Fix lower handlers pass to handle merging branches into leave_funcs. Also add tests and fix llvmpasses

Author: gbaraldi

src/llvm-lower-handlers.cpp

@@ -150,7 +150,16 @@ static bool lowerExcHandlers(Function &F) {
                 EnterDepth[CI] = Depth++;
             else if (Callee == leave_func || Callee == leave_noexcept_func) {
                 LeaveDepth[CI] = Depth;
-                Depth -= cast<ConstantInt>(CI->getArgOperand(1))->getLimitedValue();
+                if (auto change = dyn_cast<ConstantInt>(CI->getArgOperand(1)))
+                    Depth -= change->getLimitedValue();
+                else if (auto Phi = dyn_cast<PHINode>(CI->getArgOperand(1))) {
+                    //This should really do a dataflow analysis but assuming worst case means that we will always have enough space
+                    uint64_t MinPhiDepth = std::numeric_limits<uint64_t>::max();
+                    for (Value *Incoming : Phi->incoming_values()) {
+                        MinPhiDepth = std::min(MinPhiDepth, cast<ConstantInt>(Incoming)->getLimitedValue());
+                    }
+                    Depth -= MinPhiDepth;
+                }
             }
             assert(Depth >= 0);
             if (Depth > MaxDepth)
@@ -175,7 +184,7 @@ static bool lowerExcHandlers(Function &F) {
     unsigned allocaAddressSpace = F.getParent()->getDataLayout().getAllocaAddrSpace();
     for (int i = 0; i < MaxDepth; ++i) {
         auto *buff = new AllocaInst(Type::getInt8Ty(F.getContext()), allocaAddressSpace,
-                handler_sz, Align(16), "", firstInst);
+                handler_sz, Align(16), "depth" + std::to_string(i), firstInst);
         if (allocaAddressSpace) {
             AddrSpaceCastInst *buff_casted = new AddrSpaceCastInst(buff, PointerType::get(F.getContext(), AddressSpace::Generic));
             buff_casted->insertAfter(buff);
@@ -232,7 +241,18 @@ static bool lowerExcHandlers(Function &F) {
     // Insert lifetime end intrinsics after every leave.
     for (auto it : LeaveDepth) {
         int StartDepth = it.second - 1;
-        int npops = cast<ConstantInt>(it.first->getArgOperand(1))->getLimitedValue();
+        uint64_t minPops = std::numeric_limits<uint64_t>::max();
+        if (auto change = dyn_cast<ConstantInt>(it.first->getArgOperand(1)))
+            minPops = change->getLimitedValue();
+        else if (auto Phi = dyn_cast<PHINode>(it.first->getArgOperand(1))) {
+            //This should really do a dataflow analysis but assuming worst case means that we will always have enough space
+            uint64_t MinPhiDepth = std::numeric_limits<uint64_t>::max();
+            for (Value *Incoming : Phi->incoming_values()) {
+                MinPhiDepth = std::min(MinPhiDepth, cast<ConstantInt>(Incoming)->getLimitedValue());
+            }
+            minPops = MinPhiDepth;
+        }
+        int npops = minPops;
         for (int i = 0; i < npops; ++i) {
             assert(StartDepth-i >= 0);
             Value *lifetime_args[] = {

test/exceptions.jl

@@ -400,3 +400,28 @@ end
 catch
     current_exceptions()
 end) == 2
+
+# Trigger merging of branches in LLVM so we have a dynamic pop_handler
+a1234123::Int = 3
+function poll_fd2()
+    local timer
+    try
+        try
+            try
+                global a1234123 = 1
+            finally
+                global a1234123 = 2
+            end
+            return events
+        catch ex
+            return FDEvent()
+        end
+    finally
+        if @isdefined(timer)
+            global a1234123 = 2
+        else
+            global a1234123 = 4
+        end
+    end
+end
+@test_throws UndefVarError poll_fd2() #shouldn't crash

test/llvmpasses/image-codegen.jl

@@ -14,7 +14,7 @@
 # CHECK-NOT: private global
 # CHECK: jl_global
 # COM: we emit both declarations and definitions, so we may see either style in the IR
-# CHECK-SAME: = {{(external )?}}global
+# CHECK-SAME: = {{(external )?}}
 # CHECK: julia_f_
 # CHECK-NOT: internal global
 # CHECK-NOT: private global

test/llvmpasses/lower-handlers.ll

@@ -1,31 +1,88 @@
 ; This file is a part of Julia. License is MIT: https://julialang.org/license
 
 ; RUN: opt --load-pass-plugin=libjulia-codegen%shlibext -passes='function(LowerExcHandlers)' -S %s | FileCheck %s
+; ModuleID = 'lower-handlers.ll'
+; ModuleID = 'lower-handlers2.ll'
+source_filename = "lower-handlers.ll"
 
-attributes #1 = { returns_twice }
-declare {i32, i8*} @julia.except_enter({}*) #1
-declare void @ijl_pop_handler({}*, i32)
-declare i8**** @julia.ptls_states()
-declare i8**** @julia.get_pgcstack()
+declare ptr @julia.get_pgcstack()
+
+declare i64 @ijl_excstack_state(ptr)
+
+declare { i32, ptr } @julia.except_enter(ptr)
+
+declare void @ijl_pop_handler(ptr, i32)
+
+declare void @ijl_pop_handler_noexcept(ptr, i32)
+
+declare void @ijl_restore_excstack(ptr, i64)
+
+declare ptr @julia.ptls_states()
 
 define void @simple() {
 top:
-    %pgcstack = call i8**** @julia.get_pgcstack()
+  %pgcstack = call ptr @julia.get_pgcstack()
 ; CHECK: call void @llvm.lifetime.start
 ; CHECK: call void @ijl_enter_handler
 ; CHECK: setjmp
-    %rb = call {i32, i8*} @julia.except_enter({}* null)
-    %r = extractvalue {i32, i8*} %rb, 0
-    %b = extractvalue {i32, i8*} %rb, 1
-    %cmp = icmp eq i32 %r, 0
-    br i1 %cmp, label %try, label %catch
-try:
-    %lcssa = phi {i32, i8*} [ %rb, %top ]
-    br label %after
-catch:
-    br label %after
-after:
-    call void @ijl_pop_handler({}* null, i32 1)
+  %rb = call { i32, ptr } @julia.except_enter(ptr null)
+  %r = extractvalue { i32, ptr } %rb, 0
+  %b = extractvalue { i32, ptr } %rb, 1
+  %cmp = icmp eq i32 %r, 0
+  br i1 %cmp, label %try, label %catch
+
+try:                                              ; preds = %top
+  %lcssa = phi { i32, ptr } [ %rb, %top ]
+  br label %after
+
+catch:                                            ; preds = %top
+  br label %after
+
+after:                                            ; preds = %catch, %try
+  call void @ijl_pop_handler(ptr null, i32 1)
+; CHECK: llvm.lifetime.end
+  ret void
+}
+
+define ptr addrspace(10) @julia_poll_fd2_1135(){
+top:
+; CHECK: %depth0 = alloca i8, i32 256, align 16
+; CHECK: %depth1 = alloca i8, i32 256, align 16
+  %pgcstack = call ptr @julia.get_pgcstack()
+  %current_task7 = getelementptr inbounds i8, ptr %pgcstack, i64 -112
+  %0 = call i64 @ijl_excstack_state(ptr nonnull %current_task7)
+; CHECK: call void @llvm.lifetime.start
+; CHECK: call void @ijl_enter_handler
+; CHECK: setjmp
+  %1 = call { i32, ptr } @julia.except_enter(ptr nonnull %current_task7)
+  %2 = extractvalue { i32, ptr } %1, 0
+  %.not2 = icmp eq i32 %2, 0
+  br i1 %.not2, label %try, label %L50
+
+L50:                                              ; preds = %top
+  call void @ijl_pop_handler(ptr nonnull %current_task7, i32 1)
+; CHECK: llvm.lifetime.end
+  unreachable
+
+try:                                              ; preds = %top
+  %3 = call i64 @ijl_excstack_state(ptr nonnull %current_task7)
+; CHECK: call void @llvm.lifetime.start
+; CHECK: call void @ijl_enter_handler
+; CHECK: setjmp
+  %4 = call { i32, ptr } @julia.except_enter(ptr nonnull %current_task7)
+  %5 = extractvalue { i32, ptr } %4, 0
+  %6 = icmp eq i32 %5, 0
+  br i1 %6, label %common.ret, label %catch_pop14
+
+catch_pop14:                                      ; preds = %try
+  call void @ijl_pop_handler(ptr nonnull %current_task7, i32 1)
+; CHECK: llvm.lifetime.end
+  call void @ijl_restore_excstack(ptr nonnull %current_task7, i64 %3)
+  br label %common.ret
+
+common.ret:                                       ; preds = %catch_pop14, %try
+  %.sink = phi i32 [ 1, %catch_pop14 ], [ 2, %try ]
+  call void @ijl_pop_handler_noexcept(ptr nonnull %current_task7, i32 %.sink)
 ; CHECK: llvm.lifetime.end
-    ret void
+  ret ptr addrspace(10) null
 }

test/llvmpasses/pipeline-prints.ll

@@ -285,25 +285,19 @@ attributes #2 = { inaccessiblemem_or_argmemonly }
 
 ; COM: InstSimplify/InstCombine should kill this zext-trunc pair
 ; AFTEREARLYSIMPLIFICATION: [[ZEXT:%.*]] = zext i1 {{%.*}} to i8
-; AFTEREARLYSIMPLIFICATION-NEXT: trunc i8 [[ZEXT]] to i1
 
-; BEFOREEARLYOPTIMIZATION: [[ZEXT:%.*]] = zext i1 {{%.*}} to i8
-; BEFOREEARLYOPTIMIZATION-NEXT: trunc i8 [[ZEXT]] to i1
 
 ; AFTEREARLYOPTIMIZATION-NOT: zext i1 {{%.*}} to i8
-; AFTEREARLYOPTIMIZATION-NOT: trunc i8 {{%.*}} to i1
 
 ; BEFORELOOPOPTIMIZATION-NOT: zext i1 {{%.*}} to i8
-; BEFORELOOPOPTIMIZATION-NOT: trunc i8 {{%.*}} to i1
 
 ; COM: Loop simplification makes the exit condition obvious
 ; AFTERLOOPSIMPLIFICATION: L35.lr.ph:
 ; AFTERLOOPSIMPLIFICATION: add nuw nsw
 
-; COM: Scalar optimization removes the previous add from the preheader
-; AFTERSCALAROPTIMIZATION: L35.lr.ph:
-; AFTERSCALAROPTIMIZATION-NOT: add nuw nsw
-; AFTERSCALAROPTIMIZATION: br label %L35
+; COM: Scalar optimization removes the preheader
+; AFTERSCALAROPTIMIZATION: L17:
+; AFTERSCALAROPTIMIZATION: icmp eq i64 {{%.*}}, 1,
 
 ; COM: Vectorization does stuff
 ; AFTERVECTORIZATION: vector.body