avoid_recomputing_moveMask.txt

diff --git a/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs b/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs
index 163d62208c6..864bcb58b2e 100644
--- a/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs
+++ b/src/libraries/System.Private.CoreLib/src/System/Text/ASCIIUtility.cs
@@ -794,6 +794,8 @@ private static unsafe nuint GetIndexOfFirstNonAsciiChar_Sse2(char* pBuffer, nuin
             // This method is written such that control generally flows top-to-bottom, avoiding
             // jumps as much as possible in the optimistic case of "all ASCII". If we see non-ASCII
             // data, we jump out of the hot paths to targets at the end of the method.
+
+            Vector128<ushort> asciiMaskForTestZ = Vector128.Create((ushort)0xFF80); // used for PTEST on supported hardware
             Vector128<ushort> asciiMaskForAddSaturate = Vector128.Create((ushort)0x7F80); // used for PADDUSW
             const uint NonAsciiDataSeenMask = 0b_1010_1010_1010_1010; // used for determining whether 'currentMask' contains non-ASCII data
 
@@ -805,9 +807,15 @@ private static unsafe nuint GetIndexOfFirstNonAsciiChar_Sse2(char* pBuffer, nuin
 
             firstVector = Sse2.LoadVector128((ushort*)pBuffer); // unaligned load
 
-            if (VectorContainsNonAsciiChar(firstVector))
+            // The operation below forces the 0x8000 bit of each WORD to be set iff the WORD element
+            // has value >= 0x0800 (non-ASCII). Then we'll treat the vector as a BYTE vector in order
+            // to extract the mask. Reminder: the 0x0080 bit of each WORD should be ignored.
+
+            currentMask = (uint)Sse2.MoveMask(Sse2.AddSaturate(firstVector, asciiMaskForAddSaturate).AsByte());
+
+            if ((currentMask & NonAsciiDataSeenMask) != 0)
             {
-                goto FoundNonAsciiDataInFirstVector;
+                goto FoundNonAsciiDataInCurrentMask;
             }
 
             // If we have less than 32 bytes to process, just go straight to the final unaligned
@@ -880,9 +888,23 @@ private static unsafe nuint GetIndexOfFirstNonAsciiChar_Sse2(char* pBuffer, nuin
             // Remember, at this point pBuffer is still aligned.
 
             firstVector = Sse2.LoadAlignedVector128((ushort*)pBuffer);
-            if (VectorContainsNonAsciiChar(firstVector))
+            if (Sse41.IsSupported)
             {
-                goto FoundNonAsciiDataInFirstVector;
+                // If a non-ASCII bit is set in any WORD of the combined vector, we have seen non-ASCII data.
+                // Jump to the non-ASCII handler to figure out which particular vector contained non-ASCII data.
+                if (!Sse41.TestZ(firstVector, asciiMaskForTestZ))
+                {
+                    goto FoundNonAsciiDataInFirstVector;
+                }
+            }
+            else
+            {
+                // See comment earlier in the method for an explanation of how the below logic works.
+                currentMask = (uint)Sse2.MoveMask(Sse2.AddSaturate(firstVector, asciiMaskForAddSaturate).AsByte());
+                if ((currentMask & NonAsciiDataSeenMask) != 0)
+                {
+                    goto FoundNonAsciiDataInCurrentMask;
+                }
             }
 
         IncrementCurrentOffsetBeforeFinalUnalignedVectorRead:
@@ -899,9 +921,23 @@ private static unsafe nuint GetIndexOfFirstNonAsciiChar_Sse2(char* pBuffer, nuin
                 pBuffer = (char*)((byte*)pBuffer + (bufferLength & (SizeOfVector128InBytes - 1)) - SizeOfVector128InBytes);
                 firstVector = Sse2.LoadVector128((ushort*)pBuffer); // unaligned load
 
-                if (VectorContainsNonAsciiChar(firstVector))
+                if (Sse41.IsSupported)
+                {
+                    // If a non-ASCII bit is set in any WORD of the combined vector, we have seen non-ASCII data.
+                    // Jump to the non-ASCII handler to figure out which particular vector contained non-ASCII data.
+                    if (!Sse41.TestZ(firstVector, asciiMaskForTestZ))
+                    {
+                        goto FoundNonAsciiDataInFirstVector;
+                    }
+                }
+                else
                 {
-                     goto FoundNonAsciiDataInFirstVector;
+                    // See comment earlier in the method for an explanation of how the below logic works.
+                    currentMask = (uint)Sse2.MoveMask(Sse2.AddSaturate(firstVector, asciiMaskForAddSaturate).AsByte());
+                    if ((currentMask & NonAsciiDataSeenMask) != 0)
+                    {
+                        goto FoundNonAsciiDataInCurrentMask;
+                    }
                 }
 
                 pBuffer += SizeOfVector128InChars;
@@ -917,23 +953,37 @@ private static unsafe nuint GetIndexOfFirstNonAsciiChar_Sse2(char* pBuffer, nuin
             // We don't know if the first or the second vector contains non-ASCII data. Check the first
             // vector, and if that's all-ASCII then the second vector must be the culprit. Either way
             // we'll make sure the first vector local is the one that contains the non-ASCII data.
-
-            if (VectorContainsNonAsciiChar(firstVector))
+            // See comment earlier in the method for an explanation of how the below logic works.
+            if (Sse41.IsSupported)
             {
-                goto FoundNonAsciiDataInFirstVector;
+                if (!Sse41.TestZ(firstVector, asciiMaskForTestZ))
+                {
+                    goto FoundNonAsciiDataInFirstVector;
+                }
+            }
+            else
+            {
+                currentMask = (uint)Sse2.MoveMask(Sse2.AddSaturate(firstVector, asciiMaskForAddSaturate).AsByte());
+                if ((currentMask & NonAsciiDataSeenMask) != 0)
+                {
+                    goto FoundNonAsciiDataInCurrentMask;
+                }
             }
 
-            // Wasn't the first vector; must be the second.
+                // Wasn't the first vector; must be the second.
 
-            pBuffer += SizeOfVector128InChars;
+                pBuffer += SizeOfVector128InChars;
             firstVector = secondVector;
 
         FoundNonAsciiDataInFirstVector:
 
-            // The operation below forces the 0x8000 bit of each WORD to be set iff the WORD element
-            // has value >= 0x0800 (non-ASCII). Then we'll treat the vector as a BYTE vector in order
-            // to extract the mask. Reminder: the 0x0080 bit of each WORD should be ignored..
+            // See comment earlier in the method for an explanation of how the below logic works.
             currentMask = (uint)Sse2.MoveMask(Sse2.AddSaturate(firstVector, asciiMaskForAddSaturate).AsByte());
+
+        FoundNonAsciiDataInCurrentMask:
+
+            // See comment earlier in the method accounting for the 0x8000 and 0x0080 bits set after the WORD-sized operations.
+
             currentMask &= NonAsciiDataSeenMask;
 
             // Now, the mask contains - from the LSB - a 0b00 pair for each ASCII char we saw, and a 0b10 pair for each non-ASCII char.