From 13e8b7b51631de66a6ad7c56fa96a36646aa37ef Mon Sep 17 00:00:00 2001
From: MerryMage <MerryMage@users.noreply.github.com>
Date: Sun, 24 Mar 2019 15:42:14 +0000
Subject: [PATCH] emit_x64_floating_point: F16C implementation of
FPSingleToHalf
---
src/backend/x64/emit_x64_floating_point.cpp | 13 ++++++++++++-
1 file changed, 12 insertions(+), 1 deletion(-)
diff --git a/src/backend/x64/emit_x64_floating_point.cpp b/src/backend/x64/emit_x64_floating_point.cpp
index e4d97e86..5df30955 100644
--- a/src/backend/x64/emit_x64_floating_point.cpp
+++ b/src/backend/x64/emit_x64_floating_point.cpp
@@ -1133,8 +1133,19 @@ void EmitX64::EmitFPSingleToDouble(EmitContext& ctx, IR::Inst* inst) {
void EmitX64::EmitFPSingleToHalf(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const auto rounding_mode = static_cast<FP::RoundingMode>(args[1].GetImmediateU8());
+ const auto round_imm = ConvertRoundingModeToX64Immediate(rounding_mode);
- // TODO: F16C implementation requires ForceToDefaultNaN<16> to be implemented.
+ if (code.DoesCpuSupport(Xbyak::util::Cpu::tF16C) && !ctx.FPCR().AHP() && !ctx.FPCR().FZ16()) {
+ const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
+
+ if (ctx.FPCR().DN()) {
+ ForceToDefaultNaN<32>(code, result);
+ }
+ code.vcvtps2ph(result, result, static_cast<u8>(*round_imm));
+
+ ctx.reg_alloc.DefineValue(inst, result);
+ return;
+ }
ctx.reg_alloc.HostCall(inst, args[0]);
code.mov(code.ABI_PARAM2.cvt32(), ctx.FPCR().Value());