Change TM handling slighly and force no fastisel

in some targets

src/aotcompile.cpp

@@ -1001,7 +1001,7 @@ static AOTOutputs add_output_impl(Module &M, TargetMachine &SourceTM, ShardTimer
             SourceTM.getRelocationModel(),
             SourceTM.getCodeModel(),
             SourceTM.getOptLevel()));
-
+    fixupTM(*TM);
     if (unopt) {
         timers.unopt.startTimer();
         raw_svector_ostream OS(out.unopt);
@@ -1029,6 +1029,7 @@ static AOTOutputs add_output_impl(Module &M, TargetMachine &SourceTM, ShardTimer
                 SourceTM.getRelocationModel(),
                 SourceTM.getCodeModel(),
                 SourceTM.getOptLevel()));
+        fixupTM(*PMTM);
         NewPM optimizer{std::move(PMTM), getOptLevel(jl_options.opt_level), OptimizationOptions::defaults(true, true)};
         optimizer.run(M);
         assert(!verifyLLVMIR(M));
@@ -1514,6 +1515,7 @@ void jl_dump_native_impl(void *native_code,
             CMModel,
             CodeGenOpt::Aggressive // -O3 TODO: respect command -O0 flag?
             ));
+    fixupTM(*SourceTM);
     auto DL = jl_create_datalayout(*SourceTM);
     std::string StackProtectorGuard;
     unsigned OverrideStackAlignment;

src/codegen.cpp

@@ -4802,7 +4802,7 @@ static Value *global_binding_pointer(jl_codectx_t &ctx, jl_module_t *m, jl_sym_t
         // var not found. switch to delayed lookup.
         Constant *initnul = Constant::getNullValue(ctx.types().T_pjlvalue);
         GlobalVariable *bindinggv = new GlobalVariable(*ctx.f->getParent(), ctx.types().T_pjlvalue,
-                false, GlobalVariable::PrivateLinkage, initnul);
+                false, GlobalVariable::PrivateLinkage, initnul, "jl_binding_ptr"); // LLVM has bugs with nameless globals
         LoadInst *cachedval = ctx.builder.CreateAlignedLoad(ctx.types().T_pjlvalue, bindinggv, Align(sizeof(void*)));
         setName(ctx.emission_context, cachedval, jl_symbol_name(m->name) + StringRef(".") + jl_symbol_name(s) + ".cached");
         cachedval->setOrdering(AtomicOrdering::Unordered);

src/jitlayers.cpp

@@ -1198,6 +1198,12 @@ namespace {
 #else
             None;
 #endif
+        if (TheTriple.isAArch64())
+            codemodel = CodeModel::Small;
+        if (TheTriple.isPPC()) {
+        // On PPC the small model is limited to 16bit offsets
+            codemodel = CodeModel::Medium;
+        }
         auto optlevel = CodeGenOptLevelFor(jl_options.opt_level);
         auto TM = TheTarget->createTargetMachine(
                 TheTriple.getTriple(), TheCPU, FeaturesStr,
@@ -1209,11 +1215,7 @@ namespace {
                 );
         assert(TM && "Failed to select target machine -"
                      " Is the LLVM backend for this CPU enabled?");
-        if (!TheTriple.isARM() && !TheTriple.isPPC64()) {
-            // FastISel seems to be buggy for ARM. Ref #13321
-            if (jl_options.opt_level < 2)
-                TM->setFastISel(true);
-        }
+        fixupTM(*TM);
         return std::unique_ptr<TargetMachine>(TM);
     }
 } // namespace
@@ -1239,7 +1241,9 @@ namespace {
             : JTMB(createJTMBFromTM(TM, optlevel)) {}
 
         std::unique_ptr<TargetMachine> operator()() JL_NOTSAFEPOINT {
-            return cantFail(JTMB.createTargetMachine());
+            auto TM = cantFail(JTMB.createTargetMachine());
+            fixupTM(*TM);
+            return TM;
         }
     };
 
@@ -1251,7 +1255,9 @@ namespace {
             : JTMB(createJTMBFromTM(TM, optlevel)), O(getOptLevel(optlevel)), printers(printers) {}
 
         auto operator()() JL_NOTSAFEPOINT {
-            auto NPM = std::make_unique<NewPM>(cantFail(JTMB.createTargetMachine()), O);
+            auto TM = cantFail(JTMB.createTargetMachine());
+            fixupTM(*TM);
+            auto NPM = std::make_unique<NewPM>(std::move(TM), O);
             // TODO this needs to be locked, as different resource pools may add to the printer vector at the same time
             printers.push_back([NPM = NPM.get()]() JL_NOTSAFEPOINT {
                 NPM->printTimers();
@@ -1637,6 +1643,16 @@ void optimizeDLSyms(Module &M) {
     JuliaOJIT::DLSymOptimizer(true)(M);
 }
 
+void fixupTM(TargetMachine &TM){
+    auto TheTriple = TM.getTargetTriple();
+    if (jl_options.opt_level < 2) {
+        if (!TheTriple.isARM() && !TheTriple.isPPC64() && !TheTriple.isAArch64())
+            TM.setFastISel(true);
+        else    // FastISel seems to be buggy Ref #13321
+            TM.setFastISel(false);
+    }
+}
+
 llvm::DataLayout jl_create_datalayout(TargetMachine &TM) {
     // Mark our address spaces as non-integral
     auto jl_data_layout = TM.createDataLayout();
@@ -2153,7 +2169,7 @@ void jl_merge_module(orc::ThreadSafeModule &destTSM, orc::ThreadSafeModule srcTS
 
 std::unique_ptr<TargetMachine> JuliaOJIT::cloneTargetMachine() const
 {
-    return std::unique_ptr<TargetMachine>(getTarget()
+    auto NewTM = std::unique_ptr<TargetMachine>(getTarget()
         .createTargetMachine(
             getTargetTriple().str(),
             getTargetCPU(),
@@ -2162,6 +2178,8 @@ std::unique_ptr<TargetMachine> JuliaOJIT::cloneTargetMachine() const
             TM->getRelocationModel(),
             TM->getCodeModel(),
             TM->getOptLevel()));
+    fixupTM(*NewTM);
+    return NewTM;
 }
 
 const Triple& JuliaOJIT::getTargetTriple() const {

src/jitlayers.h

@@ -569,6 +569,7 @@ Module &jl_codegen_params_t::shared_module() JL_NOTSAFEPOINT {
     }
     return *_shared_module;
 }
+void fixupTM(TargetMachine &TM) JL_NOTSAFEPOINT;
 
 void optimizeDLSyms(Module &M);
 

src/pipeline.cpp

@@ -744,9 +744,9 @@ PIC.addClassToPassName(decltype(CREATE_PASS)::name(), NAME);
 }
 
 NewPM::NewPM(std::unique_ptr<TargetMachine> TM, OptimizationLevel O, OptimizationOptions options) :
-    TM(std::move(TM)), 
+    TM(std::move(TM)),
 #if JL_LLVM_VERSION < 160000
-    SI(false), 
+    SI(false),
     PIC(createPIC(SI)),
 #else
     PIC(createPIC()),