[Beignet] [PATCH 8/9 V3] Backend: Delete bswap logic in the llvm_to_gen stage.
junyan.he at inbox.com
junyan.he at inbox.com
Mon Mar 9 01:11:28 PDT 2015
From: Junyan He <junyan.he at linux.intel.com>
We move the bswap logic from llvm_to_gen to backend for
efficienc using indirect mode.
Signed-off-by: Junyan He <junyan.he at linux.intel.com>
---
backend/src/ir/instruction.hpp | 2 +
backend/src/ir/instruction.hxx | 1 +
backend/src/llvm/llvm_gen_backend.cpp | 85 +--------------------------------
3 files changed, 5 insertions(+), 83 deletions(-)
diff --git a/backend/src/ir/instruction.hpp b/backend/src/ir/instruction.hpp
index 6963111..24d27aa 100644
--- a/backend/src/ir/instruction.hpp
+++ b/backend/src/ir/instruction.hpp
@@ -620,6 +620,8 @@ namespace ir {
Instruction RNDU(Type type, Register dst, Register src);
/*! rndz.type dst src */
Instruction RNDZ(Type type, Register dst, Register src);
+ /*! bswap.type dst src */
+ Instruction BSWAP(Type type, Register dst, Register src);
/*! pow.type dst src0 src1 */
Instruction POW(Type type, Register dst, Register src0, Register src1);
/*! mul.type dst src0 src1 */
diff --git a/backend/src/ir/instruction.hxx b/backend/src/ir/instruction.hxx
index b52673e..de4abfb 100644
--- a/backend/src/ir/instruction.hxx
+++ b/backend/src/ir/instruction.hxx
@@ -40,6 +40,7 @@ DECL_INSN(RNDU, UnaryInstruction)
DECL_INSN(RNDZ, UnaryInstruction)
DECL_INSN(SIMD_ANY, UnaryInstruction)
DECL_INSN(SIMD_ALL, UnaryInstruction)
+DECL_INSN(BSWAP, UnaryInstruction)
DECL_INSN(POW, BinaryInstruction)
DECL_INSN(MUL, BinaryInstruction)
DECL_INSN(ADD, BinaryInstruction)
diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
index aad638f..74c80ee 100644
--- a/backend/src/llvm/llvm_gen_backend.cpp
+++ b/backend/src/llvm/llvm_gen_backend.cpp
@@ -2943,89 +2943,6 @@ namespace gbe
case Intrinsic::umul_with_overflow:
NOT_IMPLEMENTED;
break;
- case Intrinsic::bswap:
- {
- // FIXME, this is an unoptimized version, could be optimized by
- // leveraging GEN's register region/indirect address feature.
- Type *llvmDstType = I.getType();
- uint32_t elementSize = getTypeByteSize(unit, llvmDstType);
-
- const ir::Register dst0 = this->getRegister(&I);
- const ir::Register src0 = this->getRegister(I.getOperand(0));
- switch(elementSize)
- {
- case 2:
- {
- ir::Type srcType = getUnsignedType(ctx, llvmDstType);
- ir::Register tmp1 = ctx.reg(getFamily(srcType));
- ir::Register tmp2 = ctx.reg(getFamily(srcType));
-
- ir::Register regWMask = ctx.reg( ir::FAMILY_WORD );
- const ir::ImmediateIndex wMask = ctx.newIntegerImmediate(0x00FF, ir::TYPE_S16);
- ir::Register regShift = ctx.reg( ir::FAMILY_WORD );
- const ir::ImmediateIndex shift = ctx.newIntegerImmediate(8, ir::TYPE_S16);
-
- ctx.LOADI(ir::TYPE_S16, regWMask, wMask);
- ctx.AND(srcType, tmp1, src0, regWMask);
-
- ctx.LOADI(ir::TYPE_S16, regShift, shift);
- ctx.SHL(srcType, tmp2, tmp1, regShift);
-
- ir::Register tmp3 = ctx.reg( getFamily(srcType) );
- ctx.SHR(srcType, tmp3, src0, regShift);
-
- ctx.OR(srcType, dst0, tmp2, tmp3);
- }
- break;
- case 4:
- {
- ir::Type srcType = getType(ctx, llvmDstType);
- ir::Register tmp1 = ctx.reg(getFamily(srcType));
- ir::Register tmp2 = ctx.reg(getFamily(srcType));
- ir::Register tmp3 = ctx.reg(getFamily(srcType));
- ir::Register tmp4 = ctx.reg(getFamily(srcType));
- ir::Register tmp5 = ctx.reg(getFamily(srcType));
- ir::Register tmp6 = ctx.reg(getFamily(srcType));
-
- ir::Register regDWMask = ctx.reg( ir::FAMILY_DWORD );
- ir::Register regShift_8 = ctx.reg( ir::FAMILY_DWORD );
- ir::Register regShift_24 = ctx.reg( ir::FAMILY_DWORD );
- ir::ImmediateIndex wMask_L = ctx.newIntegerImmediate(0x0000FF00, ir::TYPE_S32);
- ir::ImmediateIndex wMask_H = ctx.newIntegerImmediate(0x00FF0000, ir::TYPE_S32);
- ir::ImmediateIndex shift_8 = ctx.newIntegerImmediate(8, ir::TYPE_S32);
- ir::ImmediateIndex shift_24 = ctx.newIntegerImmediate(24, ir::TYPE_S32);
-
- ctx.LOADI(ir::TYPE_S32, regShift_24, shift_24);
- ctx.SHL(srcType, tmp1, src0, regShift_24);
-
- ctx.LOADI(ir::TYPE_S32, regDWMask, wMask_L);
- ctx.AND(srcType, tmp2, src0, regDWMask);
- ctx.LOADI(ir::TYPE_S32, regShift_8, shift_8);
- ctx.SHL(srcType, tmp3, tmp2, regShift_8);
-
- ctx.LOADI(ir::TYPE_S32, regDWMask, wMask_H);
- ctx.AND(srcType, tmp4, src0, regDWMask);
- ctx.LOADI(ir::TYPE_S32, regShift_8, shift_8);
- ctx.SHR(makeTypeUnsigned(srcType), tmp5, tmp4, regShift_8);
-
- ctx.LOADI(ir::TYPE_S32, regShift_24, shift_24);
- ctx.SHR(makeTypeUnsigned(srcType), tmp6, src0, regShift_24);
-
- ir::Register tmp7 = ctx.reg(getFamily(srcType));
- ir::Register tmp8 = ctx.reg(getFamily(srcType));
- ctx.OR(srcType, tmp7, tmp1, tmp3);
- ctx.OR(srcType, tmp8, tmp5, tmp6);
- ctx.OR(srcType, dst0, tmp7, tmp8);
- }
- break;
- case 8:
- NOT_IMPLEMENTED;
- break;
- default:
- GBE_ASSERT(0);
- }
- }
- break;
case Intrinsic::ctlz:
{
Type *llvmDstType = I.getType();
@@ -3085,6 +3002,8 @@ namespace gbe
case Intrinsic::cos: this->emitUnaryCallInst(I,CS,ir::OP_COS); break;
case Intrinsic::log2: this->emitUnaryCallInst(I,CS,ir::OP_LOG); break;
case Intrinsic::exp2: this->emitUnaryCallInst(I,CS,ir::OP_EXP); break;
+ case Intrinsic::bswap:
+ this->emitUnaryCallInst(I,CS,ir::OP_BSWAP, getUnsignedType(ctx, I.getType())); break;
default: NOT_IMPLEMENTED;
}
} else {
--
1.7.9.5
More information about the Beignet
mailing list