From 82417da7803e2cf18efc28a1cd3f3d0a4b6045ae Mon Sep 17 00:00:00 2001
From: MerryMage <MerryMage@users.noreply.github.com>
Date: Sat, 11 Jul 2020 14:02:55 +0100
Subject: [PATCH] emit_x64{_vector}_floating_point: Add unsafe optimizations
 for RSqrtEstimate and RecipEstimate

---
 src/backend/x64/emit_x64_floating_point.cpp   | 68 +++++++++++++++----
 .../x64/emit_x64_vector_floating_point.cpp    | 58 +++++++++++++---
 2 files changed, 106 insertions(+), 20 deletions(-)

diff --git a/src/backend/x64/emit_x64_floating_point.cpp b/src/backend/x64/emit_x64_floating_point.cpp
index 2f170bbe..5efd5258 100644
--- a/src/backend/x64/emit_x64_floating_point.cpp
+++ b/src/backend/x64/emit_x64_floating_point.cpp
@@ -738,8 +738,29 @@ void EmitX64::EmitFPMulX64(EmitContext& ctx, IR::Inst* inst) {
     EmitFPMulX<64>(code, ctx, inst);
 }
 
-template<typename FPT>
+template<size_t fsize>
 static void EmitFPRecipEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
+    using FPT = mp::unsigned_integer_of_size<fsize>;
+
+    if constexpr (fsize != 16) {
+        if (ctx.UnsafeOptimizations()) {
+            auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+            const Xbyak::Xmm operand = ctx.reg_alloc.UseXmm(args[0]);
+            const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
+
+            if constexpr (fsize == 32) {
+                code.rcpss(result, operand);
+            } else {
+                code.cvtsd2ss(result, operand);
+                code.rcpss(result, result);
+                code.cvtss2sd(result, result);
+            }
+
+            ctx.reg_alloc.DefineValue(inst, result);
+            return;
+        }
+    }
+
     auto args = ctx.reg_alloc.GetArgumentInfo(inst);
     ctx.reg_alloc.HostCall(inst, args[0]);
     code.mov(code.ABI_PARAM2.cvt32(), ctx.FPCR().Value());
@@ -748,19 +769,21 @@ static void EmitFPRecipEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
 }
 
 void EmitX64::EmitFPRecipEstimate16(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRecipEstimate<u16>(code, ctx, inst);
+    EmitFPRecipEstimate<16>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPRecipEstimate32(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRecipEstimate<u32>(code, ctx, inst);
+    EmitFPRecipEstimate<32>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPRecipEstimate64(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRecipEstimate<u64>(code, ctx, inst);
+    EmitFPRecipEstimate<64>(code, ctx, inst);
 }
 
-template <typename FPT>
+template <size_t fsize>
 static void EmitFPRecipExponent(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
+    using FPT = mp::unsigned_integer_of_size<fsize>;
+
     auto args = ctx.reg_alloc.GetArgumentInfo(inst);
     ctx.reg_alloc.HostCall(inst, args[0]);
     code.mov(code.ABI_PARAM2.cvt32(), ctx.FPCR().Value());
@@ -769,15 +792,15 @@ static void EmitFPRecipExponent(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
 }
 
 void EmitX64::EmitFPRecipExponent16(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRecipExponent<u16>(code, ctx, inst);
+    EmitFPRecipExponent<16>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPRecipExponent32(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRecipExponent<u32>(code, ctx, inst);
+    EmitFPRecipExponent<32>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPRecipExponent64(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRecipExponent<u64>(code, ctx, inst);
+    EmitFPRecipExponent<64>(code, ctx, inst);
 }
 
 template<size_t fsize>
@@ -911,8 +934,29 @@ void EmitX64::EmitFPRoundInt64(EmitContext& ctx, IR::Inst* inst) {
     EmitFPRound(code, ctx, inst, 64);
 }
 
-template<typename FPT>
+template<size_t fsize>
 static void EmitFPRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
+    using FPT = mp::unsigned_integer_of_size<fsize>;
+
+    if constexpr (fsize != 16) {
+        if (ctx.UnsafeOptimizations()) {
+            auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+            const Xbyak::Xmm operand = ctx.reg_alloc.UseXmm(args[0]);
+            const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
+
+            if constexpr (fsize == 32) {
+                code.rsqrtss(result, operand);
+            } else {
+                code.cvtsd2ss(result, operand);
+                code.rsqrtss(result, result);
+                code.cvtss2sd(result, result);
+            }
+
+            ctx.reg_alloc.DefineValue(inst, result);
+            return;
+        }
+    }
+
     auto args = ctx.reg_alloc.GetArgumentInfo(inst);
     ctx.reg_alloc.HostCall(inst, args[0]);
     code.mov(code.ABI_PARAM2.cvt32(), ctx.FPCR().Value());
@@ -921,15 +965,15 @@ static void EmitFPRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
 }
 
 void EmitX64::EmitFPRSqrtEstimate16(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRSqrtEstimate<u16>(code, ctx, inst);
+    EmitFPRSqrtEstimate<16>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPRSqrtEstimate32(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRSqrtEstimate<u32>(code, ctx, inst);
+    EmitFPRSqrtEstimate<32>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPRSqrtEstimate64(EmitContext& ctx, IR::Inst* inst) {
-    EmitFPRSqrtEstimate<u64>(code, ctx, inst);
+    EmitFPRSqrtEstimate<64>(code, ctx, inst);
 }
 
 template<size_t fsize>
diff --git a/src/backend/x64/emit_x64_vector_floating_point.cpp b/src/backend/x64/emit_x64_vector_floating_point.cpp
index e1ac36a2..f5fe2c84 100644
--- a/src/backend/x64/emit_x64_vector_floating_point.cpp
+++ b/src/backend/x64/emit_x64_vector_floating_point.cpp
@@ -1178,8 +1178,29 @@ void EmitX64::EmitFPVectorPairedAddLower64(EmitContext& ctx, IR::Inst* inst) {
     });
 }
 
-template<typename FPT>
+template<size_t fsize>
 static void EmitRecipEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
+    using FPT = mp::unsigned_integer_of_size<fsize>;
+
+    if constexpr (fsize != 16) {
+        if (ctx.UnsafeOptimizations()) {
+            auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+            const Xbyak::Xmm operand = ctx.reg_alloc.UseXmm(args[0]);
+            const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
+
+            if constexpr (fsize == 32) {
+                code.rcpps(result, operand);
+            } else {
+                code.cvtpd2ps(result, operand);
+                code.rcpps(result, result);
+                code.cvtps2pd(result, result);
+            }
+
+            ctx.reg_alloc.DefineValue(inst, result);
+            return;
+        }
+    }
+
     EmitTwoOpFallback(code, ctx, inst, [](VectorArray<FPT>& result, const VectorArray<FPT>& operand, FP::FPCR fpcr, FP::FPSR& fpsr) {
         for (size_t i = 0; i < result.size(); i++) {
             result[i] = FP::FPRecipEstimate<FPT>(operand[i], fpcr, fpsr);
@@ -1188,15 +1209,15 @@ static void EmitRecipEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* ins
 }
 
 void EmitX64::EmitFPVectorRecipEstimate16(EmitContext& ctx, IR::Inst* inst) {
-    EmitRecipEstimate<u16>(code, ctx, inst);
+    EmitRecipEstimate<16>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPVectorRecipEstimate32(EmitContext& ctx, IR::Inst* inst) {
-    EmitRecipEstimate<u32>(code, ctx, inst);
+    EmitRecipEstimate<32>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPVectorRecipEstimate64(EmitContext& ctx, IR::Inst* inst) {
-    EmitRecipEstimate<u64>(code, ctx, inst);
+    EmitRecipEstimate<64>(code, ctx, inst);
 }
 
 template<size_t fsize>
@@ -1337,8 +1358,29 @@ void EmitX64::EmitFPVectorRoundInt64(EmitContext& ctx, IR::Inst* inst) {
     EmitFPVectorRoundInt<64>(code, ctx, inst);
 }
 
-template<typename FPT>
+template<size_t fsize>
 static void EmitRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
+    using FPT = mp::unsigned_integer_of_size<fsize>;
+
+    if constexpr (fsize != 16) {
+        if (ctx.UnsafeOptimizations()) {
+            auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+            const Xbyak::Xmm operand = ctx.reg_alloc.UseXmm(args[0]);
+            const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
+
+            if constexpr (fsize == 32) {
+                code.rsqrtps(result, operand);
+            } else {
+                code.cvtpd2ps(result, operand);
+                code.rsqrtps(result, result);
+                code.cvtps2pd(result, result);
+            }
+
+            ctx.reg_alloc.DefineValue(inst, result);
+            return;
+        }
+    }
+
     EmitTwoOpFallback(code, ctx, inst, [](VectorArray<FPT>& result, const VectorArray<FPT>& operand, FP::FPCR fpcr, FP::FPSR& fpsr) {
         for (size_t i = 0; i < result.size(); i++) {
             result[i] = FP::FPRSqrtEstimate<FPT>(operand[i], fpcr, fpsr);
@@ -1347,15 +1389,15 @@ static void EmitRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* ins
 }
 
 void EmitX64::EmitFPVectorRSqrtEstimate16(EmitContext& ctx, IR::Inst* inst) {
-    EmitRSqrtEstimate<u16>(code, ctx, inst);
+    EmitRSqrtEstimate<16>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPVectorRSqrtEstimate32(EmitContext& ctx, IR::Inst* inst) {
-    EmitRSqrtEstimate<u32>(code, ctx, inst);
+    EmitRSqrtEstimate<32>(code, ctx, inst);
 }
 
 void EmitX64::EmitFPVectorRSqrtEstimate64(EmitContext& ctx, IR::Inst* inst) {
-    EmitRSqrtEstimate<u64>(code, ctx, inst);
+    EmitRSqrtEstimate<64>(code, ctx, inst);
 }
 
 template<size_t fsize>