Fix performance problem in root coloring

By originally using a vector, rather than a set, neighbors would
end up getting duplicated in the list. This blew up the size of
the list, causing excessive performance problems in the ColorRoots
function.

src/codegen.cpp

@@ -64,6 +64,7 @@
 #include <llvm/Support/SourceMgr.h> // for llvmcall
 #include <llvm/Transforms/Utils/Cloning.h> // for llvmcall inlining
 #include <llvm/IR/Verifier.h> // for llvmcall validation
+#include <llvm/Bitcode/ReaderWriter.h>
 
 // C API
 #include <llvm-c/Types.h>
@@ -7145,3 +7146,15 @@ extern "C" void jl_dump_llvm_type(void *v)
 #endif
     putchar('\n');
 }
+
+extern void jl_write_bitcode_func(void *F, char *fname) {
+    std::error_code EC;
+    raw_fd_ostream OS(fname, EC, sys::fs::F_None);
+    llvm::WriteBitcodeToFile(((llvm::Function*)F)->getParent(), OS);
+}
+
+extern void jl_write_bitcode_module(void *M, char *fname) {
+    std::error_code EC;
+    raw_fd_ostream OS(fname, EC, sys::fs::F_None);
+    llvm::WriteBitcodeToFile((llvm::Module*)M, OS);
+}

src/llvm-late-gc-lowering.cpp

@@ -2,6 +2,7 @@
 
 #include <llvm/ADT/BitVector.h>
 #include <llvm/ADT/PostOrderIterator.h>
+#include <llvm/ADT/SetVector.h>
 #include "llvm/Analysis/CFG.h"
 #include <llvm/IR/Value.h>
 #include <llvm/IR/Constants.h>
@@ -248,7 +249,7 @@ struct State {
     std::map<int, Value *> ReversePtrNumbering;
     // Neighbors in the coloring interference graph. I.e. for each value, the
     // indices of other values that are used simultaneously at some safe point.
-    std::vector<std::vector<int>> Neighbors;
+    std::vector<SetVector<int>> Neighbors;
     // The result of the local analysis
     std::map<BasicBlock *, BBState> BBStates;
 
@@ -823,17 +824,19 @@ void LateLowerGCFrame::ComputeLiveSets(Function &F, State &S) {
     }
     // Compute the interference graph
     for (int i = 0; i <= S.MaxPtrNumber; ++i) {
-        std::vector<int> Neighbors;
+        SetVector<int> Neighbors;
+        BitVector NeighborBits(S.MaxPtrNumber);
         for (auto it : S.SafepointNumbering) {
             const BitVector &LS = S.LiveSets[it.second];
             if ((unsigned)i >= LS.size() || !LS[i])
                 continue;
-            for (int Idx = LS.find_first(); Idx >= 0; Idx = LS.find_next(Idx)) {
-                // We explicitly let i be a neighbor of itself, to distinguish
-                // between being the only value live at a safepoint, vs not
-                // being live at any safepoint.
-                Neighbors.push_back(Idx);
-            }
+            NeighborBits |= LS;
+        }
+        for (int Idx = NeighborBits.find_first(); Idx >= 0; Idx = NeighborBits.find_next(Idx)) {
+            // We explicitly let i be a neighbor of itself, to distinguish
+            // between being the only value live at a safepoint, vs not
+            // being live at any safepoint.
+            Neighbors.insert(Idx);
         }
         S.Neighbors.push_back(Neighbors);
     }
@@ -886,8 +889,8 @@ struct PEOIterator {
     };
     std::vector<Element> Elements;
     std::vector<std::vector<int>> Levels;
-    const std::vector<std::vector<int>> &Neighbors;
-    PEOIterator(const std::vector<std::vector<int>> &Neighbors) : Neighbors(Neighbors) {
+    const std::vector<SetVector<int>> &Neighbors;
+    PEOIterator(const std::vector<SetVector<int>> &Neighbors) : Neighbors(Neighbors) {
         // Initialize State
         std::vector<int> FirstLevel;
         for (unsigned i = 0; i < Neighbors.size(); ++i) {

test/llvmpasses/lit.cfg

@@ -7,9 +7,10 @@ import lit.util
 import lit.formats
 
 config.name = 'Julia'
-config.suffixes = ['.ll']
+config.suffixes = ['.ll','.jl']
 config.test_source_root = os.path.dirname(__file__)
-config.test_format = lit.formats.ShTest(False)
+config.test_format = lit.formats.ShTest(True)
 
-path = os.path.pathsep.join((os.path.join(os.path.dirname(__file__),"../../usr/tools"), config.environment['PATH']))
+path = os.path.pathsep.join((os.path.join(os.path.dirname(__file__),"../../usr/tools"), os.path.join(os.path.dirname(__file__),"../../usr/bin"), config.environment['PATH']))
 config.environment['PATH'] = path
+config.environment['HOME'] = "/tmp"

test/llvmpasses/safepoint_stress.jl

@@ -0,0 +1,25 @@
+# RUN: julia --startup-file=no %s | opt -load libjulia.so -LateLowerGCFrame -S - | FileCheck %s
+
+println("""
+%jl_value_t = type opaque
+declare %jl_value_t addrspace(10)* @alloc()
+declare void @one_arg_boxed(%jl_value_t addrspace(10)*)
+declare %jl_value_t*** @jl_get_ptls_states()
+
+define void @stress(i64 %a, i64 %b) {
+    %ptls = call %jl_value_t*** @jl_get_ptls_states()
+""")
+
+# CHECK: %gcframe = alloca %jl_value_t addrspace(10)*, i32 10002
+for i = 1:10000
+    println("\t%arg$i = call %jl_value_t addrspace(10)* @alloc()")
+end
+
+for i = 1:10000
+    println("\tcall void @one_arg_boxed(%jl_value_t addrspace(10)* %arg$i)")
+end
+
+println("""
+    ret void
+}
+""")