Hardware instruction set support for crossgen2

docs/design/coreclr/botr/README.md

@@ -29,6 +29,7 @@ Below is a table of contents.
 - [Cross-platform Minidumps](xplat-minidump-generation.md)
 - [Mixed Mode Assemblies](mixed-mode.md)
 - [Guide For Porting](guide-for-porting.md)
+- [Vectors and Intrinsics](vectors-and-intrinsics.md)
 
 
 It may be possible that this table is not complete.  You can get a complete list

docs/design/coreclr/botr/readytorun-format.md

@@ -246,6 +246,7 @@ fixup kind, the rest of the signature varies based on the fixup kind.
 | READYTORUN_FIXUP_DeclaringTypeHandle     |  0x2D | Dictionary lookup for method declaring type. Followed by the type signature.
 | READYTORUN_FIXUP_IndirectPInvokeTarget   |  0x2E | Target (indirect) of an inlined PInvoke. Followed by method signature.
 | READYTORUN_FIXUP_PInvokeTarget           |  0x2F | Target of an inlined PInvoke. Followed by method signature.
+| READYTORUN_FIXUP_Check_InstructionSetSupport | 0x30 | Specify the instruction sets that must be supported/unsupported to use the R2R code associated with the fixup.
 | READYTORUN_FIXUP_ModuleOverride          |  0x80 | When or-ed to the fixup ID, the fixup byte in the signature is followed by an encoded uint with assemblyref index, either within the MSIL metadata of the master context module for the signature or within the manifest metadata R2R header table (used in cases inlining brings in references to assemblies not seen in the input MSIL).
 
 #### Method Signatures

src/coreclr/src/inc/corcompile.h

@@ -691,6 +691,8 @@ enum CORCOMPILE_FIXUP_BLOB_KIND
     ENCODE_INDIRECT_PINVOKE_TARGET,                 /* For calling a pinvoke method ptr indirectly */
     ENCODE_PINVOKE_TARGET,                          /* For calling a pinvoke method ptr */
 
+    ENCODE_CHECK_INSTRUCTION_SET_SUPPORT,           /* Define the set of instruction sets that must be supported/unsupported to use the fixup */
+
     ENCODE_MODULE_HANDLE                = 0x50,     /* Module token */
     ENCODE_STATIC_FIELD_ADDRESS,                    /* For accessing a static field */
     ENCODE_MODULE_ID_FOR_STATICS,                   /* For accessing static fields */

src/coreclr/src/inc/corinfoinstructionset.h

@@ -428,4 +428,65 @@ inline const char *InstructionSetToString(CORINFO_InstructionSet instructionSet)
 #endif
 }
 
+inline CORINFO_InstructionSet InstructionSetFromR2RInstructionSet(ReadyToRunInstructionSet r2rSet)
+{
+#ifdef _MSC_VER
+#pragma warning(push)
+#pragma warning(disable: 4065) // disable warning for switch statement with only default label.
+#endif
+
+    switch (r2rSet)
+    {
+#ifdef TARGET_ARM64
+        case READYTORUN_INSTRUCTION_ArmBase: return InstructionSet_ArmBase;
+        case READYTORUN_INSTRUCTION_AdvSimd: return InstructionSet_AdvSimd;
+        case READYTORUN_INSTRUCTION_Aes: return InstructionSet_Aes;
+        case READYTORUN_INSTRUCTION_Crc32: return InstructionSet_Crc32;
+        case READYTORUN_INSTRUCTION_Sha1: return InstructionSet_Sha1;
+        case READYTORUN_INSTRUCTION_Sha256: return InstructionSet_Sha256;
+        case READYTORUN_INSTRUCTION_Atomics: return InstructionSet_Atomics;
+#endif // TARGET_ARM64
+#ifdef TARGET_AMD64
+        case READYTORUN_INSTRUCTION_Sse: return InstructionSet_SSE;
+        case READYTORUN_INSTRUCTION_Sse2: return InstructionSet_SSE2;
+        case READYTORUN_INSTRUCTION_Sse3: return InstructionSet_SSE3;
+        case READYTORUN_INSTRUCTION_Ssse3: return InstructionSet_SSSE3;
+        case READYTORUN_INSTRUCTION_Sse41: return InstructionSet_SSE41;
+        case READYTORUN_INSTRUCTION_Sse42: return InstructionSet_SSE42;
+        case READYTORUN_INSTRUCTION_Avx: return InstructionSet_AVX;
+        case READYTORUN_INSTRUCTION_Avx2: return InstructionSet_AVX2;
+        case READYTORUN_INSTRUCTION_Aes: return InstructionSet_AES;
+        case READYTORUN_INSTRUCTION_Bmi1: return InstructionSet_BMI1;
+        case READYTORUN_INSTRUCTION_Bmi2: return InstructionSet_BMI2;
+        case READYTORUN_INSTRUCTION_Fma: return InstructionSet_FMA;
+        case READYTORUN_INSTRUCTION_Lzcnt: return InstructionSet_LZCNT;
+        case READYTORUN_INSTRUCTION_Pclmulqdq: return InstructionSet_PCLMULQDQ;
+        case READYTORUN_INSTRUCTION_Popcnt: return InstructionSet_POPCNT;
+#endif // TARGET_AMD64
+#ifdef TARGET_X86
+        case READYTORUN_INSTRUCTION_Sse: return InstructionSet_SSE;
+        case READYTORUN_INSTRUCTION_Sse2: return InstructionSet_SSE2;
+        case READYTORUN_INSTRUCTION_Sse3: return InstructionSet_SSE3;
+        case READYTORUN_INSTRUCTION_Ssse3: return InstructionSet_SSSE3;
+        case READYTORUN_INSTRUCTION_Sse41: return InstructionSet_SSE41;
+        case READYTORUN_INSTRUCTION_Sse42: return InstructionSet_SSE42;
+        case READYTORUN_INSTRUCTION_Avx: return InstructionSet_AVX;
+        case READYTORUN_INSTRUCTION_Avx2: return InstructionSet_AVX2;
+        case READYTORUN_INSTRUCTION_Aes: return InstructionSet_AES;
+        case READYTORUN_INSTRUCTION_Bmi1: return InstructionSet_BMI1;
+        case READYTORUN_INSTRUCTION_Bmi2: return InstructionSet_BMI2;
+        case READYTORUN_INSTRUCTION_Fma: return InstructionSet_FMA;
+        case READYTORUN_INSTRUCTION_Lzcnt: return InstructionSet_LZCNT;
+        case READYTORUN_INSTRUCTION_Pclmulqdq: return InstructionSet_PCLMULQDQ;
+        case READYTORUN_INSTRUCTION_Popcnt: return InstructionSet_POPCNT;
+#endif // TARGET_X86
+
+        default:
+            return InstructionSet_ILLEGAL;
+    }
+#ifdef _MSC_VER
+#pragma warning(pop)
+#endif
+}
+
 #endif // CORINFOINSTRUCTIONSET_H

src/coreclr/src/inc/corjitflags.h

@@ -151,6 +151,11 @@ class CORJIT_FLAGS
         instructionSetFlags.AddInstructionSet(instructionSet);
     }
 
+    bool IsSet(CORINFO_InstructionSet instructionSet) const
+    {
+        return instructionSetFlags.HasInstructionSet(instructionSet);
+    }
+
     void Clear(CORINFO_InstructionSet instructionSet)
     {
         instructionSetFlags.RemoveInstructionSet(instructionSet);

src/coreclr/src/inc/readytorun.h

@@ -202,6 +202,8 @@ enum ReadyToRunFixupKind
 
     READYTORUN_FIXUP_IndirectPInvokeTarget      = 0x2E, /* Target (indirect) of an inlined pinvoke */
     READYTORUN_FIXUP_PInvokeTarget              = 0x2F, /* Target of an inlined pinvoke */
+
+    READYTORUN_FIXUP_Check_InstructionSetSupport= 0x30, /* Define the set of instruction sets that must be supported/unsupported to use the fixup */
 };
 
 //
@@ -359,6 +361,8 @@ enum ReadyToRunHelper
     READYTORUN_HELPER_StackProbe                = 0x111,
 };
 
+#include "readytoruninstructionset.h"
+
 //
 // Exception info
 //

src/coreclr/src/jit/codegenarm64.cpp

@@ -2723,7 +2723,7 @@ void CodeGen::genLockedInstructions(GenTreeOp* treeNode)
 
     emitAttr dataSize = emitActualTypeSize(data);
 
-    if (compiler->compSupports(InstructionSet_Atomics))
+    if (compiler->compOpportunisticallyDependsOn(InstructionSet_Atomics))
     {
         assert(!data->isContainedIntOrIImmed());
 
@@ -2860,7 +2860,7 @@ void CodeGen::genCodeForCmpXchg(GenTreeCmpXchg* treeNode)
     genConsumeRegs(data);
     genConsumeRegs(comparand);
 
-    if (compiler->compSupports(InstructionSet_Atomics))
+    if (compiler->compOpportunisticallyDependsOn(InstructionSet_Atomics))
     {
         emitAttr dataSize = emitActualTypeSize(data);
 

src/coreclr/src/jit/codegenxarch.cpp

@@ -7256,7 +7256,7 @@ void CodeGen::genSSE2BitwiseOp(GenTree* treeNode)
 void CodeGen::genSSE41RoundOp(GenTreeOp* treeNode)
 {
     // i) SSE4.1 is supported by the underlying hardware
-    assert(compiler->compSupports(InstructionSet_SSE41));
+    assert(compiler->compIsaSupportedDebugOnly(InstructionSet_SSE41));
 
     // ii) treeNode oper is a GT_INTRINSIC
     assert(treeNode->OperGet() == GT_INTRINSIC);

src/coreclr/src/jit/compiler.cpp

@@ -2194,10 +2194,11 @@ void Compiler::compSetProcessor()
 #endif // TARGET_X86
 
     CORINFO_InstructionSetFlags instructionSetFlags = jitFlags.GetInstructionSetFlags();
+    opts.compSupportsISA                            = 0;
+    opts.compSupportsISAReported                    = 0;
 
 #ifdef TARGET_XARCH
-    // Instruction set flags for Intel hardware intrinsics
-    opts.compSupportsISA = 0;
+    bool avxSupported = false;
 
     if (JitConfig.EnableHWIntrinsic())
     {

src/coreclr/src/jit/compiler.h

@@ -7635,12 +7635,12 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     SIMDLevel getSIMDSupportLevel()
     {
 #if defined(TARGET_XARCH)
-        if (compSupports(InstructionSet_AVX2))
+        if (compOpportunisticallyDependsOn(InstructionSet_AVX2))
         {
             return SIMD_AVX2_Supported;
         }
 
-        if (compSupports(InstructionSet_SSE42))
+        if (compOpportunisticallyDependsOn(InstructionSet_SSE42))
         {
             return SIMD_SSE4_Supported;
         }
@@ -7786,7 +7786,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
                     unreached();
             }
         }
-        assert(emitTypeSize(simdType) <= maxSIMDStructBytes());
+        assert(emitTypeSize(simdType) <= largestEnregisterableStructSize());
         switch (simdBaseType)
         {
             case TYP_FLOAT:
@@ -8029,26 +8029,18 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 
     GenTree* getOp1ForConstructor(OPCODE opcode, GenTree* newobjThis, CORINFO_CLASS_HANDLE clsHnd);
 
-    // Whether SIMD vector occupies part of SIMD register.
-    // SSE2: vector2f/3f are considered sub register SIMD types.
-    // AVX: vector2f, 3f and 4f are all considered sub register SIMD types.
-    bool isSubRegisterSIMDType(CORINFO_CLASS_HANDLE typeHnd)
-    {
-        unsigned  sizeBytes = 0;
-        var_types baseType  = getBaseTypeAndSizeOfSIMDType(typeHnd, &sizeBytes);
-        return (baseType == TYP_FLOAT) && (sizeBytes < getSIMDVectorRegisterByteLength());
-    }
-
     bool isSubRegisterSIMDType(GenTreeSIMD* simdNode)
     {
-        return (simdNode->gtSIMDSize < getSIMDVectorRegisterByteLength());
+        return (simdNode->gtSIMDSize < maxSIMDStructBytes());
     }
 
     // Get the type for the hardware SIMD vector.
     // This is the maximum SIMD type supported for this target.
     var_types getSIMDVectorType()
     {
 #if defined(TARGET_XARCH)
+        // Fix the exact level of simd vector support by using compExactlyDependsOn on the Avx2 isa
+        compExactlyDependsOn(InstructionSet_AVX2);
         if (getSIMDSupportLevel() == SIMD_AVX2_Supported)
         {
             return TYP_SIMD32;
@@ -8095,6 +8087,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
         else
         {
             assert(getSIMDSupportLevel() >= SIMD_SSE2_Supported);
+            compExactlyDependsOn(InstructionSet_AVX2); // Record that AVX2 isn't supported
             return XMM_REGSIZE_BYTES;
         }
 #elif defined(TARGET_ARM64)
@@ -8115,7 +8108,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     unsigned int maxSIMDStructBytes()
     {
 #if defined(FEATURE_HW_INTRINSICS) && defined(TARGET_XARCH)
-        if (compSupports(InstructionSet_AVX))
+        if (compOpportunisticallyDependsOn(InstructionSet_AVX))
         {
             return YMM_REGSIZE_BYTES;
         }
@@ -8128,6 +8121,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
         return getSIMDVectorRegisterByteLength();
 #endif
     }
+
     unsigned int minSIMDStructBytes()
     {
         return emitTypeSize(TYP_SIMD8);
@@ -8187,14 +8181,40 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     unsigned largestEnregisterableStructSize()
     {
 #ifdef FEATURE_SIMD
-        unsigned vectorRegSize = getSIMDVectorRegisterByteLength();
+#if defined(FEATURE_HW_INTRINSICS) && defined(TARGET_XARCH)
+        if (opts.IsReadyToRun())
+        {
+            // Return constant instead of maxSIMDStructBytes, as maxSIMDStructBytes performs
+            // checks that are effected by the current level of instruction set support would
+            // otherwise cause the highest level of instruction set support to be reported to crossgen2.
+            // and this api is only ever used as an optimization or assert, so no reporting should
+            // ever happen.
+            return YMM_REGSIZE_BYTES;
+        }
+#endif // defined(FEATURE_HW_INTRINSICS) && defined(TARGET_XARCH)
+        unsigned vectorRegSize = maxSIMDStructBytes();
         assert(vectorRegSize >= TARGET_POINTER_SIZE);
         return vectorRegSize;
 #else  // !FEATURE_SIMD
         return TARGET_POINTER_SIZE;
 #endif // !FEATURE_SIMD
     }
 
+    // Use to determine if a struct *might* be a SIMD type. As this function only takes a size, many
+    // structs will fit the criteria.
+    bool structSizeMightRepresentSIMDType(size_t structSize)
+    {
+#ifdef FEATURE_SIMD
+        // Do not use maxSIMDStructBytes as that api in R2R on X86 and X64 may notify the JIT
+        // about the size of a struct under the assumption that the struct size needs to be recorded.
+        // By using largestEnregisterableStructSize here, the detail of whether or not Vector256<T> is
+        // enregistered or not will not be messaged to the R2R compiler.
+        return (structSize >= minSIMDStructBytes()) && (structSize <= largestEnregisterableStructSize());
+#else
+        return false;
+#endif // FEATURE_SIMD
+    }
+
 #ifdef FEATURE_SIMD
     static bool vnEncodesResultTypeForSIMDIntrinsic(SIMDIntrinsicID intrinsicId);
 #endif // !FEATURE_SIMD
@@ -8272,21 +8292,63 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
         return false;
     }
 
-    bool compSupports(CORINFO_InstructionSet isa) const
+#ifdef DEBUG
+    // Answer the question: Is a particular ISA supported?
+    // Use this api when asking the question so that future
+    // ISA questions can be asked correctly or when asserting
+    // support/nonsupport for an instruction set
+    bool compIsaSupportedDebugOnly(CORINFO_InstructionSet isa) const
     {
 #if defined(TARGET_XARCH) || defined(TARGET_ARM64)
         return (opts.compSupportsISA & (1ULL << isa)) != 0;
 #else
         return false;
 #endif
     }
+#endif // DEBUG
 
     void notifyInstructionSetUsage(CORINFO_InstructionSet isa, bool supported) const;
 
+    // Answer the question: Is a particular ISA supported?
+    // The result of this api call will exactly match the target machine
+    // on which the function is executed (except for CoreLib, where there are special rules)
+    bool compExactlyDependsOn(CORINFO_InstructionSet isa) const
+    {
+
+#if defined(TARGET_XARCH) || defined(TARGET_ARM64)
+        uint64_t isaBit       = (1ULL << isa);
+        bool     isaSupported = (opts.compSupportsISA & (1ULL << isa)) != 0;
+        if ((opts.compSupportsISAReported & isaBit) == 0)
+        {
+            notifyInstructionSetUsage(isa, isaSupported);
+            ((Compiler*)this)->opts.compSupportsISAReported |= isaBit;
+        }
+
+        return isaSupported;
+#else
+        return false;
+#endif
+    }
+
+    // Answer the question: Is a particular ISA supported?
+    // The result of this api call will match the target machine if the result is true
+    // If the result is false, then the target machine may have support for the instruction
+    bool compOpportunisticallyDependsOn(CORINFO_InstructionSet isa) const
+    {
+        if ((opts.compSupportsISA & (1ULL << isa)) != 0)
+        {
+            return compExactlyDependsOn(isa);
+        }
+        else
+        {
+            return false;
+        }
+    }
+
     bool canUseVexEncoding() const
     {
 #ifdef TARGET_XARCH
-        return compSupports(InstructionSet_AVX);
+        return compOpportunisticallyDependsOn(InstructionSet_AVX);
 #else
         return false;
 #endif
@@ -8381,14 +8443,11 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     {
         JitFlags* jitFlags; // all flags passed from the EE
 
-#if defined(TARGET_XARCH) || defined(TARGET_ARM64)
         uint64_t compSupportsISA;
-#endif
+        uint64_t compSupportsISAReported;
         void setSupportedISAs(CORINFO_InstructionSetFlags isas)
         {
-#if defined(TARGET_XARCH) || defined(TARGET_ARM64)
             compSupportsISA = isas.GetFlagsRaw();
-#endif
         }
 
         unsigned compFlags; // method attributes

src/coreclr/src/jit/gentree.cpp

@@ -17066,7 +17066,7 @@ GenTree* Compiler::gtGetSIMDZero(var_types simdType, var_types baseType, CORINFO
         switch (simdType)
         {
             case TYP_SIMD16:
-                if (compSupports(InstructionSet_SSE))
+                if (compExactlyDependsOn(InstructionSet_SSE))
                 {
                     // We only return the HWIntrinsicNode if SSE is supported, since it is possible for
                     // the user to disable the SSE HWIntrinsic support via the COMPlus configuration knobs
@@ -17075,7 +17075,7 @@ GenTree* Compiler::gtGetSIMDZero(var_types simdType, var_types baseType, CORINFO
                 }
                 return nullptr;
             case TYP_SIMD32:
-                if (compSupports(InstructionSet_AVX))
+                if (compExactlyDependsOn(InstructionSet_AVX))
                 {
                     // We only return the HWIntrinsicNode if AVX is supported, since it is possible for
                     // the user to disable the AVX HWIntrinsic support via the COMPlus configuration knobs

src/coreclr/src/jit/hwintrinsic.cpp

@@ -309,7 +309,7 @@ NamedIntrinsic HWIntrinsicInfo::lookupId(Compiler*   comp,
         return NI_Illegal;
     }
 
-    bool isIsaSupported = comp->compSupports(isa) && comp->compSupportsHWIntrinsic(isa);
+    bool isIsaSupported = comp->compExactlyDependsOn(isa) && comp->compSupportsHWIntrinsic(isa);
 
     if (strcmp(methodName, "get_IsSupported") == 0)
     {