mirror of
https://github.com/intel/llvm.git
synced 2026-02-03 19:18:13 +08:00
[Hexagon] Make sure that offset on globals matches alignment requirements
A correctly aligned address may happen to be separated into a variable part and a constant part, where the constant part does not match the alignment needed in a load/store that uses this address. Such a constant cannot be used as an immediate offset in an indexed instruction. When lowering a global address, make sure that if there is an offset folded into the global, the offset is valid for all uses in load/store instructions. llvm-svn: 323562
This commit is contained in:
@@ -1472,32 +1472,58 @@ HexagonTargetLowering::LowerGLOBALADDRESS(SDValue Op, SelectionDAG &DAG) const {
|
||||
SDLoc dl(Op);
|
||||
auto *GAN = cast<GlobalAddressSDNode>(Op);
|
||||
auto PtrVT = getPointerTy(DAG.getDataLayout());
|
||||
auto *GV = GAN->getGlobal();
|
||||
int64_t Offset = GAN->getOffset();
|
||||
const GlobalValue *GV = GAN->getGlobal();
|
||||
int32_t Offset = GAN->getOffset();
|
||||
int32_t Addend = 0;
|
||||
|
||||
unsigned GlobAlign = GV->getAlignment();
|
||||
if (GlobAlign != 0 && Offset % GlobAlign != 0) {
|
||||
unsigned MinAlign = GlobAlign;
|
||||
for (SDNode *U : GAN->uses()) {
|
||||
if (auto *M = dyn_cast<MemSDNode>(U))
|
||||
MinAlign = std::min(MinAlign, M->getAlignment());
|
||||
}
|
||||
assert(isPowerOf2_32(MinAlign));
|
||||
if (Offset % MinAlign != 0) {
|
||||
Addend = Offset & (MinAlign-1); // Always non-negative.
|
||||
Offset -= Addend;
|
||||
}
|
||||
}
|
||||
|
||||
auto &HLOF = *HTM.getObjFileLowering();
|
||||
Reloc::Model RM = HTM.getRelocationModel();
|
||||
SDValue Res;
|
||||
|
||||
if (RM == Reloc::Static) {
|
||||
SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
|
||||
const GlobalObject *GO = GV->getBaseObject();
|
||||
if (GO && HLOF.isGlobalInSmallSection(GO, HTM))
|
||||
return DAG.getNode(HexagonISD::CONST32_GP, dl, PtrVT, GA);
|
||||
return DAG.getNode(HexagonISD::CONST32, dl, PtrVT, GA);
|
||||
Res = DAG.getNode(HexagonISD::CONST32_GP, dl, PtrVT, GA);
|
||||
else
|
||||
Res = DAG.getNode(HexagonISD::CONST32, dl, PtrVT, GA);
|
||||
} else {
|
||||
bool UsePCRel = HTM.shouldAssumeDSOLocal(*GV->getParent(), GV);
|
||||
if (UsePCRel) {
|
||||
SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset,
|
||||
HexagonII::MO_PCREL);
|
||||
Res = DAG.getNode(HexagonISD::AT_PCREL, dl, PtrVT, GA);
|
||||
} else {
|
||||
// Use GOT index.
|
||||
SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
|
||||
SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
|
||||
HexagonII::MO_GOT);
|
||||
SDValue Off = DAG.getConstant(Offset, dl, MVT::i32);
|
||||
Res = DAG.getNode(HexagonISD::AT_GOT, dl, PtrVT, GOT, GA, Off);
|
||||
}
|
||||
}
|
||||
|
||||
bool UsePCRel = getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV);
|
||||
if (UsePCRel) {
|
||||
SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset,
|
||||
HexagonII::MO_PCREL);
|
||||
return DAG.getNode(HexagonISD::AT_PCREL, dl, PtrVT, GA);
|
||||
assert(Res.getNode() != nullptr);
|
||||
if (Addend != 0) {
|
||||
SDValue A = DAG.getConstant(Addend, dl, MVT::i32);
|
||||
Res = DAG.getNode(ISD::ADD, dl, MVT::i32, Res, A);
|
||||
}
|
||||
|
||||
// Use GOT index.
|
||||
SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
|
||||
SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, HexagonII::MO_GOT);
|
||||
SDValue Off = DAG.getConstant(Offset, dl, MVT::i32);
|
||||
return DAG.getNode(HexagonISD::AT_GOT, dl, PtrVT, GOT, GA, Off);
|
||||
return Res;
|
||||
}
|
||||
|
||||
// Specifies that for loads and stores VT can be promoted to PromotedLdStVT.
|
||||
|
||||
32
llvm/test/CodeGen/Hexagon/isel-global-offset-alignment.ll
Normal file
32
llvm/test/CodeGen/Hexagon/isel-global-offset-alignment.ll
Normal file
@@ -0,0 +1,32 @@
|
||||
; RUN: llc -march=hexagon < %s | FileCheck %s
|
||||
|
||||
; This testcase checks that a valid offset is folded into a global address
|
||||
; when it's used in a load or a store instruction. The store in this code
|
||||
; is not really properly aligned (bugpoint output from a legal code), but
|
||||
; what's important is that the offset on the store instructions is a multiple
|
||||
; of the access size. In this case the actual address is @array+30, but that
|
||||
; value is not a multiple of 8, so it cannot appear as an immediate in memd.
|
||||
; Aside from the fact that @array+30 is not a valid address for memd, make
|
||||
; sure that in a memd instruction the offset field is a multiple of 8.
|
||||
|
||||
; CHECK: r[[BASE:[0-9]+]] = #6
|
||||
; CHECK-DAG: memd(r[[BASE]]+##array+24)
|
||||
; CHECK-DAG: memd(r[[BASE]]+##array+32)
|
||||
|
||||
|
||||
target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
|
||||
target triple = "hexagon"
|
||||
|
||||
@array = external global [1000000 x i16], align 8
|
||||
|
||||
define void @fred() #0 {
|
||||
b0:
|
||||
%v1 = add nsw i32 0, -1
|
||||
%v2 = getelementptr inbounds [1000000 x i16], [1000000 x i16]* @array, i32 0, i32 %v1
|
||||
%v3 = getelementptr i16, i16* %v2, i32 16
|
||||
%v4 = bitcast i16* %v3 to <8 x i16>*
|
||||
store <8 x i16> zeroinitializer, <8 x i16>* %v4, align 8
|
||||
ret void
|
||||
}
|
||||
|
||||
attributes #0 = { norecurse nounwind }
|
||||
Reference in New Issue
Block a user