ShuriZma
/
flycast
mirror of https://github.com/flyinghead/flycast.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

arm64: implement swaplb and pref. Various optimizations

This commit is contained in:
Flyinghead 2019-01-11 15:58:48 +01:00
parent ec7787c56a
commit e5ee48efa9
2 changed files with 176 additions and 91 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
core/rec-ARM64/arm64_regalloc.h
View File

@ -34,11 +34,7 @@ enum eFReg {
S17, S18, S19, S20, S21, S22, S23, S24, S25, S26, S27, S28, S29, S30, S31
};
#if HOST_OS == OS_DARWIN
#error "TODO static eReg alloc_regs[]={r5,r6,r7,r10,(eReg)-1};" // TODO
#else
static eReg alloc_regs[] = { W19, W20, W21, W22, W23, W24, W25, W26, (eReg)-1 };
#endif
static eFReg alloc_fregs[] = { S8, S9, S10, S11, S12, S13, S14, S15, (eFReg)-1 };
class Arm64Assembler;

263
core/rec-ARM64/rec_arm64.cpp
View File

@ -23,6 +23,7 @@
#include <unistd.h>
#include <sys/mman.h>
#include <map>
#include "deps/vixl/aarch64/macro-assembler-aarch64.h"
using namespace vixl::aarch64;
@ -37,6 +38,8 @@ using namespace vixl::aarch64;
#include "hw/sh4/sh4_rom.h"
#include "arm64_regalloc.h"
#undef do_sqw_nommu
struct DynaRBI : RuntimeBlockInfo
{
virtual u32 Relink() {
@ -197,9 +200,8 @@ void ngen_blockcheckfail(u32 pc) {
class Arm64Assembler : public MacroAssembler
{
typedef void (MacroAssembler::*Arm64Op)(const Register&, const Register&, const Operand&);
typedef void (MacroAssembler::*Arm64Op2)(const Register&, const Register&, const Register&);
typedef void (MacroAssembler::*Arm64Op3)(const Register&, const Register&, const Operand&, enum FlagsUpdate);
typedef void (MacroAssembler::*Arm64Op_RRO)(const Register&, const Register&, const Operand&);
typedef void (MacroAssembler::*Arm64Op_RROF)(const Register&, const Register&, const Operand&, enum FlagsUpdate);
public:
Arm64Assembler() : MacroAssembler((u8 *)emit_GetCCPtr(), 64 * 1024), regalloc(this)
@ -232,24 +234,66 @@ public:
call_fregs.push_back(&s7);
}
void ngen_BinaryOp(shil_opcode* op, Arm64Op arm_op, Arm64Op2 arm_op2, Arm64Op3 arm_op3)
void ngen_BinaryOp_RRO(shil_opcode* op, Arm64Op_RRO arm_op, Arm64Op_RROF arm_op2)
{
const Register* reg3 = &wzr;
Operand op3 = Operand(0);
if (op->rs2.is_imm())
{
Mov(w10, op->rs2._imm);
reg3 = &w10;
op3 = Operand(op->rs2._imm);
}
else if (op->rs2.is_r32i())
{
reg3 = &regalloc.MapRegister(op->rs2);
op3 = Operand(regalloc.MapRegister(op->rs2));
}
if (arm_op != NULL)
((*this).*arm_op)(regalloc.MapRegister(op->rd), regalloc.MapRegister(op->rs1), *reg3);
else if (arm_op2 != NULL)
((*this).*arm_op2)(regalloc.MapRegister(op->rd), regalloc.MapRegister(op->rs1), *reg3);
((*this).*arm_op)(regalloc.MapRegister(op->rd), regalloc.MapRegister(op->rs1), op3);
else
((*this).*arm_op3)(regalloc.MapRegister(op->rd), regalloc.MapRegister(op->rs1), *reg3, LeaveFlags);
((*this).*arm_op2)(regalloc.MapRegister(op->rd), regalloc.MapRegister(op->rs1), op3, LeaveFlags);
}
template <typename R, typename... P>
void ngen_CallRuntime(R (*function)(P...))
{
if (!frame_reg_saved)
{
Str(x30, MemOperand(sp, -16, PreIndex));
frame_reg_saved = true;
}
Literal<uintptr_t> *function_address = function_literals[(void*)function];
if (function_address == NULL)
{
function_address = new Literal<uintptr_t>(reinterpret_cast<uintptr_t>(function), GetLiteralPool(), RawLiteral::kDeletedOnPoolDestruction);
function_literals[(void*)function] = function_address;
}
Ldr(x9, function_address);
Blr(x9);
}
const Register& GenMemAddr(const shil_opcode& op, const Register* raddr = NULL)
{
const Register* ret_reg = raddr == NULL ? &w0 : raddr;
if (op.rs3.is_imm())
{
Add(*ret_reg, regalloc.MapRegister(op.rs1), op.rs3._imm);
}
else if (op.rs3.is_r32i())
{
Add(*ret_reg, regalloc.MapRegister(op.rs1), regalloc.MapRegister(op.rs3));
}
else if (!op.rs3.is_null())
{
die("invalid rs3");
}
else
{
if (raddr == NULL)
ret_reg = &regalloc.MapRegister(op.rs1);
else
Mov(*ret_reg, regalloc.MapRegister(op.rs1));
}
return *ret_reg;
}
void ngen_Compile(RuntimeBlockInfo* block, bool force_checks, bool reset, bool staging, bool optimise)
@ -258,8 +302,6 @@ public:
if (force_checks)
CheckBlock(block);
Str(x30, MemOperand(sp, -16, PreIndex));
// run register allocator
regalloc.DoAlloc(block);
@ -281,19 +323,18 @@ public:
}
Mov(*call_regs[0], op.rs3._imm);
CallRuntime(OpDesc[op.rs3._imm]->oph);
ngen_CallRuntime(OpDesc[op.rs3._imm]->oph);
break;
case shop_jcond:
case shop_jdyn:
Mov(w10, regalloc.MapRegister(op.rs1));
if (op.rs2.is_imm()) {
Mov(w9, op.rs2._imm);
Add(w10, w10, w9);
if (op.rs2.is_imm())
{
Add(w10, regalloc.MapRegister(op.rs1), op.rs2._imm);
Mov(regalloc.MapRegister(op.rd), w10);
}
Mov(regalloc.MapRegister(op.rd), w10);
else
Mov(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1));
break;
case shop_mov32:
@ -340,7 +381,7 @@ public:
if (isram)
{
Mov(x1, reinterpret_cast<uintptr_t>(ptr));
Ldr(x1, reinterpret_cast<uintptr_t>(ptr));
switch (size)
{
case 2:
@ -367,17 +408,17 @@ public:
switch(size)
{
case 1:
CallRuntime((void (*)())ptr);
ngen_CallRuntime((void (*)())ptr);
Sxtb(w0, w0);
break;
case 2:
CallRuntime((void (*)())ptr);
ngen_CallRuntime((void (*)())ptr);
Sxth(w0, w0);
break;
case 4:
CallRuntime((void (*)())ptr);
ngen_CallRuntime((void (*)())ptr);
break;
case 8:
@ -432,31 +473,26 @@ public:
}
#endif
shil_param_to_host_reg(op.rs1, *call_regs[0]);
if (!op.rs3.is_null())
{
shil_param_to_host_reg(op.rs3, w10);
Add(*call_regs[0], *call_regs[0], w10);
}
GenMemAddr(op, call_regs[0]);
switch (size)
{
case 1:
CallRuntime(ReadMem8);
ngen_CallRuntime(ReadMem8);
Sxtb(w0, w0);
break;
case 2:
CallRuntime(ReadMem16);
ngen_CallRuntime(ReadMem16);
Sxth(w0, w0);
break;
case 4:
CallRuntime(ReadMem32);
ngen_CallRuntime(ReadMem32);
break;
case 8:
CallRuntime(ReadMem64);
ngen_CallRuntime(ReadMem64);
break;
default:
@ -474,12 +510,7 @@ public:
case shop_writem:
{
shil_param_to_host_reg(op.rs1, *call_regs[0]);
if (!op.rs3.is_null())
{
shil_param_to_host_reg(op.rs3, w10);
Add(*call_regs[0], *call_regs[0], w10);
}
GenMemAddr(op, call_regs[0]);
u32 size = op.flags & 0x7f;
if (size != 8)
@ -490,19 +521,19 @@ public:
switch (size)
{
case 1:
CallRuntime(WriteMem8);
ngen_CallRuntime(WriteMem8);
break;
case 2:
CallRuntime(WriteMem16);
ngen_CallRuntime(WriteMem16);
break;
case 4:
CallRuntime(WriteMem32);
ngen_CallRuntime(WriteMem32);
break;
case 8:
CallRuntime(WriteMem64);
ngen_CallRuntime(WriteMem64);
break;
default:
@ -513,7 +544,18 @@ public:
break;
case shop_sync_sr:
CallRuntime(UpdateSR);
ngen_CallRuntime(UpdateSR);
break;
case shop_sync_fpscr:
ngen_CallRuntime(UpdateFPSCR);
break;
case shop_swaplb:
// TODO Optimize
Mov(w9, Operand(regalloc.MapRegister(op.rs1), LSR, 16));
Rev16(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1));
Bfc(regalloc.MapRegister(op.rd), 16, 16);
Orr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), Operand(w9, LSL, 16));
break;
case shop_neg:
@ -523,32 +565,44 @@ public:
Mvn(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1));
break;
case shop_shl:
ngen_BinaryOp(&op, NULL, &MacroAssembler::Lsl, NULL);
break;
case shop_shr:
ngen_BinaryOp(&op, NULL, &MacroAssembler::Lsr, NULL);
break;
case shop_sar:
ngen_BinaryOp(& op, NULL, &MacroAssembler::Asr, NULL);
break;
case shop_and:
ngen_BinaryOp(&op, &MacroAssembler::And, NULL, NULL);
ngen_BinaryOp_RRO(&op, &MacroAssembler::And, NULL);
break;
case shop_or:
ngen_BinaryOp(&op, &MacroAssembler::Orr, NULL, NULL);
ngen_BinaryOp_RRO(&op, &MacroAssembler::Orr, NULL);
break;
case shop_xor:
ngen_BinaryOp(&op, &MacroAssembler::Eor, NULL, NULL);
ngen_BinaryOp_RRO(&op, &MacroAssembler::Eor, NULL);
break;
case shop_add:
ngen_BinaryOp(&op, NULL, NULL, &MacroAssembler::Add);
ngen_BinaryOp_RRO(&op, NULL, &MacroAssembler::Add);
break;
case shop_sub:
ngen_BinaryOp(&op, NULL, NULL, &MacroAssembler::Sub);
ngen_BinaryOp_RRO(&op, NULL, &MacroAssembler::Sub);
break;
case shop_shl:
if (op.rs2.is_imm())
Lsl(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), op.rs2._imm);
else if (op.rs3.is_reg())
Lsl(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), regalloc.MapRegister(op.rs2));
break;
case shop_shr:
if (op.rs2.is_imm())
Lsr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), op.rs2._imm);
else if (op.rs3.is_reg())
Lsr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), regalloc.MapRegister(op.rs2));
break;
case shop_sar:
if (op.rs2.is_imm())
Asr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), op.rs2._imm);
else if (op.rs3.is_reg())
Asr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), regalloc.MapRegister(op.rs2));
break;
case shop_ror:
ngen_BinaryOp(&op, NULL, &MacroAssembler::Ror, NULL);
if (op.rs2.is_imm())
Ror(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), op.rs2._imm);
else if (op.rs3.is_reg())
Ror(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), regalloc.MapRegister(op.rs2));
break;
case shop_adc:
@ -563,10 +617,10 @@ public:
break;
case shop_rocr:
Ubfm(w0, regalloc.MapRegister(op.rs1), 0, 0);
Mov(regalloc.MapRegister(op.rd), Operand(regalloc.MapRegister(op.rs1), LSR, 1));
Orr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), Operand(regalloc.MapRegister(op.rs2), LSL, 31));
Mov(regalloc.MapRegister(op.rd2), w0);
Ubfm(w0, regalloc.MapRegister(op.rs1), 0, 0); // w0 = rs1[0]
Mov(regalloc.MapRegister(op.rd), Operand(regalloc.MapRegister(op.rs1), LSR, 1)); // rd = rs1 >> 1
Orr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), Operand(regalloc.MapRegister(op.rs2), LSL, 31)); // rd |= C << 31
Mov(regalloc.MapRegister(op.rd2), w0); // rd2 = w0 (new C)
break;
case shop_rocl:
Tst(regalloc.MapRegister(op.rs1), 0x80000000); // Z = ~rs1[31]
@ -575,14 +629,6 @@ public:
break;
case shop_shld:
// TODO optimize
Cmp(regalloc.MapRegister(op.rs2), 0);
Csel(w1, regalloc.MapRegister(op.rs2), wzr, ge);
Mov(w0, wzr); // wzr not supported by csneg
Csneg(w2, w0, regalloc.MapRegister(op.rs2), ge);
Lsl(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), w1);
Lsr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), w2);
break;
case shop_shad:
// TODO optimize
Cmp(regalloc.MapRegister(op.rs2), 0);
@ -590,7 +636,12 @@ public:
Mov(w0, wzr); // wzr not supported by csneg
Csneg(w2, w0, regalloc.MapRegister(op.rs2), ge);
Lsl(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1), w1);
Asr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), w2);
if (op.op == shop_shld)
// Logical shift
Lsr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), w2);
else
// Arithmetic shift
Asr(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rd), w2);
break;
case shop_test:
@ -661,6 +712,42 @@ public:
}
break;
case shop_pref:
Mov(w0, regalloc.MapRegister(op.rs1));
if (op.flags != 0x1337)
{
Lsr(w1, regalloc.MapRegister(op.rs1), 26);
Cmp(w1, 0x38);
}
if (CCN_MMUCR.AT)
{
Mov(x9, reinterpret_cast<uintptr_t>(&do_sqw_mmu));
}
else
{
verify(offsetof(Sh4RCB, cntx) - offsetof(Sh4RCB, do_sqw_nommu) > 0);
Sub(x9, x28, offsetof(Sh4RCB, cntx) - offsetof(Sh4RCB, do_sqw_nommu));
Ldr(x9, MemOperand(x9));
verify(offsetof(Sh4RCB, cntx) - offsetof(Sh4RCB, sq_buffer) > 0);
Sub(x1, x28, offsetof(Sh4RCB, cntx) - offsetof(Sh4RCB, sq_buffer));
}
if (!frame_reg_saved)
{
Str(x30, MemOperand(sp, -16, PreIndex));
frame_reg_saved = true;
}
if (op.flags == 0x1337)
Blr(x9);
else
{
Label no_branch;
B(&no_branch, ne);
Blr(x9);
Bind(&no_branch);
}
break;
case shop_ext_s8:
Sxtb(regalloc.MapRegister(op.rd), regalloc.MapRegister(op.rs1));
break;
@ -714,13 +801,14 @@ public:
case shop_fsca:
Mov(x1, reinterpret_cast<uintptr_t>(&sin_table));
Add(x1, x1, Operand(regalloc.MapRegister(op.rs1), UXTH, 3));
// TODO use regalloc
//Ldr(regalloc.MapVRegister(op.rd, 0), MemOperand(x1, 4, PostIndex));
//Ldr(regalloc.MapVRegister(op.rd, 1), MemOperand(x1));
regalloc.writeback_fpu += 2;
Ldr(w2, MemOperand(x1, 4, PostIndex));
Str(w2, sh4_context_mem_operand(op.rd.reg_ptr())); // TODO use regalloc
Str(w2, sh4_context_mem_operand(op.rd.reg_ptr()));
Ldr(w2, MemOperand(x1));
Str(w2, sh4_context_mem_operand(GetRegPtr(op.rd._reg + 1))); // TODO use regalloc
Str(w2, sh4_context_mem_operand(GetRegPtr(op.rd._reg + 1)));
break;
case shop_fipr:
@ -835,14 +923,15 @@ public:
}
Str(w10, sh4_context_mem_operand(&next_pc));
CallRuntime(UpdateINTC);
ngen_CallRuntime(UpdateINTC);
break;
default:
die("Invalid block end type");
}
Ldr(x30, MemOperand(sp, 16, PostIndex));
if (frame_reg_saved)
Ldr(x30, MemOperand(sp, 16, PostIndex));
Ret();
Label code_end;
@ -950,7 +1039,7 @@ public:
break;
}
}
CallRuntime((void (*)())function);
ngen_CallRuntime((void (*)())function);
}
MemOperand sh4_context_mem_operand(void *p)
@ -973,11 +1062,11 @@ private:
// FIXME Can a block cross a RAM / non-RAM boundary??
return;
Mov(x9, reinterpret_cast<uintptr_t>(ptr));
Ldr(x9, reinterpret_cast<uintptr_t>(ptr));
while (sz > 0)
{
if (sz >= 8 && (reinterpret_cast<uintptr_t>(ptr) & 7) == 0)
if (sz >= 8)
{
Ldr(x10, MemOperand(x9, 8, PostIndex));
Ldr(x11, *(u64*)ptr);
@ -985,7 +1074,7 @@ private:
sz -= 8;
ptr += 8;
}
else if (sz >= 4 && (reinterpret_cast<uintptr_t>(ptr) & 3) == 0)
else if (sz >= 4)
{
Ldr(w10, MemOperand(x9, 4, PostIndex));
Ldr(w11, *(u32*)ptr);
@ -1069,6 +1158,8 @@ private:
std::vector<const XRegister*> call_regs64;
std::vector<const VRegister*> call_fregs;
Arm64RegAlloc regalloc;
bool frame_reg_saved = false;
std::map<void*, Literal<uintptr_t>*> function_literals;
};
static Arm64Assembler* compiler;
@ -1134,10 +1225,8 @@ extern "C" void do_sqw_nommu_area_3(u32 dst, u8* sqb)
"add x1, x12, x1 \n\t" // SQ# selection, add to SQ ptr
"add x11, x11, #512 \n\t" // get ram ptr from x1, part 2
"add x11, x11, x0, lsl #5 \n\t" // ram + offset
"ldp x9, x10, [x1], #16 \n\t"
"stp x9, x10, [x11], #16 \n\t"
"ldp x9, x10, [x1] \n\t"
"stp x9, x10, [x11] \n\t"
"ld2 { v0.2D, v1.2D }, [x1] \n\t"
"st2 { v0.2D, v1.2D }, [x11] \n\t"
"ret \n"
: : : "memory"