Compare commits

...

8 Commits

Author SHA1 Message Date
JosJuice
9e7bb5e027
Merge 980a726313 into 2c92e5b5b3 2024-11-14 01:47:19 +00:00
OatmealDome
2c92e5b5b3
Merge pull request #13160 from cpba/flatpak-6.8-runtime
Flatpak: Upgrade kde runtime to 6.8
2024-11-12 00:30:46 -05:00
Carles Pastor
fe96bf4108 Flatpak: Upgrade kde runtime to 6.8
this version bundles SDL2-2.30.6, the temporary measure of building the
vendored version from exports is no longer necessary.
2024-11-10 12:06:06 +01:00
JosJuice
980a726313 JitArm64: Drop GetCRFieldBit's negate parameter
No caller is using it anymore.
2024-11-02 10:51:48 +01:00
JosJuice
71e9766519 JitArm64: Use BIC/EON/ORN in crXXX
This lets us save an instruction in certain scenarios.
2024-11-02 10:51:38 +01:00
JosJuice
9246bcad55 JitArm64: Add negate parameter to SetCRFieldBit
Unlike on x64, inverting EQ or GT in SetCRFieldBit saves us one
instruction. Also unlike on x64, inverting SO or LT in GetCRFieldBit
requires an extra instruction (just like in SetCRFieldBit). Due to this,
replacing an invert in GetCRFieldBit with an invert in SetCRFieldBit
when possible is either equally good or better - never worse.
2024-11-02 10:48:27 +01:00
JosJuice
7fddd39d97 JitArm64: Port some crXXX optimizations from Jit64 2024-11-02 10:24:02 +01:00
JosJuice
bbe271eec6 JitArm64: Refactor CR bit manipulation code
This brings JitArm64 more in line with Jit64, and makes the next commit
easier to implement. No functional change.
2024-11-02 10:19:31 +01:00
4 changed files with 198 additions and 182 deletions

View File

@ -1,22 +0,0 @@
{
"name": "SDL2",
"buildsystem": "autotools",
"config-opts": ["--disable-static"],
"sources": [
{
"type": "dir",
"path": "../../Externals/SDL/SDL"
}
],
"cleanup": [ "/bin/sdl2-config",
"/include",
"/lib/libSDL2.la",
"/lib/libSDL2main.a",
"/lib/libSDL2main.la",
"/lib/libSDL2_test.a",
"/lib/libSDL2_test.la",
"/lib/cmake",
"/share/aclocal",
"/lib/pkgconfig"]
}

View File

@ -1,6 +1,6 @@
app-id: org.DolphinEmu.dolphin-emu
runtime: org.kde.Platform
runtime-version: '6.7'
runtime-version: '6.8'
sdk: org.kde.Sdk
command: dolphin-emu-wrapper
rename-desktop-file: dolphin-emu.desktop
@ -47,9 +47,6 @@ modules:
url: https://github.com/Unrud/xdg-screensaver-shim/archive/0.0.2.tar.gz
sha256: 0ed2a69fe6ee6cbffd2fe16f85116db737f17fb1e79bfb812d893cf15c728399
# build the vendored SDL2 from Externals until the runtime gets 2.30.6
- SDL2/SDL2.json
- name: dolphin-emu
buildsystem: cmake-ninja
config-opts:

View File

@ -355,8 +355,13 @@ protected:
Arm64Gen::ARM64Reg exit_address_after_return_reg = Arm64Gen::ARM64Reg::INVALID_REG);
void WriteBLRExit(Arm64Gen::ARM64Reg dest);
Arm64Gen::FixupBranch JumpIfCRFieldBit(int field, int bit, bool jump_if_set);
void GetCRFieldBit(int field, int bit, Arm64Gen::ARM64Reg out);
void SetCRFieldBit(int field, int bit, Arm64Gen::ARM64Reg in, bool negate = false);
void ClearCRFieldBit(int field, int bit);
void SetCRFieldBit(int field, int bit);
void FixGTBeforeSettingCRFieldBit(Arm64Gen::ARM64Reg reg);
Arm64Gen::FixupBranch JumpIfCRFieldBit(int field, int bit, bool jump_if_set);
void UpdateFPExceptionSummary(Arm64Gen::ARM64Reg fpscr);
void UpdateRoundingMode();

View File

@ -20,6 +20,144 @@
using namespace Arm64Gen;
void JitArm64::GetCRFieldBit(int field, int bit, ARM64Reg out)
{
ARM64Reg CR = gpr.CR(field);
ARM64Reg WCR = EncodeRegTo32(CR);
switch (bit)
{
case PowerPC::CR_SO_BIT: // check bit 59 set
UBFX(out, CR, PowerPC::CR_EMU_SO_BIT, 1);
break;
case PowerPC::CR_EQ_BIT: // check bits 31-0 == 0
CMP(WCR, ARM64Reg::WZR);
CSET(out, CC_EQ);
break;
case PowerPC::CR_GT_BIT: // check val > 0
CMP(CR, ARM64Reg::ZR);
CSET(out, CC_GT);
break;
case PowerPC::CR_LT_BIT: // check bit 62 set
UBFX(out, CR, PowerPC::CR_EMU_LT_BIT, 1);
break;
default:
ASSERT_MSG(DYNA_REC, false, "Invalid CR bit");
}
}
void JitArm64::SetCRFieldBit(int field, int bit, ARM64Reg in, bool negate)
{
gpr.BindCRToRegister(field, true);
ARM64Reg CR = gpr.CR(field);
if (bit != PowerPC::CR_GT_BIT)
FixGTBeforeSettingCRFieldBit(CR);
switch (bit)
{
case PowerPC::CR_SO_BIT: // set bit 59 to input
BFI(CR, in, PowerPC::CR_EMU_SO_BIT, 1);
if (negate)
EOR(CR, CR, LogicalImm(1ULL << PowerPC::CR_EMU_SO_BIT, GPRSize::B64));
break;
case PowerPC::CR_EQ_BIT: // clear low 32 bits, set bit 0 to !input
AND(CR, CR, LogicalImm(0xFFFF'FFFF'0000'0000, GPRSize::B64));
ORR(CR, CR, in);
if (!negate)
EOR(CR, CR, LogicalImm(1ULL << 0, GPRSize::B64));
break;
case PowerPC::CR_GT_BIT: // set bit 63 to !input
BFI(CR, in, 63, 1);
if (!negate)
EOR(CR, CR, LogicalImm(1ULL << 63, GPRSize::B64));
break;
case PowerPC::CR_LT_BIT: // set bit 62 to input
BFI(CR, in, PowerPC::CR_EMU_LT_BIT, 1);
if (negate)
EOR(CR, CR, LogicalImm(1ULL << PowerPC::CR_EMU_LT_BIT, GPRSize::B64));
break;
}
ORR(CR, CR, LogicalImm(1ULL << 32, GPRSize::B64));
}
void JitArm64::ClearCRFieldBit(int field, int bit)
{
gpr.BindCRToRegister(field, true);
ARM64Reg XA = gpr.CR(field);
switch (bit)
{
case PowerPC::CR_SO_BIT:
AND(XA, XA, LogicalImm(~(u64(1) << PowerPC::CR_EMU_SO_BIT), GPRSize::B64));
break;
case PowerPC::CR_EQ_BIT:
FixGTBeforeSettingCRFieldBit(XA);
ORR(XA, XA, LogicalImm(1, GPRSize::B64));
break;
case PowerPC::CR_GT_BIT:
ORR(XA, XA, LogicalImm(u64(1) << 63, GPRSize::B64));
break;
case PowerPC::CR_LT_BIT:
AND(XA, XA, LogicalImm(~(u64(1) << PowerPC::CR_EMU_LT_BIT), GPRSize::B64));
break;
}
}
void JitArm64::SetCRFieldBit(int field, int bit)
{
gpr.BindCRToRegister(field, true);
ARM64Reg XA = gpr.CR(field);
if (bit != PowerPC::CR_GT_BIT)
FixGTBeforeSettingCRFieldBit(XA);
switch (bit)
{
case PowerPC::CR_SO_BIT:
ORR(XA, XA, LogicalImm(u64(1) << PowerPC::CR_EMU_SO_BIT, GPRSize::B64));
break;
case PowerPC::CR_EQ_BIT:
AND(XA, XA, LogicalImm(0xFFFF'FFFF'0000'0000, GPRSize::B64));
break;
case PowerPC::CR_GT_BIT:
AND(XA, XA, LogicalImm(~(u64(1) << 63), GPRSize::B64));
break;
case PowerPC::CR_LT_BIT:
ORR(XA, XA, LogicalImm(u64(1) << PowerPC::CR_EMU_LT_BIT, GPRSize::B64));
break;
}
ORR(XA, XA, LogicalImm(u64(1) << 32, GPRSize::B64));
}
void JitArm64::FixGTBeforeSettingCRFieldBit(ARM64Reg reg)
{
// GT is considered unset if the internal representation is <= 0, or in other words,
// if the internal representation either has bit 63 set or has all bits set to zero.
// If all bits are zero and we set some bit that's unrelated to GT, we need to set bit 63 so GT
// doesn't accidentally become considered set. Gross but necessary; this can break actual games.
auto WA = gpr.GetScopedReg();
ARM64Reg XA = EncodeRegTo64(WA);
ORR(XA, reg, LogicalImm(1ULL << 63, GPRSize::B64));
CMP(reg, ARM64Reg::ZR);
CSEL(reg, reg, XA, CC_NEQ);
}
FixupBranch JitArm64::JumpIfCRFieldBit(int field, int bit, bool jump_if_set)
{
ARM64Reg XA = gpr.CR(field);
@ -42,19 +180,6 @@ FixupBranch JitArm64::JumpIfCRFieldBit(int field, int bit, bool jump_if_set)
}
}
void JitArm64::FixGTBeforeSettingCRFieldBit(Arm64Gen::ARM64Reg reg)
{
// GT is considered unset if the internal representation is <= 0, or in other words,
// if the internal representation either has bit 63 set or has all bits set to zero.
// If all bits are zero and we set some bit that's unrelated to GT, we need to set bit 63 so GT
// doesn't accidentally become considered set. Gross but necessary; this can break actual games.
auto WA = gpr.GetScopedReg();
ARM64Reg XA = EncodeRegTo64(WA);
ORR(XA, reg, LogicalImm(1ULL << 63, GPRSize::B64));
CMP(reg, ARM64Reg::ZR);
CSEL(reg, reg, XA, CC_NEQ);
}
void JitArm64::UpdateFPExceptionSummary(ARM64Reg fpscr)
{
auto WA = gpr.GetScopedReg();
@ -468,72 +593,47 @@ void JitArm64::crXXX(UGeckoInstruction inst)
INSTRUCTION_START
JITDISABLE(bJITSystemRegistersOff);
// Special case: crclr
if (inst.CRBA == inst.CRBB && inst.CRBA == inst.CRBD && inst.SUBOP10 == 193)
if (inst.CRBA == inst.CRBB)
{
// Clear CR field bit
int field = inst.CRBD >> 2;
int bit = 3 - (inst.CRBD & 3);
gpr.BindCRToRegister(field, true);
ARM64Reg XA = gpr.CR(field);
switch (bit)
switch (inst.SUBOP10)
{
case PowerPC::CR_SO_BIT:
AND(XA, XA, LogicalImm(~(u64(1) << PowerPC::CR_EMU_SO_BIT), GPRSize::B64));
break;
case PowerPC::CR_EQ_BIT:
FixGTBeforeSettingCRFieldBit(XA);
ORR(XA, XA, LogicalImm(1, GPRSize::B64));
break;
case PowerPC::CR_GT_BIT:
ORR(XA, XA, LogicalImm(u64(1) << 63, GPRSize::B64));
break;
case PowerPC::CR_LT_BIT:
AND(XA, XA, LogicalImm(~(u64(1) << PowerPC::CR_EMU_LT_BIT), GPRSize::B64));
break;
// crclr
case 129: // crandc: A && ~B => 0
case 193: // crxor: A ^ B => 0
{
ClearCRFieldBit(inst.CRBD >> 2, 3 - (inst.CRBD & 3));
return;
}
// crset
case 289: // creqv: ~(A ^ B) => 1
case 417: // crorc: A || ~B => 1
{
SetCRFieldBit(inst.CRBD >> 2, 3 - (inst.CRBD & 3));
return;
}
case 257: // crand: A && B => A
case 449: // cror: A || B => A
{
auto WA = gpr.GetScopedReg();
ARM64Reg XA = EncodeRegTo64(WA);
GetCRFieldBit(inst.CRBA >> 2, 3 - (inst.CRBA & 3), XA);
SetCRFieldBit(inst.CRBD >> 2, 3 - (inst.CRBD & 3), XA, false);
return;
}
case 33: // crnor: ~(A || B) => ~A
case 225: // crnand: ~(A && B) => ~A
{
auto WA = gpr.GetScopedReg();
ARM64Reg XA = EncodeRegTo64(WA);
GetCRFieldBit(inst.CRBA >> 2, 3 - (inst.CRBA & 3), XA);
SetCRFieldBit(inst.CRBD >> 2, 3 - (inst.CRBD & 3), XA, true);
return;
}
}
return;
}
// Special case: crset
if (inst.CRBA == inst.CRBB && inst.CRBA == inst.CRBD && inst.SUBOP10 == 289)
{
// SetCRFieldBit
int field = inst.CRBD >> 2;
int bit = 3 - (inst.CRBD & 3);
gpr.BindCRToRegister(field, true);
ARM64Reg XA = gpr.CR(field);
if (bit != PowerPC::CR_GT_BIT)
FixGTBeforeSettingCRFieldBit(XA);
switch (bit)
{
case PowerPC::CR_SO_BIT:
ORR(XA, XA, LogicalImm(u64(1) << PowerPC::CR_EMU_SO_BIT, GPRSize::B64));
break;
case PowerPC::CR_EQ_BIT:
AND(XA, XA, LogicalImm(0xFFFF'FFFF'0000'0000, GPRSize::B64));
break;
case PowerPC::CR_GT_BIT:
AND(XA, XA, LogicalImm(~(u64(1) << 63), GPRSize::B64));
break;
case PowerPC::CR_LT_BIT:
ORR(XA, XA, LogicalImm(u64(1) << PowerPC::CR_EMU_LT_BIT, GPRSize::B64));
break;
}
ORR(XA, XA, LogicalImm(u64(1) << 32, GPRSize::B64));
return;
}
// crnor or crnand
const bool negate_result = inst.SUBOP10 == 33 || inst.SUBOP10 == 225;
auto WA = gpr.GetScopedReg();
ARM64Reg XA = EncodeRegTo64(WA);
@ -541,106 +641,42 @@ void JitArm64::crXXX(UGeckoInstruction inst)
auto WB = gpr.GetScopedReg();
ARM64Reg XB = EncodeRegTo64(WB);
// creqv or crnand or crnor
bool negateA = inst.SUBOP10 == 289 || inst.SUBOP10 == 225 || inst.SUBOP10 == 33;
// crandc or crorc or crnand or crnor
bool negateB =
inst.SUBOP10 == 129 || inst.SUBOP10 == 417 || inst.SUBOP10 == 225 || inst.SUBOP10 == 33;
// GetCRFieldBit
for (int i = 0; i < 2; i++)
{
int field = i ? inst.CRBB >> 2 : inst.CRBA >> 2;
int bit = i ? 3 - (inst.CRBB & 3) : 3 - (inst.CRBA & 3);
ARM64Reg out = i ? XB : XA;
bool negate = i ? negateB : negateA;
ARM64Reg XC = gpr.CR(field);
ARM64Reg WC = EncodeRegTo32(XC);
switch (bit)
{
case PowerPC::CR_SO_BIT: // check bit 59 set
UBFX(out, XC, PowerPC::CR_EMU_SO_BIT, 1);
if (negate)
EOR(out, out, LogicalImm(1, GPRSize::B64));
break;
case PowerPC::CR_EQ_BIT: // check bits 31-0 == 0
CMP(WC, ARM64Reg::WZR);
CSET(out, negate ? CC_NEQ : CC_EQ);
break;
case PowerPC::CR_GT_BIT: // check val > 0
CMP(XC, ARM64Reg::ZR);
CSET(out, negate ? CC_LE : CC_GT);
break;
case PowerPC::CR_LT_BIT: // check bit 62 set
UBFX(out, XC, PowerPC::CR_EMU_LT_BIT, 1);
if (negate)
EOR(out, out, LogicalImm(1, GPRSize::B64));
break;
default:
ASSERT_MSG(DYNA_REC, false, "Invalid CR bit");
}
}
GetCRFieldBit(inst.CRBA >> 2, 3 - (inst.CRBA & 3), XA);
GetCRFieldBit(inst.CRBB >> 2, 3 - (inst.CRBB & 3), XB);
// Compute combined bit
switch (inst.SUBOP10)
{
case 33: // crnor: ~(A || B) == (~A && ~B)
case 129: // crandc: A && ~B
case 225: // crnand: ~(A && B)
case 257: // crand: A && B
AND(XA, XA, XB);
break;
case 129: // crandc: A && ~B
BIC(XA, XA, XB);
break;
case 193: // crxor: A ^ B
case 289: // creqv: ~(A ^ B) = ~A ^ B
EOR(XA, XA, XB);
break;
case 225: // crnand: ~(A && B) == (~A || ~B)
case 417: // crorc: A || ~B
case 289: // creqv: ~(A ^ B) = A ^ ~B
EON(XA, XA, XB);
break;
case 33: // crnor: ~(A || B)
case 449: // cror: A || B
ORR(XA, XA, XB);
break;
case 417: // crorc: A || ~B
ORN(XA, XA, XB);
break;
}
}
// Store result bit in CRBD
int field = inst.CRBD >> 2;
int bit = 3 - (inst.CRBD & 3);
gpr.BindCRToRegister(field, true);
ARM64Reg CR = gpr.CR(field);
if (bit != PowerPC::CR_GT_BIT)
FixGTBeforeSettingCRFieldBit(CR);
switch (bit)
{
case PowerPC::CR_SO_BIT: // set bit 59 to input
BFI(CR, XA, PowerPC::CR_EMU_SO_BIT, 1);
break;
case PowerPC::CR_EQ_BIT: // clear low 32 bits, set bit 0 to !input
AND(CR, CR, LogicalImm(0xFFFF'FFFF'0000'0000, GPRSize::B64));
EOR(XA, XA, LogicalImm(1, GPRSize::B64));
ORR(CR, CR, XA);
break;
case PowerPC::CR_GT_BIT: // set bit 63 to !input
EOR(XA, XA, LogicalImm(1, GPRSize::B64));
BFI(CR, XA, 63, 1);
break;
case PowerPC::CR_LT_BIT: // set bit 62 to input
BFI(CR, XA, PowerPC::CR_EMU_LT_BIT, 1);
break;
}
ORR(CR, CR, LogicalImm(1ULL << 32, GPRSize::B64));
SetCRFieldBit(inst.CRBD >> 2, 3 - (inst.CRBD & 3), XA, negate_result);
}
void JitArm64::mfcr(UGeckoInstruction inst)