RPCS3/llvm
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm.git synced 2024-11-29 06:30:30 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

when we see a unaligned load from an insufficiently aligned global or

Browse Source 
alloca, increase the alignment of the load, turning it into an aligned load.

This allows us to compile:

#include <xmmintrin.h>
__m128i foo(__m128i x){
 static const unsigned int c_0[4] = { 0, 0, 0, 0 };
	  __m128i v_Zero = _mm_loadu_si128((__m128i*)c_0);
  x  = _mm_unpacklo_epi8(x,  v_Zero);
  return x;
}

into:

_foo:
	punpcklbw	_c_0.5944, %xmm0
	ret
	.data
	.lcomm	_c_0.5944,16,4		# c_0.5944

instead of:

_foo:
	movdqu	_c_0.5944, %xmm1
	punpcklbw	%xmm1, %xmm0
	ret
	.data
	.lcomm	_c_0.5944,16,2		# c_0.5944



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40971 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Chris Lattner 2007-08-09 19:05:49 +00:00
parent a333b41af9
commit f2369f2042
2 changed files with 53 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
lib/Transforms/Scalar/InstructionCombining.cpp
View File

@ -7492,13 +7492,23 @@ Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
return 0;
}
/// GetKnownAlignment - If the specified pointer has an alignment that we can
/// determine, return it, otherwise return 0.
static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
/// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
/// we can determine, return it, otherwise return 0. If PrefAlign is specified,
/// and it is more than the alignment of the ultimate object, see if we can
/// increase the alignment of the ultimate object, making this check succeed.
static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
unsigned PrefAlign = 0) {
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
unsigned Align = GV->getAlignment();
if (Align == 0 && TD)
Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
// If there is a large requested alignment and we can, bump up the alignment
// of the global.
if (PrefAlign > Align && GV->hasInitializer()) {
GV->setAlignment(PrefAlign);
Align = PrefAlign;
}
return Align;
} else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
unsigned Align = AI->getAlignment();
@ -7516,18 +7526,20 @@ static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
(unsigned)TD->getABITypeAlignment(Type::Int64Ty));
}
}
// If there is a requested alignment and if this is an alloca, round up. We
// don't do this for malloc, because some systems can't respect the request.
if (PrefAlign > Align && isa<AllocaInst>(AI)) {
AI->setAlignment(PrefAlign);
Align = PrefAlign;
}
return Align;
} else if (isa<BitCastInst>(V) ||
(isa<ConstantExpr>(V) &&
cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
User *CI = cast<User>(V);
if (isa<PointerType>(CI->getOperand(0)->getType()))
return GetKnownAlignment(CI->getOperand(0), TD);
return 0;
return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
TD, PrefAlign);
} else if (User *GEPI = dyn_castGetElementPtr(V)) {
unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
if (BaseAlignment == 0) return 0;
// If all indexes are zero, it is just the alignment of the base pointer.
bool AllZeroOperands = true;
for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
@ -7536,8 +7548,14 @@ static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
AllZeroOperands = false;
break;
}
if (AllZeroOperands)
return BaseAlignment;
if (AllZeroOperands) {
// Treat this like a bitcast.
return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
}
unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
if (BaseAlignment == 0) return 0;
// Otherwise, if the base alignment is >= the alignment we expect for the
// base pointer type, then we know that the resultant pointer is aligned at
@ -7608,15 +7626,15 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
// If we can determine a pointer alignment that is bigger than currently
// set, update the alignment.
if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
unsigned Align = std::min(Alignment1, Alignment2);
if (MI->getAlignment()->getZExtValue() < Align) {
MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
Changed = true;
}
} else if (isa<MemSetInst>(MI)) {
unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
if (MI->getAlignment()->getZExtValue() < Alignment) {
MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
Changed = true;
@ -7634,7 +7652,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
case Intrinsic::x86_sse2_loadu_dq:
// Turn PPC lvx -> load if the pointer is known aligned.
// Turn X86 loadups -> load if the pointer is known aligned.
if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
PointerType::get(II->getType()), CI);
return new LoadInst(Ptr);
@ -7643,7 +7661,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
case Intrinsic::ppc_altivec_stvx:
case Intrinsic::ppc_altivec_stvxl:
// Turn stvx -> store if the pointer is known aligned.
if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
OpPtrTy, CI);
@ -7655,7 +7673,7 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
case Intrinsic::x86_sse2_storeu_dq:
case Intrinsic::x86_sse2_storel_dq:
// Turn X86 storeu -> store if the pointer is known aligned.
if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
OpPtrTy, CI);
@ -8768,7 +8786,7 @@ Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
Value *Op = LI.getOperand(0);
// Attempt to improve the alignment.
unsigned KnownAlign = GetKnownAlignment(Op, TD);
unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
if (KnownAlign > LI.getAlignment())
LI.setAlignment(KnownAlign);
@ -8968,7 +8986,7 @@ Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
}
// Attempt to improve the alignment.
unsigned KnownAlign = GetKnownAlignment(Ptr, TD);
unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
if (KnownAlign > SI.getAlignment())
SI.setAlignment(KnownAlign);

14
test/Transforms/InstCombine/align-inc.ll Normal file
View File

@ -0,0 +1,14 @@
; RUN: llvm-as < %s | opt -instcombine | llvm-dis | grep {GLOBAL.*align 16}
; RUN: llvm-as < %s | opt -instcombine | llvm-dis | grep {tmp = load}
@GLOBAL = internal constant [4 x i32] zeroinitializer
declare <16 x i8> @llvm.x86.sse2.loadu.dq(i8*)
define <16 x i8> @foo(<2 x i64> %x) {
entry:
%tmp = tail call <16 x i8> @llvm.x86.sse2.loadu.dq( i8* bitcast ([4 x i32]* @GLOBAL to i8*) )
ret <16 x i8> %tmp
}