port fixes found in forks

This commit is contained in:
patchzyy
2026-10-22 20:12:12 +02:00
parent b1c99d4cab
commit 9d11d1ff5c
10 changed files with 341 additions and 49 deletions
+7 -12
View File
@@ -807,21 +807,17 @@ struct DlMayContainDrawVisitor {
bool OnNop(uint8_t) { return true; }
bool OnBpReg(const uint8_t*, uint32_t) { return true; }
// An array base/stride write means the list is setting up indexed geometry.
bool OnCpReg(const uint8_t*, uint8_t reg, uint32_t) { return !(reg >= 0xA0 && reg <= 0xBF); }
// CP writes must reach the HLE vertex state, and indexed XF loads need their
// array resolved, so neither may be passed through raw.
bool OnCpReg(const uint8_t*, uint8_t, uint32_t) { return false; }
bool OnXfReg(const uint8_t*, uint32_t) { return true; }
bool OnIndexedXf(const uint8_t*, uint8_t, uint32_t) { return true; }
bool OnIndexedXf(const uint8_t*, uint8_t, uint32_t) { return false; }
bool OnInvalidateVertexCache(uint8_t) { return true; }
// Draw opcodes reach this too (kHandlesDraw is false): a draw opcode, or a
// command byte this walker does not model, both mean "assume it draws".
bool OnUnknownCommand(uint8_t) { return false; }
bool OnCallDisplayList(uint32_t addr, uint32_t size, int depth) {
if (addr == 0 || size == 0) return true;
const uint8_t* nested = static_cast<const uint8_t*>(GuestToHostPtr(addr, size));
if (!nested) return false;
DlMayContainDrawVisitor child;
return WalkDisplayList(nested, size, child, depth + 1);
}
// Aurora skips nested calls, so the child has to be flattened.
bool OnCallDisplayList(uint32_t addr, uint32_t size, int) { return addr == 0 || size == 0; }
};
static bool DisplayListMayContainDraw(const uint8_t* data, uint32_t nbytes) {
@@ -1142,8 +1138,7 @@ static void ApplyAuroraVtxDesc() {
// GXGetVtxDesc only decodes GX_VA_PNMTXIDX..GX_VA_TEX7; the XF array
// pseudo-attributes above GX_VA_TEX7 have no CP descriptor bits, so
// probing them would always report GX_NONE and republish a no-op.
if (!needsPublish && !IsMatrixIndexAttr(gxAttr) && gxAttr <= GX_VA_TEX7 &&
(type == GX_INDEX8 || type == GX_INDEX16)) {
if (!needsPublish && gxAttr <= GX_VA_TEX7) {
needsPublish = !AuroraVtxDescMatchesLive(gxAttr, type);
}
if (needsPublish) {
+9 -2
View File
@@ -490,9 +490,16 @@ void HleFifoWrite(u32 val, uint32_t sizeBytes) {
const uint16_t vtxCount = ReadBE16(data + 1);
const GXVtxFmt vtxFmt = static_cast<GXVtxFmt>(cmd & GX_VAT_MASK_CMD);
const GXPrimitive prim = OpcodeToGXPrimitive(cmd);
if (vtxCount == 0) {
if (!consumeBytes(3, sink)) break;
continue;
}
uint32_t rawVertexSize = 0;
if (TryGetRawDirectFifoVertexSize(vtxFmt, rawVertexSize)) {
const uint32_t packetBytes = 3u + static_cast<uint32_t>(vtxCount) * rawVertexSize;
const uint32_t packetBytes = TryGetRawDirectFifoVertexSize(vtxFmt, rawVertexSize)
? 3u + static_cast<uint32_t>(vtxCount) * rawVertexSize
: 0u;
// A packet that cannot fit the buffer is streamed per vertex instead.
if (packetBytes != 0 && packetBytes <= g_hleGxState.fifoBytes.size()) {
if (g_hleGxState.fifoByteCount < packetBytes) {
break;
}
+5 -2
View File
@@ -703,8 +703,11 @@ extern "C" void GX__LoadTexObj_80170f2c(uint32_t oa, uint32_t tid) {
}
PPC_NATIVE_OVERRIDE_VOID(80170f2c, GX__LoadTexObj_80170f2c, (uint32_t oa, uint32_t tid), (oa, tid));
extern "C" void GX__LoadTexObjPreLoaded_80170dc8(uint32_t oa, uint32_t tid) { GX__LoadTexObj_80170f2c(oa, tid); }
PPC_NATIVE_OVERRIDE_VOID(80170dc8, GX__LoadTexObjPreLoaded_80170dc8, (uint32_t oa, uint32_t tid), (oa, tid));
extern "C" void GX__LoadTexObjPreLoaded_80170dc8(uint32_t oa, uint32_t region, uint32_t tid) {
(void)region;
GX__LoadTexObj_80170f2c(oa, tid);
}
PPC_NATIVE_OVERRIDE_VOID(80170dc8, GX__LoadTexObjPreLoaded_80170dc8, (uint32_t oa, uint32_t region, uint32_t tid), (oa, region, tid));
// ============================================================================
// Texture Object Getters
+2 -2
View File
@@ -150,7 +150,7 @@ extern "C" void GX__SetVtxAttrFmtv_8016de08(uint32_t vf, uint32_t la) {
while(true){
uint32_t a=Memory::Read32(p);
if(a==0xFFu) break;
list.push_back({a, Memory::Read32(p+4), Memory::Read32(p+8), Memory::Read32(p+12)&0xFFu});
list.push_back({a, Memory::Read32(p+4), Memory::Read32(p+8), Memory::Read8(p+12)});
p+=16;
}
GxWorker::Run([vf, list = std::move(list)] {
@@ -326,7 +326,7 @@ extern "C" void GX__Begin_8016f0f0(uint32_t t, uint32_t vf, uint32_t nv) {
static void GxBeginImpl(uint32_t t, uint32_t vf, uint32_t nv) {
ApplyAuroraVtxStateForBegin(static_cast<GXVtxFmt>(vf));
g_hleGxState.currentVtxFmt=(GXVtxFmt)vf; g_hleGxState.currentPrim=(GXPrimitive)t;
g_hleGxState.vertsRemaining=nv; g_hleGxState.inBegin=true; g_hleGxState.auroraBeginCalled=false;
g_hleGxState.vertsRemaining=nv; g_hleGxState.inBegin=nv!=0; g_hleGxState.auroraBeginCalled=false;
g_hleGxState.fifoReadOffset = 0;
g_hleGxState.fifoByteCount=0; g_hleGxState.ResetVertex();
}
+18 -7
View File
@@ -1031,16 +1031,17 @@ extern "C" int32_t DVDLowInit_80164848() {
}
PPC_NATIVE_OVERRIDE(80164848, DVDLowInit_80164848, int32_t, (), ());
extern "C" int32_t DVDLowInquiry_80165A30(uint32_t cmdBlockPtr, uint32_t callback)
// DVDLowInquiry(DVDDriveInfo* info, DVDLowCallback cb); reports a retail drive.
extern "C" int32_t DVDLowInquiry_80165A30(uint32_t driveInfoPtr, uint32_t callback)
{
// HLE: acknowledge the drive is present. Must mark the command block complete
// (State = 0) or callers polling this address hang waiting for "Busy" to clear.
if (cmdBlockPtr) {
Memory::Write32(cmdBlockPtr + DVD_CB_OFFSET_STATE, DVD_STATE_END);
if (driveInfoPtr && Memory::Contains(driveInfoPtr, 0x20)) {
Memory::Write32(driveInfoPtr + 0x0, 0x00000002); // revision level, device code
Memory::Write32(driveInfoPtr + 0x4, 0x20060526); // release date
Memory::Write32(driveInfoPtr + 0x8, 0x41000000);
}
InvokeDvdLowCallback(callback, DvdReadContract::kInterruptTransferComplete);
CompleteDvdCancelState();
return 1; // Return 1 to indicate the command was successfully issued.
return 1;
}
PPC_NATIVE_OVERRIDE(80165A30, DVDLowInquiry_80165A30, int32_t, (uint32_t b, uint32_t c), (b, c));
@@ -1051,11 +1052,21 @@ extern "C" int32_t DVDLowReadDiskID_80164AAC(uint32_t diskIdPtr, uint32_t callba
Memory::Write16(diskIdPtr + 0x04, 0x3031);
Memory::Write8 (diskIdPtr + 0x06, 0x01);
}
InvokeDvdLowCallback(callback, DvdReadContract::kInterruptTransferComplete);
CompleteDvdCancelState();
return 1; // Success
}
PPC_NATIVE_OVERRIDE(80164AAC, DVDLowReadDiskID_80164AAC, int32_t, (uint32_t p, uint32_t c), (p, c));
// Host files need no head positioning.
extern "C" int32_t DVDLowSeek_801664E0(uint32_t offset, uint32_t callback) {
(void)offset;
InvokeDvdLowCallback(callback, DvdReadContract::kInterruptTransferComplete);
CompleteDvdCancelState();
return 1;
}
PPC_NATIVE_OVERRIDE(801664E0, DVDLowSeek_801664E0, int32_t, (uint32_t o, uint32_t c), (o, c));
// 0x80166330 -> DVDLowRead (And 0x80165708 UnencryptedRead)
// The game calls this to read the Disk Header (offset 0) or raw data.
extern "C" int32_t DVDLowRead_80166330(uint32_t buffer, uint32_t length, uint32_t offset, uint32_t callback)
+1 -1
View File
@@ -964,7 +964,7 @@ static int32_t HandleIsfsReadDir(uint32_t numIn, uint32_t numOut, uint32_t vecto
}
const std::filesystem::path hostPath = TranslateNandPath(wiiPath.c_str());
if (!IsDirectory(hostPath)) {
return ISFS_ENOENT;
return PathExists(hostPath) ? ISFS_EINVAL : ISFS_ENOENT;
}
// NAND names are at most 12 characters; longer host names cannot exist on
@@ -253,6 +253,12 @@ public sealed partial class CxxLinearCodeGenerator
? index
: -1;
private static uint ClearPairedBit(uint mask, string destination)
{
var bit = PairedFloatRegisterBit(destination);
return bit >= 0 ? mask & ~(1u << bit) : mask;
}
private static HashSet<string> MaterializePairedFloatSet(uint mask, StringComparer comparer)
{
var set = new HashSet<string>(comparer);
@@ -454,11 +460,21 @@ public sealed partial class CxxLinearCodeGenerator
break;
}
case IrLoad load:
newOut = ClearPairedBit(newOut, load.Destination);
break;
case IrResolvedLoad load:
newOut = ClearPairedBit(newOut, load.Destination);
break;
case IrResolvedLoadPair pair:
newOut = ClearPairedBit(newOut, pair.FirstDestination);
newOut = ClearPairedBit(newOut, pair.SecondDestination);
break;
case IrResolvedPsqLoad load:
{
var destBit = PairedFloatRegisterBit(load.Destination);
if (destBit >= 0)
{
newOut &= ~(1u << destBit);
newOut |= 1u << destBit;
}
break;
}
@@ -70,7 +70,7 @@ internal static class JumpTableDetector
}
entriesAreRelative = HasRelativeEntryAdd(ordered, loadIndex, mtctrIndex, targetRegister, baseReg.Name);
if (!TryFindUpperBound(ordered, loadIndex, indexReg.Name, out var upperBound) || upperBound < 0 || upperBound >= MaxEntryCount)
if (!TryFindUpperBound(ordered, loadIndex, indexReg, out var upperBound) || upperBound < 0 || upperBound >= MaxEntryCount)
{
return false;
}
@@ -381,47 +381,60 @@ internal static class JumpTableDetector
return false;
}
private static bool TryFindUpperBound(IReadOnlyList<PpcInstruction> ordered, int startIndex, string register, out int upperBound)
private static bool TryFindUpperBound(IReadOnlyList<PpcInstruction> ordered, int startIndex, PpcRegisterOperand register, out int upperBound)
{
// Preferred: compare directly on the same register used for lwzx indexing.
// Preferred: the compare on the lwzx index, followed back through the copies and
// shifts that derived it. A compare older than the value's definition bounds
// something else.
var tracked = register.Number;
for (var i = startIndex - 1; i >= 0 && startIndex - i <= MaxBacktrackInstructions; i--)
{
var ins = ordered[i];
if (ins.Operands.Count < 2 || ins.Operands[0] is not PpcRegisterOperand reg ||
!string.Equals(reg.Name, register, StringComparison.OrdinalIgnoreCase) ||
ins.Operands[1] is not PpcImmediateOperand imm)
if (TryGetCompareImmediate(ins, out var compared, out var bound))
{
if (compared == tracked)
{
upperBound = bound;
return true;
}
continue;
}
if (ins.IsCall)
{
break;
}
if (!DefinesGpr(ins, tracked))
{
continue;
}
if (ins.Mnemonic is "cmplwi" or "cmpwi")
if (IsIndexDerivation(ins) &&
ins.Operands[0] is PpcRegisterOperand dest && dest.Number == tracked &&
ins.Operands[1] is PpcRegisterOperand source)
{
upperBound = imm.Value;
return true;
tracked = source.Number;
continue;
}
break;
}
// Fallback: some loops compare a logical counter (e.g. r18) distinct from the byte-offset
// register (e.g. r30) that feeds lwzx, so the bound check may not target the index register.
for (var i = startIndex - 1; i >= 0 && startIndex - i <= MaxBacktrackInstructions; i--)
{
var ins = ordered[i];
if (ins.Operands.Count < 2 || ins.Operands[1] is not PpcImmediateOperand imm)
if (ordered[i].Operands.Count > 0 && ordered[i].Operands[0] is PpcConditionRegisterOperand)
{
continue;
}
if (!TryGetCompareImmediate(ordered[i], out _, out var bound) || bound < 0 || bound >= MaxEntryCount)
{
continue;
}
if (ins.Mnemonic is not ("cmplwi" or "cmpwi"))
{
continue;
}
if (imm.Value < 0 || imm.Value >= MaxEntryCount)
{
continue;
}
upperBound = imm.Value;
upperBound = bound;
return true;
}
@@ -429,6 +442,70 @@ internal static class JumpTableDetector
return false;
}
// Copies, sign extensions and left shifts (slwi, clrlwi, clrlslwi) keep a bound on
// the source a bound on the result's entry index.
private static bool IsIndexDerivation(PpcInstruction ins) => ins.Mnemonic switch
{
"mr" or "extsb" or "extsh" => ins.Operands.Count >= 2,
"rlwinm" => ins.Operands.Count == 5 &&
ins.Operands[2] is PpcImmediateOperand sh &&
ins.Operands[4] is PpcImmediateOperand me &&
me.Value == 31 - sh.Value,
_ => false,
};
// MayWriteGpr treats both register fields of an X-form as written, which also
// counts sources (or, extsb, stwx, cmpw); read the decoded destination instead.
private static bool DefinesGpr(PpcInstruction ins, int register)
{
if (ins.Word >> 26 != 31)
{
return PpcRegisterEffects.MayWriteGpr(ins.Word, register);
}
var mnemonic = ins.Mnemonic;
if (mnemonic.EndsWith("ux", StringComparison.Ordinal) &&
ins.Operands.Count > 1 && ins.Operands[1] is PpcRegisterOperand updated && updated.Number == register)
{
return true;
}
if (mnemonic.StartsWith("st", StringComparison.Ordinal) ||
mnemonic.StartsWith("cmp", StringComparison.Ordinal) ||
mnemonic.StartsWith("mt", StringComparison.Ordinal) ||
mnemonic.StartsWith("dcb", StringComparison.Ordinal) ||
mnemonic.StartsWith("icb", StringComparison.Ordinal) ||
mnemonic.StartsWith("tw", StringComparison.Ordinal))
{
return false;
}
return ins.Operands.Count > 0 && ins.Operands[0] is PpcRegisterOperand dest && dest.Number == register;
}
private static bool TryGetCompareImmediate(PpcInstruction ins, out int register, out int immediate)
{
register = -1;
immediate = 0;
if (ins.Mnemonic is not ("cmplwi" or "cmpwi"))
{
return false;
}
var operands = ins.Operands;
var first = operands.Count > 0 && operands[0] is PpcConditionRegisterOperand ? 1 : 0;
if (operands.Count < first + 2 ||
operands[first] is not PpcRegisterOperand reg ||
operands[first + 1] is not PpcImmediateOperand imm)
{
return false;
}
register = reg.Number;
immediate = imm.Value;
return true;
}
private static bool TryReadWord(ProgramImage image, uint address, out uint value)
{
if (!image.Contains(address, sizeof(uint)))
@@ -538,4 +538,122 @@ public class FunctionTranslatorCoverageTests
Assert.Equal("// generated", result.CxxCode);
Assert.Same(metrics, result.Metrics);
}
[Fact]
public void Translate_JumpTableBoundFollowsShiftedIndexPastStaleCompare()
{
const uint entry = 0x80000000;
const uint table = 0x80000100;
const uint case0 = entry + 0x20;
const uint case1 = entry + 0x28;
const uint case2 = entry + 0x30;
var image = TranslatorCppTestHarness.CreateImage(
(entry + 0x00, 0x2C040001), // cmpwi r4, 1
(entry + 0x04, 0x28030002), // cmplwi r3, 2
(entry + 0x08, 0x5464103A), // slwi r4, r3, 2
(entry + 0x0C, 0x3D808000), // lis r12, 0x8000
(entry + 0x10, 0x398C0100), // addi r12, r12, 0x100
(entry + 0x14, 0x7C0C202E), // lwzx r0, r12, r4
(entry + 0x18, 0x7C0903A6), // mtctr r0
(entry + 0x1C, 0x4E800420), // bctr
(case0 + 0x00, 0x38600007), // li r3, 7
(case0 + 0x04, 0x4E800020), // blr
(case1 + 0x00, 0x38600009), // li r3, 9
(case1 + 0x04, 0x4E800020), // blr
(case2 + 0x00, 0x3860000B), // li r3, 11
(case2 + 0x04, 0x4E800020), // blr
(table + 0x00, case0),
(table + 0x04, case1),
(table + 0x08, case2));
var result = new FunctionTranslator(image).Translate(
entry,
TranslationOptions.Default with
{
MaxBytes = 0x80,
AllowUnsupportedInstructions = true,
KnownFunctionEntryPoints = new HashSet<uint> { entry }
});
Assert.Contains("case 0x80000030u:", result.CxxCode, StringComparison.Ordinal);
}
[Fact]
public void Translate_JumpTableBoundSurvivesCopyOfIndex()
{
const uint entry = 0x80000000;
const uint table = 0x80000100;
const uint case0 = entry + 0x28;
const uint case1 = entry + 0x30;
const uint case2 = entry + 0x38;
var image = TranslatorCppTestHarness.CreateImage(
(entry + 0x00, 0x28030002), // cmplwi r3, 2
(entry + 0x04, 0x7C7F1B78), // mr r31, r3
(entry + 0x08, 0x2C040001), // cmpwi r4, 1
(entry + 0x0C, 0x5460103A), // slwi r0, r3, 2
(entry + 0x10, 0x3D808000), // lis r12, 0x8000
(entry + 0x14, 0x398C0100), // addi r12, r12, 0x100
(entry + 0x18, 0x7C0C002E), // lwzx r0, r12, r0
(entry + 0x1C, 0x7C0903A6), // mtctr r0
(entry + 0x20, 0x4E800420), // bctr
(entry + 0x24, 0x4E800020), // blr
(case0 + 0x00, 0x38600007), // li r3, 7
(case0 + 0x04, 0x4E800020), // blr
(case1 + 0x00, 0x38600009), // li r3, 9
(case1 + 0x04, 0x4E800020), // blr
(case2 + 0x00, 0x3860000B), // li r3, 11
(case2 + 0x04, 0x4E800020), // blr
(table + 0x00, case0),
(table + 0x04, case1),
(table + 0x08, case2));
var result = new FunctionTranslator(image).Translate(
entry,
TranslationOptions.Default with
{
MaxBytes = 0x80,
AllowUnsupportedInstructions = true,
KnownFunctionEntryPoints = new HashSet<uint> { entry }
});
Assert.Contains("case 0x80000038u:", result.CxxCode, StringComparison.Ordinal);
}
[Fact]
public void Translate_JumpTableBoundAcceptsNonCr0Compare()
{
const uint entry = 0x80000000;
const uint table = 0x80000100;
const uint case0 = entry + 0x1C;
const uint case1 = entry + 0x24;
const uint case2 = entry + 0x2C;
var image = TranslatorCppTestHarness.CreateImage(
(entry + 0x00, 0x28830002), // cmplwi cr1, r3, 2
(entry + 0x04, 0x5463103A), // slwi r3, r3, 2
(entry + 0x08, 0x3D808000), // lis r12, 0x8000
(entry + 0x0C, 0x398C0100), // addi r12, r12, 0x100
(entry + 0x10, 0x7C0C182E), // lwzx r0, r12, r3
(entry + 0x14, 0x7C0903A6), // mtctr r0
(entry + 0x18, 0x4E800420), // bctr
(case0 + 0x00, 0x38600007), // li r3, 7
(case0 + 0x04, 0x4E800020), // blr
(case1 + 0x00, 0x38600009), // li r3, 9
(case1 + 0x04, 0x4E800020), // blr
(case2 + 0x00, 0x3860000B), // li r3, 11
(case2 + 0x04, 0x4E800020), // blr
(table + 0x00, case0),
(table + 0x04, case1),
(table + 0x08, case2));
var result = new FunctionTranslator(image).Translate(
entry,
TranslationOptions.Default with
{
MaxBytes = 0x80,
AllowUnsupportedInstructions = true,
KnownFunctionEntryPoints = new HashSet<uint> { entry }
});
Assert.Contains("case 0x8000002Cu:", result.CxxCode, StringComparison.Ordinal);
}
}
@@ -152,4 +152,69 @@ public sealed class PairedFlowCodeGenTests
Assert.Contains("SetCRFloatResident(cr, 0, f31.d, f1.d);", code, StringComparison.Ordinal);
Assert.DoesNotContain("SetCRFloatResident(cr, 0, PPC_PsToScalarInline(f31.d)", code, StringComparison.Ordinal);
}
[Fact]
public void ResolvedScalarLoadEndsPairedStateAcrossEdge()
{
var function = new IrFunction("resolved_load_clears_paired", "entry", new[]
{
new IrBasicBlock("entry", new IrInstruction[]
{
new IrCall("f5", "PPC_PsMul", new[] { IrValue.Register("f1"), IrValue.Register("f2") }),
new IrResolveGuestMemoryRange("range", IrValue.Register("r3"), 0, 8, true, false),
new IrResolvedLoad("f5", "range", new IrAddress("r3", 0), 0, 8),
new IrJump("next")
}),
new IrBasicBlock("next", new IrInstruction[]
{
new IrSetCrField(0, IrValue.Register("f5"), IrValue.Register("f1"), false),
new IrReturn(null)
})
});
var types = new RepresentationEnvironment(new Dictionary<string, ValueRepresentation>
{
["f1"] = ValueRepresentation.Float64,
["f2"] = ValueRepresentation.Float64,
["f5"] = ValueRepresentation.Float64,
["r3"] = ValueRepresentation.UInt32
});
var signature = new FunctionAbiClassification("resolved_load_clears_paired", ValueRepresentation.Void);
var code = new CxxLinearCodeGenerator().Emit(
0x80006008, new SsaTransformer().Convert(function), signature, types);
Assert.Contains("SetCRFloatResident(cr, 0, f5.d, f1.d);", code, StringComparison.Ordinal);
Assert.DoesNotContain("PPC_PsToScalarInline(f5.d)", code, StringComparison.Ordinal);
}
[Fact]
public void ResolvedPsqLoadStaysPairedAcrossEdge()
{
var function = new IrFunction("resolved_psq_load_paired", "entry", new[]
{
new IrBasicBlock("entry", new IrInstruction[]
{
new IrResolveGuestMemoryRange("range", IrValue.Register("r3"), 0, 8, true, false),
new IrResolvedPsqLoad("f5", "range", IrValue.Register("r3"), 0, 0, 0, 0u),
new IrJump("next")
}),
new IrBasicBlock("next", new IrInstruction[]
{
new IrSetCrField(0, IrValue.Register("f5"), IrValue.Register("f1"), false),
new IrReturn(null)
})
});
var types = new RepresentationEnvironment(new Dictionary<string, ValueRepresentation>
{
["f1"] = ValueRepresentation.Float64,
["f5"] = ValueRepresentation.Float64,
["r3"] = ValueRepresentation.UInt32
});
var signature = new FunctionAbiClassification("resolved_psq_load_paired", ValueRepresentation.Void);
var code = new CxxLinearCodeGenerator().Emit(
0x8000600C, new SsaTransformer().Convert(function), signature, types);
Assert.Contains("SetCRFloatResident(cr, 0, PPC_PsToScalarInline(f5.d), f1.d);", code, StringComparison.Ordinal);
}
}