Add JIT support for control-flow guard on x64 and arm64

docs/design/coreclr/botr/clr-abi.md

@@ -752,3 +752,43 @@ The return value is handled as follows:
 4. All other cases require the use of a return buffer, through which the value is returned.
 
 In addition, there is a guarantee that if a return buffer is used a value is stored there only upon ordinary exit from the method. The buffer is not allowed to be used for temporary storage within the method and its contents will be unaltered if an exception occurs while executing the method.
+
+# Control Flow Guard (CFG) support on Windows
+
+Control Flow Guard (CFG) is a security mitigation available in Windows.
+When CFG is enabled, the operating system maintains data structures that can be used to verify whether an address is to be considered a valid indirect call target.
+This mechanism is exposed through two different helper functions, each with different characteristics.
+
+The first mechanism is a validator that takes the target address as an argument and fails fast if the address is not an expected indirect call target; otherwise, it does nothing and returns.
+The second mechanism is a dispatcher that takes the target address in a non-standard register; on successful validation of the address, it jumps directly to the target function.
+Windows makes the dispatcher available only on ARM64 and x64, while the validator is available on all platforms.
+However, the JIT supports CFG only on ARM64 and x64, with CFG by default being disabled for these platforms.
+The expected use of the CFG feature is for NativeAOT scenarios that are running in constrained environments where CFG is required.
+
+The helpers are exposed to the JIT as standard JIT helpers `CORINFO_HELP_VALIDATE_INDIRECT_CALL` and `CORINFO_HELP_DISPATCH_INDIRECT_CALL`.
+
+To use the validator the JIT expands indirect calls into a call to the validator followed by a call to the validated address.
+For the dispatcher the JIT will transform calls to pass the target along but otherwise set up the call as normal.
+
+Note that "indirect call" here refers to any call that is not to an immediate (in the instruction stream) address.
+For example, even direct calls may emit indirect call instructions in JIT codegen due to e.g. tiering or if they have not been compiled yet; these are expanded with the CFG mechanism as well.
+
+The next sections describe the calling convention that the JIT expects from these helpers.
+
+## CFG details for ARM64
+
+On ARM64, `CORINFO_HELP_VALIDATE_INDIRECT_CALL` takes the call address in `x15`.
+In addition to the usual registers it preserves all float registers, `x0`-`x8` and `x15`.
+
+`CORINFO_HELP_DISPATCH_INDIRECT_CALL` takes the call address in `x9`.
+The JIT does not use the dispatch helper by default due to worse branch predictor performance.
+Therefore it will expand all indirect calls via the validation helper and a manual call.
+
+## CFG details for x64
+
+On x64, `CORINFO_HELP_VALIDATE_INDIRECT_CALL` takes the call address in `rcx`.
+In addition to the usual registers it also preserves all float registers and `rcx` and `r10`; furthermore, shadow stack space is not required to be allocated.
+
+`CORINFO_HELP_DISPATCH_INDIRECT_CALL` takes the call address in `rax` and it reserves the right to use and trash `r10` and `r11`.
+The JIT uses the dispatch helper on x64 whenever possible as it is expected that the code size benefits outweighs the less accurate branch prediction.
+However, note that the use of `r11` in the dispatcher makes it incompatible with VSD calls where the JIT must fall back to the validator and a manual call.

eng/pipelines/common/templates/runtimes/run-test-job.yml

@@ -197,6 +197,9 @@ jobs:
     - ${{ if in(parameters.testGroup, 'pgo') }}:
       - name: timeoutPerTestCollectionInMinutes
         value: 120
+    - ${{ if in(parameters.testGroup, 'jit-cfg') }}:
+      - name: timeoutPerTestCollectionInMinutes
+        value: 120
 
     - ${{ if eq(parameters.compositeBuildMode, true) }}:
       - name: crossgenArg
@@ -210,7 +213,7 @@ jobs:
     # TODO: update these numbers as they were determined long ago
     ${{ if eq(parameters.testGroup, 'innerloop') }}:
       timeoutInMinutes: 200
-    ${{ if in(parameters.testGroup, 'outerloop', 'jit-experimental', 'pgo') }}:
+    ${{ if in(parameters.testGroup, 'outerloop', 'jit-experimental', 'pgo', 'jit-cfg') }}:
       timeoutInMinutes: 270
     ${{ if in(parameters.testGroup, 'gc-longrunning', 'gc-simulator') }}:
       timeoutInMinutes: 480
@@ -551,6 +554,12 @@ jobs:
           - jitpartialcompilation_osr
           - jitpartialcompilation_osr_pgo
           - jitobjectstackallocation
+        ${{ if in(parameters.testGroup, 'jit-cfg') }}:
+          scenarios:
+          - jitcfg
+          - jitcfg_dispatcher_always
+          - jitcfg_dispatcher_never
+          - jitcfg_gcstress0xc
         ${{ if in(parameters.testGroup, 'ilasm') }}:
           scenarios:
           - ilasmroundtrip

eng/pipelines/coreclr/jit-cfg.yml

@@ -0,0 +1,52 @@
+trigger: none
+
+schedules:
+- cron: "0 22 * * 0,6"
+  displayName: Sun at 2:00 PM (UTC-8:00)
+  branches:
+    include:
+    - main
+  always: true
+
+jobs:
+
+- template: /eng/pipelines/common/platform-matrix.yml
+  parameters:
+    jobTemplate: /eng/pipelines/common/build-coreclr-and-libraries-job.yml
+    buildConfig: checked
+    platforms:
+    - OSX_arm64
+    - OSX_x64
+    - Linux_arm64
+    - Linux_x64
+    - windows_arm64
+    - windows_x64
+    - CoreClrTestBuildHost # Either OSX_x64 or Linux_x64
+    jobParameters:
+      testGroup: jit-cfg
+
+- template: /eng/pipelines/common/platform-matrix.yml
+  parameters:
+    jobTemplate: /eng/pipelines/common/templates/runtimes/build-test-job.yml
+    buildConfig: checked
+    platforms:
+    - CoreClrTestBuildHost # Either OSX_x64 or Linux_x64
+    jobParameters:
+      testGroup: jit-cfg
+
+- template: /eng/pipelines/common/platform-matrix.yml
+  parameters:
+    jobTemplate: /eng/pipelines/common/templates/runtimes/run-test-job.yml
+    buildConfig: checked
+    platforms:
+    - OSX_arm64
+    - OSX_x64
+    - Linux_arm64
+    - Linux_x64
+    - windows_arm64
+    - windows_x64
+    helixQueueGroup: ci
+    helixQueuesTemplate: /eng/pipelines/coreclr/templates/helix-queues-setup.yml
+    jobParameters:
+      testGroup: jit-cfg
+      liveLibrariesBuildConfig: Release

src/coreclr/inc/corinfo.h

@@ -614,6 +614,9 @@ enum CorInfoHelpFunc
     CORINFO_HELP_CLASSPROFILE64,            // Update 64-bit class profile for a call site
     CORINFO_HELP_PARTIAL_COMPILATION_PATCHPOINT,  // Notify runtime that code has reached a part of the method that wasn't originally jitted.
 
+    CORINFO_HELP_VALIDATE_INDIRECT_CALL,    // CFG: Validate function pointer
+    CORINFO_HELP_DISPATCH_INDIRECT_CALL,    // CFG: Validate and dispatch to pointer
+
     CORINFO_HELP_COUNT,
 };
 

src/coreclr/inc/corjitflags.h

@@ -32,7 +32,7 @@ class CORJIT_FLAGS
         CORJIT_FLAG_DEBUG_EnC               = 3, // We are in Edit-n-Continue mode
         CORJIT_FLAG_DEBUG_INFO              = 4, // generate line and local-var info
         CORJIT_FLAG_MIN_OPT                 = 5, // disable all jit optimizations (not necesarily debuggable code)
-        CORJIT_FLAG_UNUSED1                 = 6,
+        CORJIT_FLAG_ENABLE_CFG              = 6, // generate control-flow guard checks
         CORJIT_FLAG_MCJIT_BACKGROUND        = 7, // Calling from multicore JIT background thread, do not call JitComplete
 
     #if defined(TARGET_X86)

src/coreclr/inc/jiteeversionguid.h

@@ -43,13 +43,14 @@ typedef const GUID *LPCGUID;
 #define GUID_DEFINED
 #endif // !GUID_DEFINED
 
-constexpr GUID JITEEVersionIdentifier = { /* ccb0c159-04b3-47f6-993e-79114c9cbef8 */
-    0xccb0c159,
-    0x04b3,
-    0x47f6,
-    {0x99, 0x3e, 0x79, 0x11, 0x4c, 0x9c, 0xbe, 0xf8}
+constexpr GUID JITEEVersionIdentifier = { /* 63009f0c-662a-485b-bac1-ff67be6c7f9d */
+    0x63009f0c,
+    0x662a,
+    0x485b,
+    {0xba, 0xc1, 0xff, 0x67, 0xbe, 0x6c, 0x7f, 0x9d} 
   };
 
+
 //////////////////////////////////////////////////////////////////////////////////////////////////////////
 //
 // END JITEEVersionIdentifier

src/coreclr/inc/jithelpers.h

@@ -351,6 +351,14 @@
     JITHELPER(CORINFO_HELP_CLASSPROFILE64, JIT_ClassProfile64, CORINFO_HELP_SIG_REG_ONLY)
     JITHELPER(CORINFO_HELP_PARTIAL_COMPILATION_PATCHPOINT, JIT_PartialCompilationPatchpoint, CORINFO_HELP_SIG_REG_ONLY)
 
+#if defined(TARGET_AMD64) || defined(TARGET_ARM64)
+    JITHELPER(CORINFO_HELP_VALIDATE_INDIRECT_CALL, JIT_ValidateIndirectCall, CORINFO_HELP_SIG_REG_ONLY)
+    JITHELPER(CORINFO_HELP_DISPATCH_INDIRECT_CALL, JIT_DispatchIndirectCall, CORINFO_HELP_SIG_REG_ONLY)
+#else
+    JITHELPER(CORINFO_HELP_VALIDATE_INDIRECT_CALL, NULL, CORINFO_HELP_SIG_REG_ONLY)
+    JITHELPER(CORINFO_HELP_DISPATCH_INDIRECT_CALL, NULL, CORINFO_HELP_SIG_REG_ONLY)
+#endif
+
 #undef JITHELPER
 #undef DYNAMICJITHELPER
 #undef JITHELPER

src/coreclr/jit/codegenarmarch.cpp

@@ -3134,9 +3134,6 @@ void CodeGen::genCall(GenTreeCall* call)
     // into a volatile register that won't be restored by epilog sequence.
     if (call->IsFastTailCall())
     {
-        // Don't support fast tail calling JIT helpers
-        assert(call->gtCallType != CT_HELPER);
-
         GenTree* target = getCallTarget(call, nullptr);
 
         if (target != nullptr)
@@ -3177,22 +3174,28 @@ void CodeGen::genCall(GenTreeCall* call)
 
     genCallInstruction(call);
 
-    // if it was a pinvoke we may have needed to get the address of a label
-    if (genPendingCallLabel)
+    // for pinvoke/intrinsic/tailcalls we may have needed to get the address of
+    // a label. In case it is indirect with CFG enabled make sure we do not get
+    // the address after the validation but only after the actual call that
+    // comes after.
+    if (genPendingCallLabel && !call->IsHelperCall(compiler, CORINFO_HELP_VALIDATE_INDIRECT_CALL))
     {
         genDefineInlineTempLabel(genPendingCallLabel);
         genPendingCallLabel = nullptr;
     }
 
-    // Update GC info:
-    // All Callee arg registers are trashed and no longer contain any GC pointers.
-    // TODO-Bug?: As a matter of fact shouldn't we be killing all of callee trashed regs here?
-    // For now we will assert that other than arg regs gc ref/byref set doesn't contain any other
-    // registers from RBM_CALLEE_TRASH
-    assert((gcInfo.gcRegGCrefSetCur & (RBM_CALLEE_TRASH & ~RBM_ARG_REGS)) == 0);
-    assert((gcInfo.gcRegByrefSetCur & (RBM_CALLEE_TRASH & ~RBM_ARG_REGS)) == 0);
-    gcInfo.gcRegGCrefSetCur &= ~RBM_ARG_REGS;
-    gcInfo.gcRegByrefSetCur &= ~RBM_ARG_REGS;
+#ifdef DEBUG
+    // Killed registers should no longer contain any GC pointers.
+    regMaskTP killMask = RBM_CALLEE_TRASH;
+    if (call->IsHelperCall())
+    {
+        CorInfoHelpFunc helpFunc = compiler->eeGetHelperNum(call->gtCallMethHnd);
+        killMask                 = compiler->compHelperCallKillSet(helpFunc);
+    }
+
+    assert((gcInfo.gcRegGCrefSetCur & killMask) == 0);
+    assert((gcInfo.gcRegByrefSetCur & killMask) == 0);
+#endif
 
     var_types returnType = call->TypeGet();
     if (returnType != TYP_VOID)

src/coreclr/jit/codegencommon.cpp

@@ -661,6 +661,9 @@ regMaskTP Compiler::compHelperCallKillSet(CorInfoHelpFunc helper)
         case CORINFO_HELP_INIT_PINVOKE_FRAME:
             return RBM_INIT_PINVOKE_FRAME_TRASH;
 
+        case CORINFO_HELP_VALIDATE_INDIRECT_CALL:
+            return RBM_VALIDATE_INDIRECT_CALL_TRASH;
+
         default:
             return RBM_CALLEE_TRASH;
     }
@@ -2204,6 +2207,11 @@ void CodeGen::genGenerateMachineCode()
                    compiler->fgPgoInlineePgo, compiler->fgPgoInlineeNoPgoSingleBlock, compiler->fgPgoInlineeNoPgo);
         }
 
+        if (compiler->opts.IsCFGEnabled())
+        {
+            printf("; control-flow guard enabled\n");
+        }
+
         if (compiler->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_ALT_JIT))
         {
             printf("; invoked as altjit\n");

src/coreclr/jit/codegenxarch.cpp

@@ -5218,9 +5218,6 @@ void CodeGen::genCall(GenTreeCall* call)
     // that won't be restored by epilog sequence.
     if (call->IsFastTailCall())
     {
-        // Don't support fast tail calling JIT helpers
-        assert(call->gtCallType != CT_HELPER);
-
         GenTree* target = getCallTarget(call, nullptr);
         if (target != nullptr)
         {
@@ -5272,22 +5269,28 @@ void CodeGen::genCall(GenTreeCall* call)
 
     genCallInstruction(call X86_ARG(stackArgBytes));
 
-    // if it was a pinvoke or intrinsic we may have needed to get the address of a label
-    if (genPendingCallLabel)
+    // for pinvoke/intrinsic/tailcalls we may have needed to get the address of
+    // a label. In case it is indirect with CFG enabled make sure we do not get
+    // the address after the validation but only after the actual call that
+    // comes after.
+    if (genPendingCallLabel && !call->IsHelperCall(compiler, CORINFO_HELP_VALIDATE_INDIRECT_CALL))
     {
         genDefineInlineTempLabel(genPendingCallLabel);
         genPendingCallLabel = nullptr;
     }
 
-    // Update GC info:
-    // All Callee arg registers are trashed and no longer contain any GC pointers.
-    // TODO-XArch-Bug?: As a matter of fact shouldn't we be killing all of callee trashed regs here?
-    // For now we will assert that other than arg regs gc ref/byref set doesn't contain any other
-    // registers from RBM_CALLEE_TRASH.
-    assert((gcInfo.gcRegGCrefSetCur & (RBM_CALLEE_TRASH & ~RBM_ARG_REGS)) == 0);
-    assert((gcInfo.gcRegByrefSetCur & (RBM_CALLEE_TRASH & ~RBM_ARG_REGS)) == 0);
-    gcInfo.gcRegGCrefSetCur &= ~RBM_ARG_REGS;
-    gcInfo.gcRegByrefSetCur &= ~RBM_ARG_REGS;
+#ifdef DEBUG
+    // Killed registers should no longer contain any GC pointers.
+    regMaskTP killMask = RBM_CALLEE_TRASH;
+    if (call->IsHelperCall())
+    {
+        CorInfoHelpFunc helpFunc = compiler->eeGetHelperNum(call->gtCallMethHnd);
+        killMask                 = compiler->compHelperCallKillSet(helpFunc);
+    }
+
+    assert((gcInfo.gcRegGCrefSetCur & killMask) == 0);
+    assert((gcInfo.gcRegByrefSetCur & killMask) == 0);
+#endif
 
     var_types returnType = call->TypeGet();
     if (returnType != TYP_VOID)
@@ -5563,6 +5566,11 @@ void CodeGen::genCallInstruction(GenTreeCall* call X86_ARG(target_ssize_t stackA
 #endif
             if (target->isContainedIndir())
         {
+            // When CFG is enabled we should not be emitting any non-register indirect calls.
+            assert(!compiler->opts.IsCFGEnabled() ||
+                   call->IsHelperCall(compiler, CORINFO_HELP_VALIDATE_INDIRECT_CALL) ||
+                   call->IsHelperCall(compiler, CORINFO_HELP_DISPATCH_INDIRECT_CALL));
+
             if (target->AsIndir()->HasBase() && target->AsIndir()->Base()->isContainedIntOrIImmed())
             {
                 // Note that if gtControlExpr is an indir of an absolute address, we mark it as

src/coreclr/jit/compiler.cpp

@@ -8845,6 +8845,20 @@ void dTreeLIR(GenTree* tree)
     cTreeLIR(JitTls::GetCompiler(), tree);
 }
 
+void dTreeRange(GenTree* first, GenTree* last)
+{
+    Compiler* comp = JitTls::GetCompiler();
+    GenTree*  cur  = first;
+    while (true)
+    {
+        cTreeLIR(comp, cur);
+        if (cur == last)
+            break;
+
+        cur = cur->gtNext;
+    }
+}
+
 void dTrees()
 {
     cTrees(JitTls::GetCompiler());