GS/SW: Make local a function parameter to SetupPrim

.. and associated 32-bit removal.
This commit is contained in:
Stenzek 2023-01-25 20:51:11 +10:00 committed by refractionpcsx2
parent c8416b820b
commit 3292121e67
5 changed files with 18 additions and 33 deletions

View File

@ -1573,9 +1573,9 @@ void GSDrawScanline::CDrawScanline(int pixels, int left, int top, const GSVertex
}
#ifndef ENABLE_JIT_RASTERIZER
void GSDrawScanline::SetupPrim(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan)
void GSDrawScanline::SetupPrim(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan, GSScanlineLocalData& local)
{
CSetupPrim(vertex, index, dscan, m_local, m_global);
CSetupPrim(vertex, index, dscan, local, *local.gd);
}
void GSDrawScanline::DrawScanline(int pixels, int left, int top, const GSVertexSW& scan)
{

View File

@ -33,7 +33,7 @@ public:
GSScanlineGlobalData global;
};
typedef void (*SetupPrimPtr)(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan);
typedef void (*SetupPrimPtr)(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan, GSScanlineLocalData& local);
typedef void (*DrawScanlinePtr)(int pixels, int left, int top, const GSVertexSW& scan);
protected:
@ -69,6 +69,9 @@ public:
GSDrawScanline();
virtual ~GSDrawScanline() = default;
__forceinline GSScanlineGlobalData& GetGlobalData() { return m_global; }
__forceinline GSScanlineLocalData& GetLocalData() { return m_local; }
__forceinline bool HasEdge() const { return m_de != nullptr; }
__forceinline bool IsSolidRect() const { return m_global.sel.IsSolidRect(); }
@ -85,13 +88,13 @@ public:
#ifdef ENABLE_JIT_RASTERIZER
__forceinline void SetupPrim(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan) { m_sp(vertex, index, dscan); }
__forceinline void SetupPrim(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan, GSScanlineLocalData& local) { m_sp(vertex, index, dscan, local); }
__forceinline void DrawScanline(int pixels, int left, int top, const GSVertexSW& scan) { m_ds(pixels, left, top, scan); }
__forceinline void DrawEdge(int pixels, int left, int top, const GSVertexSW& scan) { m_de(pixels, left, top, scan); }
#else
void SetupPrim(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan);
void SetupPrim(const GSVertexSW* vertex, const u32* index, const GSVertexSW& dscan, GSScanlineLocalData& local);
void DrawScanline(int pixels, int left, int top, const GSVertexSW& scan);
void DrawEdge(int pixels, int left, int top, const GSVertexSW& scan);

View File

@ -280,7 +280,7 @@ void GSRasterizer::DrawPoint(const GSVertexSW* vertex, int vertex_count, const u
{
if (IsOneOfMyScanlines(p.y))
{
m_ds->SetupPrim(vertex, index, GSVertexSW::zero());
m_ds->SetupPrim(vertex, index, GSVertexSW::zero(), m_ds->GetLocalData());
DrawScanline(1, p.x, p.y, v);
}
@ -301,7 +301,7 @@ void GSRasterizer::DrawPoint(const GSVertexSW* vertex, int vertex_count, const u
{
if (IsOneOfMyScanlines(p.y))
{
m_ds->SetupPrim(vertex, tmp_index, GSVertexSW::zero());
m_ds->SetupPrim(vertex, tmp_index, GSVertexSW::zero(), m_ds->GetLocalData());
DrawScanline(1, p.x, p.y, v);
}
@ -369,7 +369,7 @@ void GSRasterizer::DrawLine(const GSVertexSW* vertex, const u32* index)
scan += dscan * (l - scan.p).xxxx();
m_ds->SetupPrim(vertex, index, dscan);
m_ds->SetupPrim(vertex, index, dscan, m_ds->GetLocalData());
DrawScanline(pixels, left, p.y, scan);
}
@ -864,7 +864,7 @@ void GSRasterizer::DrawSprite(const GSVertexSW* vertex, const u32* index)
scan.t = (scan.t + dt * prestep).xyzw(scan.t);
m_ds->SetupPrim(vertex, index, dscan);
m_ds->SetupPrim(vertex, index, dscan, m_ds->GetLocalData());
while (1)
{
@ -1093,7 +1093,7 @@ void GSRasterizer::Flush(const GSVertexSW* vertex, const u32* index, const GSVer
if (count > 0)
{
m_ds->SetupPrim(vertex, index, dscan);
m_ds->SetupPrim(vertex, index, dscan, m_ds->GetLocalData());
const GSVertexSW* RESTRICT e = m_edge.buff;
const GSVertexSW* RESTRICT ee = e + count;

View File

@ -20,7 +20,7 @@
MULTI_ISA_UNSHARED_IMPL;
using namespace Xbyak;
#define _rip_local(field) ((m_rip) ? ptr[rip + (char*)&m_local.field] : ptr[_m_local + OFFSETOF(GSScanlineLocalData, field)])
#define _rip_local(field) (ptr[_m_local + OFFSETOF(GSScanlineLocalData, field)])
#define _64_m_local _64_t0
@ -51,20 +51,19 @@ using namespace Xbyak;
GSSetupPrimCodeGenerator2::GSSetupPrimCodeGenerator2(Xbyak::CodeGenerator* base, const ProcessorFeatures& cpu, void* param, u64 key)
: _parent(base, cpu)
, m_local(*(GSScanlineLocalData*)param)
, m_rip(false), many_regs(false)
, many_regs(false)
// On x86 arg registers are very temporary but on x64 they aren't, so on x86 some registers overlap
#ifdef _WIN32
, _64_vertex(rcx)
, _index(rdx)
, _dscan(r8)
, _64_t0(r9), t1(r10)
, _m_local(r9), t1(r10)
#else
, _64_vertex(rdi)
, _index(rsi)
, _dscan(rdx)
, _64_t0(rcx), t1(r8)
, _m_local(rcx), t1(r8)
#endif
, _m_local(chooseLocal(&m_local, _64_m_local))
{
m_sel.key = key;
@ -98,9 +97,6 @@ void GSSetupPrimCodeGenerator2::broadcastss(const XYm& reg, const Address& mem)
void GSSetupPrimCodeGenerator2::Generate()
{
// Technically we just need the delta < 2GB
m_rip = (size_t)&m_local < 0x80000000 && (size_t)getCurr() < 0x80000000;
bool needs_shift = ((m_en.z || m_en.f) && m_sel.prim != GS_SPRITE_CLASS) || m_en.t || (m_en.c && m_sel.iip);
many_regs = isYmm && !m_sel.notest && needs_shift;
@ -116,9 +112,6 @@ void GSSetupPrimCodeGenerator2::Generate()
}
#endif
if (!m_rip)
mov(_64_m_local, (size_t)&m_local);
if (needs_shift)
{

View File

@ -39,10 +39,6 @@ class GSSetupPrimCodeGenerator2 : public GSNewCodeGenerator
using Xmm = Xbyak::Xmm;
using Ymm = Xbyak::Ymm;
/// On x86-64 we reserve a bunch of GPRs for holding addresses of locals that would otherwise be hard to reach
/// On x86-32 the same values are just raw 32-bit addresses
using LocalAddr = Choose3264<size_t, AddressReg>::type;
constexpr static bool isXmm = std::is_same<XYm, Xbyak::Xmm>::value;
constexpr static bool isYmm = std::is_same<XYm, Xbyak::Ymm>::value;
constexpr static int vecsize = isXmm ? 16 : 32;
@ -51,19 +47,12 @@ class GSSetupPrimCodeGenerator2 : public GSNewCodeGenerator
GSScanlineSelector m_sel;
GSScanlineLocalData& m_local;
bool m_rip;
bool many_regs;
struct {u32 z:1, f:1, t:1, c:1;} m_en;
const XYm xym0{0}, xym1{1}, xym2{2}, xym3{3}, xym4{4}, xym5{5}, xym6{6}, xym7{7}, xym8{8}, xym9{9}, xym10{10}, xym11{11}, xym12{12}, xym13{13}, xym14{14}, xym15{15};
const AddressReg _64_vertex, _index, _dscan, _64_t0, t1;
const LocalAddr _m_local;
/// Returns the first arg on 32-bit, second on 64-bit
static LocalAddr chooseLocal(const void* addr32, AddressReg reg64)
{
return choose3264((size_t)addr32, reg64);
}
const AddressReg _64_vertex, _index, _dscan, _m_local, t1;
public:
GSSetupPrimCodeGenerator2(Xbyak::CodeGenerator* base, const ProcessorFeatures& cpu, void* param, u64 key);