age_field,
+                        /*isVolatile*/false, Align(sizeof(size_t)), /*insertBefore*/&I);
+                    (void)store_world;
+                }
             }
+            if (&I == &prologue_end)
+                in_prologue = false;
         }
-        dbuilder.finalize();
     }
+    if (ctx.debug_enabled)
+        dbuilder.finalize();
 
     if (ctx.vaSlot > 0) {
         // remove VA allocation if we never referenced it