Add option to change SVE vector length for current and children proce…

…sses (#101295)

* Add option to change SVE vector length for current and children processes.

* Use maxVectorTBitWidth to get desired SVE length

* Use CNTB to determine current vector length

* Use ACLE to detect vector length on Linux and hardcode on Windows

* Use rdvl instead of cntb to count vector length

* Add inline assembly suitable for MSVC

* Add MSCV compatible assembly helper instead of inline assembly

* Declare GetSveLengthFromOS on Arm64

* Remove definition of GetSveLengthFromOS from .S file

* Move non-windows definition of GetSveLengthFromOS to .S file

* Apply suggestions from code review

* Move declaration of GetSveLengthFromOS to .cpp file

* Disable  temporarily for Windows on Arm machines

* Hardcode GetSveLengthFromOS() to return SVE length as 128

---------

Co-authored-by: Jan Kotas <jkotas@microsoft.com>

src/coreclr/vm/arm64/asmhelpers.S

@@ -23,6 +23,13 @@ LEAF_ENTRY GetDataCacheZeroIDReg, _TEXT
     ret     lr
 LEAF_END GetDataCacheZeroIDReg, _TEXT
 
+// uint64_t GetSveLengthFromOS(void);
+.arch_extension sve
+    LEAF_ENTRY GetSveLengthFromOS, _TEXT
+        rdvl    x0, 1
+        ret     lr
+    LEAF_END GetSveLengthFromOS, _TEXT
+
 //-----------------------------------------------------------------------------
 // This routine captures the machine state. It is used by helper method frame
 //-----------------------------------------------------------------------------

src/coreclr/vm/arm64/asmhelpers.asm

@@ -78,6 +78,14 @@
         ret     lr
     LEAF_END
 
+;; uint64_t GetSveLengthFromOS(void);
+    LEAF_ENTRY GetSveLengthFromOS
+        ;; TODO-SVE: Remove the hardcoded value 128 and uncomment once CI machines are updated to use MASM 14.4 or later
+        ;; rdvl    x0, 1
+        mov     x0, #128
+        ret     lr
+    LEAF_END
+
 ;;-----------------------------------------------------------------------------
 ;; This routine captures the machine state. It is used by helper method frame
 ;;-----------------------------------------------------------------------------

src/coreclr/vm/codeman.cpp

@@ -1243,6 +1243,7 @@ EEJitManager::EEJitManager()
 
 #ifdef TARGET_ARM64
 extern "C" DWORD64 __stdcall GetDataCacheZeroIDReg();
+extern "C" uint64_t GetSveLengthFromOS();
 #endif
 
 void EEJitManager::SetCpuInfo()
@@ -1257,12 +1258,12 @@ void EEJitManager::SetCpuInfo()
 
     int cpuFeatures = minipal_getcpufeatures();
 
-#if defined(TARGET_X86) || defined(TARGET_AMD64)
-    CPUCompileFlags.Set(InstructionSet_VectorT128);
-
     // Get the maximum bitwidth of Vector<T>, rounding down to the nearest multiple of 128-bits
     uint32_t maxVectorTBitWidth = (CLRConfig::GetConfigValue(CLRConfig::EXTERNAL_MaxVectorTBitWidth) / 128) * 128;
 
+#if defined(TARGET_X86) || defined(TARGET_AMD64)
+    CPUCompileFlags.Set(InstructionSet_VectorT128);
+
     if (((cpuFeatures & XArchIntrinsicConstants_Avx2) != 0) && ((maxVectorTBitWidth == 0) || (maxVectorTBitWidth >= 256)))
     {
         // We allow 256-bit Vector<T> by default
@@ -1514,7 +1515,14 @@ void EEJitManager::SetCpuInfo()
 
     if (((cpuFeatures & ARM64IntrinsicConstants_Sve) != 0) && CLRConfig::GetConfigValue(CLRConfig::EXTERNAL_EnableArm64Sve))
     {
-        CPUCompileFlags.Set(InstructionSet_Sve);
+        uint32_t maxVectorTLength = (maxVectorTBitWidth / 8);
+        uint64_t sveLengthFromOS = GetSveLengthFromOS();
+
+        // Do not enable SVE when the user specified vector length is smaller than the one offered by underlying OS.
+        if ((maxVectorTLength >= sveLengthFromOS) || (maxVectorTBitWidth == 0))
+        {
+            CPUCompileFlags.Set(InstructionSet_Sve);
+        }
     }
 
     // DCZID_EL0<4> (DZP) indicates whether use of DC ZVA instructions is permitted (0) or prohibited (1).