diff --git a/src/backend_x64/emit_x64_vector.cpp b/src/backend_x64/emit_x64_vector.cpp index 099cc1a1..9cc209c9 100644 --- a/src/backend_x64/emit_x64_vector.cpp +++ b/src/backend_x64/emit_x64_vector.cpp @@ -35,6 +35,18 @@ static void EmitVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* i ctx.reg_alloc.DefineValue(inst, xmm_a); } +template +static void EmitAVXVectorOperation(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Function fn) { + auto args = ctx.reg_alloc.GetArgumentInfo(inst); + + const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]); + const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]); + + (code.*fn)(xmm_a, xmm_a, xmm_b); + + ctx.reg_alloc.DefineValue(inst, xmm_a); +} + template static void EmitOneArgumentFallback(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Lambda lambda) { const auto fn = static_cast*>(lambda); @@ -1049,6 +1061,11 @@ void EmitX64::EmitVectorMaxS32(EmitContext& ctx, IR::Inst* inst) { } void EmitX64::EmitVectorMaxS64(EmitContext& ctx, IR::Inst* inst) { + if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX512VL)) { + EmitAVXVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::vpmaxsq); + return; + } + EmitTwoArgumentFallback(code, ctx, inst, [](std::array& result, const std::array& a, const std::array& b){ std::transform(a.begin(), a.end(), b.begin(), result.begin(), [](auto x, auto y) { return std::max(x, y); }); }); @@ -1086,6 +1103,11 @@ void EmitX64::EmitVectorMaxU32(EmitContext& ctx, IR::Inst* inst) { } void EmitX64::EmitVectorMaxU64(EmitContext& ctx, IR::Inst* inst) { + if (code.DoesCpuSupport(Xbyak::util::Cpu::tAVX512VL)) { + EmitAVXVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::vpmaxuq); + return; + } + EmitTwoArgumentFallback(code, ctx, inst, [](std::array& result, const std::array& a, const std::array& b){ std::transform(a.begin(), a.end(), b.begin(), result.begin(), [](auto x, auto y) { return std::max(x, y); }); });