Make X86-64 in the Large model always emit 64-bit calls.
The large code model is documented at
http://www.x86-64.org/documentation/abi.pdf and says that calls should
assume their target doesn't live within the 32-bit pc-relative offset
that fits in the call instruction.
To do this, we turn off the global-address->target-global-address
conversion in X86TargetLowering::LowerCall(). The first attempt at
this broke the lazy JIT because it can separate the movabs(imm->reg)
from the actual call instruction. The lazy JIT receives the address of
the movabs as a relocation and needs to record the return address from
the call; and then when that call happens, it needs to patch the
movabs with the newly-compiled target. We could thread the call
instruction into the relocation and record the movabs<->call mapping
explicitly, but that seems to require at least as much new
complication in the code generator as this change.
To fix this, we make lazy functions _always_ go through a call
stub. You'd think we'd only have to force lazy calls through a stub on
difficult platforms, but that turns out to break indirect calls
through a function pointer. The right fix for that is to distinguish
between calls and address-of operations on uncompiled functions, but
that's complex enough to leave for someone else to do.
Another attempt at this defined a new CALL64i pseudo-instruction,
which expanded to a 2-instruction sequence in the assembly output and
was special-cased in the X86CodeEmitter's emitInstruction()
function. That broke indirect calls in the same way as above.
This patch also removes a hack forcing Darwin to the small code model.
Without far-call-stubs, the small code model requires things of the
JITMemoryManager that the DefaultJITMemoryManager can't provide.
Thanks to echristo for lots of testing!
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@88984 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/unittests/ExecutionEngine/JIT/JITTest.cpp b/unittests/ExecutionEngine/JIT/JITTest.cpp
index 98b2922..b0c2f24 100644
--- a/unittests/ExecutionEngine/JIT/JITTest.cpp
+++ b/unittests/ExecutionEngine/JIT/JITTest.cpp
@@ -26,10 +26,22 @@
#include "llvm/Support/IRBuilder.h"
#include "llvm/Support/SourceMgr.h"
#include "llvm/Support/TypeBuilder.h"
+#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetSelect.h"
#include "llvm/Type.h"
#include <vector>
+#include <string.h>
+
+#if HAVE_ERRNO_H
+#include <errno.h>
+#endif
+#if HAVE_UNISTD_H
+#include <unistd.h>
+#endif
+#if _POSIX_MAPPED_FILES > 0
+#include <sys/mman.h>
+#endif
using namespace llvm;
@@ -177,6 +189,15 @@
}
};
+void LoadAssemblyInto(Module *M, const char *assembly) {
+ SMDiagnostic Error;
+ bool success = NULL != ParseAssemblyString(assembly, M, Error, M->getContext());
+ std::string errMsg;
+ raw_string_ostream os(errMsg);
+ Error.Print("", os);
+ ASSERT_TRUE(success) << os.str();
+}
+
class JITTest : public testing::Test {
protected:
virtual void SetUp() {
@@ -191,12 +212,7 @@
}
void LoadAssembly(const char *assembly) {
- SMDiagnostic Error;
- bool success = NULL != ParseAssemblyString(assembly, M, Error, Context);
- std::string errMsg;
- raw_string_ostream os(errMsg);
- Error.Print("", os);
- ASSERT_TRUE(success) << os.str();
+ LoadAssemblyInto(M, assembly);
}
LLVMContext Context;
@@ -498,6 +514,135 @@
}
#endif
+#if _POSIX_MAPPED_FILES > 0 && (defined (__x86_64__) || defined (_M_AMD64) || defined (_M_X64))
+class FarCallMemMgr : public RecordingJITMemoryManager {
+ void *MmapRegion;
+ size_t MmapSize;
+ uint8_t *NextStub;
+ uint8_t *NextFunction;
+
+ public:
+ FarCallMemMgr()
+ : MmapSize(16ULL << 30) { // 16GB
+ MmapRegion = mmap(NULL, MmapSize, PROT_READ | PROT_WRITE | PROT_EXEC,
+ MAP_PRIVATE | MAP_ANON, -1, 0);
+ if (MmapRegion == MAP_FAILED) {
+ ADD_FAILURE() << "mmap failed: " << strerror(errno);
+ }
+ // Set up the 16GB mapped region in several chunks:
+ // Stubs / ~5GB empty space / Function 1 / ~5GB empty space / Function 2
+ // This way no two entities can use a 32-bit relative call to reach each other.
+ NextStub = static_cast<uint8_t*>(MmapRegion);
+ NextFunction = NextStub + (5ULL << 30);
+
+ // Next, poison some of the memory so a wild call will eventually crash,
+ // even if memory was initialized by the OS to 0. We can't poison all of
+ // the memory because we want to be able to run on systems with less than
+ // 16GB of physical ram.
+ int TrapInstr = 0xCC; // INT 3
+ memset(NextStub, TrapInstr, 1<<10);
+ for (size_t Offset = 1<<30; Offset < MmapSize; Offset += 1<<30) {
+ // Fill the 2KB around each GB boundary with trap instructions. This
+ // should ensure that we can't run into emitted functions without hitting
+ // the trap.
+ memset(NextStub + Offset - (1<<10), TrapInstr, 2<<10);
+ }
+ }
+
+ ~FarCallMemMgr() {
+ EXPECT_EQ(0, munmap(MmapRegion, MmapSize));
+ }
+
+ virtual void setMemoryWritable() {}
+ virtual void setMemoryExecutable() {}
+ virtual uint8_t *startFunctionBody(const Function *F,
+ uintptr_t &ActualSize) {
+ ActualSize = 1 << 30;
+ uint8_t *Result = NextFunction;
+ NextFunction += 5ULL << 30;
+ return Result;
+ }
+ virtual void endFunctionBody(const Function*, uint8_t*, uint8_t*) {}
+ virtual uint8_t *allocateStub(const GlobalValue* F, unsigned StubSize,
+ unsigned Alignment) {
+ NextStub = reinterpret_cast<uint8_t*>(
+ uintptr_t(NextStub + Alignment - 1) &~ uintptr_t(Alignment - 1));
+ uint8_t *Result = NextStub;
+ NextStub += StubSize;
+ return Result;
+ }
+};
+
+class FarTargetTest : public ::testing::TestWithParam<CodeGenOpt::Level> {
+ protected:
+ FarTargetTest() : SavedCodeModel(TargetMachine::getCodeModel()) {}
+ ~FarTargetTest() {
+ TargetMachine::setCodeModel(SavedCodeModel);
+ }
+
+ const CodeModel::Model SavedCodeModel;
+};
+INSTANTIATE_TEST_CASE_P(CodeGenOpt,
+ FarTargetTest,
+ ::testing::Values(CodeGenOpt::None,
+ CodeGenOpt::Default));
+
+TEST_P(FarTargetTest, CallToFarTarget) {
+ // x86-64 can only make direct calls to functions within 32 bits of
+ // the current PC. To call anything farther away, we have to load
+ // the address into a register and call through the register. The
+ // old JIT did this by allocating a stub for any far call. However,
+ // that stub needed to be within 32 bits of the callsite. Here we
+ // test that the JIT correctly deals with stubs and calls more than
+ // 32 bits away from the callsite.
+
+ // Make sure the code generator is assuming code might be far away.
+ //TargetMachine::setCodeModel(CodeModel::Large);
+
+ LLVMContext Context;
+ Module *M = new Module("<main>", Context);
+ ExistingModuleProvider *MP = new ExistingModuleProvider(M);
+
+ JITMemoryManager *MemMgr = new FarCallMemMgr();
+ std::string Error;
+ OwningPtr<ExecutionEngine> JIT(EngineBuilder(MP)
+ .setEngineKind(EngineKind::JIT)
+ .setErrorStr(&Error)
+ .setJITMemoryManager(MemMgr)
+ .setOptLevel(GetParam())
+ .create());
+ ASSERT_EQ(Error, "");
+ TargetMachine::setCodeModel(CodeModel::Large);
+
+ LoadAssemblyInto(M,
+ "define i32 @test() { "
+ " ret i32 7 "
+ "} "
+ " "
+ "define i32 @test_far() { "
+ " %result = call i32 @test() "
+ " ret i32 %result "
+ "} ");
+ // First, lay out a function early in memory.
+ Function *TestFunction = M->getFunction("test");
+ int32_t (*TestFunctionPtr)() = reinterpret_cast<int32_t(*)()>(
+ (intptr_t)JIT->getPointerToFunction(TestFunction));
+ ASSERT_EQ(7, TestFunctionPtr());
+
+ // We now lay out the far-away function. This should land >4GB away from test().
+ Function *FarFunction = M->getFunction("test_far");
+ int32_t (*FarFunctionPtr)() = reinterpret_cast<int32_t(*)()>(
+ (intptr_t)JIT->getPointerToFunction(FarFunction));
+
+ EXPECT_LT(1LL << 32, llabs(intptr_t(FarFunctionPtr) - intptr_t(TestFunctionPtr)))
+ << "Functions must be >32 bits apart or the test is meaningless.";
+
+ // This used to result in a segfault in FarFunction, when its call instruction
+ // jumped to the wrong address.
+ EXPECT_EQ(7, FarFunctionPtr());
+}
+#endif // Platform has far-call problem.
+
// This code is copied from JITEventListenerTest, but it only runs once for all
// the tests in this directory. Everything seems fine, but that's strange
// behavior.