r200471 - ARM & AArch64: fully share NEON implementation of permutation intrinsics
Tim Northover
tnorthover at apple.com
Thu Jan 30 06:47:58 PST 2014
Author: tnorthover
Date: Thu Jan 30 08:47:57 2014
New Revision: 200471
URL: http://llvm.org/viewvc/llvm-project?rev=200471&view=rev
Log:
ARM & AArch64: fully share NEON implementation of permutation intrinsics
As a starting point, this moves the CodeGen for NEON permutation
instructions (vtrn, vzip, vuzp) into a new shared function.
Modified:
cfe/trunk/lib/CodeGen/CGBuiltin.cpp
cfe/trunk/lib/CodeGen/CodeGenFunction.h
Modified: cfe/trunk/lib/CodeGen/CGBuiltin.cpp
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/CGBuiltin.cpp?rev=200471&r1=200470&r2=200471&view=diff
==============================================================================
--- cfe/trunk/lib/CodeGen/CGBuiltin.cpp (original)
+++ cfe/trunk/lib/CodeGen/CGBuiltin.cpp Thu Jan 30 08:47:57 2014
@@ -1754,6 +1754,76 @@ CodeGenFunction::EmitPointerWithAlignmen
return std::make_pair(EmitScalarExpr(Addr), Align);
}
+Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(unsigned BuiltinID,
+ const CallExpr *E,
+ SmallVectorImpl<Value *> &Ops,
+ llvm::VectorType *VTy) {
+ switch (BuiltinID) {
+ default: break;
+ case NEON::BI__builtin_neon_vtrn_v:
+ case NEON::BI__builtin_neon_vtrnq_v: {
+ Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
+ Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
+ Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
+ Value *SV = 0;
+
+ for (unsigned vi = 0; vi != 2; ++vi) {
+ SmallVector<Constant*, 16> Indices;
+ for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
+ Indices.push_back(Builder.getInt32(i+vi));
+ Indices.push_back(Builder.getInt32(i+e+vi));
+ }
+ Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
+ SV = llvm::ConstantVector::get(Indices);
+ SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
+ SV = Builder.CreateStore(SV, Addr);
+ }
+ return SV;
+ }
+ case NEON::BI__builtin_neon_vuzp_v:
+ case NEON::BI__builtin_neon_vuzpq_v: {
+ Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
+ Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
+ Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
+ Value *SV = 0;
+
+ for (unsigned vi = 0; vi != 2; ++vi) {
+ SmallVector<Constant*, 16> Indices;
+ for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
+ Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
+
+ Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
+ SV = llvm::ConstantVector::get(Indices);
+ SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
+ SV = Builder.CreateStore(SV, Addr);
+ }
+ return SV;
+ }
+ case NEON::BI__builtin_neon_vzip_v:
+ case NEON::BI__builtin_neon_vzipq_v: {
+ Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
+ Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
+ Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
+ Value *SV = 0;
+
+ for (unsigned vi = 0; vi != 2; ++vi) {
+ SmallVector<Constant*, 16> Indices;
+ for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
+ Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
+ Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
+ }
+ Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
+ SV = llvm::ConstantVector::get(Indices);
+ SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
+ SV = Builder.CreateStore(SV, Addr);
+ }
+ return SV;
+ }
+ }
+
+ return 0;
+}
+
static Value *EmitAArch64ScalarBuiltinExpr(CodeGenFunction &CGF,
unsigned BuiltinID,
const CallExpr *E) {
@@ -2981,6 +3051,11 @@ Value *CodeGenFunction::EmitAArch64Built
if (!Ty)
return 0;
+ // Many NEON builtins have identical semantics and uses in ARM and
+ // AArch64. Emit these in a single function.
+ if (Value *Result = EmitCommonNeonBuiltinExpr(BuiltinID, E, Ops, VTy))
+ return Result;
+
unsigned Int;
switch (BuiltinID) {
default:
@@ -2989,18 +3064,6 @@ Value *CodeGenFunction::EmitAArch64Built
// AArch64 builtins mapping to legacy ARM v7 builtins.
// FIXME: the mapped builtins listed correspond to what has been tested
// in aarch64-neon-intrinsics.c so far.
- case NEON::BI__builtin_neon_vuzp_v:
- return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vuzp_v, E);
- case NEON::BI__builtin_neon_vuzpq_v:
- return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vuzpq_v, E);
- case NEON::BI__builtin_neon_vzip_v:
- return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vzip_v, E);
- case NEON::BI__builtin_neon_vzipq_v:
- return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vzipq_v, E);
- case NEON::BI__builtin_neon_vtrn_v:
- return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vtrn_v, E);
- case NEON::BI__builtin_neon_vtrnq_v:
- return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vtrnq_v, E);
case NEON::BI__builtin_neon_vext_v:
return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vext_v, E);
case NEON::BI__builtin_neon_vextq_v:
@@ -4213,6 +4276,11 @@ Value *CodeGenFunction::EmitARMBuiltinEx
if (!Ty)
return 0;
+ // Many NEON builtins have identical semantics and uses in ARM and
+ // AArch64. Emit these in a single function.
+ if (Value *Result = EmitCommonNeonBuiltinExpr(BuiltinID, E, Ops, VTy))
+ return Result;
+
unsigned Int;
switch (BuiltinID) {
default: return 0;
@@ -4869,65 +4937,6 @@ Value *CodeGenFunction::EmitARMBuiltinEx
ConstantAggregateZero::get(Ty));
return Builder.CreateSExt(Ops[0], Ty, "vtst");
}
- case NEON::BI__builtin_neon_vtrn_v:
- case NEON::BI__builtin_neon_vtrnq_v: {
- Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
- Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
- Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
- Value *SV = 0;
-
- for (unsigned vi = 0; vi != 2; ++vi) {
- SmallVector<Constant*, 16> Indices;
- for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
- Indices.push_back(Builder.getInt32(i+vi));
- Indices.push_back(Builder.getInt32(i+e+vi));
- }
- Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
- SV = llvm::ConstantVector::get(Indices);
- SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn");
- SV = Builder.CreateStore(SV, Addr);
- }
- return SV;
- }
- case NEON::BI__builtin_neon_vuzp_v:
- case NEON::BI__builtin_neon_vuzpq_v: {
- Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
- Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
- Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
- Value *SV = 0;
-
- for (unsigned vi = 0; vi != 2; ++vi) {
- SmallVector<Constant*, 16> Indices;
- for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
- Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi));
-
- Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
- SV = llvm::ConstantVector::get(Indices);
- SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp");
- SV = Builder.CreateStore(SV, Addr);
- }
- return SV;
- }
- case NEON::BI__builtin_neon_vzip_v:
- case NEON::BI__builtin_neon_vzipq_v: {
- Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
- Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
- Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
- Value *SV = 0;
-
- for (unsigned vi = 0; vi != 2; ++vi) {
- SmallVector<Constant*, 16> Indices;
- for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
- Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1));
- Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e));
- }
- Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi);
- SV = llvm::ConstantVector::get(Indices);
- SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip");
- SV = Builder.CreateStore(SV, Addr);
- }
- return SV;
- }
}
}
Modified: cfe/trunk/lib/CodeGen/CodeGenFunction.h
URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/CodeGenFunction.h?rev=200471&r1=200470&r2=200471&view=diff
==============================================================================
--- cfe/trunk/lib/CodeGen/CodeGenFunction.h (original)
+++ cfe/trunk/lib/CodeGen/CodeGenFunction.h Thu Jan 30 08:47:57 2014
@@ -2177,6 +2177,9 @@ public:
llvm::Value *EmitAArch64CompareBuiltinExpr(llvm::Value *Op, llvm::Type *Ty);
llvm::Value *EmitAArch64BuiltinExpr(unsigned BuiltinID, const CallExpr *E);
llvm::Value *EmitARMBuiltinExpr(unsigned BuiltinID, const CallExpr *E);
+ llvm::Value *EmitCommonNeonBuiltinExpr(unsigned BuiltinID, const CallExpr *E,
+ SmallVectorImpl<llvm::Value *> &Ops,
+ llvm::VectorType *VTy);
llvm::Value *EmitNeonCall(llvm::Function *F,
SmallVectorImpl<llvm::Value*> &O,
const char *name,
More information about the cfe-commits
mailing list