Merge pull request #9373 from MerryMage/arm64-rlwimix

JitArm64_Integer: Add optimizations for rlwimix
This commit is contained in:
Markus Wick 2021-01-31 16:50:26 +01:00 committed by GitHub
commit 27b7e5891d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 38 additions and 17 deletions

View file

@ -1658,18 +1658,26 @@ void ARM64XEmitter::UBFM(ARM64Reg Rd, ARM64Reg Rn, u32 immr, u32 imms)
void ARM64XEmitter::BFI(ARM64Reg Rd, ARM64Reg Rn, u32 lsb, u32 width)
{
u32 size = Is64Bit(Rn) ? 64 : 32;
ASSERT_MSG(DYNA_REC, (lsb + width) <= size,
ASSERT_MSG(DYNA_REC, lsb < size && width >= 1 && width <= size - lsb,
"%s passed lsb %d and width %d which is greater than the register size!", __func__,
lsb, width);
EncodeBitfieldMOVInst(1, Rd, Rn, (size - lsb) % size, width - 1);
BFM(Rd, Rn, (size - lsb) % size, width - 1);
}
void ARM64XEmitter::BFXIL(ARM64Reg Rd, ARM64Reg Rn, u32 lsb, u32 width)
{
u32 size = Is64Bit(Rn) ? 64 : 32;
ASSERT_MSG(DYNA_REC, lsb < size && width >= 1 && width <= size - lsb,
"%s passed lsb %d and width %d which is greater than the register size!", __func__,
lsb, width);
BFM(Rd, Rn, lsb, lsb + width - 1);
}
void ARM64XEmitter::UBFIZ(ARM64Reg Rd, ARM64Reg Rn, u32 lsb, u32 width)
{
u32 size = Is64Bit(Rn) ? 64 : 32;
ASSERT_MSG(DYNA_REC, (lsb + width) <= size,
ASSERT_MSG(DYNA_REC, lsb < size && width >= 1 && width <= size - lsb,
"%s passed lsb %d and width %d which is greater than the register size!", __func__,
lsb, width);
EncodeBitfieldMOVInst(2, Rd, Rn, (size - lsb) % size, width - 1);
UBFM(Rd, Rn, (size - lsb) % size, width - 1);
}
void ARM64XEmitter::EXTR(ARM64Reg Rd, ARM64Reg Rn, ARM64Reg Rm, u32 shift)
{

View file

@ -770,6 +770,7 @@ public:
void SBFM(ARM64Reg Rd, ARM64Reg Rn, u32 immr, u32 imms);
void UBFM(ARM64Reg Rd, ARM64Reg Rn, u32 immr, u32 imms);
void BFI(ARM64Reg Rd, ARM64Reg Rn, u32 lsb, u32 width);
void BFXIL(ARM64Reg Rd, ARM64Reg Rn, u32 lsb, u32 width);
void UBFIZ(ARM64Reg Rd, ARM64Reg Rn, u32 lsb, u32 width);
// Extract register (ROR with two inputs, if same then faster on A67)

View file

@ -1443,6 +1443,10 @@ void JitArm64::rlwimix(UGeckoInstruction inst)
const int a = inst.RA, s = inst.RS;
const u32 mask = MakeRotationMask(inst.MB, inst.ME);
const u32 lsb = 31 - inst.ME;
const u32 width = inst.ME - inst.MB + 1;
const u32 rot_dist = inst.SH ? 32 - inst.SH : 0;
if (gpr.IsImm(a) && gpr.IsImm(s))
{
u32 res = (gpr.GetImm(a) & ~mask) | (Common::RotateLeft(gpr.GetImm(s), inst.SH) & mask);
@ -1462,17 +1466,22 @@ void JitArm64::rlwimix(UGeckoInstruction inst)
gpr.BindToRegister(a, a == s);
if (inst.SH)
ROR(gpr.R(a), gpr.R(s), 32 - inst.SH);
ROR(gpr.R(a), gpr.R(s), rot_dist);
else if (a != s)
MOV(gpr.R(a), gpr.R(s));
}
else if (lsb == 0 && inst.MB <= inst.ME && rot_dist + width <= 32)
{
// Destination is in least significant position
// No mask inversion
// Source field pre-rotation is contiguous
gpr.BindToRegister(a, true);
BFXIL(gpr.R(a), gpr.R(s), rot_dist, width);
}
else if (inst.SH == 0 && inst.MB <= inst.ME)
{
// No rotation
// No mask inversion
u32 lsb = 31 - inst.ME;
u32 width = inst.ME - inst.MB + 1;
gpr.BindToRegister(a, true);
ARM64Reg WA = gpr.GetReg();
UBFX(WA, gpr.R(s), lsb, width);
@ -1482,15 +1491,18 @@ void JitArm64::rlwimix(UGeckoInstruction inst)
else if (inst.SH && inst.MB <= inst.ME)
{
// No mask inversion
u32 lsb = 31 - inst.ME;
u32 width = inst.ME - inst.MB + 1;
gpr.BindToRegister(a, true);
ARM64Reg WA = gpr.GetReg();
ROR(WA, gpr.R(s), 32 - inst.SH);
UBFX(WA, WA, lsb, width);
BFI(gpr.R(a), WA, lsb, width);
gpr.Unlock(WA);
if ((rot_dist + lsb) % 32 == 0)
{
BFI(gpr.R(a), gpr.R(s), lsb, width);
}
else
{
ARM64Reg WA = gpr.GetReg();
ROR(WA, gpr.R(s), (rot_dist + lsb) % 32);
BFI(gpr.R(a), WA, lsb, width);
gpr.Unlock(WA);
}
}
else
{
@ -1500,7 +1512,7 @@ void JitArm64::rlwimix(UGeckoInstruction inst)
MOVI2R(WA, mask);
BIC(WB, gpr.R(a), WA);
AND(WA, WA, gpr.R(s), ArithOption(gpr.R(s), ShiftType::ROR, 32 - inst.SH));
AND(WA, WA, gpr.R(s), ArithOption(gpr.R(s), ShiftType::ROR, rot_dist));
ORR(gpr.R(a), WB, WA);
gpr.Unlock(WA, WB);