R600 Patchset: Emit true ISA
Tom Stellard
tom at stellard.net
Fri May 3 15:53:52 PDT 2013
Hi,
The attached patches modify the CodeEmitter to emit true ISA.
Previously, we were prefixing all instructions with an instruction type
byte.
Vincent did most of the work to convert the CodeEmitter to true ISA,
these patches are just the last few cleanups that are needed to finish
the project.
Please test/review.
Thanks,
Tom
-------------- next part --------------
>From 4fc6af0637de0eae0542a987e93d467bad3a4eee Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard at amd.com>
Date: Fri, 3 May 2013 11:17:18 -0700
Subject: [PATCH 1/4] R600: Emit ISA for CALL_FS_* instructions
---
lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp | 1 -
test/CodeGen/R600/call_fs.ll | 15 +++++++++++++++
2 files changed, 15 insertions(+), 1 deletion(-)
create mode 100644 test/CodeGen/R600/call_fs.ll
diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
index 7c83d86..8261477 100644
--- a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -262,7 +262,6 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
}
case AMDGPU::CF_CALL_FS_EG:
case AMDGPU::CF_CALL_FS_R600:
- return;
case AMDGPU::CF_TC_EG:
case AMDGPU::CF_VC_EG:
case AMDGPU::CF_TC_R600:
diff --git a/test/CodeGen/R600/call_fs.ll b/test/CodeGen/R600/call_fs.ll
new file mode 100644
index 0000000..fd21b72
--- /dev/null
+++ b/test/CodeGen/R600/call_fs.ll
@@ -0,0 +1,15 @@
+
+; RUN: llc < %s -march=r600 -mcpu=redwood -show-mc-encoding -o - | FileCheck --check-prefix=EG-CHECK %s
+; RUN: llc < %s -march=r600 -mcpu=rv710 -show-mc-encoding -o - | FileCheck --check-prefix=R600-CHECK %s
+
+; EG-CHECK: @call_fs
+; EG-CHECK: CALL_FS ; encoding: [0x03,0x00,0x00,0x00,0x00,0x00,0x00,0xc0,0x84]
+; R600-CHECK: @call_fs
+; R600-CHECK:CALL_FS ; encoding: [0x03,0x00,0x00,0x00,0x00,0x00,0x00,0x80,0x89]
+
+
+define void @call_fs() #0 {
+ ret void
+}
+
+attributes #0 = { "ShaderType"="1" } ; Vertex Shader
--
1.7.11.4
-------------- next part --------------
>From 24b983199b5c81eb0362f4c583eafb842255acc5 Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard at amd.com>
Date: Fri, 3 May 2013 11:38:33 -0700
Subject: [PATCH 2/4] R600: Stop emitting the instruction type byte before
each instruction
---
lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp | 35 ++--------------------
test/CodeGen/R600/call_fs.ll | 4 +--
test/CodeGen/R600/cf_end.ll | 4 +--
3 files changed, 6 insertions(+), 37 deletions(-)
diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
index 8261477..10dee20 100644
--- a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -9,12 +9,8 @@
//
/// \file
///
-/// This code emitter outputs bytecode that is understood by the r600g driver
-/// in the Mesa [1] project. The bytecode is very similar to the hardware's ISA,
-/// but it still needs to be run through a finalizer in order to be executed
-/// by the GPU.
-///
-/// [1] http://www.mesa3d.org/
+/// \brief The R600 code emitter produces machine code that can be executed
+/// directly on the GPU device.
//
//===----------------------------------------------------------------------===//
@@ -95,16 +91,6 @@ enum RegElement {
ELEMENT_W
};
-enum InstrTypes {
- INSTR_ALU = 0,
- INSTR_TEX,
- INSTR_FC,
- INSTR_NATIVE,
- INSTR_VTX,
- INSTR_EXPORT,
- INSTR_CFALU
-};
-
enum FCInstr {
FC_IF_PREDICATE = 0,
FC_ELSE,
@@ -152,7 +138,6 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
uint64_t inst = getBinaryCodeForInstr(MI, Fixups);
- EmitByte(INSTR_NATIVE, OS);
Emit(inst, OS);
break;
}
@@ -170,9 +155,7 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
InstWord2 |= 1 << 19;
- EmitByte(INSTR_NATIVE, OS);
Emit(InstWord01, OS);
- EmitByte(INSTR_NATIVE, OS);
Emit(InstWord2, OS);
Emit((u_int32_t) 0, OS);
break;
@@ -246,9 +229,7 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
SrcSelect[ELEMENT_W] << 29 | Offsets[0] << 0 | Offsets[1] << 5 |
Offsets[2] << 10;
- EmitByte(INSTR_NATIVE, OS);
Emit(Word01, OS);
- EmitByte(INSTR_NATIVE, OS);
Emit(Word2, OS);
Emit((u_int32_t) 0, OS);
break;
@@ -256,7 +237,6 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
case AMDGPU::CF_ALU:
case AMDGPU::CF_ALU_PUSH_BEFORE: {
uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
- EmitByte(INSTR_NATIVE, OS);
Emit(Inst, OS);
break;
}
@@ -289,13 +269,11 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
case AMDGPU::CF_END_EG:
case AMDGPU::CF_END_CM: {
uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
- EmitByte(INSTR_NATIVE, OS);
Emit(Inst, OS);
break;
}
default:
uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
- EmitByte(INSTR_NATIVE, OS);
Emit(Inst, OS);
break;
}
@@ -307,9 +285,6 @@ void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
raw_ostream &OS) const {
const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
- // Emit instruction type
- EmitByte(INSTR_ALU, OS);
-
uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
//older alu have different encoding for instructions with one or two src
@@ -324,8 +299,6 @@ void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
unsigned SrcNum = MCDesc.TSFlags & R600_InstFlag::OP3 ? 3 :
MCDesc.TSFlags & R600_InstFlag::OP2 ? 2 : 1;
- EmitByte(SrcNum, OS);
-
const unsigned SrcOps[3][2] = {
{R600Operands::SRC0, R600Operands::SRC0_SEL},
{R600Operands::SRC1, R600Operands::SRC1_SEL},
@@ -335,7 +308,6 @@ void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
for (unsigned SrcIdx = 0; SrcIdx < SrcNum; ++SrcIdx) {
unsigned RegOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][0]];
unsigned SelOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][1]];
- EmitSrcISA(MI, RegOpIdx, SelOpIdx, OS);
}
Emit(InstWord01, OS);
@@ -446,9 +418,6 @@ void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned RegOpIdx,
void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
- // Emit instruction type
- EmitByte(INSTR_FC, OS);
-
// Emit SRC
unsigned NumOperands = MI.getNumOperands();
if (NumOperands > 0) {
diff --git a/test/CodeGen/R600/call_fs.ll b/test/CodeGen/R600/call_fs.ll
index fd21b72..e152bf6 100644
--- a/test/CodeGen/R600/call_fs.ll
+++ b/test/CodeGen/R600/call_fs.ll
@@ -3,9 +3,9 @@
; RUN: llc < %s -march=r600 -mcpu=rv710 -show-mc-encoding -o - | FileCheck --check-prefix=R600-CHECK %s
; EG-CHECK: @call_fs
-; EG-CHECK: CALL_FS ; encoding: [0x03,0x00,0x00,0x00,0x00,0x00,0x00,0xc0,0x84]
+; EG-CHECK: CALL_FS ; encoding: [0x00,0x00,0x00,0x00,0x00,0x00,0xc0,0x84]
; R600-CHECK: @call_fs
-; R600-CHECK:CALL_FS ; encoding: [0x03,0x00,0x00,0x00,0x00,0x00,0x00,0x80,0x89]
+; R600-CHECK:CALL_FS ; encoding: [0x00,0x00,0x00,0x00,0x00,0x00,0x80,0x89]
define void @call_fs() #0 {
diff --git a/test/CodeGen/R600/cf_end.ll b/test/CodeGen/R600/cf_end.ll
index 56cd7de..138004d 100644
--- a/test/CodeGen/R600/cf_end.ll
+++ b/test/CodeGen/R600/cf_end.ll
@@ -2,8 +2,8 @@
; RUN: llc < %s -march=r600 -mcpu=caicos --show-mc-encoding | FileCheck --check-prefix=EG-CHECK %s
; RUN: llc < %s -march=r600 -mcpu=cayman --show-mc-encoding | FileCheck --check-prefix=CM-CHECK %s
-; EG-CHECK: CF_END ; encoding: [0x03,0x00,0x00,0x00,0x00,0x00,0x00,0x20,0x80]
-; CM-CHECK: CF_END ; encoding: [0x03,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x88]
+; EG-CHECK: CF_END ; encoding: [0x00,0x00,0x00,0x00,0x00,0x00,0x20,0x80]
+; CM-CHECK: CF_END ; encoding: [0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x88]
define void @eop() {
ret void
}
--
1.7.11.4
-------------- next part --------------
>From 295bda0baafcf653d4146fa0ba612ef54f98386a Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard at amd.com>
Date: Fri, 3 May 2013 13:12:47 -0700
Subject: [PATCH 3/4] R600: Emit config values in register / value pairs
---
lib/Target/R600/AMDGPUAsmPrinter.cpp | 32 +++++++++++++++++++++++++++++---
lib/Target/R600/R600Defines.h | 26 ++++++++++++++++++++++++++
test/CodeGen/R600/elf.r600.ll | 5 +++--
3 files changed, 58 insertions(+), 5 deletions(-)
diff --git a/lib/Target/R600/AMDGPUAsmPrinter.cpp b/lib/Target/R600/AMDGPUAsmPrinter.cpp
index c915f50..4c35ecf 100644
--- a/lib/Target/R600/AMDGPUAsmPrinter.cpp
+++ b/lib/Target/R600/AMDGPUAsmPrinter.cpp
@@ -22,6 +22,7 @@
#include "SIDefines.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
+#include "R600Defines.h"
#include "R600MachineFunctionInfo.h"
#include "R600RegisterInfo.h"
#include "llvm/MC/MCContext.h"
@@ -78,6 +79,7 @@ void AMDGPUAsmPrinter::EmitProgramInfoR600(MachineFunction &MF) {
const R600RegisterInfo * RI =
static_cast<const R600RegisterInfo*>(TM.getRegisterInfo());
R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
+ const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
BB != BB_E; ++BB) {
@@ -101,9 +103,33 @@ void AMDGPUAsmPrinter::EmitProgramInfoR600(MachineFunction &MF) {
}
}
}
- OutStreamer.EmitIntValue(MaxGPR + 1, 4);
- OutStreamer.EmitIntValue(MFI->StackSize, 4);
- OutStreamer.EmitIntValue(killPixel, 4);
+
+ unsigned RsrcReg;
+ if (STM.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX) {
+ // Evergreen / Northern Islands
+ switch (MFI->ShaderType) {
+ default: // Fall through
+ case ShaderType::COMPUTE: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break;
+ case ShaderType::GEOMETRY: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break;
+ case ShaderType::PIXEL: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break;
+ case ShaderType::VERTEX: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break;
+ }
+ } else {
+ // R600 / R700
+ switch (MFI->ShaderType) {
+ default: // Fall through
+ case ShaderType::GEOMETRY: // Fall through
+ case ShaderType::COMPUTE: // Fall through
+ case ShaderType::VERTEX: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break;
+ case ShaderType::PIXEL: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break;
+ }
+ }
+
+ OutStreamer.EmitIntValue(RsrcReg, 4);
+ OutStreamer.EmitIntValue(S_NUM_GPRS(MaxGPR + 1) |
+ S_STACK_SIZE(MFI->StackSize), 4);
+ OutStreamer.EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4);
+ OutStreamer.EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4);
}
void AMDGPUAsmPrinter::EmitProgramInfoSI(MachineFunction &MF) {
diff --git a/lib/Target/R600/R600Defines.h b/lib/Target/R600/R600Defines.h
index 303ca73..4041550 100644
--- a/lib/Target/R600/R600Defines.h
+++ b/lib/Target/R600/R600Defines.h
@@ -97,4 +97,30 @@ namespace R600Operands {
}
+//===----------------------------------------------------------------------===//
+// Config register definitions
+//===----------------------------------------------------------------------===//
+
+#define R_02880C_DB_SHADER_CONTROL 0x02880C
+#define S_02880C_KILL_ENABLE(x) (((x) & 0x1) << 6)
+
+// These fields are the same for all shader types and families.
+#define S_NUM_GPRS(x) (((x) & 0xFF) << 0)
+#define S_STACK_SIZE(x) (((x) & 0xFF) << 8)
+//===----------------------------------------------------------------------===//
+// R600, R700 Registers
+//===----------------------------------------------------------------------===//
+
+#define R_028850_SQ_PGM_RESOURCES_PS 0x028850
+#define R_028868_SQ_PGM_RESOURCES_VS 0x028868
+
+//===----------------------------------------------------------------------===//
+// Evergreen, Northern Islands Registers
+//===----------------------------------------------------------------------===//
+
+#define R_028844_SQ_PGM_RESOURCES_PS 0x028844
+#define R_028860_SQ_PGM_RESOURCES_VS 0x028860
+#define R_028878_SQ_PGM_RESOURCES_GS 0x028878
+#define R_0288D4_SQ_PGM_RESOURCES_LS 0x0288d4
+
#endif // R600DEFINES_H_
diff --git a/test/CodeGen/R600/elf.r600.ll b/test/CodeGen/R600/elf.r600.ll
index 31ba84a..0590efb 100644
--- a/test/CodeGen/R600/elf.r600.ll
+++ b/test/CodeGen/R600/elf.r600.ll
@@ -5,8 +5,9 @@
; ELF-CHECK: Name: .AMDGPU.config
; CONFIG-CHECK: .section .AMDGPU.config
-; CONFIG-CHECK-NEXT: .long 2
-; CONFIG-CHECK-NEXT: .long 1
+; CONFIG-CHECK-NEXT: .long 166100
+; CONFIG-CHECK-NEXT: .long 258
+; CONFIG-CHECK-NEXT: .long 165900
; CONFIG-CHECK-NEXT: .long 0
define void @test(float addrspace(1)* %out, i32 %p) {
%i = add i32 %p, 2
--
1.7.11.4
-------------- next part --------------
>From 46cff1c52c9d7bd77f007a4f3efd9931ada8d17d Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard at amd.com>
Date: Fri, 3 May 2013 15:27:23 -0700
Subject: [PATCH 4/4] R600: Remove dead code from the CodeEmitter
---
lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp | 310 +--------------------
1 file changed, 1 insertion(+), 309 deletions(-)
diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
index 10dee20..ac33697 100644
--- a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -26,9 +26,6 @@
#include "llvm/Support/raw_ostream.h"
#include <stdio.h>
-#define SRC_BYTE_COUNT 11
-#define DST_BYTE_COUNT 5
-
using namespace llvm;
namespace {
@@ -56,30 +53,14 @@ public:
SmallVectorImpl<MCFixup> &Fixups) const;
private:
- void EmitALUInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups,
- raw_ostream &OS) const;
- void EmitSrc(const MCInst &MI, unsigned OpIdx, raw_ostream &OS) const;
- void EmitSrcISA(const MCInst &MI, unsigned RegOpIdx, unsigned SelOpIdx,
- raw_ostream &OS) const;
- void EmitDst(const MCInst &MI, raw_ostream &OS) const;
- void EmitFCInstr(const MCInst &MI, raw_ostream &OS) const;
-
- void EmitNullBytes(unsigned int byteCount, raw_ostream &OS) const;
-
void EmitByte(unsigned int byte, raw_ostream &OS) const;
- void EmitTwoBytes(uint32_t bytes, raw_ostream &OS) const;
-
void Emit(uint32_t value, raw_ostream &OS) const;
void Emit(uint64_t value, raw_ostream &OS) const;
unsigned getHWRegChan(unsigned reg) const;
unsigned getHWReg(unsigned regNo) const;
- bool isFCOp(unsigned opcode) const;
- bool isTexOp(unsigned opcode) const;
- bool isFlagSet(const MCInst &MI, unsigned Operand, unsigned Flag) const;
-
};
} // End anonymous namespace
@@ -125,9 +106,7 @@ MCCodeEmitter *llvm::createR600MCCodeEmitter(const MCInstrInfo &MCII,
void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
SmallVectorImpl<MCFixup> &Fixups) const {
- if (isFCOp(MI.getOpcode())){
- EmitFCInstr(MI, OS);
- } else if (MI.getOpcode() == AMDGPU::RETURN ||
+ if (MI.getOpcode() == AMDGPU::RETURN ||
MI.getOpcode() == AMDGPU::FETCH_CLAUSE ||
MI.getOpcode() == AMDGPU::ALU_CLAUSE ||
MI.getOpcode() == AMDGPU::BUNDLE ||
@@ -135,12 +114,6 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
return;
} else {
switch(MI.getOpcode()) {
- case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
- case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
- uint64_t inst = getBinaryCodeForInstr(MI, Fixups);
- Emit(inst, OS);
- break;
- }
case AMDGPU::CONSTANT_LOAD_eg:
case AMDGPU::VTX_READ_PARAM_8_eg:
case AMDGPU::VTX_READ_PARAM_16_eg:
@@ -234,44 +207,6 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
Emit((u_int32_t) 0, OS);
break;
}
- case AMDGPU::CF_ALU:
- case AMDGPU::CF_ALU_PUSH_BEFORE: {
- uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
- Emit(Inst, OS);
- break;
- }
- case AMDGPU::CF_CALL_FS_EG:
- case AMDGPU::CF_CALL_FS_R600:
- case AMDGPU::CF_TC_EG:
- case AMDGPU::CF_VC_EG:
- case AMDGPU::CF_TC_R600:
- case AMDGPU::CF_VC_R600:
- case AMDGPU::WHILE_LOOP_EG:
- case AMDGPU::END_LOOP_EG:
- case AMDGPU::LOOP_BREAK_EG:
- case AMDGPU::CF_CONTINUE_EG:
- case AMDGPU::CF_JUMP_EG:
- case AMDGPU::CF_ELSE_EG:
- case AMDGPU::POP_EG:
- case AMDGPU::WHILE_LOOP_R600:
- case AMDGPU::END_LOOP_R600:
- case AMDGPU::LOOP_BREAK_R600:
- case AMDGPU::CF_CONTINUE_R600:
- case AMDGPU::CF_JUMP_R600:
- case AMDGPU::CF_ELSE_R600:
- case AMDGPU::POP_R600:
- case AMDGPU::EG_ExportSwz:
- case AMDGPU::R600_ExportSwz:
- case AMDGPU::EG_ExportBuf:
- case AMDGPU::R600_ExportBuf:
- case AMDGPU::PAD:
- case AMDGPU::CF_END_R600:
- case AMDGPU::CF_END_EG:
- case AMDGPU::CF_END_CM: {
- uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
- Emit(Inst, OS);
- break;
- }
default:
uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
Emit(Inst, OS);
@@ -280,202 +215,10 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
}
}
-void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
- SmallVectorImpl<MCFixup> &Fixups,
- raw_ostream &OS) const {
- const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
-
- uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
-
- //older alu have different encoding for instructions with one or two src
- //parameters.
- if ((STI.getFeatureBits() & AMDGPU::FeatureR600ALUInst) &&
- !(MCDesc.TSFlags & R600_InstFlag::OP3)) {
- uint64_t ISAOpCode = InstWord01 & (0x3FFULL << 39);
- InstWord01 &= ~(0x3FFULL << 39);
- InstWord01 |= ISAOpCode << 1;
- }
-
- unsigned SrcNum = MCDesc.TSFlags & R600_InstFlag::OP3 ? 3 :
- MCDesc.TSFlags & R600_InstFlag::OP2 ? 2 : 1;
-
- const unsigned SrcOps[3][2] = {
- {R600Operands::SRC0, R600Operands::SRC0_SEL},
- {R600Operands::SRC1, R600Operands::SRC1_SEL},
- {R600Operands::SRC2, R600Operands::SRC2_SEL}
- };
-
- for (unsigned SrcIdx = 0; SrcIdx < SrcNum; ++SrcIdx) {
- unsigned RegOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][0]];
- unsigned SelOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][1]];
- }
-
- Emit(InstWord01, OS);
- return;
-}
-
-void R600MCCodeEmitter::EmitSrc(const MCInst &MI, unsigned OpIdx,
- raw_ostream &OS) const {
- const MCOperand &MO = MI.getOperand(OpIdx);
- union {
- float f;
- uint32_t i;
- } Value;
- Value.i = 0;
- // Emit the source select (2 bytes). For GPRs, this is the register index.
- // For other potential instruction operands, (e.g. constant registers) the
- // value of the source select is defined in the r600isa docs.
- if (MO.isReg()) {
- unsigned reg = MO.getReg();
- EmitTwoBytes(getHWReg(reg), OS);
- if (reg == AMDGPU::ALU_LITERAL_X) {
- unsigned ImmOpIndex = MI.getNumOperands() - 1;
- MCOperand ImmOp = MI.getOperand(ImmOpIndex);
- if (ImmOp.isFPImm()) {
- Value.f = ImmOp.getFPImm();
- } else {
- assert(ImmOp.isImm());
- Value.i = ImmOp.getImm();
- }
- }
- } else {
- // XXX: Handle other operand types.
- EmitTwoBytes(0, OS);
- }
-
- // Emit the source channel (1 byte)
- if (MO.isReg()) {
- EmitByte(getHWRegChan(MO.getReg()), OS);
- } else {
- EmitByte(0, OS);
- }
-
- // XXX: Emit isNegated (1 byte)
- if ((!(isFlagSet(MI, OpIdx, MO_FLAG_ABS)))
- && (isFlagSet(MI, OpIdx, MO_FLAG_NEG) ||
- (MO.isReg() &&
- (MO.getReg() == AMDGPU::NEG_ONE || MO.getReg() == AMDGPU::NEG_HALF)))){
- EmitByte(1, OS);
- } else {
- EmitByte(0, OS);
- }
-
- // Emit isAbsolute (1 byte)
- if (isFlagSet(MI, OpIdx, MO_FLAG_ABS)) {
- EmitByte(1, OS);
- } else {
- EmitByte(0, OS);
- }
-
- // XXX: Emit relative addressing mode (1 byte)
- EmitByte(0, OS);
-
- // Emit kc_bank, This will be adjusted later by r600_asm
- EmitByte(0, OS);
-
- // Emit the literal value, if applicable (4 bytes).
- Emit(Value.i, OS);
-
-}
-
-void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned RegOpIdx,
- unsigned SelOpIdx, raw_ostream &OS) const {
- const MCOperand &RegMO = MI.getOperand(RegOpIdx);
- const MCOperand &SelMO = MI.getOperand(SelOpIdx);
-
- union {
- float f;
- uint32_t i;
- } InlineConstant;
- InlineConstant.i = 0;
- // Emit source type (1 byte) and source select (4 bytes). For GPRs type is 0
- // and select is 0 (GPR index is encoded in the instr encoding. For constants
- // type is 1 and select is the original const select passed from the driver.
- unsigned Reg = RegMO.getReg();
- if (Reg == AMDGPU::ALU_CONST) {
- EmitByte(1, OS);
- uint32_t Sel = SelMO.getImm();
- Emit(Sel, OS);
- } else {
- EmitByte(0, OS);
- Emit((uint32_t)0, OS);
- }
-
- if (Reg == AMDGPU::ALU_LITERAL_X) {
- unsigned ImmOpIndex = MI.getNumOperands() - 2;
- MCOperand ImmOp = MI.getOperand(ImmOpIndex);
- if (ImmOp.isFPImm()) {
- InlineConstant.f = ImmOp.getFPImm();
- } else {
- assert(ImmOp.isImm());
- InlineConstant.i = ImmOp.getImm();
- }
- }
-
- // Emit the literal value, if applicable (4 bytes).
- Emit(InlineConstant.i, OS);
-}
-
-void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
-
- // Emit SRC
- unsigned NumOperands = MI.getNumOperands();
- if (NumOperands > 0) {
- assert(NumOperands == 1);
- EmitSrc(MI, 0, OS);
- } else {
- EmitNullBytes(SRC_BYTE_COUNT, OS);
- }
-
- // Emit FC Instruction
- enum FCInstr instr;
- switch (MI.getOpcode()) {
- case AMDGPU::PREDICATED_BREAK:
- instr = FC_BREAK_PREDICATE;
- break;
- case AMDGPU::CONTINUE:
- instr = FC_CONTINUE;
- break;
- case AMDGPU::IF_PREDICATE_SET:
- instr = FC_IF_PREDICATE;
- break;
- case AMDGPU::ELSE:
- instr = FC_ELSE;
- break;
- case AMDGPU::ENDIF:
- instr = FC_ENDIF;
- break;
- case AMDGPU::ENDLOOP:
- instr = FC_ENDLOOP;
- break;
- case AMDGPU::WHILELOOP:
- instr = FC_BGNLOOP;
- break;
- default:
- abort();
- break;
- }
- EmitByte(instr, OS);
-}
-
-void R600MCCodeEmitter::EmitNullBytes(unsigned int ByteCount,
- raw_ostream &OS) const {
-
- for (unsigned int i = 0; i < ByteCount; i++) {
- EmitByte(0, OS);
- }
-}
-
void R600MCCodeEmitter::EmitByte(unsigned int Byte, raw_ostream &OS) const {
OS.write((uint8_t) Byte & 0xff);
}
-void R600MCCodeEmitter::EmitTwoBytes(unsigned int Bytes,
- raw_ostream &OS) const {
- OS.write((uint8_t) (Bytes & 0xff));
- OS.write((uint8_t) ((Bytes >> 8) & 0xff));
-}
-
void R600MCCodeEmitter::Emit(uint32_t Value, raw_ostream &OS) const {
for (unsigned i = 0; i < 4; i++) {
OS.write((uint8_t) ((Value >> (8 * i)) & 0xff));
@@ -513,55 +256,4 @@ uint64_t R600MCCodeEmitter::getMachineOpValue(const MCInst &MI,
}
}
-//===----------------------------------------------------------------------===//
-// Encoding helper functions
-//===----------------------------------------------------------------------===//
-
-bool R600MCCodeEmitter::isFCOp(unsigned opcode) const {
- switch(opcode) {
- default: return false;
- case AMDGPU::PREDICATED_BREAK:
- case AMDGPU::CONTINUE:
- case AMDGPU::IF_PREDICATE_SET:
- case AMDGPU::ELSE:
- case AMDGPU::ENDIF:
- case AMDGPU::ENDLOOP:
- case AMDGPU::WHILELOOP:
- return true;
- }
-}
-
-bool R600MCCodeEmitter::isTexOp(unsigned opcode) const {
- switch(opcode) {
- default: return false;
- case AMDGPU::TEX_LD:
- case AMDGPU::TEX_GET_TEXTURE_RESINFO:
- case AMDGPU::TEX_SAMPLE:
- case AMDGPU::TEX_SAMPLE_C:
- case AMDGPU::TEX_SAMPLE_L:
- case AMDGPU::TEX_SAMPLE_C_L:
- case AMDGPU::TEX_SAMPLE_LB:
- case AMDGPU::TEX_SAMPLE_C_LB:
- case AMDGPU::TEX_SAMPLE_G:
- case AMDGPU::TEX_SAMPLE_C_G:
- case AMDGPU::TEX_GET_GRADIENTS_H:
- case AMDGPU::TEX_GET_GRADIENTS_V:
- case AMDGPU::TEX_SET_GRADIENTS_H:
- case AMDGPU::TEX_SET_GRADIENTS_V:
- return true;
- }
-}
-
-bool R600MCCodeEmitter::isFlagSet(const MCInst &MI, unsigned Operand,
- unsigned Flag) const {
- const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
- unsigned FlagIndex = GET_FLAG_OPERAND_IDX(MCDesc.TSFlags);
- if (FlagIndex == 0) {
- return false;
- }
- assert(MI.getOperand(FlagIndex).isImm());
- return !!((MI.getOperand(FlagIndex).getImm() >>
- (NUM_MO_FLAGS * Operand)) & Flag);
-}
-
#include "AMDGPUGenMCCodeEmitter.inc"
--
1.7.11.4
More information about the llvm-commits
mailing list