Hello_User/XenonRecomp
1
0
Fork 0
mirror of https://github.com/hedge-dev/XenonRecomp.git synced 2025-11-04 06:47:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Handling float rounding.

Browse Source
This commit is contained in:
Skyth
2024-09-22 00:07:15 +03:00
parent 4d6eac1c24
commit 1bfeb0a7c4
3 changed files with 85 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

132
PowerRecomp/recompiler.cpp
View File

@@ -407,147 +407,147 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_FABS: case PPC_INST_FABS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = fabs(ctx.f{}.f64);", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = fabs(ctx.f{}.f64);", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FADD: case PPC_INST_FADD:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 + ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = ctx.f{}.f64 + ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FADDS: case PPC_INST_FADDS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64 + ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = float(ctx.f{}.f64 + ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FCFID: case PPC_INST_FCFID:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = double(ctx.f{}.s64);", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = double(ctx.f{}.s64);", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FCMPU: case PPC_INST_FCMPU:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.cr{}.compare(ctx.f{}.f64, ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.cr{}.compare(ctx.f{}.f64, ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FCTID: case PPC_INST_FCTID:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.s64 = int64_t(round(ctx.f{}.f64));", insn.operands[0], insn.operands[1]); println("\tctx.f{}.s64 = _mm_cvtsd_si64(_mm_load1_pd(&ctx.f{}.f64));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FCTIDZ: case PPC_INST_FCTIDZ:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.s64 = int64_t(trunc(ctx.f{}.f64));", insn.operands[0], insn.operands[1]); println("\tctx.f{}.s64 = _mm_cvttsd_si64(_mm_load1_pd(&ctx.f{}.f64));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FCTIWZ: case PPC_INST_FCTIWZ:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.s64 = int32_t(trunc(ctx.f{}.f64));", insn.operands[0], insn.operands[1]); println("\tctx.f{}.s64 = _mm_cvttsd_si32(_mm_load1_pd(&ctx.f{}.f64));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FDIV: case PPC_INST_FDIV:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 / ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = ctx.f{}.f64 / ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FDIVS: case PPC_INST_FDIVS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64 / ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = float(ctx.f{}.f64 / ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FMADD: case PPC_INST_FMADD:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 * ctx.f{}.f64 + ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = ctx.f{}.f64 * ctx.f{}.f64 + ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FMADDS: case PPC_INST_FMADDS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64 * ctx.f{}.f64 + ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = float(ctx.f{}.f64 * ctx.f{}.f64 + ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FMR: case PPC_INST_FMR:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64;", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = ctx.f{}.f64;", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FMSUB: case PPC_INST_FMSUB:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FMSUBS: case PPC_INST_FMSUBS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = float(ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FMUL: case PPC_INST_FMUL:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 * ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = ctx.f{}.f64 * ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FMULS: case PPC_INST_FMULS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64 * ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = float(ctx.f{}.f64 * ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FNABS: case PPC_INST_FNABS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = -fabs(ctx.f{}.f64);", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = -fabs(ctx.f{}.f64);", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FNEG: case PPC_INST_FNEG:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = -ctx.f{}.f64;", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = -ctx.f{}.f64;", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FNMADDS: case PPC_INST_FNMADDS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(-(ctx.f{}.f64 * ctx.f{}.f64 + ctx.f{}.f64));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = float(-(ctx.f{}.f64 * ctx.f{}.f64 + ctx.f{}.f64));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FNMSUB: case PPC_INST_FNMSUB:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = -(ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = -(ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FNMSUBS: case PPC_INST_FNMSUBS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(-(ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = float(-(ctx.f{}.f64 * ctx.f{}.f64 - ctx.f{}.f64));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FRES: case PPC_INST_FRES:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = 1.0f / float(ctx.f{}.f64);", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = 1.0f / float(ctx.f{}.f64);", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FRSP: case PPC_INST_FRSP:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64);", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = float(ctx.f{}.f64);", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FSEL: case PPC_INST_FSEL:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 >= 0.0 ? ctx.f{}.f64 : ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\tctx.f{}.f64 = ctx.f{}.f64 >= 0.0 ? ctx.f{}.f64 : ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_FSQRT: case PPC_INST_FSQRT:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = sqrt(ctx.f{}.f64);", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = sqrt(ctx.f{}.f64);", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FSQRTS: case PPC_INST_FSQRTS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(sqrt(ctx.f{}.f64));", insn.operands[0], insn.operands[1]); println("\tctx.f{}.f64 = float(sqrt(ctx.f{}.f64));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_FSUB: case PPC_INST_FSUB:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = ctx.f{}.f64 - ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = ctx.f{}.f64 - ctx.f{}.f64;", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_FSUBS: case PPC_INST_FSUBS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\tctx.f{}.f64 = float(ctx.f{}.f64 - ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]); println("\tctx.f{}.f64 = float(ctx.f{}.f64 - ctx.f{}.f64);", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
@@ -600,7 +600,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_LFD: case PPC_INST_LFD:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\tctx.f{}.u64 = PPC_LOAD_U64(", insn.operands[0]); print("\tctx.f{}.u64 = PPC_LOAD_U64(", insn.operands[0]);
if (insn.operands[2] != 0) if (insn.operands[2] != 0)
print("ctx.r{}.u32 + ", insn.operands[2]); print("ctx.r{}.u32 + ", insn.operands[2]);
@@ -608,7 +608,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_LFDX: case PPC_INST_LFDX:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\tctx.f{}.u64 = PPC_LOAD_U64(", insn.operands[0]); print("\tctx.f{}.u64 = PPC_LOAD_U64(", insn.operands[0]);
if (insn.operands[1] != 0) if (insn.operands[1] != 0)
print("ctx.r{}.u32 + ", insn.operands[1]); print("ctx.r{}.u32 + ", insn.operands[1]);
@@ -616,7 +616,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_LFS: case PPC_INST_LFS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\ttemp.u32 = PPC_LOAD_U32("); print("\ttemp.u32 = PPC_LOAD_U32(");
if (insn.operands[2] != 0) if (insn.operands[2] != 0)
print("ctx.r{}.u32 + ", insn.operands[2]); print("ctx.r{}.u32 + ", insn.operands[2]);
@@ -625,7 +625,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_LFSX: case PPC_INST_LFSX:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\ttemp.u32 = PPC_LOAD_U32("); print("\ttemp.u32 = PPC_LOAD_U32(");
if (insn.operands[1] != 0) if (insn.operands[1] != 0)
print("ctx.r{}.u32 + ", insn.operands[1]); print("ctx.r{}.u32 + ", insn.operands[1]);
@@ -777,7 +777,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_MFFS: case PPC_INST_MFFS:
println("\tctx.f{}.u64 = ctx.fpscr;", insn.operands[0]); println("\tctx.f{}.u64 = ctx.fpscr.loadFromHost();", insn.operands[0]);
break; break;
case PPC_INST_MFLR: case PPC_INST_MFLR:
@@ -816,7 +816,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_MTFSF: case PPC_INST_MTFSF:
println("\tctx.fpscr = ctx.f{}.u32;", insn.operands[1]); println("\tctx.fpscr.storeFromGuest(ctx.f{}.u32);", insn.operands[1]);
break; break;
case PPC_INST_MTLR: case PPC_INST_MTLR:
@@ -1041,7 +1041,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_STFD: case PPC_INST_STFD:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\tPPC_STORE_U64("); print("\tPPC_STORE_U64(");
if (insn.operands[2] != 0) if (insn.operands[2] != 0)
print("ctx.r{}.u32 + ", insn.operands[2]); print("ctx.r{}.u32 + ", insn.operands[2]);
@@ -1049,7 +1049,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_STFDX: case PPC_INST_STFDX:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\tPPC_STORE_U64("); print("\tPPC_STORE_U64(");
if (insn.operands[1] != 0) if (insn.operands[1] != 0)
print("ctx.r{}.u32 + ", insn.operands[1]); print("ctx.r{}.u32 + ", insn.operands[1]);
@@ -1057,7 +1057,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_STFIWX: case PPC_INST_STFIWX:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
print("\tPPC_STORE_U32("); print("\tPPC_STORE_U32(");
if (insn.operands[1] != 0) if (insn.operands[1] != 0)
print("ctx.r{}.u32 + ", insn.operands[1]); print("ctx.r{}.u32 + ", insn.operands[1]);
@@ -1065,7 +1065,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_STFS: case PPC_INST_STFS:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\ttemp.f32 = ctx.f{}.f64;", insn.operands[0]); println("\ttemp.f32 = ctx.f{}.f64;", insn.operands[0]);
print("\tPPC_STORE_U32("); print("\tPPC_STORE_U32(");
if (insn.operands[2] != 0) if (insn.operands[2] != 0)
@@ -1074,7 +1074,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
break; break;
case PPC_INST_STFSX: case PPC_INST_STFSX:
println("\tctx.csr.setFlushMode(false);"); println("\tctx.fpscr.setFlushMode(false);");
println("\ttemp.f32 = ctx.f{}.f64;", insn.operands[0]); println("\ttemp.f32 = ctx.f{}.f64;", insn.operands[0]);
print("\tPPC_STORE_U32("); print("\tPPC_STORE_U32(");
if (insn.operands[1] != 0) if (insn.operands[1] != 0)
@@ -1253,7 +1253,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VADDFP: case PPC_INST_VADDFP:
case PPC_INST_VADDFP128: case PPC_INST_VADDFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_add_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_add_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
@@ -1304,7 +1304,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VCTSXS: case PPC_INST_VCTSXS:
case PPC_INST_VCFPSXWS128: case PPC_INST_VCFPSXWS128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_si128((__m128i*)ctx.v{}.s32, _mm_vctsxs(_mm_mul_ps(_mm_load_ps(ctx.v{}.f32), _mm_set1_ps({}))));", insn.operands[0], insn.operands[1], 1u << insn.operands[2]); println("\t_mm_store_si128((__m128i*)ctx.v{}.s32, _mm_vctsxs(_mm_mul_ps(_mm_load_ps(ctx.v{}.f32), _mm_set1_ps({}))));", insn.operands[0], insn.operands[1], 1u << insn.operands[2]);
break; break;
@@ -1313,7 +1313,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
{ {
const float v = ldexp(1.0f, -int32_t(insn.operands[2])); const float v = ldexp(1.0f, -int32_t(insn.operands[2]));
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_mul_ps(_mm_cvtepi32_ps(_mm_load_si128((__m128i*)ctx.v{}.u32)), _mm_castsi128_ps(_mm_set1_epi32(int(0x{:X})))));", insn.operands[0], insn.operands[1], *reinterpret_cast<const uint32_t*>(&v)); println("\t_mm_store_ps(ctx.v{}.f32, _mm_mul_ps(_mm_cvtepi32_ps(_mm_load_si128((__m128i*)ctx.v{}.u32)), _mm_castsi128_ps(_mm_set1_epi32(int(0x{:X})))));", insn.operands[0], insn.operands[1], *reinterpret_cast<const uint32_t*>(&v));
break; break;
} }
@@ -1323,7 +1323,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
{ {
const float v = ldexp(1.0f, -int32_t(insn.operands[2])); const float v = ldexp(1.0f, -int32_t(insn.operands[2]));
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_mul_ps(_mm_cvtepu32_ps_(_mm_load_si128((__m128i*)ctx.v{}.u32)), _mm_castsi128_ps(_mm_set1_epi32(int(0x{:X})))));", insn.operands[0], insn.operands[1], *reinterpret_cast<const uint32_t*>(&v)); println("\t_mm_store_ps(ctx.v{}.f32, _mm_mul_ps(_mm_cvtepu32_ps_(_mm_load_si128((__m128i*)ctx.v{}.u32)), _mm_castsi128_ps(_mm_set1_epi32(int(0x{:X})))));", insn.operands[0], insn.operands[1], *reinterpret_cast<const uint32_t*>(&v));
break; break;
} }
@@ -1334,7 +1334,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VCMPEQFP: case PPC_INST_VCMPEQFP:
case PPC_INST_VCMPEQFP128: case PPC_INST_VCMPEQFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_cmpeq_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_cmpeq_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
if (strchr(insn.opcode->name, '.')) if (strchr(insn.opcode->name, '.'))
println("\tctx.cr6.setFromMask(_mm_load_ps(ctx.v{}.f32), 0xF);", insn.operands[0]); println("\tctx.cr6.setFromMask(_mm_load_ps(ctx.v{}.f32), 0xF);", insn.operands[0]);
@@ -1355,7 +1355,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VCMPGEFP: case PPC_INST_VCMPGEFP:
case PPC_INST_VCMPGEFP128: case PPC_INST_VCMPGEFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_cmpge_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_cmpge_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
if (strchr(insn.opcode->name, '.')) if (strchr(insn.opcode->name, '.'))
println("\tctx.cr6.setFromMask(_mm_load_ps(ctx.v{}.f32), 0xF);", insn.operands[0]); println("\tctx.cr6.setFromMask(_mm_load_ps(ctx.v{}.f32), 0xF);", insn.operands[0]);
@@ -1363,7 +1363,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VCMPGTFP: case PPC_INST_VCMPGTFP:
case PPC_INST_VCMPGTFP128: case PPC_INST_VCMPGTFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_cmpgt_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_cmpgt_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
if (strchr(insn.opcode->name, '.')) if (strchr(insn.opcode->name, '.'))
println("\tctx.cr6.setFromMask(_mm_load_ps(ctx.v{}.f32), 0xF);", insn.operands[0]); println("\tctx.cr6.setFromMask(_mm_load_ps(ctx.v{}.f32), 0xF);", insn.operands[0]);
@@ -1379,14 +1379,14 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VEXPTEFP128: case PPC_INST_VEXPTEFP128:
// TODO: vectorize // TODO: vectorize
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
for (size_t i = 0; i < 4; i++) for (size_t i = 0; i < 4; i++)
println("\tctx.v{}.f32[{}] = exp2f(ctx.v{}.f32[{}]);", insn.operands[0], i, insn.operands[1], i); println("\tctx.v{}.f32[{}] = exp2f(ctx.v{}.f32[{}]);", insn.operands[0], i, insn.operands[1], i);
break; break;
case PPC_INST_VLOGEFP128: case PPC_INST_VLOGEFP128:
// TODO: vectorize // TODO: vectorize
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
for (size_t i = 0; i < 4; i++) for (size_t i = 0; i < 4; i++)
println("\tctx.v{}.f32[{}] = log2f(ctx.v{}.f32[{}]);", insn.operands[0], i, insn.operands[1], i); println("\tctx.v{}.f32[{}] = log2f(ctx.v{}.f32[{}]);", insn.operands[0], i, insn.operands[1], i);
break; break;
@@ -1394,13 +1394,13 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VMADDCFP128: case PPC_INST_VMADDCFP128:
case PPC_INST_VMADDFP: case PPC_INST_VMADDFP:
case PPC_INST_VMADDFP128: case PPC_INST_VMADDFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_fmadd_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_fmadd_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
case PPC_INST_VMAXFP: case PPC_INST_VMAXFP:
case PPC_INST_VMAXFP128: case PPC_INST_VMAXFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_max_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_max_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
@@ -1410,7 +1410,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VMINFP: case PPC_INST_VMINFP:
case PPC_INST_VMINFP128: case PPC_INST_VMINFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_min_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_min_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
@@ -1442,23 +1442,23 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VMSUM3FP128: case PPC_INST_VMSUM3FP128:
// NOTE: accounting for full vector reversal here. should dot product yzw instead of xyz // NOTE: accounting for full vector reversal here. should dot product yzw instead of xyz
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_dp_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), 0xEF));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_dp_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), 0xEF));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_VMSUM4FP128: case PPC_INST_VMSUM4FP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_dp_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), 0xFF));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_dp_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), 0xFF));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_VMULFP128: case PPC_INST_VMULFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_mul_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_mul_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;
case PPC_INST_VNMSUBFP: case PPC_INST_VNMSUBFP:
case PPC_INST_VNMSUBFP128: case PPC_INST_VNMSUBFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_fnmadd_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_fnmadd_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2], insn.operands[3]);
break; break;
@@ -1487,7 +1487,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VPKD3D128: case PPC_INST_VPKD3D128:
// TODO: vectorize somehow? // TODO: vectorize somehow?
// NOTE: handling vector reversal here too // NOTE: handling vector reversal here too
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
switch (insn.operands[2]) switch (insn.operands[2])
{ {
case 0: // D3D color case 0: // D3D color
@@ -1514,23 +1514,23 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VREFP: case PPC_INST_VREFP:
case PPC_INST_VREFP128: case PPC_INST_VREFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_rcp_ps(_mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_rcp_ps(_mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_VRFIM128: case PPC_INST_VRFIM128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_round_ps(_mm_load_ps(ctx.v{}.f32), _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC));", insn.operands[0], insn.operands[1]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_round_ps(_mm_load_ps(ctx.v{}.f32), _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_VRFIN: case PPC_INST_VRFIN:
case PPC_INST_VRFIN128: case PPC_INST_VRFIN128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_round_ps(_mm_load_ps(ctx.v{}.f32), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));", insn.operands[0], insn.operands[1]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_round_ps(_mm_load_ps(ctx.v{}.f32), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));", insn.operands[0], insn.operands[1]);
break; break;
case PPC_INST_VRFIZ128: case PPC_INST_VRFIZ128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_round_ps(_mm_load_ps(ctx.v{}.f32), _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC));", insn.operands[0], insn.operands[1]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_round_ps(_mm_load_ps(ctx.v{}.f32), _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC));", insn.operands[0], insn.operands[1]);
break; break;
@@ -1543,7 +1543,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VRSQRTEFP: case PPC_INST_VRSQRTEFP:
case PPC_INST_VRSQRTEFP128: case PPC_INST_VRSQRTEFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_rsqrt_ps(_mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_rsqrt_ps(_mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1]);
break; break;
@@ -1623,7 +1623,7 @@ bool Recompiler::Recompile(const Function& fn, uint32_t base, const ppc_insn& in
case PPC_INST_VSUBFP: case PPC_INST_VSUBFP:
case PPC_INST_VSUBFP128: case PPC_INST_VSUBFP128:
println("\tctx.csr.setFlushMode(true);"); println("\tctx.fpscr.setFlushMode(true);");
println("\t_mm_store_ps(ctx.v{}.f32, _mm_sub_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]); println("\t_mm_store_ps(ctx.v{}.f32, _mm_sub_ps(_mm_load_ps(ctx.v{}.f32), _mm_load_ps(ctx.v{}.f32)));", insn.operands[0], insn.operands[1], insn.operands[2]);
break; break;

2
PowerRecomp/test_recompiler.cpp
View File

@@ -132,7 +132,7 @@ void TestRecompiler::RecompileTests(const char* srcDirectoryPath, const char* ds
std::println(file, "PPC_FUNC({});\n", symbol->second); std::println(file, "PPC_FUNC({});\n", symbol->second);
std::println(file, "void {}(uint8_t* base) {{", name); std::println(file, "void {}(uint8_t* base) {{", name);
std::println(file, "\tPPCContext ctx{{}};"); std::println(file, "\tPPCContext ctx{{}};");
std::println(file, "\tctx.csr.storeValue();"); std::println(file, "\tctx.fpscr.loadFromGuest();");
while (getline() && !str.empty() && str[0] == '#') while (getline() && !str.empty() && str[0] == '#')
{ {

29
PowerUtils/ppc_context.h
View File

@@ -128,24 +128,32 @@ struct alignas(0x10) PPCVRegister
}; };
}; };
struct CSRRegister struct PPCFPSCRRegister
{ {
uint32_t value; uint32_t csr;
void storeValue() uint32_t loadFromHost()
{ {
value = _mm_getcsr(); csr = _mm_getcsr();
return (0x6C >> ((csr & _MM_ROUND_MASK) >> 12)) & 3;
}
void storeFromGuest(uint32_t value)
{
csr &= ~_MM_ROUND_MASK;
csr |= ((0x6C >> (2 * (value & 3))) & 3) << 13;
_mm_setcsr(csr);
} }
void setFlushMode(bool enable) void setFlushMode(bool enable)
{ {
uint32_t mask = _MM_FLUSH_ZERO_MASK | _MM_DENORMALS_ZERO_MASK; constexpr uint32_t mask = _MM_FLUSH_ZERO_MASK | _MM_DENORMALS_ZERO_MASK;
uint32_t newValue = enable ? (value | mask) : (value & ~mask); uint32_t value = enable ? (csr | mask) : (csr & ~mask);
if (value != newValue) if (csr != value)
{ {
_mm_setcsr(newValue); _mm_setcsr(value);
value = newValue; csr = value;
} }
} }
}; };
@@ -158,8 +166,7 @@ struct PPCContext
PPCXERRegister xer; PPCXERRegister xer;
PPCRegister reserved; PPCRegister reserved;
uint32_t msr; uint32_t msr;
uint32_t fpscr; PPCFPSCRRegister fpscr;
CSRRegister csr;
union union
{ {