diff --git a/src/sys/driver.h b/src/sys/driver.h
index f8cfd27d..905dd6f5 100644
--- a/src/sys/driver.h
+++ b/src/sys/driver.h
@@ -566,7 +566,7 @@ BOOLEAN FspFileNameIsPrefix(
 #else
 #define FspFileNameUpcase(D,S,U)        (ASSERT(0 == (U)), RtlUpcaseUnicodeString(D,S,FALSE))
 #define FspEaNameUpcase(D,S,U)          (ASSERT(0 == (U)), RtlUpperString(D,S))
-#define FspFileNameCompare(N1,N2,I,U)   (ASSERT(0 == (U)), RtlCompareUnicodeString(N1,N2,I))
+#define FspFileNameCompare(N1,N2,I,U)   (ASSERT(0 == (U)), FspCompareUnicodeString(N1,N2,I))
 #define FspFileNameIsPrefix(N1,N2,I,U)  (ASSERT(0 == (U)), RtlPrefixUnicodeString(N1,N2,I))
 #endif
 NTSTATUS FspFileNameInExpression(
@@ -733,6 +733,12 @@ VOID FspIrpHookReset(PIRP Irp);
 PVOID FspIrpHookContext(PVOID Context);
 NTSTATUS FspIrpHookNext(PDEVICE_OBJECT DeviceObject, PIRP Irp, PVOID Context);
 
+/* utility: string compare */
+LONG FspCompareUnicodeString(
+    PCUNICODE_STRING String1,
+    PCUNICODE_STRING String2,
+    BOOLEAN CaseInsensitive);
+
 /* silos */
 typedef struct
 {
diff --git a/src/sys/util.c b/src/sys/util.c
index 96f63363..fda82b68 100644
--- a/src/sys/util.c
+++ b/src/sys/util.c
@@ -131,6 +131,10 @@ NTSTATUS FspIrpHook(PIRP Irp, PIO_COMPLETION_ROUTINE CompletionRoutine, PVOID Ow
 VOID FspIrpHookReset(PIRP Irp);
 PVOID FspIrpHookContext(PVOID Context);
 NTSTATUS FspIrpHookNext(PDEVICE_OBJECT DeviceObject, PIRP Irp, PVOID Context);
+LONG FspCompareUnicodeString(
+    PCUNICODE_STRING String1,
+    PCUNICODE_STRING String2,
+    BOOLEAN CaseInsensitive);
 
 #ifdef ALLOC_PRAGMA
 #pragma alloc_text(PAGE, FspIsNtDdiVersionAvailable)
@@ -174,6 +178,7 @@ NTSTATUS FspIrpHookNext(PDEVICE_OBJECT DeviceObject, PIRP Irp, PVOID Context);
 #pragma alloc_text(PAGE, FspSafeMdlDelete)
 #pragma alloc_text(PAGE, FspIrpHook)
 #pragma alloc_text(PAGE, FspIrpHookReset)
+#pragma alloc_text(PAGE, FspCompareUnicodeString)
 #endif
 
 static const LONG Delays[] =
@@ -1493,3 +1498,82 @@ NTSTATUS FspIrpHookNext(PDEVICE_OBJECT DeviceObject, PIRP Irp, PVOID Context)
 
     return Result;
 }
+
+static inline
+USHORT FspUpcaseAscii(USHORT c)
+{
+    /*
+     * Bit-twiddling upper case char:
+     *
+     * - Let signbit(x) = x & 0x100 (treat bit 0x100 as "signbit").
+     * - 'A' <= c && c <= 'Z' <=> s = signbit(c - 'A') ^ signbit(c - ('Z' + 1)) == 1
+     *     - c >= 'A' <=> c - 'A' >= 0      <=> signbit(c - 'A') = 0
+     *     - c <= 'Z' <=> c - ('Z' + 1) < 0 <=> signbit(c - ('Z' + 1)) = 1
+     * - Bit 0x20 = 0x100 >> 3 toggles uppercase to lowercase and vice-versa.
+     *
+     * This is actually faster than `(c - 'a' <= 'z' - 'a') ? (c & ~0x20) : c`, even
+     * when compiled using cmov conditional moves at least on this system (i7-1065G7).
+     *
+     * See https://godbolt.org/z/ebv131Wrh
+     */
+    USHORT s = ((c - 'a') ^ (c - ('z' + 1))) & 0x100;
+    return c & ~(s >> 3);
+}
+
+#if DBG
+static
+LONG FspCompareUnicodeStringReal(
+    PCUNICODE_STRING S1,
+    PCUNICODE_STRING S2,
+    BOOLEAN CaseInsensitive)
+#else
+LONG FspCompareUnicodeString(
+    PCUNICODE_STRING S1,
+    PCUNICODE_STRING S2,
+    BOOLEAN CaseInsensitive)
+#endif
+{
+    PAGED_CODE();
+
+    LONG LResult = S1->Length - S2->Length;
+    PWCH P1 = S1->Buffer;
+    PWCH P2 = S2->Buffer;
+    PWCH EndP1 = P1 + (0 >= LResult ? S1->Length : S2->Length) / sizeof(WCHAR);
+
+    if (CaseInsensitive)
+    {
+        for (; EndP1 != P1; ++P1, ++P2)
+        {
+            USHORT C1 = *P1, C2 = *P2;
+            if (0xff80 & (C1 | C2))
+                return RtlCompareUnicodeString(S1, S2, TRUE);
+            C1 = FspUpcaseAscii(C1);
+            C2 = FspUpcaseAscii(C2);
+            if (C1 != C2)
+                return C1 - C2;
+        }
+    }
+    else
+    {
+        for (; EndP1 != P1; ++P1, ++P2)
+        {
+            USHORT C1 = *P1, C2 = *P2;
+            if (C1 != C2)
+                return C1 - C2;
+        }
+    }
+
+    return LResult;
+}
+#if DBG
+LONG FspCompareUnicodeString(
+    PCUNICODE_STRING S1,
+    PCUNICODE_STRING S2,
+    BOOLEAN CaseInsensitive)
+{
+    LONG Result0 = FspCompareUnicodeStringReal(S1, S2, CaseInsensitive);
+    LONG Result1 = RtlCompareUnicodeString(S1, S2, CaseInsensitive);
+    ASSERT((0 < Result0) - (0 > Result0) == (0 < Result1) - (0 > Result1));
+    return Result0;
+}
+#endif
diff --git a/tools/run-perf-tests.bat b/tools/run-perf-tests.bat
index 0c3409b0..13e5e506 100755
--- a/tools/run-perf-tests.bat
+++ b/tools/run-perf-tests.bat
@@ -30,25 +30,25 @@ pushd fsbench
 set OptFiles=1000 2000 3000 4000 5000
 if X%2==Xbaseline set OptFiles=10000
 for %%a in (%OptFiles%) do (
-    call :csv %%a "%fsbench% --empty-cache --files=%%a file_*"
+    call :csv %%a "%fsbench% --empty-cache=C --files=%%a file_*"
 )
 
 set OptRdwrCc=100 200 300 400 500
 if X%2==Xbaseline set OptRdwrCc=1000
 for %%a in (%OptRdwrCc%) do (
-    call :csv %%a "%fsbench% --empty-cache --rdwr-cc=%%a rdwr_cc_*"
+    call :csv %%a "%fsbench% --empty-cache=C --rdwr-cc=%%a rdwr_cc_*"
 )
 
 set OptRdwrNc=100 200 300 400 500
 if X%2==Xbaseline set OptRdwrNc=100
 for %%a in (%OptRdwrNc%) do (
-    call :csv %%a "%fsbench% --empty-cache --rdwr-nc=%%a rdwr_nc_*"
+    call :csv %%a "%fsbench% --empty-cache=C --rdwr-nc=%%a rdwr_nc_*"
 )
 
 set OptMmap=100 200 300 400 500
 if X%2==Xbaseline set OptMmap=1000
 for %%a in (%OptMmap%) do (
-    call :csv %%a "%fsbench% --empty-cache --mmap=%%a mmap_*"
+    call :csv %%a "%fsbench% --empty-cache=C --mmap=%%a mmap_*"
 )
 
 popd