zdoom/gzdoom-gles
0
0
Fork 0
mirror of https://github.com/ZDoom/gzdoom-gles.git synced 2024-11-11 07:12:16 +00:00
Code Issues Projects Releases Packages Wiki Activity

- don't use SSE for integer min/max

Browse source
This commit is contained in:
Magnus Norddahl 2018-12-02 09:58:08 +01:00
parent 706a923884
commit edc05111cd
1 changed files with 26 additions and 32 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

58
src/scripting/vm/jit_math.cpp
View file

@ -447,46 +447,40 @@ void JitCompiler::EmitXOR_RK()
void JitCompiler::EmitMIN_RR() void JitCompiler::EmitMIN_RR()
{ {
auto tmp0 = newTempXmmSs(); auto rc = CheckRegD(C, A);
auto tmp1 = newTempXmmSs(); if (A != B)
cc.movd(tmp0, regD[B]); cc.mov(regD[A], regD[B]);
cc.movd(tmp1, regD[C]); cc.cmp(rc, regD[A]);
cc.pminsd(tmp0, tmp1); cc.cmovl(regD[A], rc);
cc.movd(regD[A], tmp0);
} }
void JitCompiler::EmitMIN_RK() void JitCompiler::EmitMIN_RK()
{ {
auto tmp0 = newTempXmmSs(); auto rc = newTempInt32();
auto tmp1 = newTempXmmSs(); if (A != B)
auto konstTmp = newTempIntPtr(); cc.mov(regD[A], regD[B]);
cc.mov(konstTmp, asmjit::imm_ptr(&konstd[C])); cc.mov(rc, asmjit::imm(konstd[C]));
cc.movd(tmp0, regD[B]); cc.cmp(rc, regD[A]);
cc.movss(tmp1, asmjit::x86::dword_ptr(konstTmp)); cc.cmovl(regD[A], rc);
cc.pminsd(tmp0, tmp1);
cc.movd(regD[A], tmp0);
} }
void JitCompiler::EmitMAX_RR() void JitCompiler::EmitMAX_RR()
{ {
auto tmp0 = newTempXmmSs(); auto rc = CheckRegD(C, A);
auto tmp1 = newTempXmmSs(); if (A != B)
cc.movd(tmp0, regD[B]); cc.mov(regD[A], regD[B]);
cc.movd(tmp1, regD[C]); cc.cmp(rc, regD[A]);
cc.pmaxsd(tmp0, tmp1); cc.cmovg(regD[A], rc);
cc.movd(regD[A], tmp0);
} }
void JitCompiler::EmitMAX_RK() void JitCompiler::EmitMAX_RK()
{ {
auto tmp0 = newTempXmmSs(); auto rc = newTempInt32();
auto tmp1 = newTempXmmSs(); if (A != B)
auto konstTmp = newTempIntPtr(); cc.mov(regD[A], regD[B]);
cc.mov(konstTmp, asmjit::imm_ptr(&konstd[C])); cc.mov(rc, asmjit::imm(konstd[C]));
cc.movd(tmp0, regD[B]); cc.cmp(rc, regD[A]);
cc.movss(tmp1, asmjit::x86::dword_ptr(konstTmp)); cc.cmovg(regD[A], rc);
cc.pmaxsd(tmp0, tmp1);
cc.movd(regD[A], tmp0);
} }
void JitCompiler::EmitABS() void JitCompiler::EmitABS()
@ -852,7 +846,7 @@ void JitCompiler::EmitMINF_RR()
auto rc = CheckRegF(C, A); auto rc = CheckRegF(C, A);
if (A != B) if (A != B)
cc.movsd(regF[A], regF[B]); cc.movsd(regF[A], regF[B]);
cc.minpd(regF[A], rc); // minsd required SSE 4.1 cc.minpd(regF[A], rc); // minsd requires SSE 4.1
} }
void JitCompiler::EmitMINF_RK() void JitCompiler::EmitMINF_RK()
@ -861,7 +855,7 @@ void JitCompiler::EmitMINF_RK()
auto tmp = newTempIntPtr(); auto tmp = newTempIntPtr();
cc.mov(tmp, asmjit::imm_ptr(&konstf[C])); cc.mov(tmp, asmjit::imm_ptr(&konstf[C]));
cc.movsd(regF[A], asmjit::x86::qword_ptr(tmp)); cc.movsd(regF[A], asmjit::x86::qword_ptr(tmp));
cc.minpd(regF[A], rb); // minsd required SSE 4.1 cc.minpd(regF[A], rb); // minsd requires SSE 4.1
} }
void JitCompiler::EmitMAXF_RR() void JitCompiler::EmitMAXF_RR()
@ -869,7 +863,7 @@ void JitCompiler::EmitMAXF_RR()
auto rc = CheckRegF(C, A); auto rc = CheckRegF(C, A);
if (A != B) if (A != B)
cc.movsd(regF[A], regF[B]); cc.movsd(regF[A], regF[B]);
cc.maxpd(regF[A], rc); // maxsd required SSE 4.1 cc.maxpd(regF[A], rc); // maxsd requires SSE 4.1
} }
void JitCompiler::EmitMAXF_RK() void JitCompiler::EmitMAXF_RK()
@ -878,7 +872,7 @@ void JitCompiler::EmitMAXF_RK()
auto tmp = newTempIntPtr(); auto tmp = newTempIntPtr();
cc.mov(tmp, asmjit::imm_ptr(&konstf[C])); cc.mov(tmp, asmjit::imm_ptr(&konstf[C]));
cc.movsd(regF[A], asmjit::x86::qword_ptr(tmp)); cc.movsd(regF[A], asmjit::x86::qword_ptr(tmp));
cc.maxpd(regF[A], rb); // maxsd required SSE 4.1 cc.maxpd(regF[A], rb); // maxsd requires SSE 4.1
} }
void JitCompiler::EmitATAN2() void JitCompiler::EmitATAN2()