Adding support for X86Base.Pause() and ArmBase.Yield() (#61065)

* Adding support for X86Base.Pause() and ArmBase.Yield() * Applying formatting patch * Ensure NI_ArmBase_Yield actually gets through to codegen on arm64
dotnet · Nov 15, 2021 · fdafc7c · fdafc7c
1 parent a93e0d2
commit fdafc7c
Show file tree

Hide file tree

Showing 25 changed files with 273 additions and 38 deletions.
diff --git a/src/coreclr/jit/compiler.h b/src/coreclr/jit/compiler.h
@@ -3318,6 +3318,7 @@ class Compiler
                                unsigned    simdSize,
                                bool        isSimdAsHWIntrinsic);
 
+    GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types type, NamedIntrinsic hwIntrinsicID);
     GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types type, GenTree* op1, NamedIntrinsic hwIntrinsicID);
     GenTreeHWIntrinsic* gtNewScalarHWIntrinsicNode(var_types      type,
                                                    GenTree*       op1,

diff --git a/src/coreclr/jit/emit.h b/src/coreclr/jit/emit.h
@@ -1240,21 +1240,22 @@ class emitter
 
 #define PERFSCORE_THROUGHPUT_1C 1.0f // Single Issue
 
-#define PERFSCORE_THROUGHPUT_2C 2.0f   // slower - 2 cycles
-#define PERFSCORE_THROUGHPUT_3C 3.0f   // slower - 3 cycles
-#define PERFSCORE_THROUGHPUT_4C 4.0f   // slower - 4 cycles
-#define PERFSCORE_THROUGHPUT_5C 5.0f   // slower - 5 cycles
-#define PERFSCORE_THROUGHPUT_6C 6.0f   // slower - 6 cycles
-#define PERFSCORE_THROUGHPUT_7C 7.0f   // slower - 7 cycles
-#define PERFSCORE_THROUGHPUT_8C 8.0f   // slower - 8 cycles
-#define PERFSCORE_THROUGHPUT_9C 9.0f   // slower - 9 cycles
-#define PERFSCORE_THROUGHPUT_10C 10.0f // slower - 10 cycles
-#define PERFSCORE_THROUGHPUT_13C 13.0f // slower - 13 cycles
-#define PERFSCORE_THROUGHPUT_19C 19.0f // slower - 19 cycles
-#define PERFSCORE_THROUGHPUT_25C 25.0f // slower - 25 cycles
-#define PERFSCORE_THROUGHPUT_33C 33.0f // slower - 33 cycles
-#define PERFSCORE_THROUGHPUT_52C 52.0f // slower - 52 cycles
-#define PERFSCORE_THROUGHPUT_57C 57.0f // slower - 57 cycles
+#define PERFSCORE_THROUGHPUT_2C 2.0f     // slower - 2 cycles
+#define PERFSCORE_THROUGHPUT_3C 3.0f     // slower - 3 cycles
+#define PERFSCORE_THROUGHPUT_4C 4.0f     // slower - 4 cycles
+#define PERFSCORE_THROUGHPUT_5C 5.0f     // slower - 5 cycles
+#define PERFSCORE_THROUGHPUT_6C 6.0f     // slower - 6 cycles
+#define PERFSCORE_THROUGHPUT_7C 7.0f     // slower - 7 cycles
+#define PERFSCORE_THROUGHPUT_8C 8.0f     // slower - 8 cycles
+#define PERFSCORE_THROUGHPUT_9C 9.0f     // slower - 9 cycles
+#define PERFSCORE_THROUGHPUT_10C 10.0f   // slower - 10 cycles
+#define PERFSCORE_THROUGHPUT_13C 13.0f   // slower - 13 cycles
+#define PERFSCORE_THROUGHPUT_19C 19.0f   // slower - 19 cycles
+#define PERFSCORE_THROUGHPUT_25C 25.0f   // slower - 25 cycles
+#define PERFSCORE_THROUGHPUT_33C 33.0f   // slower - 33 cycles
+#define PERFSCORE_THROUGHPUT_52C 52.0f   // slower - 52 cycles
+#define PERFSCORE_THROUGHPUT_57C 57.0f   // slower - 57 cycles
+#define PERFSCORE_THROUGHPUT_140C 140.0f // slower - 140 cycles
 
 #define PERFSCORE_LATENCY_ILLEGAL -1024.0f
 
@@ -1281,6 +1282,7 @@ class emitter
 #define PERFSCORE_LATENCY_26C 26.0f
 #define PERFSCORE_LATENCY_62C 62.0f
 #define PERFSCORE_LATENCY_69C 69.0f
+#define PERFSCORE_LATENCY_140C 140.0f
 #define PERFSCORE_LATENCY_400C 400.0f // Intel microcode issue with these instuctions
 
 #define PERFSCORE_LATENCY_BRANCH_DIRECT 1.0f   // cost of an unconditional branch

diff --git a/src/coreclr/jit/emitarm64.cpp b/src/coreclr/jit/emitarm64.cpp
@@ -14588,6 +14588,12 @@ emitter::insExecutionCharacteristics emitter::getInsExecutionCharacteristics(ins
                 result.insThroughput = PERFSCORE_THROUGHPUT_ZERO;
                 result.insLatency    = PERFSCORE_LATENCY_ZERO;
             }
+            else if (ins == INS_yield)
+            {
+                // @ToDo - find out the actual latency, match x86/x64 for now
+                result.insThroughput = PERFSCORE_THROUGHPUT_140C;
+                result.insLatency    = PERFSCORE_LATENCY_140C;
+            }
             else
             {
                 result.insThroughput = PERFSCORE_THROUGHPUT_2X;

diff --git a/src/coreclr/jit/emitxarch.cpp b/src/coreclr/jit/emitxarch.cpp
@@ -2898,7 +2898,8 @@ void emitter::emitIns(instruction ins)
              ins == INS_r_movsp || ins == INS_r_stosb || ins == INS_r_stosd || ins == INS_r_stosp || ins == INS_ret ||
              ins == INS_sahf || ins == INS_stosb || ins == INS_stosd || ins == INS_stosp
              // These instructions take zero operands
-             || ins == INS_vzeroupper || ins == INS_lfence || ins == INS_mfence || ins == INS_sfence);
+             || ins == INS_vzeroupper || ins == INS_lfence || ins == INS_mfence || ins == INS_sfence ||
+             ins == INS_pause);
 
         assert(assertCond);
     }
@@ -12333,8 +12334,8 @@ BYTE* emitter::emitOutputRR(BYTE* dst, instrDesc* id)
                         // Due to elided register moves, we can't have the following assert.
                         // For example, consider:
                         //    t85 = LCL_VAR byref V01 arg1 rdx (last use) REG rdx
-                        //        /--*  t85    byref                                                       
-                        //        *  STORE_LCL_VAR byref  V40 tmp31 rdx REG rdx                 
+                        //        /--*  t85    byref
+                        //        *  STORE_LCL_VAR byref  V40 tmp31 rdx REG rdx
                         // Here, V01 is type `long` on entry, then is stored as a byref. But because
                         // the register allocator assigned the same register, no instruction was
                         // generated, and we only (currently) make gcref/byref changes in emitter GC info
@@ -16104,6 +16105,13 @@ emitter::insExecutionCharacteristics emitter::getInsExecutionCharacteristics(ins
             result.insThroughput = PERFSCORE_THROUGHPUT_2X;
             break;
 
+        case INS_pause:
+        {
+            result.insLatency    = PERFSCORE_LATENCY_140C;
+            result.insThroughput = PERFSCORE_THROUGHPUT_140C;
+            break;
+        }
+
         default:
             // unhandled instruction insFmt combination
             perfScoreUnhandledInstruction(id, &result);

diff --git a/src/coreclr/jit/gentree.cpp b/src/coreclr/jit/gentree.cpp
@@ -21798,6 +21798,12 @@ GenTree* Compiler::gtNewSimdZeroNode(var_types   type,
     return gtNewSimdHWIntrinsicNode(type, intrinsic, simdBaseJitType, simdSize, isSimdAsHWIntrinsic);
 }
 
+GenTreeHWIntrinsic* Compiler::gtNewScalarHWIntrinsicNode(var_types type, NamedIntrinsic hwIntrinsicID)
+{
+    return new (this, GT_HWINTRINSIC)
+        GenTreeHWIntrinsic(type, hwIntrinsicID, CORINFO_TYPE_UNDEF, 0, /* isSimdAsHWIntrinsic */ false);
+}
+
 GenTreeHWIntrinsic* Compiler::gtNewScalarHWIntrinsicNode(var_types type, GenTree* op1, NamedIntrinsic hwIntrinsicID)
 {
     SetOpLclRelatedToSIMDIntrinsic(op1);

diff --git a/src/coreclr/jit/hwintrinsic.h b/src/coreclr/jit/hwintrinsic.h
@@ -804,7 +804,7 @@ struct HWIntrinsic final
 
         if (baseType == TYP_UNKNOWN)
         {
-            assert(category == HW_Category_Scalar);
+            assert((category == HW_Category_Scalar) || (category == HW_Category_Special));
 
             if (HWIntrinsicInfo::BaseTypeFromFirstArg(id))
             {

diff --git a/src/coreclr/jit/hwintrinsicarm64.cpp b/src/coreclr/jit/hwintrinsicarm64.cpp
@@ -308,17 +308,23 @@ GenTree* Compiler::impSpecialIntrinsic(NamedIntrinsic        intrinsic,
                                        var_types             retType,
                                        unsigned              simdSize)
 {
-    HWIntrinsicCategory category     = HWIntrinsicInfo::lookupCategory(intrinsic);
-    int                 numArgs      = sig->numArgs;
-    var_types           simdBaseType = JitType2PreciseVarType(simdBaseJitType);
+    HWIntrinsicCategory category = HWIntrinsicInfo::lookupCategory(intrinsic);
+    int                 numArgs  = sig->numArgs;
 
     if (!featureSIMD || !IsBaselineSimdIsaSupported())
     {
         return nullptr;
     }
 
     assert(numArgs >= 0);
-    assert(varTypeIsArithmetic(simdBaseType));
+
+    var_types simdBaseType = TYP_UNKNOWN;
+
+    if (intrinsic != NI_ArmBase_Yield)
+    {
+        simdBaseType = JitType2PreciseVarType(simdBaseJitType);
+        assert(varTypeIsArithmetic(simdBaseType));
+    }
 
     GenTree* retNode = nullptr;
     GenTree* op1     = nullptr;
@@ -327,6 +333,16 @@ GenTree* Compiler::impSpecialIntrinsic(NamedIntrinsic        intrinsic,
 
     switch (intrinsic)
     {
+        case NI_ArmBase_Yield:
+        {
+            assert(sig->numArgs == 0);
+            assert(JITtype2varType(sig->retType) == TYP_VOID);
+            assert(simdSize == 0);
+
+            retNode = gtNewScalarHWIntrinsicNode(TYP_VOID, intrinsic);
+            break;
+        }
+
         case NI_Vector64_Abs:
         case NI_Vector128_Abs:
         {

diff --git a/src/coreclr/jit/hwintrinsiccodegenarm64.cpp b/src/coreclr/jit/hwintrinsiccodegenarm64.cpp
@@ -255,6 +255,13 @@ void CodeGen::genHWIntrinsic(GenTreeHWIntrinsic* node)
         emitSize = emitActualTypeSize(intrin.baseType);
         opt      = INS_OPTS_NONE;
     }
+    else if (intrin.category == HW_Category_Special)
+    {
+        assert(intrin.id == NI_ArmBase_Yield);
+
+        emitSize = EA_UNKNOWN;
+        opt      = INS_OPTS_NONE;
+    }
     else
     {
         emitSize = emitActualTypeSize(Compiler::getSIMDTypeForSize(node->GetSimdSize()));
@@ -443,6 +450,12 @@ void CodeGen::genHWIntrinsic(GenTreeHWIntrinsic* node)
                 }
                 break;
 
+            case NI_ArmBase_Yield:
+            {
+                ins = INS_yield;
+                break;
+            }
+
             default:
                 ins = HWIntrinsicInfo::lookupIns(intrin.id, intrin.baseType);
                 break;
@@ -735,6 +748,12 @@ void CodeGen::genHWIntrinsic(GenTreeHWIntrinsic* node)
                 }
                 break;
 
+            case NI_ArmBase_Yield:
+            {
+                GetEmitter()->emitIns(ins);
+                break;
+            }
+
             // mvni doesn't support the range of element types, so hard code the 'opts' value.
             case NI_Vector64_get_Zero:
             case NI_Vector64_get_AllBitsSet:

diff --git a/src/coreclr/jit/hwintrinsiccodegenxarch.cpp b/src/coreclr/jit/hwintrinsiccodegenxarch.cpp
@@ -1447,6 +1447,8 @@ void CodeGen::genX86BaseIntrinsic(GenTreeHWIntrinsic* node)
 {
     NamedIntrinsic intrinsicId = node->gtHWIntrinsicId;
 
+    genConsumeOperands(node);
+
     switch (intrinsicId)
     {
         case NI_X86Base_BitScanForward:
@@ -1459,16 +1461,25 @@ void CodeGen::genX86BaseIntrinsic(GenTreeHWIntrinsic* node)
             var_types   targetType = node->TypeGet();
             instruction ins        = HWIntrinsicInfo::lookupIns(intrinsicId, targetType);
 
-            genConsumeOperands(node);
             genHWIntrinsic_R_RM(node, ins, emitTypeSize(targetType), targetReg, op1);
-            genProduceReg(node);
+            break;
+        }
+
+        case NI_X86Base_Pause:
+        {
+            assert(node->GetSimdBaseType() == TYP_UNKNOWN);
+            assert(node->gtGetOp1() == nullptr);
+            assert(node->gtGetOp2() == nullptr);
+            GetEmitter()->emitIns(INS_pause);
             break;
         }
 
         default:
             unreached();
             break;
     }
+
+    genProduceReg(node);
 }
 
 //------------------------------------------------------------------------
@@ -1532,7 +1543,7 @@ void CodeGen::genSSEIntrinsic(GenTreeHWIntrinsic* node)
 
         case NI_SSE_StoreFence:
         {
-            assert(baseType == TYP_VOID);
+            assert(baseType == TYP_UNKNOWN);
             assert(op1 == nullptr);
             assert(op2 == nullptr);
             emit->emitIns(INS_sfence);
@@ -1617,7 +1628,7 @@ void CodeGen::genSSE2Intrinsic(GenTreeHWIntrinsic* node)
 
         case NI_SSE2_LoadFence:
         {
-            assert(baseType == TYP_VOID);
+            assert(baseType == TYP_UNKNOWN);
             assert(op1 == nullptr);
             assert(op2 == nullptr);
             emit->emitIns(INS_lfence);
@@ -1626,7 +1637,7 @@ void CodeGen::genSSE2Intrinsic(GenTreeHWIntrinsic* node)
 
         case NI_SSE2_MemoryFence:
         {
-            assert(baseType == TYP_VOID);
+            assert(baseType == TYP_UNKNOWN);
             assert(op1 == nullptr);
             assert(op2 == nullptr);
             emit->emitIns(INS_mfence);

diff --git a/src/coreclr/jit/hwintrinsiclistarm64.h b/src/coreclr/jit/hwintrinsiclistarm64.h
@@ -627,6 +627,7 @@ HARDWARE_INTRINSIC(Aes,           PolynomialMultiplyWideningUpper,
 //  Base Intrinsics
 HARDWARE_INTRINSIC(ArmBase,       LeadingZeroCount,                                                  0,      1,     {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_clz,            INS_clz,            INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid},     HW_Category_Scalar,                HW_Flag_BaseTypeFromFirstArg|HW_Flag_NoFloatingPointUsed)
 HARDWARE_INTRINSIC(ArmBase,       ReverseElementBits,                                                0,      1,     {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_rbit,           INS_rbit,           INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid},     HW_Category_Scalar,                HW_Flag_NoFloatingPointUsed)
+HARDWARE_INTRINSIC(ArmBase,       Yield,                                                             0,      0,     {INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid,        INS_invalid},     HW_Category_Special,               HW_Flag_SpecialCodeGen|HW_Flag_SpecialImport)
 
 // ***************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************
 //                 ISA            Function name                                              SIMD size  Number of arguments                                                                                            Instructions                                                                                        Category                           Flags

diff --git a/src/coreclr/jit/hwintrinsiclistxarch.h b/src/coreclr/jit/hwintrinsiclistxarch.h
@@ -199,6 +199,7 @@ HARDWARE_INTRINSIC(Vector256,       Xor,
 //  X86Base Intrinsics
 HARDWARE_INTRINSIC(X86Base,         BitScanForward,                             0,              1,      {INS_invalid,           INS_invalid,            INS_invalid,            INS_invalid,            INS_bsf,                INS_bsf,                INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid},           HW_Category_Scalar,                 HW_Flag_NoFloatingPointUsed|HW_Flag_NoRMWSemantics)
 HARDWARE_INTRINSIC(X86Base,         BitScanReverse,                             0,              1,      {INS_invalid,           INS_invalid,            INS_invalid,            INS_invalid,            INS_bsr,                INS_bsr,                INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid},           HW_Category_Scalar,                 HW_Flag_NoFloatingPointUsed|HW_Flag_NoRMWSemantics)
+HARDWARE_INTRINSIC(X86Base,         Pause,                                      0,              0,      {INS_invalid,           INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid,            INS_invalid},           HW_Category_Special,                HW_Flag_NoContainment|HW_Flag_NoRMWSemantics)
 
 // ***************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************
 //                 ISA              Function name                               SIMD size       NumArg                                                                                                         Instructions                                                                                                                             Category                            Flags

diff --git a/src/coreclr/jit/hwintrinsicxarch.cpp b/src/coreclr/jit/hwintrinsicxarch.cpp
@@ -491,6 +491,7 @@ GenTree* Compiler::impSpecialIntrinsic(NamedIntrinsic        intrinsic,
     {
         case InstructionSet_Vector256:
         case InstructionSet_Vector128:
+        case InstructionSet_X86Base:
             return impBaseIntrinsic(intrinsic, clsHnd, method, sig, simdBaseJitType, retType, simdSize);
         case InstructionSet_SSE:
             return impSSEIntrinsic(intrinsic, method, sig);
@@ -548,8 +549,13 @@ GenTree* Compiler::impBaseIntrinsic(NamedIntrinsic        intrinsic,
         return nullptr;
     }
 
-    var_types simdBaseType = JitType2PreciseVarType(simdBaseJitType);
-    assert(varTypeIsArithmetic(simdBaseType));
+    var_types simdBaseType = TYP_UNKNOWN;
+
+    if (intrinsic != NI_X86Base_Pause)
+    {
+        simdBaseType = JitType2PreciseVarType(simdBaseJitType);
+        assert(varTypeIsArithmetic(simdBaseType));
+    }
 
     switch (intrinsic)
     {
@@ -1532,6 +1538,16 @@ GenTree* Compiler::impBaseIntrinsic(NamedIntrinsic        intrinsic,
             break;
         }
 
+        case NI_X86Base_Pause:
+        {
+            assert(sig->numArgs == 0);
+            assert(JITtype2varType(sig->retType) == TYP_VOID);
+            assert(simdSize == 0);
+
+            retNode = gtNewScalarHWIntrinsicNode(TYP_VOID, intrinsic);
+            break;
+        }
+
         default:
         {
             return nullptr;
@@ -1604,7 +1620,7 @@ GenTree* Compiler::impSSEIntrinsic(NamedIntrinsic intrinsic, CORINFO_METHOD_HAND
         case NI_SSE_StoreFence:
             assert(sig->numArgs == 0);
             assert(JITtype2varType(sig->retType) == TYP_VOID);
-            retNode = gtNewSimdHWIntrinsicNode(TYP_VOID, intrinsic, CORINFO_TYPE_VOID, 0);
+            retNode = gtNewScalarHWIntrinsicNode(TYP_VOID, intrinsic);
             break;
 
         default:
@@ -1667,7 +1683,7 @@ GenTree* Compiler::impSSE2Intrinsic(NamedIntrinsic intrinsic, CORINFO_METHOD_HAN
             assert(JITtype2varType(sig->retType) == TYP_VOID);
             assert(simdSize == 0);
 
-            retNode = gtNewSimdHWIntrinsicNode(TYP_VOID, intrinsic, CORINFO_TYPE_VOID, simdSize);
+            retNode = gtNewScalarHWIntrinsicNode(TYP_VOID, intrinsic);
             break;
         }
 

diff --git a/src/coreclr/jit/instrsarm64.h b/src/coreclr/jit/instrsarm64.h
@@ -1563,6 +1563,9 @@ INST1(uxth,        "uxth",         0,      IF_DR_2H,  0x53003C00)
 INST1(nop,         "nop",          0,      IF_SN_0A,  0xD503201F)
                                    //  nop                          SN_0A  1101010100000011 0010000000011111   D503 201F
 
+INST1(yield,       "yield",        0,      IF_SN_0A,  0xD503203F)
+                                   //  yield                        SN_0A  1101010100000011 0010000000111111   D503 203F
+
 INST1(bkpt,        "bkpt",         0,      IF_SN_0A,  0xD43E0000)
                                    //  brpt                         SN_0A  1101010000111110 0000000000000000   D43E 0000   0xF000