Luke Drummond 19459580af Add IR fixups for RenderScript ABI mismatch between ARMV7 frontend and x86 backend
Expression evaluation for function calls to certain public RenderScript
API functions in libRSCPURef can segfault.

`slang`,
the compiler frontend for RenderScript embeds an ARM specific triple in
IR that is shipped in the app, after generating IR that has some
assumptions that an ARM device is the target.
As the IR is then compiled on a device of unknown (at time the IR was
generated at least) architecture, when calling RenderScript API function
as part of debugger expressions, we have to perform a fixup pass that
removes those assumptions right before the module is sent to be
generated by the llvm backend.

This issue is caused by multiple problems with the ARMv7-specific
assumptions encoded in the LLVM IR. x86 large value returns use a hidden
first argument (mapping to llvm::Attribute::StructRet), which can't be
picked up by the JIT due to the mismatch between IR generated by the
slang frontend and llvm backend. This means that code generated by bcc
did not necessarily match the default SysV Linux/Android ABI used by the
LLDB JIT

- Original Authors: Luke Drummond (@ldrumm), Function declarations fixed by Aidan Dodds (@ADodds)

Subscribers: lldb-commits

Differential Revision: https://reviews.llvm.org/D18059

llvm-svn: 276976
2016-07-28 14:21:07 +00:00

287 lines
11 KiB
C++

//===-- x86ABIFixups.cpp ----------------------------------------*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
// C Includes
// C++ Includes
#include <set>
// Other libraries and framework includes
#include "llvm/ADT/StringRef.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CallSite.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Module.h"
#include "llvm/IRReader/IRReader.h"
#include "llvm/Pass.h"
// Project includes
#include "lldb/Core/Log.h"
#include "lldb/Target/Process.h"
using namespace lldb_private;
namespace
{
bool
isRSAPICall(llvm::Module &module, llvm::CallInst *call_inst)
{
// TODO get the list of renderscript modules from lldb and check if
// this llvm::Module calls into any of them.
(void)module;
const auto func_name = call_inst->getCalledFunction()->getName();
if (func_name.startswith("llvm") || func_name.startswith("lldb"))
return false;
if (call_inst->getCalledFunction()->isIntrinsic())
return false;
return true;
}
bool
isRSLargeReturnCall(llvm::Module &module, llvm::CallInst *call_inst)
{
// i686 and x86_64 returns for large vectors in the RenderScript API are not handled as normal
// register pairs, but as a hidden sret type. This is not reflected in the debug info or mangled
// symbol name, and the android ABI for x86 and x86_64, (as well as the emulators) specifies there is
// no AVX, so bcc generates an sret function because we cannot natively return 256 bit vectors.
// This function simply checks whether a function has a > 128bit return type. It is perhaps an
// unreliable heuristic, and relies on bcc not generating AVX code, so if the android ABI one day
// provides for AVX, this function may go out of fashion.
(void)module;
if (!call_inst || !call_inst->getCalledFunction())
return false;
return call_inst->getCalledFunction()->getReturnType()->getPrimitiveSizeInBits() > 128;
}
bool
isRSAllocationPtrTy(const llvm::Type *type)
{
if (!type->isPointerTy())
return false;
auto ptr_type = type->getPointerElementType();
return ptr_type->isStructTy() && ptr_type->getStructName().startswith("struct.rs_allocation");
}
bool
isRSAllocationTyCallSite(llvm::Module &module, llvm::CallInst *call_inst)
{
(void)module;
if (!call_inst->hasByValArgument())
return false;
for (const auto &param : call_inst->operand_values())
if (isRSAllocationPtrTy(param->getType()))
return true;
return false;
}
llvm::FunctionType *
cloneToStructRetFnTy(llvm::CallInst *call_inst)
{
// on x86 StructReturn functions return a pointer to the return value, rather than the return
// value itself [ref](http://www.agner.org/optimize/calling_conventions.pdf section 6).
// We create a return type by getting the pointer type of the old return type, and inserting a new
// initial argument of pointer type of the original return type.
Log *log(GetLogIfAnyCategoriesSet(LIBLLDB_LOG_LANGUAGE | LIBLLDB_LOG_EXPRESSIONS));
assert(call_inst && "no CallInst");
llvm::Function *orig = call_inst->getCalledFunction();
assert(orig && "CallInst has no called function");
llvm::FunctionType *orig_type = orig->getFunctionType();
auto name = orig->getName();
if (log)
log->Printf("%s - cloning to StructRet function for '%s'", __FUNCTION__, name.str().c_str());
std::vector<llvm::Type *> new_params{orig_type->getNumParams() + 1, nullptr};
unsigned num_params = orig_type->getNumParams();
std::vector<llvm::Type *> params{orig_type->param_begin(), orig_type->param_end()};
// This may not work if the function is somehow declared void as llvm is strongly typed
// and represents void* with i8*
assert(!orig_type->getReturnType()->isVoidTy() && "Cannot add StructRet attribute to void function");
llvm::PointerType *return_type_ptr_type = llvm::PointerType::getUnqual(orig->getReturnType());
assert(return_type_ptr_type && "failed to get function return type PointerType");
if (!return_type_ptr_type)
return nullptr;
if (log)
log->Printf("%s - return type pointer type for StructRet clone @ '0x%p':\n", __FUNCTION__,
(void *)return_type_ptr_type);
// put the the sret pointer argument in place at the beginning of the argument list.
params.emplace(params.begin(), return_type_ptr_type);
assert(params.size() == num_params + 1);
return llvm::FunctionType::get(return_type_ptr_type, params, orig->isVarArg());
}
bool
findRSCallSites(llvm::Module &module, std::set<llvm::CallInst *> &rs_callsites,
bool (*predicate)(llvm::Module &, llvm::CallInst *))
{
bool found = false;
for (auto &func : module.getFunctionList())
for (auto &block : func.getBasicBlockList())
for (auto &inst : block)
{
llvm::CallInst *call_inst = llvm::dyn_cast_or_null<llvm::CallInst>(&inst);
if (!call_inst || !call_inst->getCalledFunction())
// This is not the call-site you are looking for...
continue;
if (isRSAPICall(module, call_inst) && predicate(module, call_inst))
{
rs_callsites.insert(call_inst);
found = true;
}
}
return found;
}
bool
fixupX86StructRetCalls(llvm::Module &module)
{
bool changed = false;
// changing a basic block while iterating over it seems to have some undefined behaviour
// going on so we find all RS callsites first, then fix them up after consuming
// the iterator.
std::set<llvm::CallInst *> rs_callsites;
if (!findRSCallSites(module, rs_callsites, isRSLargeReturnCall))
return false;
for (auto call_inst : rs_callsites)
{
llvm::FunctionType *new_func_type = cloneToStructRetFnTy(call_inst);
assert(new_func_type && "failed to clone functionType for Renderscript ABI fixup");
llvm::CallSite call_site(call_inst);
llvm::Function *func = call_inst->getCalledFunction();
assert(func && "cannot resolve function in RenderScriptRuntime");
// Copy the original call arguments
std::vector<llvm::Value *> new_call_args(call_site.arg_begin(), call_site.arg_end());
// Allocate enough space to store the return value of the original function
// we pass a pointer to this allocation as the StructRet param, and then copy its
// value into the lldb return value
llvm::AllocaInst *return_value_alloc =
new llvm::AllocaInst(func->getReturnType(), "var_vector_return_alloc", call_inst);
// use the new allocation as the new first argument
new_call_args.emplace(new_call_args.begin(), llvm::cast<llvm::Value>(return_value_alloc));
llvm::PointerType *new_func_ptr_type = llvm::PointerType::get(new_func_type, 0);
// Create the type cast from the old function type to the new one
llvm::Constant *new_func_cast =
llvm::ConstantExpr::getCast(llvm::Instruction::BitCast, func, new_func_ptr_type);
// create an allocation for a new function pointer
llvm::AllocaInst *new_func_ptr = new llvm::AllocaInst(new_func_ptr_type, "new_func_ptr", call_inst);
// store the new_func_cast to the newly allocated space
(void)new llvm::StoreInst(new_func_cast, new_func_ptr, "new_func_ptr_load_cast", call_inst);
// load the new function address ready for a jump
llvm::LoadInst *new_func_addr_load = new llvm::LoadInst(new_func_ptr, "load_func_pointer", call_inst);
// and create a callinstruction from it
llvm::CallInst *new_call_inst =
llvm::CallInst::Create(new_func_addr_load, new_call_args, "new_func_call", call_inst);
new_call_inst->setCallingConv(call_inst->getCallingConv());
new_call_inst->setTailCall(call_inst->isTailCall());
llvm::LoadInst *lldb_save_result_address = new llvm::LoadInst(return_value_alloc, "save_return_val", call_inst);
// Now remove the old broken call
call_inst->replaceAllUsesWith(lldb_save_result_address);
call_inst->eraseFromParent();
changed = true;
}
return changed;
}
bool
fixupRSAllocationStructByValCalls(llvm::Module &module)
{
// On x86_64, calls to functions in the RS runtime that take an `rs_allocation` type argument
// are actually handled as by-ref params by bcc, but appear to be passed by value by lldb (the callsite all use
// `struct byval`).
// On x86_64 Linux, struct arguments are transferred in registers if the struct size is no bigger than
// 128bits [ref](http://www.agner.org/optimize/calling_conventions.pdf) section 7.1 "Passing and returning objects"
// otherwise passed on the stack.
// an object of type `rs_allocation` is actually 256bits, so should be passed on the stack. However, code generated
// by bcc actually treats formal params of type `rs_allocation` as `rs_allocation *` so we need to convert the
// calling convention to pass by reference, and remove any hint of byval from formal parameters.
bool changed = false;
std::set<llvm::CallInst *> rs_callsites;
if (!findRSCallSites(module, rs_callsites, isRSAllocationTyCallSite))
return false;
std::set<llvm::Function *> rs_functions;
// for all call instructions
for (auto call_inst : rs_callsites)
{
// add the called function to a set so that we can strip its byval attributes in another pass
rs_functions.insert(call_inst->getCalledFunction());
// get the function attributes
llvm::AttributeSet call_attribs = call_inst->getAttributes();
// iterate over the argument attributes
for (size_t i = 1; i <= call_attribs.getNumSlots(); ++i)
{
// if this argument is passed by val
if (call_attribs.hasAttribute(i, llvm::Attribute::ByVal))
{
// strip away the byval attribute
call_inst->removeAttribute(i, llvm::Attribute::get(module.getContext(), llvm::Attribute::ByVal));
changed = true;
}
}
}
llvm::AttributeSet attr_byval = llvm::AttributeSet::get(module.getContext(), 1u, llvm::Attribute::ByVal);
// for all called function decls
for (auto func : rs_functions)
{
// inspect all of the arguments in the call
llvm::SymbolTableList<llvm::Argument> &argList = func->getArgumentList();
for (auto &arg : argList)
{
if (arg.hasByValAttr())
{
arg.removeAttr(attr_byval);
changed = true;
}
}
}
return changed;
}
} // end anonymous namespace
namespace lldb_private
{
namespace lldb_renderscript
{
bool
fixupX86FunctionCalls(llvm::Module &module)
{
return fixupX86StructRetCalls(module);
}
bool
fixupX86_64FunctionCalls(llvm::Module &module)
{
bool changed = false;
changed |= fixupX86StructRetCalls(module);
changed |= fixupRSAllocationStructByValCalls(module);
return changed;
}
} // end namespace lldb_renderscript
} // end namespace lldb_private