This pass lowers the pseudo control flow instructions to real machine instructions. More...

#include "AMDGPU.h"
#include "GCNSubtarget.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/Target/TargetMachine.h"

Macros
#define	DEBUG_TYPE "si-lower-control-flow"

Functions
	INITIALIZE_PASS (SILowerControlFlow, DEBUG_TYPE, "SI lower control flow", false, false) static void setImpSCCDefDead(MachineInstr &MI

	assert (ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())

ImpDefSCC	setIsDead (IsDead)

static bool	isSimpleIf (const MachineInstr &MI, const MachineRegisterInfo *MRI)

Variables
static cl::opt< bool >	RemoveRedundantEndcf ("amdgpu-remove-redundant-endcf", cl::init(true), cl::ReallyHidden)

bool	IsDead

Detailed Description

This pass lowers the pseudo control flow instructions to real machine instructions.

All control flow is handled using predicated instructions and a predicate stack. Each Scalar ALU controls the operations of 64 Vector ALUs. The Scalar ALU can update the predicate for any of the Vector ALUs by writing to the 64-bit EXEC register (each bit corresponds to a single vector ALU). Typically, for predicates, a vector ALU will write to its bit of the VCC register (like EXEC VCC is 64-bits, one for each Vector ALU) and then the ScalarALU will AND the VCC register with the EXEC to update the predicates.

For example: vcc = V_CMP_GT_F32 vgpr1, vgpr2 sgpr0 = SI_IF vcc vgpr0 = V_ADD_F32 vgpr0, vgpr0 sgpr0 = SI_ELSE sgpr0 vgpr0 = V_SUB_F32 vgpr0, vgpr0 SI_END_CF sgpr0

becomes:

sgpr0 = S_AND_SAVEEXEC_B64 vcc // Save and update the exec mask sgpr0 = S_XOR_B64 sgpr0, exec // Clear live bits from saved exec mask S_CBRANCH_EXECZ label0 // This instruction is an optional // optimization which allows us to // branch if all the bits of // EXEC are zero. vgpr0 = V_ADD_F32 vgpr0, vgpr0 // Do the IF block of the branch

label0: sgpr0 = S_OR_SAVEEXEC_B64 sgpr0 // Restore the exec mask for the Then // block exec = S_XOR_B64 sgpr0, exec // Update the exec mask S_CBRANCH_EXECZ label1 // Use our branch optimization // instruction again. vgpr0 = V_SUB_F32 vgpr0, vgpr // Do the ELSE block label1: exec = S_OR_B64 exec, sgpr0 // Re-enable saved exec mask bits

Definition in file SILowerControlFlow.cpp.

Macro Definition Documentation

◆ DEBUG_TYPE

#define DEBUG_TYPE "si-lower-control-flow"

Definition at line 63 of file SILowerControlFlow.cpp.

Function Documentation

◆ assert()

assert ( ImpDefSCC. getReg() = =AMDGPU::SCC &&ImpDefSCC.isDef() )

◆ INITIALIZE_PASS()

INITIALIZE_PASS	(	SILowerControlFlow	,
		DEBUG_TYPE	,
		"SI lower control flow"	,
		false	,
		false
	)		&

◆ isSimpleIf()

static bool isSimpleIf	(	const MachineInstr &	MI,
		const MachineRegisterInfo *	MRI
	)

static

Definition at line 195 of file SILowerControlFlow.cpp.

References MI, and MRI.

◆ setIsDead()

ImpDefSCC setIsDead ( IsDead )

Variable Documentation

◆ IsDead

bool IsDead

Initial value:

{

MachineOperand &ImpDefSCC = MI.getOperand(3)

MI

IRTranslator LLVM IR MI

Definition: IRTranslator.cpp:112

llvm::MachineOperand

MachineOperand class - Representation of each machine instruction operand.

Definition: MachineOperand.h:48

Definition at line 167 of file SILowerControlFlow.cpp.

Referenced by llvm::rdf::DeadCodeElimination::collect(), llvm::ReachingDefAnalysis::collectKilledOperands(), llvm::rdf::Liveness::computePhiInfo(), llvm::rdf::Liveness::getAllReachedUses(), rematerializeCheapDef(), llvm::objcopy::macho::Object::removeSections(), and swapRegAndNonRegOperand().

◆ RemoveRedundantEndcf

cl::opt< bool > RemoveRedundantEndcf("amdgpu-remove-redundant-endcf", cl::init(true), cl::ReallyHidden)	(	"amdgpu-remove-redundant-endcf"	,
		cl::init(true)	,
		cl::ReallyHidden
	)