1 //===-- NVPTXTargetTransformInfo.cpp - NVPTX specific TTI -----------------===//
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
26   switch (II->getIntrinsicID()) {  in readsThreadIndex()
36   return II->getIntrinsicID() == Intrinsic::nvvm_read_ptx_sreg_laneid;  in readsLaneId()
41   switch (II->getIntrinsicID()) {  in isNVVMAtomic()
73   // Without inter-procedural analysis, we conservatively assume that arguments  in isSourceOfDivergence()
76     return !isKernelFunction(*Arg->getParent());  in isSourceOfDivergence()
82       unsigned AS = LI->getPointerAddressSpace();  in isSourceOfDivergence()
94     if (I->isAtomic())  in isSourceOfDivergence()
107     // inter-procedural analysis.  in isSourceOfDivergence()
115 // Convert NVVM intrinsics to target-generic LLVM code where possible.
116 static Instruction *simplifyNvvmIntrinsic(IntrinsicInst *II, InstCombiner &IC) {  in simplifyNvvmIntrinsic()  argument
122   //  * ad-hoc LLVM IR for the particular operation.  in simplifyNvvmIntrinsic()
125   // flush-denormals-to-zero (ftz) setting is true/false, whereas other  in simplifyNvvmIntrinsic()
132   // Classes of NVVM intrinsics that can't be replaced one-to-one with a  in simplifyNvvmIntrinsic()
133   // target-generic intrinsic, cast op, or binary op but that we can nonetheless  in simplifyNvvmIntrinsic()
139   // SimplifyAction is a poor-man's variant (plus an additional flag) that  in simplifyNvvmIntrinsic()
140   // represents how to replace an NVVM intrinsic with target-generic LLVM IR.  in simplifyNvvmIntrinsic()
150     // type (denormal-fp-math vs denormal-fp-math-f32), take note of halfs.  in simplifyNvvmIntrinsic()
171   // IntrinsicInstr with target-generic LLVM IR.  in simplifyNvvmIntrinsic()
172   const SimplifyAction Action = [II]() -> SimplifyAction {  in simplifyNvvmIntrinsic()
173     switch (II->getIntrinsicID()) {  in simplifyNvvmIntrinsic()
267       // ftz version, and foo_f is the non-ftz version.  But nvvm_sqrt_f adopts  in simplifyNvvmIntrinsic()
268       // the ftz-ness of the surrounding code.  sqrt_rn_f and sqrt_rn_ftz_f are  in simplifyNvvmIntrinsic()
269       // the versions with explicit ftz-ness.  in simplifyNvvmIntrinsic()
280     // Note that llvm's target-generic conversion operators correspond to the rz  in simplifyNvvmIntrinsic()
319       //   - nvvm_cos_approx_{f,ftz_f}  in simplifyNvvmIntrinsic()
320       //   - nvvm_ex2_approx_{d,f,ftz_f}  in simplifyNvvmIntrinsic()
321       //   - nvvm_lg2_approx_{d,f,ftz_f}  in simplifyNvvmIntrinsic()
322       //   - nvvm_sin_approx_{f,ftz_f}  in simplifyNvvmIntrinsic()
323       //   - nvvm_sqrt_approx_{f,ftz_f}  in simplifyNvvmIntrinsic()
324       //   - nvvm_rsqrt_approx_{d,f,ftz_f}  in simplifyNvvmIntrinsic()
325       //   - nvvm_div_approx_{ftz_d,ftz_f,f}  in simplifyNvvmIntrinsic()
326       //   - nvvm_rcp_approx_ftz_d  in simplifyNvvmIntrinsic()
347     DenormalMode Mode = II->getFunction()->getDenormalMode(  in simplifyNvvmIntrinsic()
355   // Simplify to target-generic intrinsic.  in simplifyNvvmIntrinsic()
357     SmallVector<Value *, 4> Args(II->args());  in simplifyNvvmIntrinsic()
358     // All the target-generic intrinsics currently of interest to us have one  in simplifyNvvmIntrinsic()
360     Type *Tys[] = {II->getArgOperand(0)->getType()};  in simplifyNvvmIntrinsic()
362         Intrinsic::getDeclaration(II->getModule(), *Action.IID, Tys), Args);  in simplifyNvvmIntrinsic()
365   // Simplify to target-generic binary op.  in simplifyNvvmIntrinsic()
367     return BinaryOperator::Create(*Action.BinaryOp, II->getArgOperand(0),  in simplifyNvvmIntrinsic()
368                                   II->getArgOperand(1), II->getName());  in simplifyNvvmIntrinsic()
370   // Simplify to target-generic cast op.  in simplifyNvvmIntrinsic()
372     return CastInst::Create(*Action.CastOp, II->getArgOperand(0), II->getType(),  in simplifyNvvmIntrinsic()
373                             II->getName());  in simplifyNvvmIntrinsic()
383         Instruction::FDiv, ConstantFP::get(II->getArgOperand(0)->getType(), 1),  in simplifyNvvmIntrinsic()
384         II->getArgOperand(0), II->getName());  in simplifyNvvmIntrinsic()
390 NVPTXTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {  in instCombineIntrinsic()  argument
391   if (Instruction *I = simplifyNvvmIntrinsic(&II, IC)) {  in instCombineIntrinsic()
405   int ISD = TLI->InstructionOpcodeToISD(Opcode);  in getArithmeticInstrCost()