citra-emu
/
citra-canary
Archived
1
0
Fork 0
Code

Merge pull request #5546 from FearlessTobi/port-5524 

Port yuzu-emu/yuzu#4086 and yuzu-emu/yuzu#4611: Xbyak cleanups
This commit is contained in:
Marshall Mohror 2021-02-08 16:00:18 -06:00 committed by GitHub
parent e770fd6e95 9c08409e0e
commit 5776bdda82
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 79 additions and 70 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
externals/xbyak vendored

@ -1 +1 @@
Subproject commit 18c9caaa0a3ed5706c39f5aa86cce0db6e65b174 Subproject commit c306b8e5786eeeb87b8925a8af5c3bf057ff5a90

108
src/common/x64/xbyak_abi.h
View File

@ -4,14 +4,14 @@
#pragma once #pragma once
#include <bitset>
#include <initializer_list> #include <initializer_list>
#include <xbyak.h> #include <xbyak.h>
#include "common/assert.h" #include "common/assert.h"
#include "common/bit_set.h"
namespace Common::X64 { namespace Common::X64 {
inline int RegToIndex(const Xbyak::Reg& reg) { constexpr std::size_t RegToIndex(const Xbyak::Reg& reg) {
using Kind = Xbyak::Reg::Kind; using Kind = Xbyak::Reg::Kind;
ASSERT_MSG((reg.getKind() & (Kind::REG | Kind::XMM)) != 0, ASSERT_MSG((reg.getKind() & (Kind::REG | Kind::XMM)) != 0,
"RegSet only support GPRs and XMM registers."); "RegSet only support GPRs and XMM registers.");
@ -19,17 +19,17 @@ inline int RegToIndex(const Xbyak::Reg& reg) {
return reg.getIdx() + (reg.getKind() == Kind::REG ? 0 : 16); return reg.getIdx() + (reg.getKind() == Kind::REG ? 0 : 16);
} }
inline Xbyak::Reg64 IndexToReg64(int reg_index) { constexpr Xbyak::Reg64 IndexToReg64(std::size_t reg_index) {
ASSERT(reg_index < 16); ASSERT(reg_index < 16);
return Xbyak::Reg64(reg_index); return Xbyak::Reg64(static_cast<int>(reg_index));
} }
inline Xbyak::Xmm IndexToXmm(int reg_index) { constexpr Xbyak::Xmm IndexToXmm(std::size_t reg_index) {
ASSERT(reg_index >= 16 && reg_index < 32); ASSERT(reg_index >= 16 && reg_index < 32);
return Xbyak::Xmm(reg_index - 16); return Xbyak::Xmm(static_cast<int>(reg_index - 16));
} }
inline Xbyak::Reg IndexToReg(int reg_index) { constexpr Xbyak::Reg IndexToReg(std::size_t reg_index) {
if (reg_index < 16) { if (reg_index < 16) {
return IndexToReg64(reg_index); return IndexToReg64(reg_index);
} else { } else {
@ -37,27 +37,27 @@ inline Xbyak::Reg IndexToReg(int reg_index) {
} }
} }
inline BitSet32 BuildRegSet(std::initializer_list<Xbyak::Reg> regs) { inline std::bitset<32> BuildRegSet(std::initializer_list<Xbyak::Reg> regs) {
BitSet32 bits; std::bitset<32> bits;
for (const Xbyak::Reg& reg : regs) { for (const Xbyak::Reg& reg : regs) {
bits[RegToIndex(reg)] = true; bits[RegToIndex(reg)] = true;
} }
return bits; return bits;
} }
const BitSet32 ABI_ALL_GPRS(0x0000FFFF); constexpr inline std::bitset<32> ABI_ALL_GPRS(0x0000FFFF);
const BitSet32 ABI_ALL_XMMS(0xFFFF0000); constexpr inline std::bitset<32> ABI_ALL_XMMS(0xFFFF0000);
#ifdef _WIN32 #ifdef _WIN32
// Microsoft x64 ABI // Microsoft x64 ABI
const Xbyak::Reg ABI_RETURN = Xbyak::util::rax; constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax;
const Xbyak::Reg ABI_PARAM1 = Xbyak::util::rcx; constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rcx;
const Xbyak::Reg ABI_PARAM2 = Xbyak::util::rdx; constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rdx;
const Xbyak::Reg ABI_PARAM3 = Xbyak::util::r8; constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::r8;
const Xbyak::Reg ABI_PARAM4 = Xbyak::util::r9; constexpr inline Xbyak::Reg ABI_PARAM4 = Xbyak::util::r9;
const BitSet32 ABI_ALL_CALLER_SAVED = BuildRegSet({ const std::bitset<32> ABI_ALL_CALLER_SAVED = BuildRegSet({
// GPRs // GPRs
Xbyak::util::rcx, Xbyak::util::rcx,
Xbyak::util::rdx, Xbyak::util::rdx,
@ -74,7 +74,7 @@ const BitSet32 ABI_ALL_CALLER_SAVED = BuildRegSet({
Xbyak::util::xmm5, Xbyak::util::xmm5,
}); });
const BitSet32 ABI_ALL_CALLEE_SAVED = BuildRegSet({ const std::bitset<32> ABI_ALL_CALLEE_SAVED = BuildRegSet({
// GPRs // GPRs
Xbyak::util::rbx, Xbyak::util::rbx,
Xbyak::util::rsi, Xbyak::util::rsi,
@ -102,13 +102,13 @@ constexpr std::size_t ABI_SHADOW_SPACE = 0x20;
#else #else
// System V x86-64 ABI // System V x86-64 ABI
const Xbyak::Reg ABI_RETURN = Xbyak::util::rax; constexpr inline Xbyak::Reg ABI_RETURN = Xbyak::util::rax;
const Xbyak::Reg ABI_PARAM1 = Xbyak::util::rdi; constexpr inline Xbyak::Reg ABI_PARAM1 = Xbyak::util::rdi;
const Xbyak::Reg ABI_PARAM2 = Xbyak::util::rsi; constexpr inline Xbyak::Reg ABI_PARAM2 = Xbyak::util::rsi;
const Xbyak::Reg ABI_PARAM3 = Xbyak::util::rdx; constexpr inline Xbyak::Reg ABI_PARAM3 = Xbyak::util::rdx;
const Xbyak::Reg ABI_PARAM4 = Xbyak::util::rcx; constexpr inline Xbyak::Reg ABI_PARAM4 = Xbyak::util::rcx;
const BitSet32 ABI_ALL_CALLER_SAVED = BuildRegSet({ const std::bitset<32> ABI_ALL_CALLER_SAVED = BuildRegSet({
// GPRs // GPRs
Xbyak::util::rcx, Xbyak::util::rcx,
Xbyak::util::rdx, Xbyak::util::rdx,
@ -137,7 +137,7 @@ const BitSet32 ABI_ALL_CALLER_SAVED = BuildRegSet({
Xbyak::util::xmm15, Xbyak::util::xmm15,
}); });
const BitSet32 ABI_ALL_CALLEE_SAVED = BuildRegSet({ const std::bitset<32> ABI_ALL_CALLEE_SAVED = BuildRegSet({
// GPRs // GPRs
Xbyak::util::rbx, Xbyak::util::rbx,
Xbyak::util::rbp, Xbyak::util::rbp,
@ -151,13 +151,17 @@ constexpr std::size_t ABI_SHADOW_SPACE = 0;
#endif #endif
inline void ABI_CalculateFrameSize(BitSet32 regs, std::size_t rsp_alignment, struct ABIFrameInfo {
std::size_t needed_frame_size, s32* out_subtraction, s32 subtraction;
s32* out_xmm_offset) { s32 xmm_offset;
int count = (regs & ABI_ALL_GPRS).Count(); };
inline ABIFrameInfo ABI_CalculateFrameSize(std::bitset<32> regs, std::size_t rsp_alignment,
std::size_t needed_frame_size) {
int count = (regs & ABI_ALL_GPRS).count();
rsp_alignment -= count * 8; rsp_alignment -= count * 8;
std::size_t subtraction = 0; std::size_t subtraction = 0;
int xmm_count = (regs & ABI_ALL_XMMS).Count(); int xmm_count = (regs & ABI_ALL_XMMS).count();
if (xmm_count) { if (xmm_count) {
// If we have any XMMs to save, we must align the stack here. // If we have any XMMs to save, we must align the stack here.
subtraction = rsp_alignment & 0xF; subtraction = rsp_alignment & 0xF;
@ -170,45 +174,49 @@ inline void ABI_CalculateFrameSize(BitSet32 regs, std::size_t rsp_alignment,
rsp_alignment -= subtraction; rsp_alignment -= subtraction;
subtraction += rsp_alignment & 0xF; subtraction += rsp_alignment & 0xF;
*out_subtraction = (s32)subtraction; return ABIFrameInfo{static_cast<s32>(subtraction),
*out_xmm_offset = (s32)(subtraction - xmm_base_subtraction); static_cast<s32>(subtraction - xmm_base_subtraction)};
} }
inline std::size_t ABI_PushRegistersAndAdjustStack(Xbyak::CodeGenerator& code, BitSet32 regs, inline std::size_t ABI_PushRegistersAndAdjustStack(Xbyak::CodeGenerator& code, std::bitset<32> regs,
std::size_t rsp_alignment, std::size_t rsp_alignment,
std::size_t needed_frame_size = 0) { std::size_t needed_frame_size = 0) {
s32 subtraction, xmm_offset; auto frame_info = ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size);
ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size, &subtraction, &xmm_offset);
for (int reg_index : (regs & ABI_ALL_GPRS)) { for (std::size_t i = 0; i < regs.size(); ++i) {
code.push(IndexToReg64(reg_index)); if (regs[i] && ABI_ALL_GPRS[i]) {
code.push(IndexToReg64(i));
}
} }
if (subtraction != 0) { if (frame_info.subtraction != 0) {
code.sub(code.rsp, subtraction); code.sub(code.rsp, frame_info.subtraction);
} }
for (int reg_index : (regs & ABI_ALL_XMMS)) { for (std::size_t i = 0; i < regs.size(); ++i) {
code.movaps(code.xword[code.rsp + xmm_offset], IndexToXmm(reg_index)); if (regs[i] && ABI_ALL_XMMS[i]) {
xmm_offset += 0x10; code.movaps(code.xword[code.rsp + frame_info.xmm_offset], IndexToXmm(i));
frame_info.xmm_offset += 0x10;
}
} }
return ABI_SHADOW_SPACE; return ABI_SHADOW_SPACE;
} }
inline void ABI_PopRegistersAndAdjustStack(Xbyak::CodeGenerator& code, BitSet32 regs, inline void ABI_PopRegistersAndAdjustStack(Xbyak::CodeGenerator& code, std::bitset<32> regs,
std::size_t rsp_alignment, std::size_t rsp_alignment,
std::size_t needed_frame_size = 0) { std::size_t needed_frame_size = 0) {
s32 subtraction, xmm_offset; auto frame_info = ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size);
ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size, &subtraction, &xmm_offset);
for (int reg_index : (regs & ABI_ALL_XMMS)) { for (std::size_t i = 0; i < regs.size(); ++i) {
code.movaps(IndexToXmm(reg_index), code.xword[code.rsp + xmm_offset]); if (regs[i] && ABI_ALL_XMMS[i]) {
xmm_offset += 0x10; code.movaps(IndexToXmm(i), code.xword[code.rsp + frame_info.xmm_offset]);
frame_info.xmm_offset += 0x10;
}
} }
if (subtraction != 0) { if (frame_info.subtraction != 0) {
code.add(code.rsp, subtraction); code.add(code.rsp, frame_info.subtraction);
} }
// GPRs need to be popped in reverse order // GPRs need to be popped in reverse order

36
src/video_core/shader/shader_jit_x64_compiler.cpp
View File

@ -102,40 +102,40 @@ const JitFunction instr_table[64] = {
// purposes, as documented below: // purposes, as documented below:
/// Pointer to the uniform memory /// Pointer to the uniform memory
static const Reg64 UNIFORMS = r9; constexpr Reg64 UNIFORMS = r9;
/// The two 32-bit VS address offset registers set by the MOVA instruction /// The two 32-bit VS address offset registers set by the MOVA instruction
static const Reg64 ADDROFFS_REG_0 = r10; constexpr Reg64 ADDROFFS_REG_0 = r10;
static const Reg64 ADDROFFS_REG_1 = r11; constexpr Reg64 ADDROFFS_REG_1 = r11;
/// VS loop count register (Multiplied by 16) /// VS loop count register (Multiplied by 16)
static const Reg32 LOOPCOUNT_REG = r12d; constexpr Reg32 LOOPCOUNT_REG = r12d;
/// Current VS loop iteration number (we could probably use LOOPCOUNT_REG, but this quicker) /// Current VS loop iteration number (we could probably use LOOPCOUNT_REG, but this quicker)
static const Reg32 LOOPCOUNT = esi; constexpr Reg32 LOOPCOUNT = esi;
/// Number to increment LOOPCOUNT_REG by on each loop iteration (Multiplied by 16) /// Number to increment LOOPCOUNT_REG by on each loop iteration (Multiplied by 16)
static const Reg32 LOOPINC = edi; constexpr Reg32 LOOPINC = edi;
/// Result of the previous CMP instruction for the X-component comparison /// Result of the previous CMP instruction for the X-component comparison
static const Reg64 COND0 = r13; constexpr Reg64 COND0 = r13;
/// Result of the previous CMP instruction for the Y-component comparison /// Result of the previous CMP instruction for the Y-component comparison
static const Reg64 COND1 = r14; constexpr Reg64 COND1 = r14;
/// Pointer to the UnitState instance for the current VS unit /// Pointer to the UnitState instance for the current VS unit
static const Reg64 STATE = r15; constexpr Reg64 STATE = r15;
/// SIMD scratch register /// SIMD scratch register
static const Xmm SCRATCH = xmm0; constexpr Xmm SCRATCH = xmm0;
/// Loaded with the first swizzled source register, otherwise can be used as a scratch register /// Loaded with the first swizzled source register, otherwise can be used as a scratch register
static const Xmm SRC1 = xmm1; constexpr Xmm SRC1 = xmm1;
/// Loaded with the second swizzled source register, otherwise can be used as a scratch register /// Loaded with the second swizzled source register, otherwise can be used as a scratch register
static const Xmm SRC2 = xmm2; constexpr Xmm SRC2 = xmm2;
/// Loaded with the third swizzled source register, otherwise can be used as a scratch register /// Loaded with the third swizzled source register, otherwise can be used as a scratch register
static const Xmm SRC3 = xmm3; constexpr Xmm SRC3 = xmm3;
/// Additional scratch register /// Additional scratch register
static const Xmm SCRATCH2 = xmm4; constexpr Xmm SCRATCH2 = xmm4;
/// Constant vector of [1.0f, 1.0f, 1.0f, 1.0f], used to efficiently set a vector to one /// Constant vector of [1.0f, 1.0f, 1.0f, 1.0f], used to efficiently set a vector to one
static const Xmm ONE = xmm14; constexpr Xmm ONE = xmm14;
/// Constant vector of [-0.f, -0.f, -0.f, -0.f], used to efficiently negate a vector with XOR /// Constant vector of [-0.f, -0.f, -0.f, -0.f], used to efficiently negate a vector with XOR
static const Xmm NEGBIT = xmm15; constexpr Xmm NEGBIT = xmm15;
// State registers that must not be modified by external functions calls // State registers that must not be modified by external functions calls
// Scratch registers, e.g., SRC1 and SCRATCH, have to be saved on the side if needed // Scratch registers, e.g., SRC1 and SCRATCH, have to be saved on the side if needed
static const BitSet32 persistent_regs = BuildRegSet({ static const std::bitset<32> persistent_regs = BuildRegSet({
// Pointers to register blocks // Pointers to register blocks
UNIFORMS, UNIFORMS,
STATE, STATE,
@ -356,7 +356,7 @@ void JitShader::Compile_UniformCondition(Instruction instr) {
cmp(byte[UNIFORMS + offset], 0); cmp(byte[UNIFORMS + offset], 0);
} }
BitSet32 JitShader::PersistentCallerSavedRegs() { std::bitset<32> JitShader::PersistentCallerSavedRegs() {
return persistent_regs & ABI_ALL_CALLER_SAVED; return persistent_regs & ABI_ALL_CALLER_SAVED;
} }

3
src/video_core/shader/shader_jit_x64_compiler.h
View File

@ -5,6 +5,7 @@
#pragma once #pragma once
#include <array> #include <array>
#include <bitset>
#include <cstddef> #include <cstddef>
#include <optional> #include <optional>
#include <utility> #include <utility>
@ -91,7 +92,7 @@ private:
*/ */
void Compile_Return(); void Compile_Return();
BitSet32 PersistentCallerSavedRegs(); std::bitset<32> PersistentCallerSavedRegs();
/** /**
* Assertion evaluated at compile-time, but only triggered if executed at runtime. * Assertion evaluated at compile-time, but only triggered if executed at runtime.