[NEON] Support vldNq intrinsics in AArch32 (Clang part)
This patch reworks the support for dup NEON intrinsics as
described in D48439.
Differential Revision: https://reviews.llvm.org/D48440
llvm-svn: 335734
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 09edfa7..3095284 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -4081,16 +4081,22 @@
NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0),
NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0),
NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0),
+ NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0),
NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
NEONMAP1(vld2_v, arm_neon_vld2, 0),
+ NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0),
NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
NEONMAP1(vld2q_v, arm_neon_vld2, 0),
+ NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0),
NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
NEONMAP1(vld3_v, arm_neon_vld3, 0),
+ NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0),
NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
NEONMAP1(vld3q_v, arm_neon_vld3, 0),
+ NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0),
NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
NEONMAP1(vld4_v, arm_neon_vld4, 0),
+ NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0),
NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
NEONMAP1(vld4q_v, arm_neon_vld4, 0),
NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
@@ -4980,7 +4986,13 @@
case NEON::BI__builtin_neon_vld3_v:
case NEON::BI__builtin_neon_vld3q_v:
case NEON::BI__builtin_neon_vld4_v:
- case NEON::BI__builtin_neon_vld4q_v: {
+ case NEON::BI__builtin_neon_vld4q_v:
+ case NEON::BI__builtin_neon_vld2_dup_v:
+ case NEON::BI__builtin_neon_vld2q_dup_v:
+ case NEON::BI__builtin_neon_vld3_dup_v:
+ case NEON::BI__builtin_neon_vld3q_dup_v:
+ case NEON::BI__builtin_neon_vld4_dup_v:
+ case NEON::BI__builtin_neon_vld4q_dup_v: {
llvm::Type *Tys[] = {Ty, Int8PtrTy};
Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
Value *Align = getAlignmentValue32(PtrOp1);
@@ -5866,8 +5878,11 @@
case NEON::BI__builtin_neon_vld4_lane_v:
case NEON::BI__builtin_neon_vld4q_lane_v:
case NEON::BI__builtin_neon_vld2_dup_v:
+ case NEON::BI__builtin_neon_vld2q_dup_v:
case NEON::BI__builtin_neon_vld3_dup_v:
+ case NEON::BI__builtin_neon_vld3q_dup_v:
case NEON::BI__builtin_neon_vld4_dup_v:
+ case NEON::BI__builtin_neon_vld4q_dup_v:
// Get the alignment for the argument in addition to the value;
// we'll use it later.
PtrOp1 = EmitPointerWithAlignment(E->getArg(1));
@@ -6044,68 +6059,6 @@
Value *Ld = Builder.CreateLoad(PtrOp0);
return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
}
- case NEON::BI__builtin_neon_vld2_dup_v:
- case NEON::BI__builtin_neon_vld3_dup_v:
- case NEON::BI__builtin_neon_vld4_dup_v: {
- // Handle 64-bit elements as a special-case. There is no "dup" needed.
- if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) {
- switch (BuiltinID) {
- case NEON::BI__builtin_neon_vld2_dup_v:
- Int = Intrinsic::arm_neon_vld2;
- break;
- case NEON::BI__builtin_neon_vld3_dup_v:
- Int = Intrinsic::arm_neon_vld3;
- break;
- case NEON::BI__builtin_neon_vld4_dup_v:
- Int = Intrinsic::arm_neon_vld4;
- break;
- default: llvm_unreachable("unknown vld_dup intrinsic?");
- }
- llvm::Type *Tys[] = {Ty, Int8PtrTy};
- Function *F = CGM.getIntrinsic(Int, Tys);
- llvm::Value *Align = getAlignmentValue32(PtrOp1);
- Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup");
- Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
- Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
- return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
- }
- switch (BuiltinID) {
- case NEON::BI__builtin_neon_vld2_dup_v:
- Int = Intrinsic::arm_neon_vld2lane;
- break;
- case NEON::BI__builtin_neon_vld3_dup_v:
- Int = Intrinsic::arm_neon_vld3lane;
- break;
- case NEON::BI__builtin_neon_vld4_dup_v:
- Int = Intrinsic::arm_neon_vld4lane;
- break;
- default: llvm_unreachable("unknown vld_dup intrinsic?");
- }
- llvm::Type *Tys[] = {Ty, Int8PtrTy};
- Function *F = CGM.getIntrinsic(Int, Tys);
- llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType());
-
- SmallVector<Value*, 6> Args;
- Args.push_back(Ops[1]);
- Args.append(STy->getNumElements(), UndefValue::get(Ty));
-
- llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
- Args.push_back(CI);
- Args.push_back(getAlignmentValue32(PtrOp1));
-
- Ops[1] = Builder.CreateCall(F, Args, "vld_dup");
- // splat lane 0 to all elts in each vector of the result.
- for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
- Value *Val = Builder.CreateExtractValue(Ops[1], i);
- Value *Elt = Builder.CreateBitCast(Val, Ty);
- Elt = EmitNeonSplat(Elt, CI);
- Elt = Builder.CreateBitCast(Elt, Val->getType());
- Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i);
- }
- Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
- Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
- return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
- }
case NEON::BI__builtin_neon_vqrshrn_n_v:
Int =
usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;