// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
+// This pass merges globals with internal linkage into one. This way all the
+// globals which were merged into a biggest one can be addressed using offsets
+// from the same base pointer (no need for separate base pointer for each of the
+// global). Such a transformation can significantly reduce the register pressure
+// when many globals are involved.
//
-//===----------------------------------------------------------------------===//
+// For example, consider the code which touches several global variables at once:
+//
+// static int foo[N], bar[N], baz[N];
+//
+// for (i = 0; i < N; ++i) {
+// foo[i] = bar[i] * baz[i];
+// }
+//
+// On ARM the addresses of 3 arrays should be kept in the registers, thus
+// this code has quite large register pressure (loop body):
+//
+// ldr r1, [r5], #4
+// ldr r2, [r6], #4
+// mul r1, r2, r1
+// str r1, [r0], #4
+//
+// Pass converts the code to something like:
+//
+// static struct {
+// int foo[N];
+// int bar[N];
+// int baz[N];
+// } merged;
+//
+// for (i = 0; i < N; ++i) {
+// merged.foo[i] = merged.bar[i] * merged.baz[i];
+// }
+//
+// and in ARM code this becomes:
+//
+// ldr r0, [r5, #40]
+// ldr r1, [r5, #80]
+// mul r0, r1, r0
+// str r0, [r5], #4
+//
+// note that we saved 2 registers here almostly "for free".
+// ===----------------------------------------------------------------------===//
#define DEBUG_TYPE "arm-global-merge"
#include "ARM.h"
using namespace llvm;
namespace {
- class VISIBILITY_HIDDEN ARMGlobalMerge : public FunctionPass {
+ class LLVM_LIBRARY_VISIBILITY ARMGlobalMerge : public FunctionPass {
/// TLI - Keep a pointer of a TargetLowering to consult for determining
/// target type sizes.
const TargetLowering *TLI;
- std::vector<GlobalVariable*> InternalGlobals;
+ bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
+ Module &M, bool) const;
public:
static char ID; // Pass identification, replacement for typeid.
const TargetData *TD;
GlobalCmp(const TargetData *td):
- TD(td) { };
+ TD(td) { }
bool operator() (const GlobalVariable* GV1,
const GlobalVariable* GV2) {
const Type* Ty1 = cast<PointerType>(GV1->getType())->getElementType();
const Type* Ty2 = cast<PointerType>(GV2->getType())->getElementType();
- return (TD->getTypeAllocSize(Ty1) <
- TD->getTypeAllocSize(Ty2));
+ return (TD->getTypeAllocSize(Ty1) < TD->getTypeAllocSize(Ty2));
}
};
};
char ARMGlobalMerge::ID = 0;
-#define MAX_OFFSET 4095
-
-bool ARMGlobalMerge::doInitialization(Module& M) {
+bool ARMGlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
+ Module &M, bool isConst) const {
const TargetData *TD = TLI->getTargetData();
- for (Module::global_iterator I = M.global_begin(),
- E = M.global_end(); I != E; ++I) {
- // FIXME: Can we just grab all 'local' vars here?
- // Won't we break some semantics?
- if (I->hasInternalLinkage() &&
- TD->getTypeAllocSize(I->getType()) < MAX_OFFSET)
- InternalGlobals.push_back(I);
- }
+ // FIXME: Infer the maximum possible offset depending on the actual users
+ // (these max offsets are different for the users inside Thumb or ARM
+ // functions)
+ unsigned MaxOffset = TLI->getMaximalGlobalOffset();
// FIXME: Find better heuristics
- std::stable_sort(InternalGlobals.begin(), InternalGlobals.end(),
- GlobalCmp(TD));
+ std::stable_sort(Globals.begin(), Globals.end(), GlobalCmp(TD));
const Type *Int32Ty = Type::getInt32Ty(M.getContext());
- for (size_t i = 0, e = InternalGlobals.size(); i != e; ) {
+ for (size_t i = 0, e = Globals.size(); i != e; ) {
size_t j = 0;
uint64_t MergedSize = 0;
std::vector<const Type*> Tys;
std::vector<Constant*> Inits;
- for (j = i; MergedSize < MAX_OFFSET && j != e; ++j) {
- const Type* Ty =
- cast<PointerType>(InternalGlobals[j]->getType())->getElementType();
+ for (j = i; MergedSize < MaxOffset && j != e; ++j) {
+ const Type* Ty = Globals[j]->getType()->getElementType();
Tys.push_back(Ty);
- Inits.push_back(InternalGlobals[j]->getInitializer());
+ Inits.push_back(Globals[j]->getInitializer());
MergedSize += TD->getTypeAllocSize(Ty);
}
StructType* MergedTy = StructType::get(M.getContext(), Tys);
Constant* MergedInit = ConstantStruct::get(MergedTy, Inits);
- // FIXME: Should we handle constants and 'normal' globals separately?
- GlobalVariable* MergedGV = new GlobalVariable(M, MergedTy, false,
+ GlobalVariable* MergedGV = new GlobalVariable(M, MergedTy, isConst,
GlobalValue::InternalLinkage,
MergedInit, "merged");
for (size_t k = i; k < j; ++k) {
ConstantExpr::getInBoundsGetElementPtr(MergedGV,
&Idx[0], Idx.size());
- InternalGlobals[k]->replaceAllUsesWith(GEP);
+ Globals[k]->replaceAllUsesWith(GEP);
+ Globals[k]->eraseFromParent();
}
i = j;
}
return true;
}
+
+bool ARMGlobalMerge::doInitialization(Module& M) {
+ SmallVector<GlobalVariable*, 16> Globals, ConstGlobals;
+ const TargetData *TD = TLI->getTargetData();
+ unsigned MaxOffset = TLI->getMaximalGlobalOffset();
+ bool Changed = false;
+
+ // Grab all non-const globals.
+ for (Module::global_iterator I = M.global_begin(),
+ E = M.global_end(); I != E; ++I) {
+ // Merge is safe for "normal" internal globals only
+ if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
+ continue;
+
+ // Ignore fancy-aligned globals for now.
+ if (I->getAlignment() != 0)
+ continue;
+
+ // Ignore all 'special' globals.
+ if (I->getName().startswith("llvm.") ||
+ I->getName().startswith(".llvm."))
+ continue;
+
+ if (TD->getTypeAllocSize(I->getType()) < MaxOffset) {
+ if (I->isConstant())
+ ConstGlobals.push_back(I);
+ else
+ Globals.push_back(I);
+ }
+ }
+
+ if (Globals.size() > 1)
+ Changed |= doMerge(Globals, M, false);
+ // FIXME: This currently breaks the EH processing due to way how the
+ // typeinfo detection works. We might want to detect the TIs and ignore
+ // them in the future.
+
+ // if (ConstGlobals.size() > 1)
+ // Changed |= doMerge(ConstGlobals, M, true);
+
+ return Changed;
+}
+
bool ARMGlobalMerge::runOnFunction(Function& F) {
return false;
}