summaryrefslogtreecommitdiffstats
path: root/src/gallium
diff options
context:
space:
mode:
authorTom Stellard <thomas.stellard@amd.com>2012-08-28 15:25:35 -0400
committerTom Stellard <thomas.stellard@amd.com>2012-09-17 21:09:43 +0000
commitbfd55711c1eb32ffbfceb9d566abae98f0015f23 (patch)
treedb80e06a376175080bd0362dec090027e4309672 /src/gallium
parentef0d7e13d7633b2d855e95dc7de2bb833592ec70 (diff)
radeon/llvm: Inital flow control support for SI
This adds basic flow control support for If-Then-Else blocks using predicates (stored in the EXEC register) and a predicate stack for nested flow control.
Diffstat (limited to 'src/gallium')
-rw-r--r--src/gallium/drivers/radeon/AMDGPU.h1
-rw-r--r--src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp1
-rw-r--r--src/gallium/drivers/radeon/AMDILCFGStructurizer.cpp2
-rw-r--r--src/gallium/drivers/radeon/AMDILInstrInfo.td2
-rw-r--r--src/gallium/drivers/radeon/Makefile.sources1
-rw-r--r--src/gallium/drivers/radeon/SIISelLowering.cpp2
-rw-r--r--src/gallium/drivers/radeon/SILowerFlowControl.cpp161
7 files changed, 168 insertions, 2 deletions
diff --git a/src/gallium/drivers/radeon/AMDGPU.h b/src/gallium/drivers/radeon/AMDGPU.h
index fe36545a165..f484caa9a3e 100644
--- a/src/gallium/drivers/radeon/AMDGPU.h
+++ b/src/gallium/drivers/radeon/AMDGPU.h
@@ -25,6 +25,7 @@ FunctionPass *createR600ExpandSpecialInstrsPass(TargetMachine &tm);
// SI Passes
FunctionPass *createSIAssignInterpRegsPass(TargetMachine &tm);
+FunctionPass *createSILowerFlowControlPass(TargetMachine &tm);
FunctionPass *createSICodeEmitterPass(formatted_raw_ostream &OS);
FunctionPass *createSILowerLiteralConstantsPass(TargetMachine &tm);
diff --git a/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp b/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp
index c1b68403379..0b70191b83b 100644
--- a/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp
+++ b/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp
@@ -134,6 +134,7 @@ bool AMDGPUPassConfig::addPreEmitPass() {
addPass(FinalizeMachineBundlesID);
} else {
PM->add(createSILowerLiteralConstantsPass(*TM));
+ PM->add(createSILowerFlowControlPass(*TM));
}
return false;
diff --git a/src/gallium/drivers/radeon/AMDILCFGStructurizer.cpp b/src/gallium/drivers/radeon/AMDILCFGStructurizer.cpp
index b167d62fa0a..20e27ef1132 100644
--- a/src/gallium/drivers/radeon/AMDILCFGStructurizer.cpp
+++ b/src/gallium/drivers/radeon/AMDILCFGStructurizer.cpp
@@ -2892,6 +2892,8 @@ struct CFGStructTraits<AMDGPUCFGStructurizer>
switch (instr->getOpcode()) {
case AMDGPU::JUMP:
return instr->getOperand(instr->findFirstPredOperandIdx()).getReg() == 0;
+ case AMDGPU::BRANCH:
+ return true;
default:
return false;
}
diff --git a/src/gallium/drivers/radeon/AMDILInstrInfo.td b/src/gallium/drivers/radeon/AMDILInstrInfo.td
index f8771afa180..050a5bd874f 100644
--- a/src/gallium/drivers/radeon/AMDILInstrInfo.td
+++ b/src/gallium/drivers/radeon/AMDILInstrInfo.td
@@ -211,7 +211,7 @@ include "AMDILIntrinsics.td"
// Custom Inserter for Branches and returns, this eventually will be a
// seperate pass
//===---------------------------------------------------------------------===//
-let isTerminator = 1, usesCustomInserter = 1 in {
+let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
def BRANCH : ILFormat<(outs), (ins brtarget:$target),
"; Pseudo unconditional branch instruction",
[(br bb:$target)]>;
diff --git a/src/gallium/drivers/radeon/Makefile.sources b/src/gallium/drivers/radeon/Makefile.sources
index 4b0699b9e16..c5d1207451a 100644
--- a/src/gallium/drivers/radeon/Makefile.sources
+++ b/src/gallium/drivers/radeon/Makefile.sources
@@ -71,6 +71,7 @@ CPP_SOURCES := \
SIInstrInfo.cpp \
SIISelLowering.cpp \
SILowerLiteralConstants.cpp \
+ SILowerFlowControl.cpp \
SIMachineFunctionInfo.cpp \
SIRegisterInfo.cpp \
InstPrinter/AMDGPUInstPrinter.cpp \
diff --git a/src/gallium/drivers/radeon/SIISelLowering.cpp b/src/gallium/drivers/radeon/SIISelLowering.cpp
index 3f23949ae90..7c2739cf5b6 100644
--- a/src/gallium/drivers/radeon/SIISelLowering.cpp
+++ b/src/gallium/drivers/radeon/SIISelLowering.cpp
@@ -78,7 +78,7 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
switch (MI->getOpcode()) {
default:
return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
-
+ case AMDGPU::BRANCH: return BB;
case AMDGPU::CLAMP_SI:
BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
.addOperand(MI->getOperand(0))
diff --git a/src/gallium/drivers/radeon/SILowerFlowControl.cpp b/src/gallium/drivers/radeon/SILowerFlowControl.cpp
new file mode 100644
index 00000000000..bf5192efe3d
--- /dev/null
+++ b/src/gallium/drivers/radeon/SILowerFlowControl.cpp
@@ -0,0 +1,161 @@
+//===-- SILowerFlowControl.cpp - Use predicates for flow control ----------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass lowers the pseudo flow control instructions (SI_IF_NZ, ELSE, ENDIF)
+// to predicated instructions.
+//
+// All flow control (except loops) 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 writting 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
+// SI_IF_NZ %VCC
+// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0
+// ELSE
+// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR0
+// ENDIF
+//
+// becomes:
+//
+// %SGPR0 = S_MOV_B64 %EXEC // Save the current exec mask
+// %EXEC = S_AND_B64 %VCC, %EXEC // Update the exec mask
+// S_CBRANCH_EXECZ label0 // This instruction is an
+// // 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:
+// %EXEC = S_NOT_B64 %EXEC // Invert the exec mask for the
+// // Then block.
+// %EXEC = S_AND_B64 %SGPR0, %EXEC
+// S_BRANCH_EXECZ label1 // Use our branch optimization
+// // instruction again.
+// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR // Do the THEN block
+// label1:
+// S_MOV_B64 // Restore the old EXEC value
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "SIInstrInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+
+namespace {
+
+class SILowerFlowControlPass : public MachineFunctionPass {
+
+private:
+ static char ID;
+ const TargetInstrInfo *TII;
+ std::vector<unsigned> PredicateStack;
+ std::vector<unsigned> UnusedRegisters;
+
+ void pushExecMask(MachineBasicBlock &MBB, MachineBasicBlock::iterator I);
+ void popExecMask(MachineBasicBlock &MBB, MachineBasicBlock::iterator I);
+
+public:
+ SILowerFlowControlPass(TargetMachine &tm) :
+ MachineFunctionPass(ID), TII(tm.getInstrInfo()) { }
+
+ virtual bool runOnMachineFunction(MachineFunction &MF);
+
+ const char *getPassName() const {
+ return "SI Lower flow control instructions";
+ }
+
+};
+
+} // End anonymous namespace
+
+char SILowerFlowControlPass::ID = 0;
+
+FunctionPass *llvm::createSILowerFlowControlPass(TargetMachine &tm) {
+ return new SILowerFlowControlPass(tm);
+}
+
+bool SILowerFlowControlPass::runOnMachineFunction(MachineFunction &MF) {
+
+ // Find all the unused registers that can be used for the predicate stack.
+ for (TargetRegisterClass::iterator S = AMDGPU::SReg_64RegClass.begin(),
+ I = AMDGPU::SReg_64RegClass.end();
+ I != S; --I) {
+ unsigned Reg = *I;
+ if (!MF.getRegInfo().isPhysRegOrOverlapUsed(Reg)) {
+ UnusedRegisters.push_back(Reg);
+ }
+ }
+
+ for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+ BB != BB_E; ++BB) {
+ MachineBasicBlock &MBB = *BB;
+ for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
+ I != MBB.end(); I = Next, Next = llvm::next(I)) {
+ MachineInstr &MI = *I;
+ switch (MI.getOpcode()) {
+ default: break;
+ case AMDGPU::SI_IF_NZ:
+ pushExecMask(MBB, I);
+ BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_AND_B64),
+ AMDGPU::EXEC)
+ .addOperand(MI.getOperand(0)) // VCC
+ .addReg(AMDGPU::EXEC);
+ MI.eraseFromParent();
+ break;
+ case AMDGPU::ELSE:
+ BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_NOT_B64),
+ AMDGPU::EXEC)
+ .addReg(AMDGPU::EXEC);
+ BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_AND_B64),
+ AMDGPU::EXEC)
+ .addReg(PredicateStack.back())
+ .addReg(AMDGPU::EXEC);
+ MI.eraseFromParent();
+ break;
+ case AMDGPU::ENDIF:
+ popExecMask(MBB, I);
+ MI.eraseFromParent();
+ break;
+ }
+ }
+ }
+ return false;
+}
+
+void SILowerFlowControlPass::pushExecMask(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I) {
+
+ assert(!UnusedRegisters.empty() && "Ran out of registers for predicate stack");
+ unsigned StackReg = UnusedRegisters.back();
+ UnusedRegisters.pop_back();
+ PredicateStack.push_back(StackReg);
+ BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B64),
+ StackReg)
+ .addReg(AMDGPU::EXEC);
+}
+
+void SILowerFlowControlPass::popExecMask(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I) {
+ unsigned StackReg = PredicateStack.back();
+ PredicateStack.pop_back();
+ UnusedRegisters.push_back(StackReg);
+ BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B64),
+ AMDGPU::EXEC)
+ .addReg(StackReg);
+}