1 //===----------------------- AMDGPUFrameLowering.cpp ----------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //==-----------------------------------------------------------------------===//
9 //
10 // Interface to describe a layout of a stack frame on a AMDIL target machine
11 //
12 //===----------------------------------------------------------------------===//
13 #include "AMDGPUFrameLowering.h"
14 #include "AMDGPURegisterInfo.h"
15 #include "R600MachineFunctionInfo.h"
16 #include "llvm/CodeGen/MachineFrameInfo.h"
17 #include "llvm/CodeGen/MachineRegisterInfo.h"
18 #include "llvm/IR/Instructions.h"
19
20 using namespace llvm;
AMDGPUFrameLowering(StackDirection D,unsigned StackAl,int LAO,unsigned TransAl)21 AMDGPUFrameLowering::AMDGPUFrameLowering(StackDirection D, unsigned StackAl,
22 int LAO, unsigned TransAl)
23 : TargetFrameLowering(D, StackAl, LAO, TransAl) { }
24
~AMDGPUFrameLowering()25 AMDGPUFrameLowering::~AMDGPUFrameLowering() { }
26
getStackWidth(const MachineFunction & MF) const27 unsigned AMDGPUFrameLowering::getStackWidth(const MachineFunction &MF) const {
28
29 // XXX: Hardcoding to 1 for now.
30 //
31 // I think the StackWidth should stored as metadata associated with the
32 // MachineFunction. This metadata can either be added by a frontend, or
33 // calculated by a R600 specific LLVM IR pass.
34 //
35 // The StackWidth determines how stack objects are laid out in memory.
36 // For a vector stack variable, like: int4 stack[2], the data will be stored
37 // in the following ways depending on the StackWidth.
38 //
39 // StackWidth = 1:
40 //
41 // T0.X = stack[0].x
42 // T1.X = stack[0].y
43 // T2.X = stack[0].z
44 // T3.X = stack[0].w
45 // T4.X = stack[1].x
46 // T5.X = stack[1].y
47 // T6.X = stack[1].z
48 // T7.X = stack[1].w
49 //
50 // StackWidth = 2:
51 //
52 // T0.X = stack[0].x
53 // T0.Y = stack[0].y
54 // T1.X = stack[0].z
55 // T1.Y = stack[0].w
56 // T2.X = stack[1].x
57 // T2.Y = stack[1].y
58 // T3.X = stack[1].z
59 // T3.Y = stack[1].w
60 //
61 // StackWidth = 4:
62 // T0.X = stack[0].x
63 // T0.Y = stack[0].y
64 // T0.Z = stack[0].z
65 // T0.W = stack[0].w
66 // T1.X = stack[1].x
67 // T1.Y = stack[1].y
68 // T1.Z = stack[1].z
69 // T1.W = stack[1].w
70 return 1;
71 }
72
73 /// \returns The number of registers allocated for \p FI.
getFrameIndexOffset(const MachineFunction & MF,int FI) const74 int AMDGPUFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
75 int FI) const {
76 const MachineFrameInfo *MFI = MF.getFrameInfo();
77 // Start the offset at 2 so we don't overwrite work group information.
78 // XXX: We should only do this when the shader actually uses this
79 // information.
80 unsigned OffsetBytes = 2 * (getStackWidth(MF) * 4);
81 int UpperBound = FI == -1 ? MFI->getNumObjects() : FI;
82
83 for (int i = MFI->getObjectIndexBegin(); i < UpperBound; ++i) {
84 OffsetBytes = RoundUpToAlignment(OffsetBytes, MFI->getObjectAlignment(i));
85 OffsetBytes += MFI->getObjectSize(i);
86 // Each register holds 4 bytes, so we must always align the offset to at
87 // least 4 bytes, so that 2 frame objects won't share the same register.
88 OffsetBytes = RoundUpToAlignment(OffsetBytes, 4);
89 }
90
91 if (FI != -1)
92 OffsetBytes = RoundUpToAlignment(OffsetBytes, MFI->getObjectAlignment(FI));
93
94 return OffsetBytes / (getStackWidth(MF) * 4);
95 }
96
97 const TargetFrameLowering::SpillSlot *
getCalleeSavedSpillSlots(unsigned & NumEntries) const98 AMDGPUFrameLowering::getCalleeSavedSpillSlots(unsigned &NumEntries) const {
99 NumEntries = 0;
100 return nullptr;
101 }
102 void
emitPrologue(MachineFunction & MF) const103 AMDGPUFrameLowering::emitPrologue(MachineFunction &MF) const {
104 }
105 void
emitEpilogue(MachineFunction & MF,MachineBasicBlock & MBB) const106 AMDGPUFrameLowering::emitEpilogue(MachineFunction &MF,
107 MachineBasicBlock &MBB) const {
108 }
109
110 bool
hasFP(const MachineFunction & MF) const111 AMDGPUFrameLowering::hasFP(const MachineFunction &MF) const {
112 return false;
113 }
114