[AArch64] Implement GCS ACLE intrinsics (#96903)

john-brawn-arm · web-flow · commit 3a14ffbae39b · 2024-07-11T14:09:36.000+01:00
This adds intrinsics defined in ARM-software/acle#260 Doing this requires some changes to the GCS instruction definitions, as these intrinsics make use of how some instructions don't modify the input register when GCS is disabled, and they need to be correctly marked with mayLoad/mayStore/hasSideEffects for instruction selection to work.
diff --git a/clang/include/clang/Basic/BuiltinsAArch64.def b/clang/include/clang/Basic/BuiltinsAArch64.def
@@ -49,6 +49,7 @@ BUILTIN(__builtin_arm_wfe, "v", "")
 BUILTIN(__builtin_arm_wfi, "v", "")
 BUILTIN(__builtin_arm_sev, "v", "")
 BUILTIN(__builtin_arm_sevl, "v", "")
+BUILTIN(__builtin_arm_chkfeat, "WUiWUi", "")
 
 // Like __builtin_trap but provide an 16-bit immediate reason code (which goes into `brk #N`).
 BUILTIN(__builtin_arm_trap, "vUIs", "nr")
@@ -136,6 +137,10 @@ TARGET_BUILTIN(__builtin_arm_st64b, "vv*WUiC*", "n", "ls64")
 TARGET_BUILTIN(__builtin_arm_st64bv, "WUiv*WUiC*", "n", "ls64")
 TARGET_BUILTIN(__builtin_arm_st64bv0, "WUiv*WUiC*", "n", "ls64")
 
+// Armv9.3-A Guarded Control Stack
+TARGET_BUILTIN(__builtin_arm_gcspopm, "WUiWUi", "n", "gcs")
+TARGET_BUILTIN(__builtin_arm_gcsss, "vC*vC*", "n", "gcs")
+
 TARGET_HEADER_BUILTIN(_BitScanForward, "UcUNi*UNi", "nh", INTRIN_H, ALL_MS_LANGUAGES, "")
 TARGET_HEADER_BUILTIN(_BitScanReverse, "UcUNi*UNi", "nh", INTRIN_H, ALL_MS_LANGUAGES, "")
 TARGET_HEADER_BUILTIN(_BitScanForward64, "UcUNi*ULLi", "nh", INTRIN_H, ALL_MS_LANGUAGES, "")
diff --git a/clang/lib/Headers/arm_acle.h b/clang/lib/Headers/arm_acle.h
@@ -75,6 +75,14 @@ static __inline__ void __attribute__((__always_inline__, __nodebug__)) __yield(v
 #define __dbg(t) __builtin_arm_dbg(t)
 #endif
 
+#if defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE
+#define _CHKFEAT_GCS 1
+static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__))
+__chkfeat(uint64_t __features) {
+  return __builtin_arm_chkfeat(__features) ^ __features;
+}
+#endif
+
 /* 7.5 Swap */
 static __inline__ uint32_t __attribute__((__always_inline__, __nodebug__))
 __swp(uint32_t __x, volatile uint32_t *__p) {
@@ -855,6 +863,24 @@ __rndrrs(uint64_t *__p) {
 }
 #endif
 
+/* 11.2 Guarded Control Stack intrinsics */
+#if defined(__ARM_64BIT_STATE) && __ARM_64BIT_STATE
+static __inline__ void * __attribute__((__always_inline__, __nodebug__))
+__gcspr() {
+  return (void *)__builtin_arm_rsr64("gcspr_el0");
+}
+
+static __inline__ uint64_t __attribute__((__always_inline__, __nodebug__, target("gcs")))
+__gcspopm() {
+  return __builtin_arm_gcspopm(0);
+}
+
+static __inline__ const void * __attribute__((__always_inline__, __nodebug__, target("gcs")))
+__gcsss(const void *__stack) {
+  return __builtin_arm_gcsss(__stack);
+}
+#endif
+
 #if defined(__cplusplus)
 }
 #endif
diff --git a/clang/test/CodeGen/aarch64-gcs.c b/clang/test/CodeGen/aarch64-gcs.c
@@ -0,0 +1,56 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 2
+// RUN: %clang_cc1 -triple aarch64-eabi -target-feature +gcs -emit-llvm %s -o - | FileCheck %s
+
+#include <arm_acle.h>
+
+// CHECK-LABEL: define dso_local i64 @test_chkfeat
+// CHECK-SAME: () #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[__FEATURES_ADDR_I:%.*]] = alloca i64, align 8
+// CHECK-NEXT:    store i64 1, ptr [[__FEATURES_ADDR_I]], align 8
+// CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr [[__FEATURES_ADDR_I]], align 8
+// CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.aarch64.chkfeat(i64 [[TMP0]])
+// CHECK-NEXT:    [[TMP2:%.*]] = load i64, ptr [[__FEATURES_ADDR_I]], align 8
+// CHECK-NEXT:    [[XOR_I:%.*]] = xor i64 [[TMP1]], [[TMP2]]
+// CHECK-NEXT:    ret i64 [[XOR_I]]
+//
+uint64_t test_chkfeat() {
+  return __chkfeat(_CHKFEAT_GCS);
+}
+
+// CHECK-LABEL: define dso_local ptr @test_gcspr
+// CHECK-SAME: () #[[ATTR0]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.read_volatile_register.i64(metadata [[META2:![0-9]+]])
+// CHECK-NEXT:    [[TMP1:%.*]] = inttoptr i64 [[TMP0]] to ptr
+// CHECK-NEXT:    ret ptr [[TMP1]]
+//
+void *test_gcspr() {
+  return __gcspr();
+}
+
+// CHECK-LABEL: define dso_local i64 @test_gcspopm
+// CHECK-SAME: () #[[ATTR0]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.aarch64.gcspopm(i64 0)
+// CHECK-NEXT:    ret i64 [[TMP0]]
+//
+uint64_t test_gcspopm() {
+  return __gcspopm();
+}
+
+// CHECK-LABEL: define dso_local ptr @test_gcsss
+// CHECK-SAME: (ptr noundef [[P:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[__STACK_ADDR_I:%.*]] = alloca ptr, align 8
+// CHECK-NEXT:    [[P_ADDR:%.*]] = alloca ptr, align 8
+// CHECK-NEXT:    store ptr [[P]], ptr [[P_ADDR]], align 8
+// CHECK-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[P_ADDR]], align 8
+// CHECK-NEXT:    store ptr [[TMP0]], ptr [[__STACK_ADDR_I]], align 8
+// CHECK-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__STACK_ADDR_I]], align 8
+// CHECK-NEXT:    [[TMP2:%.*]] = call ptr @llvm.aarch64.gcsss(ptr [[TMP1]])
+// CHECK-NEXT:    ret ptr [[TMP2]]
+//
+const void *test_gcsss(const void *p) {
+  return __gcsss(p);
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -92,6 +92,23 @@ def int_aarch64_isb : ClangBuiltin<"__builtin_arm_isb">, MSBuiltin<"__isb">,
 // ordering during ISel.
 def int_aarch64_space : DefaultAttrsIntrinsic<[llvm_i64_ty], [llvm_i32_ty, llvm_i64_ty], []>;
 
+//===----------------------------------------------------------------------===//
+// Guarded Control Stack
+
+def int_aarch64_chkfeat : ClangBuiltin<"__builtin_arm_chkfeat">,
+                          DefaultAttrsIntrinsic<[llvm_i64_ty], [llvm_i64_ty],
+                                                [IntrNoMem]>;
+
+// FIXME: This should be marked as [IntrReadMem, IntrHasSideEffects], as it has
+// the side-effect of updating gcspr, but this combination doesn't work
+// correctly.
+def int_aarch64_gcspopm : ClangBuiltin<"__builtin_arm_gcspopm">,
+                          DefaultAttrsIntrinsic<[llvm_i64_ty], [llvm_i64_ty],
+                                                []>;
+
+def int_aarch64_gcsss : ClangBuiltin<"__builtin_arm_gcsss">,
+                        DefaultAttrsIntrinsic<[llvm_ptr_ty], [llvm_ptr_ty], []>;
+
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp b/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp
@@ -4587,6 +4587,18 @@ void AArch64DAGToDAGISel::Select(SDNode *Node) {
     switch (IntNo) {
     default:
       break;
+    case Intrinsic::aarch64_gcsss: {
+      SDLoc DL(Node);
+      SDValue Chain = Node->getOperand(0);
+      SDValue Val = Node->getOperand(2);
+      SDValue Zero = CurDAG->getCopyFromReg(Chain, DL, AArch64::XZR, MVT::i64);
+      SDNode *SS1 =
+          CurDAG->getMachineNode(AArch64::GCSSS1, DL, MVT::Other, Val, Chain);
+      SDNode *SS2 = CurDAG->getMachineNode(AArch64::GCSSS2, DL, MVT::i64,
+                                           MVT::Other, Zero, SDValue(SS1, 0));
+      ReplaceNode(Node, SS2);
+      return;
+    }
     case Intrinsic::aarch64_ldaxp:
     case Intrinsic::aarch64_ldxp: {
       unsigned Op =
diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.td b/llvm/lib/Target/AArch64/AArch64InstrInfo.td
@@ -1267,23 +1267,34 @@ class GCSRtIn<bits<3> op1, bits<3> op2, string mnemonic,
   let Inst{15-8} = 0b01110111;
   let Inst{7-5} = op2;
   let Predicates = [HasGCS];
+  let hasSideEffects = 1;
 }
 
+let mayStore = 1, mayLoad = 1 in
 def GCSSS1   : GCSRtIn<0b011, 0b010, "gcsss1">;
+let mayStore = 1 in
 def GCSPUSHM : GCSRtIn<0b011, 0b000, "gcspushm">;
 
 class GCSRtOut<bits<3> op1, bits<3> op2, string mnemonic,
             list<dag> pattern = []>
-    : RtSystemI<1, (outs GPR64:$Rt), (ins), mnemonic, "\t$Rt", pattern> {
+    : RtSystemI<1, (outs GPR64:$Rt), (ins GPR64:$src), mnemonic, "\t$Rt", pattern> {
   let Inst{20-19} = 0b01;
   let Inst{18-16} = op1;
   let Inst{15-8} = 0b01110111;
   let Inst{7-5} = op2;
   let Predicates = [HasGCS];
+  let hasSideEffects = 1;
+  // The input register is unchanged when GCS is disabled, so we need it as
+  // both an input and output operand.
+  let Constraints = "$src = $Rt";
 }
 
+let mayStore = 1, mayLoad = 1 in
 def GCSSS2  : GCSRtOut<0b011, 0b011, "gcsss2">;
-def GCSPOPM : GCSRtOut<0b011, 0b001, "gcspopm">;
+// FIXME: mayStore = 1 only needed to match the intrinsic definition
+let mayStore = 1, mayLoad = 1 in
+def GCSPOPM : GCSRtOut<0b011, 0b001, "gcspopm",
+                       [(set GPR64:$Rt, (int_aarch64_gcspopm GPR64:$src))]>;
 def GCSPOPM_NoOp : InstAlias<"gcspopm", (GCSPOPM XZR)>, Requires<[HasGCS]>; // Rt defaults to XZR if absent
 
 def GCSB_DSYNC_disable : InstAlias<"gcsb\tdsync", (HINT 19), 0>;
@@ -1292,7 +1303,8 @@ def GCSB_DSYNC         : InstAlias<"gcsb\tdsync", (HINT 19), 1>, Requires<[HasGC
 def : TokenAlias<"DSYNC", "dsync">;
 
 let Uses = [X16], Defs = [X16], CRm = 0b0101 in {
-  def CHKFEAT   : SystemNoOperands<0b000, "hint\t#40">;
+  def CHKFEAT   : SystemNoOperands<0b000, "hint\t#40",
+                                   [(set X16, (int_aarch64_chkfeat X16))]>;
 }
 def : InstAlias<"chkfeat\tx16", (CHKFEAT), 0>;
 def : InstAlias<"chkfeat\tx16", (CHKFEAT), 1>, Requires<[HasCHK]>;
@@ -1311,7 +1323,6 @@ class GCSSt<string mnemonic, bits<3> op>
 def GCSSTR  : GCSSt<"gcsstr",  0b000>;
 def GCSSTTR : GCSSt<"gcssttr", 0b001>;
 
-
 // ARMv8.2-A Dot Product
 let Predicates = [HasDotProd] in {
 defm SDOT : SIMDThreeSameVectorDot<0, 0, "sdot", AArch64sdot>;
diff --git a/llvm/test/CodeGen/AArch64/gcs-intrinsics.ll b/llvm/test/CodeGen/AArch64/gcs-intrinsics.ll
@@ -0,0 +1,54 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64 -mattr=+gcs -verify-machineinstrs -o - %s | FileCheck %s
+
+; We call each intrinsic twice, once with the result being unused and once with
+; it being used, to check that dead code elimination is being done correctly.
+; chkfeat does not have side effects so can be eliminated, but the others do and
+; can't be eliminated.
+
+define i64 @test_chkfeat(i64 %arg) {
+; CHECK-LABEL: test_chkfeat:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    mov x16, x0
+; CHECK-NEXT:    chkfeat x16
+; CHECK-NEXT:    mov x0, x16
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i64 @llvm.aarch64.chkfeat(i64 %arg)
+  %1 = call i64 @llvm.aarch64.chkfeat(i64 %arg)
+  ret i64 %1
+}
+
+define i64 @test_gcspopm(i64 %arg) {
+; CHECK-LABEL: test_gcspopm:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    mov x8, x0
+; CHECK-NEXT:    gcspopm x8
+; CHECK-NEXT:    gcspopm x0
+; CHECK-NEXT:    ret
+entry:
+  %0 = call i64 @llvm.aarch64.gcspopm(i64 %arg)
+  %1 = call i64 @llvm.aarch64.gcspopm(i64 %arg)
+  ret i64 %1
+}
+
+define ptr @test_gcsss(ptr %p) {
+; CHECK-LABEL: test_gcsss:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    mov x9, xzr
+; CHECK-NEXT:    gcsss1 x0
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:    gcsss2 x9
+; CHECK-NEXT:    gcsss1 x0
+; CHECK-NEXT:    gcsss2 x8
+; CHECK-NEXT:    mov x0, x8
+; CHECK-NEXT:    ret
+entry:
+  %0 = call ptr @llvm.aarch64.gcsss(ptr %p)
+  %1 = call ptr @llvm.aarch64.gcsss(ptr %p)
+  ret ptr %1
+}
+
+declare i64 @llvm.aarch64.chkfeat(i64)
+declare i64 @llvm.aarch64.gcspopm(i64)
+declare ptr @llvm.aarch64.gcsss(ptr)
