From 81bc8107cf57fb3945d3b1f100fb35661642c945 Mon Sep 17 00:00:00 2001 From: "Unknown W. Brackets" Date: Sat, 27 Jun 2015 22:32:17 -0700 Subject: [PATCH 1/4] arm64: Use UBFX, not LSR, for slti sign check. This is about 22% faster on the A57 (for just this one instruction, so not a huge impact overall.) Makes sense that it would be since not arith. --- Core/MIPS/ARM64/Arm64CompALU.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Core/MIPS/ARM64/Arm64CompALU.cpp b/Core/MIPS/ARM64/Arm64CompALU.cpp index bbebd2f141..261c33a7ef 100644 --- a/Core/MIPS/ARM64/Arm64CompALU.cpp +++ b/Core/MIPS/ARM64/Arm64CompALU.cpp @@ -100,8 +100,8 @@ void Arm64Jit::Comp_IType(MIPSOpcode op) { break; } else if (simm == 0) { gpr.MapDirtyIn(rt, rs); - // Shift to get the sign bit only (for < 0.) - LSR(gpr.R(rt), gpr.R(rs), 31); + // Grab the sign bit (< 0) as 1/0. Slightly faster than a shift. + UBFX(gpr.R(rt), gpr.R(rs), 31, 1); break; } gpr.MapDirtyIn(rt, rs); From 223e55a453e9556d4db9742eb06149951ef547bc Mon Sep 17 00:00:00 2001 From: "Unknown W. Brackets" Date: Sat, 27 Jun 2015 22:33:15 -0700 Subject: [PATCH 2/4] arm64: Undisable clz/clo, they work. Also, avoid a temp in clo. It's the tiniest bit faster on A57, though we'll see how it works out elsewhere. A bit clearer without the temp imho. --- Core/MIPS/ARM64/Arm64CompALU.cpp | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/Core/MIPS/ARM64/Arm64CompALU.cpp b/Core/MIPS/ARM64/Arm64CompALU.cpp index 261c33a7ef..e1b88632fe 100644 --- a/Core/MIPS/ARM64/Arm64CompALU.cpp +++ b/Core/MIPS/ARM64/Arm64CompALU.cpp @@ -138,8 +138,6 @@ void Arm64Jit::Comp_IType(MIPSOpcode op) { void Arm64Jit::Comp_RType2(MIPSOpcode op) { CONDITIONAL_DISABLE; - DISABLE; - MIPSGPReg rs = _RS; MIPSGPReg rd = _RD; @@ -176,8 +174,8 @@ void Arm64Jit::Comp_RType2(MIPSOpcode op) { break; } gpr.MapDirtyIn(rd, rs); - MVN(SCRATCH1, gpr.R(rs)); - CLZ(gpr.R(rd), SCRATCH1); + MVN(gpr.R(rd), gpr.R(rs)); + CLZ(gpr.R(rd), gpr.R(rd)); break; default: DISABLE; From 11a851a1393f81b19dcf3d1a685d384c73129244 Mon Sep 17 00:00:00 2001 From: "Unknown W. Brackets" Date: Sat, 27 Jun 2015 23:05:51 -0700 Subject: [PATCH 3/4] arm64: Enable movz/movn. --- Core/MIPS/ARM64/Arm64CompALU.cpp | 2 -- 1 file changed, 2 deletions(-) diff --git a/Core/MIPS/ARM64/Arm64CompALU.cpp b/Core/MIPS/ARM64/Arm64CompALU.cpp index e1b88632fe..6a2cb204e1 100644 --- a/Core/MIPS/ARM64/Arm64CompALU.cpp +++ b/Core/MIPS/ARM64/Arm64CompALU.cpp @@ -223,13 +223,11 @@ void Arm64Jit::Comp_RType3(MIPSOpcode op) { switch (op & 63) { case 10: //if (!R(rt)) R(rd) = R(rs); break; //movz - DISABLE; gpr.MapDirtyInIn(rd, rt, rs); CMP(gpr.R(rt), 0); CSEL(gpr.R(rd), gpr.R(rs), gpr.R(rd), CC_EQ); break; case 11:// if (R(rt)) R(rd) = R(rs); break; //movn - DISABLE; gpr.MapDirtyInIn(rd, rt, rs); CMP(gpr.R(rt), 0); CSEL(gpr.R(rd), gpr.R(rs), gpr.R(rd), CC_NEQ); From daddb73f228bf051f0ba4a1c7a568f7172f59ff7 Mon Sep 17 00:00:00 2001 From: "Unknown W. Brackets" Date: Sat, 27 Jun 2015 23:54:43 -0700 Subject: [PATCH 4/4] arm64: Implement nor. --- Core/MIPS/ARM64/Arm64CompALU.cpp | 28 +++++++++++++++++++++++++++- 1 file changed, 27 insertions(+), 1 deletion(-) diff --git a/Core/MIPS/ARM64/Arm64CompALU.cpp b/Core/MIPS/ARM64/Arm64CompALU.cpp index 6a2cb204e1..45c9dd9e9f 100644 --- a/Core/MIPS/ARM64/Arm64CompALU.cpp +++ b/Core/MIPS/ARM64/Arm64CompALU.cpp @@ -264,7 +264,33 @@ void Arm64Jit::Comp_RType3(MIPSOpcode op) { break; case 39: // R(rd) = ~(R(rs) | R(rt)); break; //nor - DISABLE; + if (gpr.IsImm(rs) && gpr.IsImm(rt)) { + gpr.SetImm(rd, ~(gpr.GetImm(rs) | gpr.GetImm(rt))); + } else if (gpr.IsImm(rs) || gpr.IsImm(rt)) { + MIPSGPReg lhs = gpr.IsImm(rs) ? rt : rs; + MIPSGPReg rhs = gpr.IsImm(rs) ? rs : rt; + u32 rhsImm = gpr.GetImm(rhs); + if (rhsImm == 0) { + gpr.MapDirtyIn(rd, lhs); + MVN(gpr.R(rd), gpr.R(lhs)); + } else { + // Ignored, just for IsImmLogical. + unsigned int n, imm_s, imm_r; + if (IsImmLogical(rhsImm, 32, &n, &imm_s, &imm_r)) { + // Great, we can avoid flushing a reg. + gpr.MapDirtyIn(rd, lhs); + ORRI2R(gpr.R(rd), gpr.R(lhs), rhsImm); + } else { + gpr.MapDirtyInIn(rd, rs, rt); + ORR(gpr.R(rd), gpr.R(rs), gpr.R(rt)); + } + MVN(gpr.R(rd), gpr.R(rd)); + } + } else { + gpr.MapDirtyInIn(rd, rs, rt); + ORR(gpr.R(rd), gpr.R(rs), gpr.R(rt)); + MVN(gpr.R(rd), gpr.R(rd)); + } break; case 42: //R(rd) = (int)R(rs) < (int)R(rt); break; //slt