void VertexShaderManager::Init()
{
// Initialize state tracking variables
nTransformMatricesChanged[0] = -1;
nTransformMatricesChanged[1] = -1;
nNormalMatricesChanged[0] = -1;
nNormalMatricesChanged[1] = -1;
nPostTransformMatricesChanged[0] = -1;
nPostTransformMatricesChanged[1] = -1;
nLightsChanged[0] = -1;
nLightsChanged[1] = -1;
nMaterialsChanged = BitSet32(0);
bTexMatricesChanged[0] = false;
bTexMatricesChanged[1] = false;
bPosNormalMatrixChanged = false;
bProjectionChanged = true;
bViewportChanged = false;
memset(&xfmem, 0, sizeof(xfmem));
memset(&constants, 0, sizeof(constants));
ResetView();
// TODO: should these go inside ResetView()?
Matrix44::LoadIdentity(s_viewportCorrection);
memset(g_fProjectionMatrix, 0, sizeof(g_fProjectionMatrix));
for (int i = 0; i < 4; ++i)
g_fProjectionMatrix[i*5] = 1.0f;
dirty = true;
}
TEST(BitSet, Count)
{
u32 random_numbers[] = {
0x2cb0b5f3, 0x81ab32a6, 0xd9030dc5, 0x325ffe26, 0xb2fcaee3,
0x4ccf188a, 0xf8be36dc, 0xb2fcecd5, 0xb750c2e5, 0x31d19074,
0xf267644a, 0xac00a719, 0x6d45f19b, 0xf7e91c5b, 0xf687e694,
0x9057c24e, 0x5eb65c39, 0x85d3038b, 0x101f4e66, 0xc202d136
};
u32 counts[] = {
17, 14, 14, 19, 20, 14, 20, 20, 16, 13, 16, 12, 18, 20, 18, 14, 18, 14, 14, 12
};
for (size_t i = 0; i < 20; i++)
{
EXPECT_EQ(counts[i], BitSet32(random_numbers[i]).Count());
}
u64 random_numbers_64[] = {
0xf86cd6f6ef09d7d4ULL, 0x6f2d8533255ead3cULL, 0x9da7941e0e52b345ULL,
0x06e4189be67d2b17ULL, 0x3eb0681f65cb6d25ULL, 0xccab8a7c74a51203ULL,
0x09d470516694c64bULL, 0x38cd077e075c778fULL, 0xd69ebfa6355ebfdeULL
};
u32 counts_64[] = {
39, 34, 31, 32, 33, 29, 27, 35, 43
};
for (size_t i = 0; i < 9; i++)
{
EXPECT_EQ(counts_64[i], BitSet64(random_numbers_64[i]).Count());
}
}
TEST(BitSet, Basics)
{
BitSet32 bs;
BitSet64 bs2(1);
BitSet64 bs3(2);
EXPECT_EQ(true, !!bs2);
EXPECT_EQ(false, !!bs);
EXPECT_EQ(bs2, bs2);
EXPECT_NE(bs2, bs3);
EXPECT_EQ(BitSet32(0xfff), BitSet32::AllTrue(12));
EXPECT_EQ(BitSet64(0xffffffffffffffff), BitSet64::AllTrue(64));
}
TEST(BitSet, BitOps)
{
BitSet32 a(3), b(5), c;
EXPECT_EQ(BitSet32(7), a | b);
EXPECT_EQ(BitSet32(6), a ^ b);
EXPECT_EQ(BitSet32(1), a & b);
EXPECT_EQ(BitSet32(0xfffffffc), ~a);
c = a; c |= b; EXPECT_EQ(BitSet32(7), c);
c = a; c ^= b; EXPECT_EQ(BitSet32(6), c);
c = a; c &= b; EXPECT_EQ(BitSet32(1), c);
}
__forceinline void WriteToHardware(u32 em_address, const T data)
{
int segment = em_address >> 28;
// Quick check for an address that can't meet any of the following conditions,
// to speed up the MMU path.
if (!BitSet32(0xCFC)[segment])
{
// First, let's check for FIFO writes, since they are probably the most common
// reason we end up in this function:
if ((em_address & 0xFFFFF000) == 0xCC008000)
{
switch (sizeof(T))
{
case 1: GPFifo::Write8((u8)data, em_address); return;
case 2: GPFifo::Write16((u16)data, em_address); return;
case 4: GPFifo::Write32((u32)data, em_address); return;
case 8: GPFifo::Write64((u64)data, em_address); return;
}
}
if ((em_address & 0xC8000000) == 0xC8000000)
{
if (em_address < 0xcc000000)
{
int x = (em_address & 0xfff) >> 2;
int y = (em_address >> 12) & 0x3ff;
// TODO figure out a way to send data without falling into the template trap
if (em_address & 0x00400000)
{
g_video_backend->Video_AccessEFB(POKE_Z, x, y, (u32)data);
DEBUG_LOG(MEMMAP, "EFB Z Write %08x @ %i, %i", (u32)data, x, y);
}
else
{
g_video_backend->Video_AccessEFB(POKE_COLOR, x, y, (u32)data);
DEBUG_LOG(MEMMAP, "EFB Color Write %08x @ %i, %i", (u32)data, x, y);
}
return;
}
else
{
mmio_mapping->Write(em_address, data);
return;
}
}
// Syncs the shader constant buffers with xfmem
// TODO: A cleaner way to control the matrices without making a mess in the parameters field
void VertexShaderManager::SetConstants()
{
if (nTransformMatricesChanged[0] >= 0)
{
int startn = nTransformMatricesChanged[0] / 4;
int endn = (nTransformMatricesChanged[1] + 3) / 4;
memcpy(constants.transformmatrices[startn], &xfmem.posMatrices[startn * 4], (endn - startn) * 16);
dirty = true;
nTransformMatricesChanged[0] = nTransformMatricesChanged[1] = -1;
}
if (nNormalMatricesChanged[0] >= 0)
{
int startn = nNormalMatricesChanged[0] / 3;
int endn = (nNormalMatricesChanged[1] + 2) / 3;
for (int i=startn; i<endn; i++)
{
memcpy(constants.normalmatrices[i], &xfmem.normalMatrices[3*i], 12);
}
dirty = true;
nNormalMatricesChanged[0] = nNormalMatricesChanged[1] = -1;
}
if (nPostTransformMatricesChanged[0] >= 0)
{
int startn = nPostTransformMatricesChanged[0] / 4;
int endn = (nPostTransformMatricesChanged[1] + 3 ) / 4;
memcpy(constants.posttransformmatrices[startn], &xfmem.postMatrices[startn * 4], (endn - startn) * 16);
dirty = true;
nPostTransformMatricesChanged[0] = nPostTransformMatricesChanged[1] = -1;
}
if (nLightsChanged[0] >= 0)
{
// TODO: Outdated comment
// lights don't have a 1 to 1 mapping, the color component needs to be converted to 4 floats
int istart = nLightsChanged[0] / 0x10;
int iend = (nLightsChanged[1] + 15) / 0x10;
for (int i = istart; i < iend; ++i)
{
const Light& light = xfmem.lights[i];
VertexShaderConstants::Light& dstlight = constants.lights[i];
// xfmem.light.color is packed as abgr in u8[4], so we have to swap the order
dstlight.color[0] = light.color[3];
dstlight.color[1] = light.color[2];
dstlight.color[2] = light.color[1];
dstlight.color[3] = light.color[0];
dstlight.cosatt[0] = light.cosatt[0];
dstlight.cosatt[1] = light.cosatt[1];
dstlight.cosatt[2] = light.cosatt[2];
if (fabs(light.distatt[0]) < 0.00001f &&
fabs(light.distatt[1]) < 0.00001f &&
fabs(light.distatt[2]) < 0.00001f)
{
// dist attenuation, make sure not equal to 0!!!
dstlight.distatt[0] = .00001f;
}
else
{
dstlight.distatt[0] = light.distatt[0];
}
dstlight.distatt[1] = light.distatt[1];
dstlight.distatt[2] = light.distatt[2];
dstlight.pos[0] = light.dpos[0];
dstlight.pos[1] = light.dpos[1];
dstlight.pos[2] = light.dpos[2];
double norm = double(light.ddir[0]) * double(light.ddir[0]) +
double(light.ddir[1]) * double(light.ddir[1]) +
double(light.ddir[2]) * double(light.ddir[2]);
norm = 1.0 / sqrt(norm);
float norm_float = static_cast<float>(norm);
dstlight.dir[0] = light.ddir[0] * norm_float;
dstlight.dir[1] = light.ddir[1] * norm_float;
dstlight.dir[2] = light.ddir[2] * norm_float;
}
dirty = true;
nLightsChanged[0] = nLightsChanged[1] = -1;
}
for (int i : nMaterialsChanged)
{
u32 data = i >= 2 ? xfmem.matColor[i - 2] : xfmem.ambColor[i];
constants.materials[i][0] = (data >> 24) & 0xFF;
constants.materials[i][1] = (data >> 16) & 0xFF;
constants.materials[i][2] = (data >> 8) & 0xFF;
constants.materials[i][3] = data & 0xFF;
dirty = true;
}
nMaterialsChanged = BitSet32(0);
if (bPosNormalMatrixChanged)
{
bPosNormalMatrixChanged = false;
const float *pos = (const float *)xfmem.posMatrices + g_main_cp_state.matrix_index_a.PosNormalMtxIdx * 4;
const float *norm = (const float *)xfmem.normalMatrices + 3 * (g_main_cp_state.matrix_index_a.PosNormalMtxIdx & 31);
memcpy(constants.posnormalmatrix, pos, 3*16);
memcpy(constants.posnormalmatrix[3], norm, 12);
memcpy(constants.posnormalmatrix[4], norm+3, 12);
memcpy(constants.posnormalmatrix[5], norm+6, 12);
dirty = true;
}
if (bTexMatricesChanged[0])
{
bTexMatricesChanged[0] = false;
const float *fptrs[] =
{
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex0MtxIdx * 4],
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex1MtxIdx * 4],
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex2MtxIdx * 4],
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex3MtxIdx * 4]
};
for (int i = 0; i < 4; ++i)
{
memcpy(constants.texmatrices[3*i], fptrs[i], 3*16);
}
dirty = true;
}
if (bTexMatricesChanged[1])
{
bTexMatricesChanged[1] = false;
const float *fptrs[] = {
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex4MtxIdx * 4],
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex5MtxIdx * 4],
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex6MtxIdx * 4],
(const float *)&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex7MtxIdx * 4]
};
for (int i = 0; i < 4; ++i)
{
memcpy(constants.texmatrices[3*i+12], fptrs[i], 3*16);
}
dirty = true;
}
if (bViewportChanged)
{
bViewportChanged = false;
// The console GPU places the pixel center at 7/12 unless antialiasing
// is enabled, while D3D and OpenGL place it at 0.5. See the comment
// in VertexShaderGen.cpp for details.
// NOTE: If we ever emulate antialiasing, the sample locations set by
// BP registers 0x01-0x04 need to be considered here.
const float pixel_center_correction = 7.0f / 12.0f - 0.5f;
const float pixel_size_x = 2.f / Renderer::EFBToScaledXf(2.f * xfmem.viewport.wd);
const float pixel_size_y = 2.f / Renderer::EFBToScaledXf(2.f * xfmem.viewport.ht);
constants.pixelcentercorrection[0] = pixel_center_correction * pixel_size_x;
constants.pixelcentercorrection[1] = pixel_center_correction * pixel_size_y;
dirty = true;
// This is so implementation-dependent that we can't have it here.
g_renderer->SetViewport();
// Update projection if the viewport isn't 1:1 useable
if (!g_ActiveConfig.backend_info.bSupportsOversizedViewports)
{
ViewportCorrectionMatrix(s_viewportCorrection);
bProjectionChanged = true;
}
}
if (bProjectionChanged)
{
bProjectionChanged = false;
float *rawProjection = xfmem.projection.rawProjection;
switch (xfmem.projection.type)
{
case GX_PERSPECTIVE:
g_fProjectionMatrix[0] = rawProjection[0] * g_ActiveConfig.fAspectRatioHackW;
g_fProjectionMatrix[1] = 0.0f;
g_fProjectionMatrix[2] = rawProjection[1];
g_fProjectionMatrix[3] = 0.0f;
g_fProjectionMatrix[4] = 0.0f;
g_fProjectionMatrix[5] = rawProjection[2] * g_ActiveConfig.fAspectRatioHackH;
g_fProjectionMatrix[6] = rawProjection[3];
g_fProjectionMatrix[7] = 0.0f;
g_fProjectionMatrix[8] = 0.0f;
g_fProjectionMatrix[9] = 0.0f;
g_fProjectionMatrix[10] = rawProjection[4];
g_fProjectionMatrix[11] = rawProjection[5];
g_fProjectionMatrix[12] = 0.0f;
g_fProjectionMatrix[13] = 0.0f;
// donkopunchstania suggested the GC GPU might round differently
// He had thus changed this to -(1 + epsilon) to fix clipping issues.
// I (neobrain) don't think his conjecture is true and thus reverted his change.
g_fProjectionMatrix[14] = -1.0f;
g_fProjectionMatrix[15] = 0.0f;
SETSTAT_FT(stats.gproj_0, g_fProjectionMatrix[0]);
SETSTAT_FT(stats.gproj_1, g_fProjectionMatrix[1]);
SETSTAT_FT(stats.gproj_2, g_fProjectionMatrix[2]);
SETSTAT_FT(stats.gproj_3, g_fProjectionMatrix[3]);
SETSTAT_FT(stats.gproj_4, g_fProjectionMatrix[4]);
SETSTAT_FT(stats.gproj_5, g_fProjectionMatrix[5]);
SETSTAT_FT(stats.gproj_6, g_fProjectionMatrix[6]);
SETSTAT_FT(stats.gproj_7, g_fProjectionMatrix[7]);
SETSTAT_FT(stats.gproj_8, g_fProjectionMatrix[8]);
SETSTAT_FT(stats.gproj_9, g_fProjectionMatrix[9]);
SETSTAT_FT(stats.gproj_10, g_fProjectionMatrix[10]);
SETSTAT_FT(stats.gproj_11, g_fProjectionMatrix[11]);
SETSTAT_FT(stats.gproj_12, g_fProjectionMatrix[12]);
SETSTAT_FT(stats.gproj_13, g_fProjectionMatrix[13]);
SETSTAT_FT(stats.gproj_14, g_fProjectionMatrix[14]);
SETSTAT_FT(stats.gproj_15, g_fProjectionMatrix[15]);
break;
case GX_ORTHOGRAPHIC:
g_fProjectionMatrix[0] = rawProjection[0];
g_fProjectionMatrix[1] = 0.0f;
g_fProjectionMatrix[2] = 0.0f;
g_fProjectionMatrix[3] = rawProjection[1];
g_fProjectionMatrix[4] = 0.0f;
g_fProjectionMatrix[5] = rawProjection[2];
g_fProjectionMatrix[6] = 0.0f;
g_fProjectionMatrix[7] = rawProjection[3];
g_fProjectionMatrix[8] = 0.0f;
g_fProjectionMatrix[9] = 0.0f;
g_fProjectionMatrix[10] = (g_ProjHack1.value + rawProjection[4]) * ((g_ProjHack1.sign == 0) ? 1.0f : g_ProjHack1.sign);
g_fProjectionMatrix[11] = (g_ProjHack2.value + rawProjection[5]) * ((g_ProjHack2.sign == 0) ? 1.0f : g_ProjHack2.sign);
g_fProjectionMatrix[12] = 0.0f;
g_fProjectionMatrix[13] = 0.0f;
g_fProjectionMatrix[14] = 0.0f;
g_fProjectionMatrix[15] = 1.0f + FLT_EPSILON; // hack to fix depth clipping precision issues (such as Sonic Unleashed UI)
SETSTAT_FT(stats.g2proj_0, g_fProjectionMatrix[0]);
SETSTAT_FT(stats.g2proj_1, g_fProjectionMatrix[1]);
SETSTAT_FT(stats.g2proj_2, g_fProjectionMatrix[2]);
SETSTAT_FT(stats.g2proj_3, g_fProjectionMatrix[3]);
SETSTAT_FT(stats.g2proj_4, g_fProjectionMatrix[4]);
SETSTAT_FT(stats.g2proj_5, g_fProjectionMatrix[5]);
SETSTAT_FT(stats.g2proj_6, g_fProjectionMatrix[6]);
SETSTAT_FT(stats.g2proj_7, g_fProjectionMatrix[7]);
SETSTAT_FT(stats.g2proj_8, g_fProjectionMatrix[8]);
SETSTAT_FT(stats.g2proj_9, g_fProjectionMatrix[9]);
SETSTAT_FT(stats.g2proj_10, g_fProjectionMatrix[10]);
SETSTAT_FT(stats.g2proj_11, g_fProjectionMatrix[11]);
SETSTAT_FT(stats.g2proj_12, g_fProjectionMatrix[12]);
SETSTAT_FT(stats.g2proj_13, g_fProjectionMatrix[13]);
SETSTAT_FT(stats.g2proj_14, g_fProjectionMatrix[14]);
SETSTAT_FT(stats.g2proj_15, g_fProjectionMatrix[15]);
SETSTAT_FT(stats.proj_0, rawProjection[0]);
SETSTAT_FT(stats.proj_1, rawProjection[1]);
SETSTAT_FT(stats.proj_2, rawProjection[2]);
SETSTAT_FT(stats.proj_3, rawProjection[3]);
SETSTAT_FT(stats.proj_4, rawProjection[4]);
SETSTAT_FT(stats.proj_5, rawProjection[5]);
break;
default:
ERROR_LOG(VIDEO, "Unknown projection type: %d", xfmem.projection.type);
}
PRIM_LOG("Projection: %f %f %f %f %f %f\n", rawProjection[0], rawProjection[1], rawProjection[2], rawProjection[3], rawProjection[4], rawProjection[5]);
if (g_ActiveConfig.bFreeLook && xfmem.projection.type == GX_PERSPECTIVE)
{
Matrix44 mtxA;
Matrix44 mtxB;
Matrix44 viewMtx;
Matrix44::Translate(mtxA, s_fViewTranslationVector);
Matrix44::LoadMatrix33(mtxB, s_viewRotationMatrix);
Matrix44::Multiply(mtxB, mtxA, viewMtx); // view = rotation x translation
Matrix44::Set(mtxB, g_fProjectionMatrix);
Matrix44::Multiply(mtxB, viewMtx, mtxA); // mtxA = projection x view
Matrix44::Multiply(s_viewportCorrection, mtxA, mtxB); // mtxB = viewportCorrection x mtxA
memcpy(constants.projection, mtxB.data, 4*16);
}
else
{
Matrix44 projMtx;
Matrix44::Set(projMtx, g_fProjectionMatrix);
Matrix44 correctedMtx;
Matrix44::Multiply(s_viewportCorrection, projMtx, correctedMtx);
memcpy(constants.projection, correctedMtx.data, 4*16);
}
dirty = true;
}
}
__forceinline T ReadFromHardware(const u32 em_address)
{
int segment = em_address >> 28;
// Quick check for an address that can't meet any of the following conditions,
// to speed up the MMU path.
if (!BitSet32(0xCFC)[segment])
{
// TODO: Figure out the fastest order of tests for both read and write (they are probably different).
if ((em_address & 0xC8000000) == 0xC8000000)
{
if (em_address < 0xcc000000)
return EFB_Read(em_address);
else
return (T)mmio_mapping->Read<typename std::make_unsigned<T>::type>(em_address);
}
else if (segment == 0x8 || segment == 0xC || segment == 0x0)
{
return bswap((*(const T*)&m_pRAM[em_address & RAM_MASK]));
}
else if (m_pEXRAM && (segment == 0x9 || segment == 0xD || segment == 0x1))
{
return bswap((*(const T*)&m_pEXRAM[em_address & EXRAM_MASK]));
}
else if (segment == 0xE && (em_address < (0xE0000000 + L1_CACHE_SIZE)))
{
return bswap((*(const T*)&m_pL1Cache[em_address & L1_CACHE_MASK]));
}
}
if (bFakeVMEM && (segment == 0x7 || segment == 0x4))
{
// fake VMEM
return bswap((*(const T*)&m_pFakeVMEM[em_address & FAKEVMEM_MASK]));
}
// MMU: Do page table translation
u32 tlb_addr = TranslateAddress<flag>(em_address);
if (tlb_addr == 0)
{
if (flag == FLAG_READ)
GenerateDSIException(em_address, false);
return 0;
}
// Handle loads that cross page boundaries (ewwww)
// The alignment check isn't strictly necessary, but since this is a rare slow path, it provides a faster
// (1 instruction on x86) bailout.
if (sizeof(T) > 1 && (em_address & (sizeof(T) - 1)) && (em_address & (HW_PAGE_SIZE - 1)) > HW_PAGE_SIZE - sizeof(T))
{
// This could be unaligned down to the byte level... hopefully this is rare, so doing it this
// way isn't too terrible.
// TODO: floats on non-word-aligned boundaries should technically cause alignment exceptions.
// Note that "word" means 32-bit, so paired singles or doubles might still be 32-bit aligned!
u32 em_address_next_page = (em_address + sizeof(T) - 1) & ~(HW_PAGE_SIZE - 1);
u32 tlb_addr_next_page = TranslateAddress<flag>(em_address_next_page);
if (tlb_addr == 0 || tlb_addr_next_page == 0)
{
if (flag == FLAG_READ)
GenerateDSIException(em_address_next_page, false);
return 0;
}
T var = 0;
for (u32 addr = em_address; addr < em_address + sizeof(T); addr++, tlb_addr++)
{
if (addr == em_address_next_page)
tlb_addr = tlb_addr_next_page;
var = (var << 8) | Memory::base[tlb_addr];
}
return var;
}
// The easy case!
return bswap(*(const T*)&Memory::base[tlb_addr]);
}
// Syncs the shader constant buffers with xfmem
// TODO: A cleaner way to control the matrices without making a mess in the parameters field
void VertexShaderManager::SetConstants()
{
if (nTransformMatricesChanged[0] >= 0)
{
int startn = nTransformMatricesChanged[0] / 4;
int endn = (nTransformMatricesChanged[1] + 3) / 4;
memcpy(constants.transformmatrices[startn].data(), &xfmem.posMatrices[startn * 4],
(endn - startn) * sizeof(float4));
dirty = true;
nTransformMatricesChanged[0] = nTransformMatricesChanged[1] = -1;
}
if (nNormalMatricesChanged[0] >= 0)
{
int startn = nNormalMatricesChanged[0] / 3;
int endn = (nNormalMatricesChanged[1] + 2) / 3;
for (int i = startn; i < endn; i++)
{
memcpy(constants.normalmatrices[i].data(), &xfmem.normalMatrices[3 * i], 12);
}
dirty = true;
nNormalMatricesChanged[0] = nNormalMatricesChanged[1] = -1;
}
if (nPostTransformMatricesChanged[0] >= 0)
{
int startn = nPostTransformMatricesChanged[0] / 4;
int endn = (nPostTransformMatricesChanged[1] + 3) / 4;
memcpy(constants.posttransformmatrices[startn].data(), &xfmem.postMatrices[startn * 4],
(endn - startn) * sizeof(float4));
dirty = true;
nPostTransformMatricesChanged[0] = nPostTransformMatricesChanged[1] = -1;
}
if (nLightsChanged[0] >= 0)
{
// TODO: Outdated comment
// lights don't have a 1 to 1 mapping, the color component needs to be converted to 4 floats
int istart = nLightsChanged[0] / 0x10;
int iend = (nLightsChanged[1] + 15) / 0x10;
for (int i = istart; i < iend; ++i)
{
const Light& light = xfmem.lights[i];
VertexShaderConstants::Light& dstlight = constants.lights[i];
// xfmem.light.color is packed as abgr in u8[4], so we have to swap the order
dstlight.color[0] = light.color[3];
dstlight.color[1] = light.color[2];
dstlight.color[2] = light.color[1];
dstlight.color[3] = light.color[0];
dstlight.cosatt[0] = light.cosatt[0];
dstlight.cosatt[1] = light.cosatt[1];
dstlight.cosatt[2] = light.cosatt[2];
if (fabs(light.distatt[0]) < 0.00001f && fabs(light.distatt[1]) < 0.00001f &&
fabs(light.distatt[2]) < 0.00001f)
{
// dist attenuation, make sure not equal to 0!!!
dstlight.distatt[0] = .00001f;
}
else
{
dstlight.distatt[0] = light.distatt[0];
}
dstlight.distatt[1] = light.distatt[1];
dstlight.distatt[2] = light.distatt[2];
dstlight.pos[0] = light.dpos[0];
dstlight.pos[1] = light.dpos[1];
dstlight.pos[2] = light.dpos[2];
double norm = double(light.ddir[0]) * double(light.ddir[0]) +
double(light.ddir[1]) * double(light.ddir[1]) +
double(light.ddir[2]) * double(light.ddir[2]);
norm = 1.0 / sqrt(norm);
float norm_float = static_cast<float>(norm);
dstlight.dir[0] = light.ddir[0] * norm_float;
dstlight.dir[1] = light.ddir[1] * norm_float;
dstlight.dir[2] = light.ddir[2] * norm_float;
}
dirty = true;
nLightsChanged[0] = nLightsChanged[1] = -1;
}
for (int i : nMaterialsChanged)
{
u32 data = i >= 2 ? xfmem.matColor[i - 2] : xfmem.ambColor[i];
constants.materials[i][0] = (data >> 24) & 0xFF;
constants.materials[i][1] = (data >> 16) & 0xFF;
constants.materials[i][2] = (data >> 8) & 0xFF;
constants.materials[i][3] = data & 0xFF;
dirty = true;
}
nMaterialsChanged = BitSet32(0);
if (bPosNormalMatrixChanged)
{
bPosNormalMatrixChanged = false;
const float* pos = &xfmem.posMatrices[g_main_cp_state.matrix_index_a.PosNormalMtxIdx * 4];
const float* norm =
&xfmem.normalMatrices[3 * (g_main_cp_state.matrix_index_a.PosNormalMtxIdx & 31)];
memcpy(constants.posnormalmatrix.data(), pos, 3 * sizeof(float4));
memcpy(constants.posnormalmatrix[3].data(), norm, 3 * sizeof(float));
memcpy(constants.posnormalmatrix[4].data(), norm + 3, 3 * sizeof(float));
memcpy(constants.posnormalmatrix[5].data(), norm + 6, 3 * sizeof(float));
dirty = true;
}
if (bTexMatricesChanged[0])
{
bTexMatricesChanged[0] = false;
const float* pos_matrix_ptrs[] = {
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex0MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex1MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex2MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex3MtxIdx * 4]};
for (size_t i = 0; i < ArraySize(pos_matrix_ptrs); ++i)
{
memcpy(constants.texmatrices[3 * i].data(), pos_matrix_ptrs[i], 3 * sizeof(float4));
}
dirty = true;
}
if (bTexMatricesChanged[1])
{
bTexMatricesChanged[1] = false;
const float* pos_matrix_ptrs[] = {
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex4MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex5MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex6MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex7MtxIdx * 4]};
for (size_t i = 0; i < ArraySize(pos_matrix_ptrs); ++i)
{
memcpy(constants.texmatrices[3 * i + 12].data(), pos_matrix_ptrs[i], 3 * sizeof(float4));
}
dirty = true;
}
if (bViewportChanged)
{
bViewportChanged = false;
// The console GPU places the pixel center at 7/12 unless antialiasing
// is enabled, while D3D and OpenGL place it at 0.5. See the comment
// in VertexShaderGen.cpp for details.
// NOTE: If we ever emulate antialiasing, the sample locations set by
// BP registers 0x01-0x04 need to be considered here.
const float pixel_center_correction = 7.0f / 12.0f - 0.5f;
const bool bUseVertexRounding = g_ActiveConfig.bVertexRounding && g_ActiveConfig.iEFBScale != 1;
const float viewport_width = bUseVertexRounding ?
(2.f * xfmem.viewport.wd) :
g_renderer->EFBToScaledXf(2.f * xfmem.viewport.wd);
const float viewport_height = bUseVertexRounding ?
(2.f * xfmem.viewport.ht) :
g_renderer->EFBToScaledXf(2.f * xfmem.viewport.ht);
const float pixel_size_x = 2.f / viewport_width;
const float pixel_size_y = 2.f / viewport_height;
constants.pixelcentercorrection[0] = pixel_center_correction * pixel_size_x;
constants.pixelcentercorrection[1] = pixel_center_correction * pixel_size_y;
// By default we don't change the depth value at all in the vertex shader.
constants.pixelcentercorrection[2] = 1.0f;
constants.pixelcentercorrection[3] = 0.0f;
constants.viewport[0] = (2.f * xfmem.viewport.wd);
constants.viewport[1] = (2.f * xfmem.viewport.ht);
if (g_renderer->UseVertexDepthRange())
{
// Oversized depth ranges are handled in the vertex shader. We need to reverse
// the far value to use the reversed-Z trick.
if (g_ActiveConfig.backend_info.bSupportsReversedDepthRange)
{
// Sometimes the console also tries to use the reversed-Z trick. We can only do
// that with the expected accuracy if the backend can reverse the depth range.
constants.pixelcentercorrection[2] = fabs(xfmem.viewport.zRange) / 16777215.0f;
if (xfmem.viewport.zRange < 0.0f)
constants.pixelcentercorrection[3] = xfmem.viewport.farZ / 16777215.0f;
else
constants.pixelcentercorrection[3] = 1.0f - xfmem.viewport.farZ / 16777215.0f;
}
else
{
// For backends that don't support reversing the depth range we can still render
// cases where the console uses the reversed-Z trick. But we simply can't provide
// the expected accuracy, which might result in z-fighting.
constants.pixelcentercorrection[2] = xfmem.viewport.zRange / 16777215.0f;
constants.pixelcentercorrection[3] = 1.0f - xfmem.viewport.farZ / 16777215.0f;
}
}
dirty = true;
BPFunctions::SetViewport();
// Update projection if the viewport isn't 1:1 useable
if (!g_ActiveConfig.backend_info.bSupportsOversizedViewports)
{
ViewportCorrectionMatrix(s_viewportCorrection);
bProjectionChanged = true;
}
}
if (bProjectionChanged)
{
bProjectionChanged = false;
float* rawProjection = xfmem.projection.rawProjection;
switch (xfmem.projection.type)
{
case GX_PERSPECTIVE:
g_fProjectionMatrix[0] = rawProjection[0] * g_ActiveConfig.fAspectRatioHackW;
g_fProjectionMatrix[1] = 0.0f;
g_fProjectionMatrix[2] = rawProjection[1] * g_ActiveConfig.fAspectRatioHackW;
g_fProjectionMatrix[3] = 0.0f;
g_fProjectionMatrix[4] = 0.0f;
g_fProjectionMatrix[5] = rawProjection[2] * g_ActiveConfig.fAspectRatioHackH;
g_fProjectionMatrix[6] = rawProjection[3] * g_ActiveConfig.fAspectRatioHackH;
g_fProjectionMatrix[7] = 0.0f;
g_fProjectionMatrix[8] = 0.0f;
g_fProjectionMatrix[9] = 0.0f;
g_fProjectionMatrix[10] = rawProjection[4];
g_fProjectionMatrix[11] = rawProjection[5];
g_fProjectionMatrix[12] = 0.0f;
g_fProjectionMatrix[13] = 0.0f;
g_fProjectionMatrix[14] = -1.0f;
g_fProjectionMatrix[15] = 0.0f;
SETSTAT_FT(stats.gproj_0, g_fProjectionMatrix[0]);
SETSTAT_FT(stats.gproj_1, g_fProjectionMatrix[1]);
SETSTAT_FT(stats.gproj_2, g_fProjectionMatrix[2]);
SETSTAT_FT(stats.gproj_3, g_fProjectionMatrix[3]);
SETSTAT_FT(stats.gproj_4, g_fProjectionMatrix[4]);
SETSTAT_FT(stats.gproj_5, g_fProjectionMatrix[5]);
SETSTAT_FT(stats.gproj_6, g_fProjectionMatrix[6]);
SETSTAT_FT(stats.gproj_7, g_fProjectionMatrix[7]);
SETSTAT_FT(stats.gproj_8, g_fProjectionMatrix[8]);
SETSTAT_FT(stats.gproj_9, g_fProjectionMatrix[9]);
SETSTAT_FT(stats.gproj_10, g_fProjectionMatrix[10]);
SETSTAT_FT(stats.gproj_11, g_fProjectionMatrix[11]);
SETSTAT_FT(stats.gproj_12, g_fProjectionMatrix[12]);
SETSTAT_FT(stats.gproj_13, g_fProjectionMatrix[13]);
SETSTAT_FT(stats.gproj_14, g_fProjectionMatrix[14]);
SETSTAT_FT(stats.gproj_15, g_fProjectionMatrix[15]);
break;
case GX_ORTHOGRAPHIC:
g_fProjectionMatrix[0] = rawProjection[0];
g_fProjectionMatrix[1] = 0.0f;
g_fProjectionMatrix[2] = 0.0f;
g_fProjectionMatrix[3] = rawProjection[1];
g_fProjectionMatrix[4] = 0.0f;
g_fProjectionMatrix[5] = rawProjection[2];
g_fProjectionMatrix[6] = 0.0f;
g_fProjectionMatrix[7] = rawProjection[3];
g_fProjectionMatrix[8] = 0.0f;
g_fProjectionMatrix[9] = 0.0f;
g_fProjectionMatrix[10] = rawProjection[4];
g_fProjectionMatrix[11] = rawProjection[5];
g_fProjectionMatrix[12] = 0.0f;
g_fProjectionMatrix[13] = 0.0f;
g_fProjectionMatrix[14] = 0.0f;
g_fProjectionMatrix[15] = 1.0f;
SETSTAT_FT(stats.g2proj_0, g_fProjectionMatrix[0]);
SETSTAT_FT(stats.g2proj_1, g_fProjectionMatrix[1]);
SETSTAT_FT(stats.g2proj_2, g_fProjectionMatrix[2]);
SETSTAT_FT(stats.g2proj_3, g_fProjectionMatrix[3]);
SETSTAT_FT(stats.g2proj_4, g_fProjectionMatrix[4]);
SETSTAT_FT(stats.g2proj_5, g_fProjectionMatrix[5]);
SETSTAT_FT(stats.g2proj_6, g_fProjectionMatrix[6]);
SETSTAT_FT(stats.g2proj_7, g_fProjectionMatrix[7]);
SETSTAT_FT(stats.g2proj_8, g_fProjectionMatrix[8]);
SETSTAT_FT(stats.g2proj_9, g_fProjectionMatrix[9]);
SETSTAT_FT(stats.g2proj_10, g_fProjectionMatrix[10]);
SETSTAT_FT(stats.g2proj_11, g_fProjectionMatrix[11]);
SETSTAT_FT(stats.g2proj_12, g_fProjectionMatrix[12]);
SETSTAT_FT(stats.g2proj_13, g_fProjectionMatrix[13]);
SETSTAT_FT(stats.g2proj_14, g_fProjectionMatrix[14]);
SETSTAT_FT(stats.g2proj_15, g_fProjectionMatrix[15]);
SETSTAT_FT(stats.proj_0, rawProjection[0]);
SETSTAT_FT(stats.proj_1, rawProjection[1]);
SETSTAT_FT(stats.proj_2, rawProjection[2]);
SETSTAT_FT(stats.proj_3, rawProjection[3]);
SETSTAT_FT(stats.proj_4, rawProjection[4]);
SETSTAT_FT(stats.proj_5, rawProjection[5]);
break;
default:
ERROR_LOG(VIDEO, "Unknown projection type: %d", xfmem.projection.type);
}
PRIM_LOG("Projection: %f %f %f %f %f %f", rawProjection[0], rawProjection[1], rawProjection[2],
rawProjection[3], rawProjection[4], rawProjection[5]);
if (g_ActiveConfig.bFreeLook && xfmem.projection.type == GX_PERSPECTIVE)
{
Matrix44 mtxA;
Matrix44 mtxB;
Matrix44 viewMtx;
Matrix44::Translate(mtxA, s_fViewTranslationVector);
Matrix44::LoadMatrix33(mtxB, s_viewRotationMatrix);
Matrix44::Multiply(mtxB, mtxA, viewMtx); // view = rotation x translation
Matrix44::Set(mtxB, g_fProjectionMatrix);
Matrix44::Multiply(mtxB, viewMtx, mtxA); // mtxA = projection x view
Matrix44::Multiply(s_viewportCorrection, mtxA, mtxB); // mtxB = viewportCorrection x mtxA
memcpy(constants.projection.data(), mtxB.data, 4 * sizeof(float4));
}
else
{
Matrix44 projMtx;
Matrix44::Set(projMtx, g_fProjectionMatrix);
Matrix44 correctedMtx;
Matrix44::Multiply(s_viewportCorrection, projMtx, correctedMtx);
memcpy(constants.projection.data(), correctedMtx.data, 4 * sizeof(float4));
}
dirty = true;
}
if (bTexMtxInfoChanged)
{
bTexMtxInfoChanged = false;
constants.xfmem_dualTexInfo = xfmem.dualTexTrans.enabled;
for (size_t i = 0; i < ArraySize(xfmem.texMtxInfo); i++)
constants.xfmem_pack1[i][0] = xfmem.texMtxInfo[i].hex;
for (size_t i = 0; i < ArraySize(xfmem.postMtxInfo); i++)
constants.xfmem_pack1[i][1] = xfmem.postMtxInfo[i].hex;
dirty = true;
}
if (bLightingConfigChanged)
{
bLightingConfigChanged = false;
for (size_t i = 0; i < 2; i++)
{
constants.xfmem_pack1[i][2] = xfmem.color[i].hex;
constants.xfmem_pack1[i][3] = xfmem.alpha[i].hex;
}
constants.xfmem_numColorChans = xfmem.numChan.numColorChans;
dirty = true;
}
}
__forceinline static void WriteToHardware(u32 em_address, const T data)
{
int segment = em_address >> 28;
// Quick check for an address that can't meet any of the following conditions,
// to speed up the MMU path.
bool performTranslation = UReg_MSR(MSR).DR;
if (!BitSet32(0xCFC)[segment] && performTranslation)
{
// First, let's check for FIFO writes, since they are probably the most common
// reason we end up in this function.
// Note that we must mask the address to correctly emulate certain games;
// Pac-Man World 3 in particular is affected by this.
if (flag == FLAG_WRITE && (em_address & 0xFFFFF000) == 0xCC008000)
{
switch (sizeof(T))
{
case 1: GPFifo::Write8((u8)data); return;
case 2: GPFifo::Write16((u16)data); return;
case 4: GPFifo::Write32((u32)data); return;
case 8: GPFifo::Write64((u64)data); return;
}
}
if (flag == FLAG_WRITE && (em_address & 0xF8000000) == 0xC8000000)
{
if (em_address < 0xcc000000)
{
// TODO: This only works correctly for 32-bit writes.
EFB_Write((u32)data, em_address);
return;
}
else
{
Memory::mmio_mapping->Write(em_address & 0x0FFFFFFF, data);
return;
}
}
if (segment == 0x0 || segment == 0x8 || segment == 0xC)
{
// Handle RAM; the masking intentionally discards bits (essentially creating
// mirrors of memory).
// TODO: Only the first REALRAM_SIZE is supposed to be backed by actual memory.
*(T*)&Memory::m_pRAM[em_address & Memory::RAM_MASK] = bswap(data);
return;
}
if (Memory::m_pEXRAM && (segment == 0x9 || segment == 0xD) && (em_address & 0x0FFFFFFF) < Memory::EXRAM_SIZE)
{
// Handle EXRAM.
// TODO: Is this supposed to be mirrored like main RAM?
*(T*)&Memory::m_pEXRAM[em_address & 0x0FFFFFFF] = bswap(data);
return;
}
if (segment == 0xE && (em_address < (0xE0000000 + Memory::L1_CACHE_SIZE)))
{
*(T*)&Memory::m_pL1Cache[em_address & 0x0FFFFFFF] = bswap(data);
return;
}
}
if (Memory::bFakeVMEM && performTranslation && (segment == 0x7 || segment == 0x4))
{
// fake VMEM
*(T*)&Memory::m_pFakeVMEM[em_address & Memory::FAKEVMEM_MASK] = bswap(data);
return;
}
if (!performTranslation)
{
if (flag == FLAG_WRITE && (em_address & 0xFFFFF000) == 0x0C008000)
{
switch (sizeof(T))
{
case 1: GPFifo::Write8((u8)data); return;
case 2: GPFifo::Write16((u16)data); return;
case 4: GPFifo::Write32((u32)data); return;
case 8: GPFifo::Write64((u64)data); return;
}
}
if (flag == FLAG_WRITE && (em_address & 0xF8000000) == 0x08000000)
{
if (em_address < 0x0c000000)
{
// TODO: This only works correctly for 32-bit writes.
EFB_Write((u32)data, em_address);
return;
}
else
{
Memory::mmio_mapping->Write(em_address, data);
return;
}
}
if (segment == 0x0)
{
// Handle RAM; the masking intentionally discards bits (essentially creating
// mirrors of memory).
// TODO: Only the first REALRAM_SIZE is supposed to be backed by actual memory.
*(T*)&Memory::m_pRAM[em_address & Memory::RAM_MASK] = bswap(data);
return;
}
if (Memory::m_pEXRAM && segment == 0x1 && (em_address & 0x0FFFFFFF) < Memory::EXRAM_SIZE)
{
*(T*)&Memory::m_pEXRAM[em_address & 0x0FFFFFFF] = bswap(data);
return;
}
PanicAlert("Unable to resolve write address %x PC %x", em_address, PC);
return;
}
// MMU: Do page table translation
u32 tlb_addr = TranslateAddress<flag>(em_address);
if (tlb_addr == 0)
{
if (flag == FLAG_WRITE)
GenerateDSIException(em_address, true);
return;
}
// Handle stores that cross page boundaries (ewwww)
if (sizeof(T) > 1 && (em_address & (sizeof(T) - 1)) && (em_address & (HW_PAGE_SIZE - 1)) > HW_PAGE_SIZE - sizeof(T))
{
T val = bswap(data);
// We need to check both addresses before writing in case there's a DSI.
u32 em_address_next_page = (em_address + sizeof(T) - 1) & ~(HW_PAGE_SIZE - 1);
u32 tlb_addr_next_page = TranslateAddress<flag>(em_address_next_page);
if (tlb_addr_next_page == 0)
{
if (flag == FLAG_WRITE)
GenerateDSIException(em_address_next_page, true);
return;
}
for (u32 addr = em_address; addr < em_address + sizeof(T); addr++, tlb_addr++, val >>= 8)
{
if (addr == em_address_next_page)
tlb_addr = tlb_addr_next_page;
Memory::physical_base[tlb_addr] = (u8)val;
}
return;
}
__forceinline static T ReadFromHardware(const u32 em_address)
{
int segment = em_address >> 28;
bool performTranslation = UReg_MSR(MSR).DR;
// Quick check for an address that can't meet any of the following conditions,
// to speed up the MMU path.
if (!BitSet32(0xCFC)[segment] && performTranslation)
{
// TODO: Figure out the fastest order of tests for both read and write (they are probably different).
if (flag == FLAG_READ && (em_address & 0xF8000000) == 0xC8000000)
{
if (em_address < 0xcc000000)
return EFB_Read(em_address);
else
return (T)Memory::mmio_mapping->Read<typename std::make_unsigned<T>::type>(em_address & 0x0FFFFFFF);
}
if (segment == 0x0 || segment == 0x8 || segment == 0xC)
{
// Handle RAM; the masking intentionally discards bits (essentially creating
// mirrors of memory).
// TODO: Only the first REALRAM_SIZE is supposed to be backed by actual memory.
return bswap((*(const T*)&Memory::m_pRAM[em_address & Memory::RAM_MASK]));
}
if (Memory::m_pEXRAM && (segment == 0x9 || segment == 0xD) && (em_address & 0x0FFFFFFF) < Memory::EXRAM_SIZE)
{
// Handle EXRAM.
// TODO: Is this supposed to be mirrored like main RAM?
return bswap((*(const T*)&Memory::m_pEXRAM[em_address & 0x0FFFFFFF]));
}
if (segment == 0xE && (em_address < (0xE0000000 + Memory::L1_CACHE_SIZE)))
{
return bswap((*(const T*)&Memory::m_pL1Cache[em_address & 0x0FFFFFFF]));
}
}
if (Memory::bFakeVMEM && performTranslation && (segment == 0x7 || segment == 0x4))
{
// fake VMEM
return bswap((*(const T*)&Memory::m_pFakeVMEM[em_address & Memory::FAKEVMEM_MASK]));
}
if (!performTranslation)
{
if (flag == FLAG_READ && (em_address & 0xF8000000) == 0x08000000)
{
if (em_address < 0x0c000000)
return EFB_Read(em_address);
else
return (T)Memory::mmio_mapping->Read<typename std::make_unsigned<T>::type>(em_address);
}
if (segment == 0x0)
{
// Handle RAM; the masking intentionally discards bits (essentially creating
// mirrors of memory).
// TODO: Only the first REALRAM_SIZE is supposed to be backed by actual memory.
return bswap((*(const T*)&Memory::m_pRAM[em_address & Memory::RAM_MASK]));
}
if (Memory::m_pEXRAM && segment == 0x1 && (em_address & 0x0FFFFFFF) < Memory::EXRAM_SIZE)
{
return bswap((*(const T*)&Memory::m_pEXRAM[em_address & 0x0FFFFFFF]));
}
PanicAlert("Unable to resolve read address %x PC %x", em_address, PC);
return 0;
}
// MMU: Do page table translation
u32 tlb_addr = TranslateAddress<flag>(em_address);
if (tlb_addr == 0)
{
if (flag == FLAG_READ)
GenerateDSIException(em_address, false);
return 0;
}
// Handle loads that cross page boundaries (ewwww)
// The alignment check isn't strictly necessary, but since this is a rare slow path, it provides a faster
// (1 instruction on x86) bailout.
if (sizeof(T) > 1 && (em_address & (sizeof(T) - 1)) && (em_address & (HW_PAGE_SIZE - 1)) > HW_PAGE_SIZE - sizeof(T))
{
// This could be unaligned down to the byte level... hopefully this is rare, so doing it this
// way isn't too terrible.
// TODO: floats on non-word-aligned boundaries should technically cause alignment exceptions.
// Note that "word" means 32-bit, so paired singles or doubles might still be 32-bit aligned!
u32 em_address_next_page = (em_address + sizeof(T) - 1) & ~(HW_PAGE_SIZE - 1);
u32 tlb_addr_next_page = TranslateAddress<flag>(em_address_next_page);
if (tlb_addr == 0 || tlb_addr_next_page == 0)
{
if (flag == FLAG_READ)
GenerateDSIException(em_address_next_page, false);
return 0;
}
T var = 0;
for (u32 addr = em_address; addr < em_address + sizeof(T); addr++, tlb_addr++)
{
if (addr == em_address_next_page)
tlb_addr = tlb_addr_next_page;
var = (var << 8) | Memory::physical_base[tlb_addr];
}
return var;
}
// The easy case!
return bswap(*(const T*)&Memory::physical_base[tlb_addr]);
}
// Syncs the shader constant buffers with xfmem
// TODO: A cleaner way to control the matrices without making a mess in the parameters field
void VertexShaderManager::SetConstants()
{
if (nTransformMatricesChanged[0] >= 0)
{
int startn = nTransformMatricesChanged[0] / 4;
int endn = (nTransformMatricesChanged[1] + 3) / 4;
memcpy(constants.transformmatrices[startn], &xfmem.posMatrices[startn * 4],
(endn - startn) * sizeof(float4));
dirty = true;
nTransformMatricesChanged[0] = nTransformMatricesChanged[1] = -1;
}
if (nNormalMatricesChanged[0] >= 0)
{
int startn = nNormalMatricesChanged[0] / 3;
int endn = (nNormalMatricesChanged[1] + 2) / 3;
for (int i = startn; i < endn; i++)
{
memcpy(constants.normalmatrices[i], &xfmem.normalMatrices[3 * i], 12);
}
dirty = true;
nNormalMatricesChanged[0] = nNormalMatricesChanged[1] = -1;
}
if (nPostTransformMatricesChanged[0] >= 0)
{
int startn = nPostTransformMatricesChanged[0] / 4;
int endn = (nPostTransformMatricesChanged[1] + 3) / 4;
memcpy(constants.posttransformmatrices[startn], &xfmem.postMatrices[startn * 4],
(endn - startn) * sizeof(float4));
dirty = true;
nPostTransformMatricesChanged[0] = nPostTransformMatricesChanged[1] = -1;
}
if (nLightsChanged[0] >= 0)
{
// TODO: Outdated comment
// lights don't have a 1 to 1 mapping, the color component needs to be converted to 4 floats
int istart = nLightsChanged[0] / 0x10;
int iend = (nLightsChanged[1] + 15) / 0x10;
for (int i = istart; i < iend; ++i)
{
const Light& light = xfmem.lights[i];
VertexShaderConstants::Light& dstlight = constants.lights[i];
// xfmem.light.color is packed as abgr in u8[4], so we have to swap the order
dstlight.color[0] = light.color[3];
dstlight.color[1] = light.color[2];
dstlight.color[2] = light.color[1];
dstlight.color[3] = light.color[0];
dstlight.cosatt[0] = light.cosatt[0];
dstlight.cosatt[1] = light.cosatt[1];
dstlight.cosatt[2] = light.cosatt[2];
if (fabs(light.distatt[0]) < 0.00001f && fabs(light.distatt[1]) < 0.00001f &&
fabs(light.distatt[2]) < 0.00001f)
{
// dist attenuation, make sure not equal to 0!!!
dstlight.distatt[0] = .00001f;
}
else
{
dstlight.distatt[0] = light.distatt[0];
}
dstlight.distatt[1] = light.distatt[1];
dstlight.distatt[2] = light.distatt[2];
dstlight.pos[0] = light.dpos[0];
dstlight.pos[1] = light.dpos[1];
dstlight.pos[2] = light.dpos[2];
double norm = double(light.ddir[0]) * double(light.ddir[0]) +
double(light.ddir[1]) * double(light.ddir[1]) +
double(light.ddir[2]) * double(light.ddir[2]);
norm = 1.0 / sqrt(norm);
float norm_float = static_cast<float>(norm);
dstlight.dir[0] = light.ddir[0] * norm_float;
dstlight.dir[1] = light.ddir[1] * norm_float;
dstlight.dir[2] = light.ddir[2] * norm_float;
}
dirty = true;
nLightsChanged[0] = nLightsChanged[1] = -1;
}
for (int i : nMaterialsChanged)
{
u32 data = i >= 2 ? xfmem.matColor[i - 2] : xfmem.ambColor[i];
constants.materials[i][0] = (data >> 24) & 0xFF;
constants.materials[i][1] = (data >> 16) & 0xFF;
constants.materials[i][2] = (data >> 8) & 0xFF;
constants.materials[i][3] = data & 0xFF;
dirty = true;
}
nMaterialsChanged = BitSet32(0);
if (bPosNormalMatrixChanged)
{
bPosNormalMatrixChanged = false;
const float* pos = &xfmem.posMatrices[g_main_cp_state.matrix_index_a.PosNormalMtxIdx * 4];
const float* norm =
&xfmem.normalMatrices[3 * (g_main_cp_state.matrix_index_a.PosNormalMtxIdx & 31)];
memcpy(constants.posnormalmatrix, pos, 3 * sizeof(float4));
memcpy(constants.posnormalmatrix[3], norm, 3 * sizeof(float));
memcpy(constants.posnormalmatrix[4], norm + 3, 3 * sizeof(float));
memcpy(constants.posnormalmatrix[5], norm + 6, 3 * sizeof(float));
dirty = true;
}
if (bTexMatricesChanged[0])
{
bTexMatricesChanged[0] = false;
const float* pos_matrix_ptrs[] = {
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex0MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex1MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex2MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_a.Tex3MtxIdx * 4]};
for (size_t i = 0; i < ArraySize(pos_matrix_ptrs); ++i)
{
memcpy(constants.texmatrices[3 * i], pos_matrix_ptrs[i], 3 * sizeof(float4));
}
dirty = true;
}
if (bTexMatricesChanged[1])
{
bTexMatricesChanged[1] = false;
const float* pos_matrix_ptrs[] = {
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex4MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex5MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex6MtxIdx * 4],
&xfmem.posMatrices[g_main_cp_state.matrix_index_b.Tex7MtxIdx * 4]};
for (size_t i = 0; i < ArraySize(pos_matrix_ptrs); ++i)
{
memcpy(constants.texmatrices[3 * i + 12], pos_matrix_ptrs[i], 3 * sizeof(float4));
}
dirty = true;
}
if (bViewportChanged)
{
bViewportChanged = false;
// The console GPU places the pixel center at 7/12 unless antialiasing
// is enabled, while D3D and OpenGL place it at 0.5. See the comment
// in VertexShaderGen.cpp for details.
// NOTE: If we ever emulate antialiasing, the sample locations set by
// BP registers 0x01-0x04 need to be considered here.
const float pixel_center_correction = 7.0f / 12.0f - 0.5f;
const float pixel_size_x = 2.f / Renderer::EFBToScaledXf(2.f * xfmem.viewport.wd);
const float pixel_size_y = 2.f / Renderer::EFBToScaledXf(2.f * xfmem.viewport.ht);
constants.pixelcentercorrection[0] = pixel_center_correction * pixel_size_x;
constants.pixelcentercorrection[1] = pixel_center_correction * pixel_size_y;
// By default we don't change the depth value at all in the vertex shader.
constants.pixelcentercorrection[2] = 1.0f;
constants.pixelcentercorrection[3] = 0.0f;
if (g_ActiveConfig.backend_info.bSupportsDepthClamp)
{
// Oversized depth ranges are handled in the vertex shader. We need to reverse
// the far value to get a reversed depth range mapping. This is necessary
// because the standard depth range equation pushes all depth values towards
// the back of the depth buffer where conventionally depth buffers have the
// least precision.
if (g_ActiveConfig.backend_info.bSupportsReversedDepthRange)
{
if (fabs(xfmem.viewport.zRange) > 16777215.0f || fabs(xfmem.viewport.farZ) > 16777215.0f)
{
// For backends that support reversing the depth range we also support cases
// where the console also uses reversed depth with the same accuracy. We need
// to make sure the depth range is positive here and then reverse the depth in
// the backend viewport.
constants.pixelcentercorrection[2] = fabs(xfmem.viewport.zRange) / 16777215.0f;
if (xfmem.viewport.zRange < 0.0f)
constants.pixelcentercorrection[3] = xfmem.viewport.farZ / 16777215.0f;
else
constants.pixelcentercorrection[3] = 1.0f - xfmem.viewport.farZ / 16777215.0f;
}
}
else
{
if (xfmem.viewport.zRange < 0.0f || xfmem.viewport.zRange > 16777215.0f ||
fabs(xfmem.viewport.farZ) > 16777215.0f)
{
// For backends that don't support reversing the depth range we can still render
// cases where the console uses reversed depth correctly. But we simply can't
// provide the same accuracy as the console.
constants.pixelcentercorrection[2] = xfmem.viewport.zRange / 16777215.0f;
constants.pixelcentercorrection[3] = 1.0f - xfmem.viewport.farZ / 16777215.0f;
}
}
}
dirty = true;
// This is so implementation-dependent that we can't have it here.
g_renderer->SetViewport();
// Update projection if the viewport isn't 1:1 useable
if (!g_ActiveConfig.backend_info.bSupportsOversizedViewports)
{
ViewportCorrectionMatrix(s_viewportCorrection);
bProjectionChanged = true;
}
}
if (bProjectionChanged)
{
bProjectionChanged = false;
float* rawProjection = xfmem.projection.rawProjection;
switch (xfmem.projection.type)
{
case GX_PERSPECTIVE:
g_fProjectionMatrix[0] = rawProjection[0] * g_ActiveConfig.fAspectRatioHackW;
g_fProjectionMatrix[1] = 0.0f;
g_fProjectionMatrix[2] = rawProjection[1];
g_fProjectionMatrix[3] = 0.0f;
g_fProjectionMatrix[4] = 0.0f;
g_fProjectionMatrix[5] = rawProjection[2] * g_ActiveConfig.fAspectRatioHackH;
g_fProjectionMatrix[6] = rawProjection[3];
g_fProjectionMatrix[7] = 0.0f;
g_fProjectionMatrix[8] = 0.0f;
g_fProjectionMatrix[9] = 0.0f;
g_fProjectionMatrix[10] = rawProjection[4];
g_fProjectionMatrix[11] = rawProjection[5];
g_fProjectionMatrix[12] = 0.0f;
g_fProjectionMatrix[13] = 0.0f;
g_fProjectionMatrix[14] = -1.0f;
g_fProjectionMatrix[15] = 0.0f;
// Heuristic to detect if a GameCube game is in 16:9 anamorphic widescreen mode.
if (!SConfig::GetInstance().bWii)
{
bool viewport_is_4_3 = AspectIs4_3(xfmem.viewport.wd, xfmem.viewport.ht);
if (AspectIs16_9(rawProjection[2], rawProjection[0]) && viewport_is_4_3)
Core::g_aspect_wide = true; // Projection is 16:9 and viewport is 4:3, we are rendering
// an anamorphic widescreen picture
else if (AspectIs4_3(rawProjection[2], rawProjection[0]) && viewport_is_4_3)
Core::g_aspect_wide =
false; // Project and viewports are both 4:3, we are rendering a normal image.
}
SETSTAT_FT(stats.gproj_0, g_fProjectionMatrix[0]);
SETSTAT_FT(stats.gproj_1, g_fProjectionMatrix[1]);
SETSTAT_FT(stats.gproj_2, g_fProjectionMatrix[2]);
SETSTAT_FT(stats.gproj_3, g_fProjectionMatrix[3]);
SETSTAT_FT(stats.gproj_4, g_fProjectionMatrix[4]);
SETSTAT_FT(stats.gproj_5, g_fProjectionMatrix[5]);
SETSTAT_FT(stats.gproj_6, g_fProjectionMatrix[6]);
SETSTAT_FT(stats.gproj_7, g_fProjectionMatrix[7]);
SETSTAT_FT(stats.gproj_8, g_fProjectionMatrix[8]);
SETSTAT_FT(stats.gproj_9, g_fProjectionMatrix[9]);
SETSTAT_FT(stats.gproj_10, g_fProjectionMatrix[10]);
SETSTAT_FT(stats.gproj_11, g_fProjectionMatrix[11]);
SETSTAT_FT(stats.gproj_12, g_fProjectionMatrix[12]);
SETSTAT_FT(stats.gproj_13, g_fProjectionMatrix[13]);
SETSTAT_FT(stats.gproj_14, g_fProjectionMatrix[14]);
SETSTAT_FT(stats.gproj_15, g_fProjectionMatrix[15]);
break;
case GX_ORTHOGRAPHIC:
g_fProjectionMatrix[0] = rawProjection[0];
g_fProjectionMatrix[1] = 0.0f;
g_fProjectionMatrix[2] = 0.0f;
g_fProjectionMatrix[3] = rawProjection[1];
g_fProjectionMatrix[4] = 0.0f;
g_fProjectionMatrix[5] = rawProjection[2];
g_fProjectionMatrix[6] = 0.0f;
g_fProjectionMatrix[7] = rawProjection[3];
g_fProjectionMatrix[8] = 0.0f;
g_fProjectionMatrix[9] = 0.0f;
g_fProjectionMatrix[10] = (g_ProjHack1.value + rawProjection[4]) *
((g_ProjHack1.sign == 0) ? 1.0f : g_ProjHack1.sign);
g_fProjectionMatrix[11] = (g_ProjHack2.value + rawProjection[5]) *
((g_ProjHack2.sign == 0) ? 1.0f : g_ProjHack2.sign);
g_fProjectionMatrix[12] = 0.0f;
g_fProjectionMatrix[13] = 0.0f;
g_fProjectionMatrix[14] = 0.0f;
g_fProjectionMatrix[15] = 1.0f;
SETSTAT_FT(stats.g2proj_0, g_fProjectionMatrix[0]);
SETSTAT_FT(stats.g2proj_1, g_fProjectionMatrix[1]);
SETSTAT_FT(stats.g2proj_2, g_fProjectionMatrix[2]);
SETSTAT_FT(stats.g2proj_3, g_fProjectionMatrix[3]);
SETSTAT_FT(stats.g2proj_4, g_fProjectionMatrix[4]);
SETSTAT_FT(stats.g2proj_5, g_fProjectionMatrix[5]);
SETSTAT_FT(stats.g2proj_6, g_fProjectionMatrix[6]);
SETSTAT_FT(stats.g2proj_7, g_fProjectionMatrix[7]);
SETSTAT_FT(stats.g2proj_8, g_fProjectionMatrix[8]);
SETSTAT_FT(stats.g2proj_9, g_fProjectionMatrix[9]);
SETSTAT_FT(stats.g2proj_10, g_fProjectionMatrix[10]);
SETSTAT_FT(stats.g2proj_11, g_fProjectionMatrix[11]);
SETSTAT_FT(stats.g2proj_12, g_fProjectionMatrix[12]);
SETSTAT_FT(stats.g2proj_13, g_fProjectionMatrix[13]);
SETSTAT_FT(stats.g2proj_14, g_fProjectionMatrix[14]);
SETSTAT_FT(stats.g2proj_15, g_fProjectionMatrix[15]);
SETSTAT_FT(stats.proj_0, rawProjection[0]);
SETSTAT_FT(stats.proj_1, rawProjection[1]);
SETSTAT_FT(stats.proj_2, rawProjection[2]);
SETSTAT_FT(stats.proj_3, rawProjection[3]);
SETSTAT_FT(stats.proj_4, rawProjection[4]);
SETSTAT_FT(stats.proj_5, rawProjection[5]);
break;
default:
ERROR_LOG(VIDEO, "Unknown projection type: %d", xfmem.projection.type);
}
PRIM_LOG("Projection: %f %f %f %f %f %f", rawProjection[0], rawProjection[1], rawProjection[2],
rawProjection[3], rawProjection[4], rawProjection[5]);
if (g_ActiveConfig.bFreeLook && xfmem.projection.type == GX_PERSPECTIVE)
{
Matrix44 mtxA;
Matrix44 mtxB;
Matrix44 viewMtx;
Matrix44::Translate(mtxA, s_fViewTranslationVector);
Matrix44::LoadMatrix33(mtxB, s_viewRotationMatrix);
Matrix44::Multiply(mtxB, mtxA, viewMtx); // view = rotation x translation
Matrix44::Set(mtxB, g_fProjectionMatrix);
Matrix44::Multiply(mtxB, viewMtx, mtxA); // mtxA = projection x view
Matrix44::Multiply(s_viewportCorrection, mtxA, mtxB); // mtxB = viewportCorrection x mtxA
memcpy(constants.projection, mtxB.data, 4 * sizeof(float4));
}
else
{
Matrix44 projMtx;
Matrix44::Set(projMtx, g_fProjectionMatrix);
Matrix44 correctedMtx;
Matrix44::Multiply(s_viewportCorrection, projMtx, correctedMtx);
memcpy(constants.projection, correctedMtx.data, 4 * sizeof(float4));
}
dirty = true;
}
}
u32 PPCAnalyzer::Analyze(u32 address, CodeBlock *block, CodeBuffer *buffer, u32 blockSize)
{
// Clear block stats
memset(block->m_stats, 0, sizeof(BlockStats));
// Clear register stats
block->m_gpa->any = true;
block->m_fpa->any = false;
block->m_gpa->Clear();
block->m_fpa->Clear();
// Set the blocks start address
block->m_address = address;
// Reset our block state
block->m_broken = false;
block->m_memory_exception = false;
block->m_num_instructions = 0;
if (address == 0)
{
// Memory exception occurred during instruction fetch
block->m_memory_exception = true;
return address;
}
if (SConfig::GetInstance().m_LocalCoreStartupParameter.bMMU && (address & JIT_ICACHE_VMEM_BIT))
{
if (!Memory::TranslateAddress(address, Memory::FLAG_NO_EXCEPTION))
{
// Memory exception occurred during instruction fetch
block->m_memory_exception = true;
return address;
}
}
CodeOp *code = buffer->codebuffer;
bool found_exit = false;
u32 return_address = 0;
u32 numFollows = 0;
u32 num_inst = 0;
for (u32 i = 0; i < blockSize; ++i)
{
UGeckoInstruction inst = JitInterface::ReadOpcodeJIT(address);
if (inst.hex != 0)
{
num_inst++;
memset(&code[i], 0, sizeof(CodeOp));
GekkoOPInfo *opinfo = GetOpInfo(inst);
code[i].opinfo = opinfo;
code[i].address = address;
code[i].inst = inst;
code[i].branchTo = -1;
code[i].branchToIndex = -1;
code[i].skip = false;
block->m_stats->numCycles += opinfo->numCycles;
SetInstructionStats(block, &code[i], opinfo, i);
bool follow = false;
u32 destination = 0;
bool conditional_continue = false;
// Do we inline leaf functions?
if (HasOption(OPTION_LEAF_INLINE))
{
if (inst.OPCD == 18 && blockSize > 1)
{
//Is bx - should we inline? yes!
if (inst.AA)
destination = SignExt26(inst.LI << 2);
else
destination = address + SignExt26(inst.LI << 2);
if (destination != block->m_address)
follow = true;
}
else if (inst.OPCD == 19 && inst.SUBOP10 == 16 &&
(inst.BO & (1 << 4)) && (inst.BO & (1 << 2)) &&
return_address != 0)
{
// bclrx with unconditional branch = return
follow = true;
destination = return_address;
return_address = 0;
if (inst.LK)
return_address = address + 4;
}
else if (inst.OPCD == 31 && inst.SUBOP10 == 467)
{
// mtspr
const u32 index = (inst.SPRU << 5) | (inst.SPRL & 0x1F);
if (index == SPR_LR)
{
// We give up to follow the return address
// because we have to check the register usage.
return_address = 0;
}
}
// TODO: Find the optimal value for FUNCTION_FOLLOWING_THRESHOLD.
// If it is small, the performance will be down.
// If it is big, the size of generated code will be big and
// cache clearning will happen many times.
// TODO: Investivate the reason why
// "0" is fastest in some games, MP2 for example.
if (numFollows > FUNCTION_FOLLOWING_THRESHOLD)
follow = false;
}
if (HasOption(OPTION_CONDITIONAL_CONTINUE))
{
if (inst.OPCD == 16 &&
((inst.BO & BO_DONT_DECREMENT_FLAG) == 0 || (inst.BO & BO_DONT_CHECK_CONDITION) == 0))
{
// bcx with conditional branch
conditional_continue = true;
}
else if (inst.OPCD == 19 && inst.SUBOP10 == 16 &&
((inst.BO & BO_DONT_DECREMENT_FLAG) == 0 || (inst.BO & BO_DONT_CHECK_CONDITION) == 0))
{
// bclrx with conditional branch
conditional_continue = true;
}
else if (inst.OPCD == 3 ||
(inst.OPCD == 31 && inst.SUBOP10 == 4))
{
// tw/twi tests and raises an exception
conditional_continue = true;
}
else if (inst.OPCD == 19 && inst.SUBOP10 == 528 &&
(inst.BO_2 & BO_DONT_CHECK_CONDITION) == 0)
{
// Rare bcctrx with conditional branch
// Seen in NES games
conditional_continue = true;
}
}
if (!follow)
{
address += 4;
if (!conditional_continue && opinfo->flags & FL_ENDBLOCK) //right now we stop early
{
found_exit = true;
break;
}
}
// XXX: We don't support inlining yet.
#if 0
else
{
numFollows++;
// We don't "code[i].skip = true" here
// because bx may store a certain value to the link register.
// Instead, we skip a part of bx in Jit**::bx().
address = destination;
merged_addresses[size_of_merged_addresses++] = address;
}
#endif
}
else
{
// ISI exception or other critical memory exception occured (game over)
ERROR_LOG(DYNA_REC, "Instruction hex was 0!");
break;
}
}
block->m_num_instructions = num_inst;
if (block->m_num_instructions > 1)
ReorderInstructions(block->m_num_instructions, code);
if ((!found_exit && num_inst > 0) || blockSize == 1)
{
// We couldn't find an exit
block->m_broken = true;
}
// Scan for flag dependencies; assume the next block (or any branch that can leave the block)
// wants flags, to be safe.
bool wantsCR0 = true, wantsCR1 = true, wantsFPRF = true, wantsCA = true;
BitSet32 fprInUse, gprInUse, gprInReg, fprInXmm;
for (int i = block->m_num_instructions - 1; i >= 0; i--)
{
bool opWantsCR0 = code[i].wantsCR0;
bool opWantsCR1 = code[i].wantsCR1;
bool opWantsFPRF = code[i].wantsFPRF;
bool opWantsCA = code[i].wantsCA;
code[i].wantsCR0 = wantsCR0 || code[i].canEndBlock;
code[i].wantsCR1 = wantsCR1 || code[i].canEndBlock;
code[i].wantsFPRF = wantsFPRF || code[i].canEndBlock;
code[i].wantsCA = wantsCA || code[i].canEndBlock;
wantsCR0 |= opWantsCR0 || code[i].canEndBlock;
wantsCR1 |= opWantsCR1 || code[i].canEndBlock;
wantsFPRF |= opWantsFPRF || code[i].canEndBlock;
wantsCA |= opWantsCA || code[i].canEndBlock;
wantsCR0 &= !code[i].outputCR0 || opWantsCR0;
wantsCR1 &= !code[i].outputCR1 || opWantsCR1;
wantsFPRF &= !code[i].outputFPRF || opWantsFPRF;
wantsCA &= !code[i].outputCA || opWantsCA;
code[i].gprInUse = gprInUse;
code[i].fprInUse = fprInUse;
code[i].gprInReg = gprInReg;
code[i].fprInXmm = fprInXmm;
// TODO: if there's no possible endblocks or exceptions in between, tell the regcache
// we can throw away a register if it's going to be overwritten later.
gprInUse |= code[i].regsIn;
gprInReg |= code[i].regsIn;
fprInUse |= code[i].fregsIn;
if (strncmp(code[i].opinfo->opname, "stfd", 4))
fprInXmm |= code[i].fregsIn;
// For now, we need to count output registers as "used" though; otherwise the flush
// will result in a redundant store (e.g. store to regcache, then store again to
// the same location later).
gprInUse |= code[i].regsOut;
if (code[i].fregOut >= 0)
fprInUse[code[i].fregOut] = true;
}
// Forward scan, for flags that need the other direction for calculation.
BitSet32 fprIsSingle, fprIsDuplicated, fprIsStoreSafe;
for (u32 i = 0; i < block->m_num_instructions; i++)
{
code[i].fprIsSingle = fprIsSingle;
code[i].fprIsDuplicated = fprIsDuplicated;
code[i].fprIsStoreSafe = fprIsStoreSafe;
if (code[i].fregOut >= 0)
{
fprIsSingle[code[i].fregOut] = false;
fprIsDuplicated[code[i].fregOut] = false;
fprIsStoreSafe[code[i].fregOut] = false;
// Single, duplicated, and doesn't need PPC_FP.
if (code[i].opinfo->type == OPTYPE_SINGLEFP)
{
fprIsSingle[code[i].fregOut] = true;
fprIsDuplicated[code[i].fregOut] = true;
fprIsStoreSafe[code[i].fregOut] = true;
}
// Single and duplicated, but might be a denormal (not safe to skip PPC_FP).
// TODO: if we go directly from a load to store, skip conversion entirely?
// TODO: if we go directly from a load to a float instruction, and the value isn't used
// for anything else, we can skip PPC_FP on a load too.
if (!strncmp(code[i].opinfo->opname, "lfs", 3))
{
fprIsSingle[code[i].fregOut] = true;
fprIsDuplicated[code[i].fregOut] = true;
}
// Paired are still floats, but the top/bottom halves may differ.
if (code[i].opinfo->type == OPTYPE_PS || code[i].opinfo->type == OPTYPE_LOADPS)
{
fprIsSingle[code[i].fregOut] = true;
fprIsStoreSafe[code[i].fregOut] = true;
}
// Careful: changing the float mode in a block breaks this optimization, since
// a previous float op might have had had FTZ off while the later store has FTZ
// on. So, discard all information we have.
if (!strncmp(code[i].opinfo->opname, "mtfs", 4))
fprIsStoreSafe = BitSet32(0);
}
}
return address;
}
void PPCAnalyzer::SetInstructionStats(CodeBlock *block, CodeOp *code, GekkoOPInfo *opinfo, u32 index)
{
code->wantsCR0 = false;
code->wantsCR1 = false;
if (opinfo->flags & FL_USE_FPU)
block->m_fpa->any = true;
if (opinfo->flags & FL_TIMER)
block->m_gpa->anyTimer = true;
// Does the instruction output CR0?
if (opinfo->flags & FL_RC_BIT)
code->outputCR0 = code->inst.hex & 1; //todo fix
else if ((opinfo->flags & FL_SET_CRn) && code->inst.CRFD == 0)
code->outputCR0 = true;
else
code->outputCR0 = (opinfo->flags & FL_SET_CR0) ? true : false;
// Does the instruction output CR1?
if (opinfo->flags & FL_RC_BIT_F)
code->outputCR1 = code->inst.hex & 1; //todo fix
else if ((opinfo->flags & FL_SET_CRn) && code->inst.CRFD == 1)
code->outputCR1 = true;
else
code->outputCR1 = (opinfo->flags & FL_SET_CR1) ? true : false;
code->wantsFPRF = (opinfo->flags & FL_READ_FPRF) ? true : false;
code->outputFPRF = (opinfo->flags & FL_SET_FPRF) ? true : false;
code->canEndBlock = (opinfo->flags & FL_ENDBLOCK) ? true : false;
code->wantsCA = (opinfo->flags & FL_READ_CA) ? true : false;
code->outputCA = (opinfo->flags & FL_SET_CA) ? true : false;
// We're going to try to avoid storing carry in XER if we can avoid it -- keep it in the x86 carry flag!
// If the instruction reads CA but doesn't write it, we still need to store CA in XER; we can't
// leave it in flags.
if (HasOption(OPTION_CARRY_MERGE))
code->wantsCAInFlags = code->wantsCA && code->outputCA && opinfo->type == OPTYPE_INTEGER;
else
code->wantsCAInFlags = false;
// mfspr/mtspr can affect/use XER, so be super careful here
// we need to note specifically that mfspr needs CA in XER, not in the x86 carry flag
if (code->inst.OPCD == 31 && code->inst.SUBOP10 == 339) // mfspr
code->wantsCA = ((code->inst.SPRU << 5) | (code->inst.SPRL & 0x1F)) == SPR_XER;
if (code->inst.OPCD == 31 && code->inst.SUBOP10 == 467) // mtspr
code->outputCA = ((code->inst.SPRU << 5) | (code->inst.SPRL & 0x1F)) == SPR_XER;
code->regsIn = BitSet32(0);
code->regsOut = BitSet32(0);
if (opinfo->flags & FL_OUT_A)
{
code->regsOut[code->inst.RA] = true;
block->m_gpa->SetOutputRegister(code->inst.RA, index);
}
if (opinfo->flags & FL_OUT_D)
{
code->regsOut[code->inst.RD] = true;
block->m_gpa->SetOutputRegister(code->inst.RD, index);
}
if (opinfo->flags & FL_OUT_S)
{
code->regsOut[code->inst.RS] = true;
block->m_gpa->SetOutputRegister(code->inst.RS, index);
}
if ((opinfo->flags & FL_IN_A) || ((opinfo->flags & FL_IN_A0) && code->inst.RA != 0))
{
code->regsIn[code->inst.RA] = true;
block->m_gpa->SetInputRegister(code->inst.RA, index);
}
if (opinfo->flags & FL_IN_B)
{
code->regsIn[code->inst.RB] = true;
block->m_gpa->SetInputRegister(code->inst.RB, index);
}
if (opinfo->flags & FL_IN_C)
{
code->regsIn[code->inst.RC] = true;
block->m_gpa->SetInputRegister(code->inst.RC, index);
}
if (opinfo->flags & FL_IN_S)
{
code->regsIn[code->inst.RS] = true;
block->m_gpa->SetInputRegister(code->inst.RS, index);
}
if (code->inst.OPCD == 46) // lmw
{
for (int iReg = code->inst.RD; iReg < 32; ++iReg)
{
code->regsOut[iReg] = true;
block->m_gpa->SetOutputRegister(iReg, index);
}
}
else if (code->inst.OPCD == 47) //stmw
{
for (int iReg = code->inst.RS; iReg < 32; ++iReg)
{
code->regsIn[iReg] = true;
block->m_gpa->SetInputRegister(iReg, index);
}
}
code->fregOut = -1;
if (opinfo->flags & FL_OUT_FLOAT_D)
code->fregOut = code->inst.FD;
else if (opinfo->flags & FL_OUT_FLOAT_S)
code->fregOut = code->inst.FS;
code->fregsIn = BitSet32(0);
if (opinfo->flags & FL_IN_FLOAT_A)
code->fregsIn[code->inst.FA] = true;
if (opinfo->flags & FL_IN_FLOAT_B)
code->fregsIn[code->inst.FB] = true;
if (opinfo->flags & FL_IN_FLOAT_C)
code->fregsIn[code->inst.FC] = true;
if (opinfo->flags & FL_IN_FLOAT_D)
code->fregsIn[code->inst.FD] = true;
if (opinfo->flags & FL_IN_FLOAT_S)
code->fregsIn[code->inst.FS] = true;
switch (opinfo->type)
{
case OPTYPE_INTEGER:
case OPTYPE_LOAD:
case OPTYPE_STORE:
case OPTYPE_LOADFP:
case OPTYPE_STOREFP:
break;
case OPTYPE_SINGLEFP:
case OPTYPE_DOUBLEFP:
break;
case OPTYPE_BRANCH:
if (code->inst.hex == 0x4e800020)
{
// For analysis purposes, we can assume that blr eats opinfo->flags.
code->outputCR0 = true;
code->outputCR1 = true;
}
break;
case OPTYPE_SYSTEM:
case OPTYPE_SYSTEMFP:
break;
}
}
