[clang] [aarch64] Add more MSVC intrinsics (PR #194632)
Adhemerval Zanella via cfe-commits
cfe-commits at lists.llvm.org
Tue Apr 28 06:50:02 PDT 2026
https://github.com/zatrazz created https://github.com/llvm/llvm-project/pull/194632
This patchset adds more intrinsics from MSVC build 2026 [1]. I focused on the one documented by Microsoft, although intrin.h and related headers do export more symbols.
This is split from https://github.com/llvm/llvm-project/pull/193221 , where I only added the non-controversial ones.
>From 29f66c50d6632b941946bf48faab42fc4dbd200b Mon Sep 17 00:00:00 2001
From: Adhemerval Zanella <zatrazz at gmail.com>
Date: Thu, 9 Apr 2026 16:24:34 -0300
Subject: [PATCH 1/6] [aarch64] Add support for the _CountTrailingZeros{64} MS
intrinsics
Adds support for the following MSVC intrinsics:
* `_CountTrailingZeros`
* `_CountTrailingZeros64`
These are documented at:
<https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-180>
---
clang/include/clang/Basic/BuiltinsAArch64.td | 2 ++
.../lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp | 4 +++-
clang/lib/CodeGen/TargetBuiltins/ARM.cpp | 12 ++++++++++
clang/lib/Headers/intrin.h | 2 ++
.../test/CodeGen/arm64-microsoft-intrinsics.c | 23 +++++++++++++++++++
5 files changed, 42 insertions(+), 1 deletion(-)
diff --git a/clang/include/clang/Basic/BuiltinsAArch64.td b/clang/include/clang/Basic/BuiltinsAArch64.td
index 8e5dbb5f4172d..502652e208dff 100644
--- a/clang/include/clang/Basic/BuiltinsAArch64.td
+++ b/clang/include/clang/Basic/BuiltinsAArch64.td
@@ -394,6 +394,8 @@ let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES",
def _CountLeadingZeros64 : AArch64NoPrefixTargetLibBuiltin<"unsigned int (unsigned long long int)">;
def _CountOneBits : AArch64NoPrefixTargetLibBuiltin<"unsigned int (msuint32_t)">;
def _CountOneBits64 : AArch64NoPrefixTargetLibBuiltin<"unsigned int (unsigned long long int)">;
+ def _CountTrailingZeros : AArch64NoPrefixTargetLibBuiltin<"unsigned int (msuint32_t)">;
+ def _CountTrailingZeros64: AArch64NoPrefixTargetLibBuiltin<"unsigned int (unsigned long long int)">;
}
let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES", Header = "intrin.h" in {
diff --git a/clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp b/clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp
index cac5f8eced8a7..d563ccbf1a38e 100644
--- a/clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp
@@ -2064,7 +2064,9 @@ CIRGenFunction::emitAArch64BuiltinExpr(unsigned builtinID, const CallExpr *expr,
}
if (builtinID == AArch64::BI_CountOneBits ||
- builtinID == AArch64::BI_CountOneBits64) {
+ builtinID == AArch64::BI_CountOneBits64 ||
+ builtinID == AArch64::BI_CountTrailingZeros ||
+ builtinID == AArch64::BI_CountTrailingZeros64) {
cgm.errorNYI(expr->getSourceRange(),
std::string("unimplemented AArch64 builtin call: ") +
getContext().BuiltinInfo.getName(builtinID));
diff --git a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
index f8d383b15313f..c2744ede7e647 100644
--- a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
@@ -5235,6 +5235,18 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
return Result;
}
+ if (BuiltinID == AArch64::BI_CountTrailingZeros ||
+ BuiltinID == AArch64::BI_CountTrailingZeros64) {
+ Value *ArgValue = EmitScalarExpr(E->getArg(0));
+ llvm::Type *ArgType = ArgValue->getType();
+ Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
+
+ Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getInt1(false)});
+ if (BuiltinID == AArch64::BI_CountTrailingZeros64)
+ Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
+ return Result;
+ }
+
if (BuiltinID == AArch64::BI__prefetch) {
Value *Address = EmitScalarExpr(E->getArg(0));
Value *RW = llvm::ConstantInt::get(Int32Ty, 0);
diff --git a/clang/lib/Headers/intrin.h b/clang/lib/Headers/intrin.h
index 210ed0c1f773b..7b5698763d1ca 100644
--- a/clang/lib/Headers/intrin.h
+++ b/clang/lib/Headers/intrin.h
@@ -438,6 +438,8 @@ unsigned int _CountLeadingSigns(long);
unsigned int _CountLeadingSigns64(__int64);
unsigned int _CountOneBits(unsigned long);
unsigned int _CountOneBits64(unsigned __int64);
+unsigned int _CountTrailingZeros(unsigned long);
+unsigned int _CountTrailingZeros64(unsigned __int64);
unsigned int __hlt(unsigned int, ...);
diff --git a/clang/test/CodeGen/arm64-microsoft-intrinsics.c b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
index 2f5ab50d6c848..c3efe232420e2 100644
--- a/clang/test/CodeGen/arm64-microsoft-intrinsics.c
+++ b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
@@ -584,6 +584,29 @@ unsigned int check_CountOneBits64(unsigned __int64 arg1) {
// CHECK-MSCOMPAT: ret i32 %[[VAR2]]
// CHECK-LINUX: error: call to undeclared function '_CountOneBits64'
+unsigned int check_CountTrailingZeros(unsigned LONG arg1) {
+ return _CountTrailingZeros(arg1);
+}
+
+// CHECK-MSCOMPAT: %[[ARG1:.*]].addr = alloca i32, align 4
+// CHECK-MSCOMPAT: store i32 %[[ARG1]], ptr %[[ARG1]].addr, align 4
+// CHECK-MSCOMPAT: %[[VAR0:.*]] = load i32, ptr %[[ARG1]].addr, align 4
+// CHECK-MSCOMPAT: %[[VAR1:.*]] = call i32 @llvm.cttz.i32(i32 %0, i1 false)
+// CHECK-MSCOMPAT: ret i32 %[[VAR1]]
+// CHECK-LINUX: error: call to undeclared function '_CountTrailingZeros'
+
+unsigned int check_CountTrailingZeros64(unsigned __int64 arg1) {
+ return _CountTrailingZeros64(arg1);
+}
+
+// CHECK-MSCOMPAT: %[[ARG1:.*]].addr = alloca i64, align 8
+// CHECK-MSCOMPAT: store i64 %[[ARG1]], ptr %[[ARG1]].addr, align 8
+// CHECK-MSCOMPAT: %[[VAR0:.*]] = load i64, ptr %[[ARG1]].addr, align 8
+// CHECK-MSCOMPAT: %[[VAR1:.*]] = call i64 @llvm.cttz.i64(i64 %0, i1 false)
+// CHECK-MSCOMPAT: %[[VAR2:.*]] = trunc i64 %1 to i32
+// CHECK-MSCOMPAT: ret i32 %[[VAR2]]
+// CHECK-LINUX: error: call to undeclared function '_CountTrailingZeros64'
+
void check__prefetch(void *arg1) {
return __prefetch(arg1);
}
>From b9482457371757dc5c7270e5673bd0c9f4d91caf Mon Sep 17 00:00:00 2001
From: Adhemerval Zanella <zatrazz at gmail.com>
Date: Sat, 18 Apr 2026 08:37:49 -0300
Subject: [PATCH 2/6] [aarch64] Add missing _ReadStatusReg/_WriteStatusReg on
intrin.h
Add ARM64_SYSREG, ARM64_FPCR, and ARM64_FPSR definitions. These
are provided by MSVC as a way to encode a register to be used
along the _ReadStatusReg/_WriteStatusReg intrisincs.
The ARM64_SYSREG works with other ARM64_* registers defined in other
MSVC headers (like windows.h).
---
clang/lib/Headers/arm64intr.h | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/clang/lib/Headers/arm64intr.h b/clang/lib/Headers/arm64intr.h
index 4943b2db69d02..53a3d57a6e9d1 100644
--- a/clang/lib/Headers/arm64intr.h
+++ b/clang/lib/Headers/arm64intr.h
@@ -15,6 +15,16 @@
#ifndef __ARM64INTR_H
#define __ARM64INTR_H
+/* Encode an AArch64 system register for use with
+ _ReadStatusReg/_WriteStatusReg. op0 must be 2 or 3; only the low bit is
+ stored. */
+#define ARM64_SYSREG(op0, op1, CRn, CRm, op2) \
+ ((((op0) & 0x1) << 14) | (((op1) & 0x7) << 11) | (((CRn) & 0xF) << 7) | \
+ (((CRm) & 0xF) << 3) | ((op2) & 0x7))
+
+#define ARM64_FPCR ARM64_SYSREG(3, 3, 4, 4, 0)
+#define ARM64_FPSR ARM64_SYSREG(3, 3, 4, 4, 1)
+
typedef enum
{
_ARM64_BARRIER_SY = 0xF,
>From 91b960ef5d2e75b13ecf936d8d11ccce8c0be36d Mon Sep 17 00:00:00 2001
From: Adhemerval Zanella <zatrazz at gmail.com>
Date: Sat, 18 Apr 2026 09:13:27 -0300
Subject: [PATCH 3/6] [aarch64] Add support for the __setReg MS intrinsics
The builtin write a 64-bit value directly into a specific hardware
processor register, identified by an integer index.
It is ARM64 specific and it is documented at:
<https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-180>
---
clang/include/clang/Basic/BuiltinsAArch64.td | 1 +
clang/lib/CodeGen/TargetBuiltins/ARM.cpp | 16 ++++++++++++----
clang/lib/Headers/intrin.h | 1 +
clang/lib/Sema/SemaARM.cpp | 2 +-
clang/test/CodeGen/arm64-microsoft-intrinsics.c | 11 +++++++++++
5 files changed, 26 insertions(+), 5 deletions(-)
diff --git a/clang/include/clang/Basic/BuiltinsAArch64.td b/clang/include/clang/Basic/BuiltinsAArch64.td
index 502652e208dff..793388b22c139 100644
--- a/clang/include/clang/Basic/BuiltinsAArch64.td
+++ b/clang/include/clang/Basic/BuiltinsAArch64.td
@@ -335,6 +335,7 @@ let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES",
let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES", Header = "intrin.h" in {
def _ReadWriteBarrier : AArch64NoPrefixTargetLibBuiltin<"void ()">;
def __getReg : AArch64NoPrefixTargetLibBuiltin<"unsigned long long int (int)">;
+ def __setReg : AArch64NoPrefixTargetLibBuiltin<"void (int, unsigned long long int)">;
def _ReadStatusReg : AArch64NoPrefixTargetLibBuiltin<"long long int (int)">;
def _WriteStatusReg : AArch64NoPrefixTargetLibBuiltin<"void (int, long long int)">;
def __sys : AArch64NoPrefixTargetLibBuiltin<"unsigned int (int, long long int)">;
diff --git a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
index c2744ede7e647..c3394544b3c84 100644
--- a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
@@ -4810,7 +4810,8 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
return CI;
}
- if (BuiltinID == clang::AArch64::BI__getReg) {
+ if (BuiltinID == clang::AArch64::BI__getReg ||
+ BuiltinID == clang::AArch64::BI__setReg) {
Expr::EvalResult Result;
if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
llvm_unreachable("Sema will ensure that the parameter is constant");
@@ -4823,9 +4824,16 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
- llvm::Function *F =
- CGM.getIntrinsic(Intrinsic::read_register, {Int64Ty});
- return Builder.CreateCall(F, Metadata);
+ CallInst *CI;
+ if (BuiltinID == clang::AArch64::BI__getReg) {
+ llvm::Function *F = CGM.getIntrinsic(Intrinsic::read_register, {Int64Ty});
+ CI = Builder.CreateCall(F, Metadata);
+ } else {
+ llvm::Function *F =
+ CGM.getIntrinsic(Intrinsic::write_register, {Int64Ty});
+ CI = Builder.CreateCall(F, {Metadata, EmitScalarExpr(E->getArg(1))});
+ }
+ return CI;
}
if (BuiltinID == clang::AArch64::BI__break) {
diff --git a/clang/lib/Headers/intrin.h b/clang/lib/Headers/intrin.h
index 7b5698763d1ca..b0c82acf927fb 100644
--- a/clang/lib/Headers/intrin.h
+++ b/clang/lib/Headers/intrin.h
@@ -374,6 +374,7 @@ static __inline__ void __DEFAULT_FN_ATTRS __nop(void) {
\*----------------------------------------------------------------------------*/
#if defined(__aarch64__) || defined(__arm64ec__)
unsigned __int64 __getReg(int);
+void __setReg(int, unsigned __int64);
unsigned char _interlockedbittestandreset_acq(long volatile *, long);
unsigned char _interlockedbittestandreset_nf(long volatile *, long);
unsigned char _interlockedbittestandreset_rel(long volatile *, long);
diff --git a/clang/lib/Sema/SemaARM.cpp b/clang/lib/Sema/SemaARM.cpp
index 7871fd624a467..f69375f3b3edd 100644
--- a/clang/lib/Sema/SemaARM.cpp
+++ b/clang/lib/Sema/SemaARM.cpp
@@ -1169,7 +1169,7 @@ bool SemaARM::CheckAArch64BuiltinFunctionCall(const TargetInfo &TI,
if (BuiltinID == AArch64::BI__sys)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 0x3fff);
- if (BuiltinID == AArch64::BI__getReg)
+ if (BuiltinID == AArch64::BI__getReg || BuiltinID == AArch64::BI__setReg)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 31);
if (BuiltinID == AArch64::BI__break)
diff --git a/clang/test/CodeGen/arm64-microsoft-intrinsics.c b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
index c3efe232420e2..703dbc1c27f67 100644
--- a/clang/test/CodeGen/arm64-microsoft-intrinsics.c
+++ b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
@@ -157,6 +157,17 @@ unsigned __int64 check__getReg(void) {
// CHECK-MSCOMPAT: call i64 @llvm.read_register.i64(metadata ![[MD2:.*]])
// CHECK-MSCOMPAT: call i64 @llvm.read_register.i64(metadata ![[MD3:.*]])
+void test__setReg(unsigned __int64 v)
+{
+ __setReg(18, v);
+ __setReg(31, v);
+}
+
+// CHECK-MSCOMPAT-LABEL: define{{.*}}void @test__setReg(i64{{.*}}%v){{.*}}{
+// CHECK-MSCOMPAT: call void @llvm.write_register.i64(metadata ![[MD2]], i64
+// CHECK-MSCOMPAT: call void @llvm.write_register.i64(metadata ![[MD3]], i64
+// CHECK-LINUX: error: call to undeclared function '__setReg'
+
#ifdef __LP64__
#define LONG __int32
#else
>From 42a64545bdc0d80a07709aca1e5cfb5cb3000c6c Mon Sep 17 00:00:00 2001
From: Adhemerval Zanella <zatrazz at gmail.com>
Date: Sun, 19 Apr 2026 09:41:12 -0300
Subject: [PATCH 4/6] [aarch64] Add support for the __getRegFp MS intrinsicsA
The builtin reads from a hardware floating-point register using an integer index.
It is ARM64 specific and it is documented at:
<https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-180>
---
clang/include/clang/Basic/BuiltinsAArch64.td | 1 +
clang/lib/CodeGen/TargetBuiltins/ARM.cpp | 18 ++++++++++++++++++
clang/lib/Headers/intrin.h | 1 +
clang/lib/Sema/SemaARM.cpp | 3 ++-
.../test/CodeGen/arm64-microsoft-intrinsics.c | 11 +++++++++++
5 files changed, 33 insertions(+), 1 deletion(-)
diff --git a/clang/include/clang/Basic/BuiltinsAArch64.td b/clang/include/clang/Basic/BuiltinsAArch64.td
index 793388b22c139..9e670f559d81f 100644
--- a/clang/include/clang/Basic/BuiltinsAArch64.td
+++ b/clang/include/clang/Basic/BuiltinsAArch64.td
@@ -335,6 +335,7 @@ let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES",
let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES", Header = "intrin.h" in {
def _ReadWriteBarrier : AArch64NoPrefixTargetLibBuiltin<"void ()">;
def __getReg : AArch64NoPrefixTargetLibBuiltin<"unsigned long long int (int)">;
+ def __getRegFp : AArch64NoPrefixTargetLibBuiltin<"double (int)">;
def __setReg : AArch64NoPrefixTargetLibBuiltin<"void (int, unsigned long long int)">;
def _ReadStatusReg : AArch64NoPrefixTargetLibBuiltin<"long long int (int)">;
def _WriteStatusReg : AArch64NoPrefixTargetLibBuiltin<"void (int, long long int)">;
diff --git a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
index c3394544b3c84..d4609363e4f18 100644
--- a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
@@ -4836,6 +4836,24 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
return CI;
}
+ if (BuiltinID == clang::AArch64::BI__getRegFp) {
+ Expr::EvalResult Result;
+ if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
+ llvm_unreachable("Sema will ensure that the parameter is constant");
+
+ llvm::APSInt Value = Result.Val.getInt();
+ LLVMContext &Context = CGM.getLLVMContext();
+ std::string Reg = "d" + toString(Value, 10);
+
+ llvm::Metadata *Ops[] = {llvm::MDString::get(Context, Reg)};
+ llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
+ llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
+
+ llvm::Function *F = CGM.getIntrinsic(Intrinsic::read_register, {Int64Ty});
+ llvm::Value *Bits = Builder.CreateCall(F, Metadata);
+ return Builder.CreateBitCast(Bits, llvm::Type::getDoubleTy(Context));
+ }
+
if (BuiltinID == clang::AArch64::BI__break) {
Expr::EvalResult Result;
if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
diff --git a/clang/lib/Headers/intrin.h b/clang/lib/Headers/intrin.h
index b0c82acf927fb..040dd0b734a18 100644
--- a/clang/lib/Headers/intrin.h
+++ b/clang/lib/Headers/intrin.h
@@ -374,6 +374,7 @@ static __inline__ void __DEFAULT_FN_ATTRS __nop(void) {
\*----------------------------------------------------------------------------*/
#if defined(__aarch64__) || defined(__arm64ec__)
unsigned __int64 __getReg(int);
+double __getRegFp(int _Reg);
void __setReg(int, unsigned __int64);
unsigned char _interlockedbittestandreset_acq(long volatile *, long);
unsigned char _interlockedbittestandreset_nf(long volatile *, long);
diff --git a/clang/lib/Sema/SemaARM.cpp b/clang/lib/Sema/SemaARM.cpp
index f69375f3b3edd..3a8a17c11bfb2 100644
--- a/clang/lib/Sema/SemaARM.cpp
+++ b/clang/lib/Sema/SemaARM.cpp
@@ -1169,7 +1169,8 @@ bool SemaARM::CheckAArch64BuiltinFunctionCall(const TargetInfo &TI,
if (BuiltinID == AArch64::BI__sys)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 0x3fff);
- if (BuiltinID == AArch64::BI__getReg || BuiltinID == AArch64::BI__setReg)
+ if (BuiltinID == AArch64::BI__getReg || BuiltinID == AArch64::BI__setReg ||
+ BuiltinID == AArch64::BI__getRegFp)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 31);
if (BuiltinID == AArch64::BI__break)
diff --git a/clang/test/CodeGen/arm64-microsoft-intrinsics.c b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
index 703dbc1c27f67..622714a00031b 100644
--- a/clang/test/CodeGen/arm64-microsoft-intrinsics.c
+++ b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
@@ -168,6 +168,16 @@ void test__setReg(unsigned __int64 v)
// CHECK-MSCOMPAT: call void @llvm.write_register.i64(metadata ![[MD3]], i64
// CHECK-LINUX: error: call to undeclared function '__setReg'
+double test__getRegFp(void)
+{
+ return __getRegFp(5);
+}
+
+// CHECK-MSCOMPAT-LABEL: define{{.*}}double @test__getRegFp(){{.*}}{
+// CHECK-MSCOMPAT: [[BITS:%.*]] = call i64 @llvm.read_register.i64(metadata ![[MD4:.*]])
+// CHECK-MSCOMPAT: bitcast i64 [[BITS]] to double
+// CHECK-LINUX: error: call to undeclared function '__getRegFp'
+
#ifdef __LP64__
#define LONG __int32
#else
@@ -631,3 +641,4 @@ void check__prefetch(void *arg1) {
// CHECK-MSCOMPAT: ![[MD2]] = !{!"x18"}
// CHECK-MSCOMPAT: ![[MD3]] = !{!"sp"}
+// CHECK-MSCOMPAT: ![[MD4]] = !{!"d5"}
>From e6eba857f1e3fe039ec43ceb693ac76dcd73c85e Mon Sep 17 00:00:00 2001
From: Adhemerval Zanella <zatrazz at gmail.com>
Date: Sun, 19 Apr 2026 10:03:37 -0300
Subject: [PATCH 5/6] [aarch64] Add support for the __setRegFp MS intrinsics
The builtin writes to a hardware floating-point register using an integer index.
It is ARM64 specific and it is documented at:
<https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-180>
---
clang/include/clang/Basic/BuiltinsAArch64.td | 1 +
clang/lib/CodeGen/TargetBuiltins/ARM.cpp | 19 +++++++++++++++----
clang/lib/Headers/intrin.h | 1 +
clang/lib/Sema/SemaARM.cpp | 2 +-
.../test/CodeGen/arm64-microsoft-intrinsics.c | 9 +++++++++
5 files changed, 27 insertions(+), 5 deletions(-)
diff --git a/clang/include/clang/Basic/BuiltinsAArch64.td b/clang/include/clang/Basic/BuiltinsAArch64.td
index 9e670f559d81f..1a04f86f602fc 100644
--- a/clang/include/clang/Basic/BuiltinsAArch64.td
+++ b/clang/include/clang/Basic/BuiltinsAArch64.td
@@ -336,6 +336,7 @@ let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES",
def _ReadWriteBarrier : AArch64NoPrefixTargetLibBuiltin<"void ()">;
def __getReg : AArch64NoPrefixTargetLibBuiltin<"unsigned long long int (int)">;
def __getRegFp : AArch64NoPrefixTargetLibBuiltin<"double (int)">;
+ def __setRegFp : AArch64NoPrefixTargetLibBuiltin<"void (int, double)">;
def __setReg : AArch64NoPrefixTargetLibBuiltin<"void (int, unsigned long long int)">;
def _ReadStatusReg : AArch64NoPrefixTargetLibBuiltin<"long long int (int)">;
def _WriteStatusReg : AArch64NoPrefixTargetLibBuiltin<"void (int, long long int)">;
diff --git a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
index d4609363e4f18..751d9f8613de1 100644
--- a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
@@ -4836,7 +4836,8 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
return CI;
}
- if (BuiltinID == clang::AArch64::BI__getRegFp) {
+ if (BuiltinID == clang::AArch64::BI__getRegFp ||
+ BuiltinID == clang::AArch64::BI__setRegFp) {
Expr::EvalResult Result;
if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
llvm_unreachable("Sema will ensure that the parameter is constant");
@@ -4849,9 +4850,19 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
- llvm::Function *F = CGM.getIntrinsic(Intrinsic::read_register, {Int64Ty});
- llvm::Value *Bits = Builder.CreateCall(F, Metadata);
- return Builder.CreateBitCast(Bits, llvm::Type::getDoubleTy(Context));
+ llvm::Value *Ret;
+ if (BuiltinID == clang::AArch64::BI__getRegFp) {
+ llvm::Function *F = CGM.getIntrinsic(Intrinsic::read_register, {Int64Ty});
+ llvm::Value *Bits = Builder.CreateCall(F, Metadata);
+ Ret = Builder.CreateBitCast(Bits, llvm::Type::getDoubleTy(Context));
+ } else {
+ llvm::Value *Val = EmitScalarExpr(E->getArg(1));
+ llvm::Value *Bits = Builder.CreateBitCast(Val, Int64Ty);
+ llvm::Function *F =
+ CGM.getIntrinsic(Intrinsic::write_register, {Int64Ty});
+ Ret = Builder.CreateCall(F, {Metadata, Bits});
+ }
+ return Ret;
}
if (BuiltinID == clang::AArch64::BI__break) {
diff --git a/clang/lib/Headers/intrin.h b/clang/lib/Headers/intrin.h
index 040dd0b734a18..039846bafacf5 100644
--- a/clang/lib/Headers/intrin.h
+++ b/clang/lib/Headers/intrin.h
@@ -376,6 +376,7 @@ static __inline__ void __DEFAULT_FN_ATTRS __nop(void) {
unsigned __int64 __getReg(int);
double __getRegFp(int _Reg);
void __setReg(int, unsigned __int64);
+void __setRegFp(int, double);
unsigned char _interlockedbittestandreset_acq(long volatile *, long);
unsigned char _interlockedbittestandreset_nf(long volatile *, long);
unsigned char _interlockedbittestandreset_rel(long volatile *, long);
diff --git a/clang/lib/Sema/SemaARM.cpp b/clang/lib/Sema/SemaARM.cpp
index 3a8a17c11bfb2..d3bfd75223220 100644
--- a/clang/lib/Sema/SemaARM.cpp
+++ b/clang/lib/Sema/SemaARM.cpp
@@ -1170,7 +1170,7 @@ bool SemaARM::CheckAArch64BuiltinFunctionCall(const TargetInfo &TI,
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 0x3fff);
if (BuiltinID == AArch64::BI__getReg || BuiltinID == AArch64::BI__setReg ||
- BuiltinID == AArch64::BI__getRegFp)
+ BuiltinID == AArch64::BI__getRegFp || BuiltinID == AArch64::BI__setRegFp)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 31);
if (BuiltinID == AArch64::BI__break)
diff --git a/clang/test/CodeGen/arm64-microsoft-intrinsics.c b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
index 622714a00031b..4d4ea29fc125a 100644
--- a/clang/test/CodeGen/arm64-microsoft-intrinsics.c
+++ b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
@@ -178,6 +178,15 @@ double test__getRegFp(void)
// CHECK-MSCOMPAT: bitcast i64 [[BITS]] to double
// CHECK-LINUX: error: call to undeclared function '__getRegFp'
+void test__setRegFp(double v)
+{
+ __setRegFp(5, v);
+}
+// CHECK-MSCOMPAT-LABEL: define{{.*}}void @test__setRegFp(double{{.*}}%v){{.*}}{
+// CHECK-MSCOMPAT: [[BITS:%.*]] = bitcast double {{.*}} to i64
+// CHECK-MSCOMPAT: call void @llvm.write_register.i64(metadata ![[MD4:.*]], i64 [[BITS]])
+// CHECK-LINUX: error: call to undeclared function '__setRegFp'
+
#ifdef __LP64__
#define LONG __int32
#else
>From 194c8168befffa65dcdc1f789175afaa2d6691fb Mon Sep 17 00:00:00 2001
From: Adhemerval Zanella <zatrazz at gmail.com>
Date: Mon, 20 Apr 2026 13:11:40 -0300
Subject: [PATCH 6/6] [aarch64] Add support for the __prefetch2 MS intrinsic
This is similar to already supported __prefetch builtin, with an extra 8-bit
argument that defines the prefetch operation hint. The MSVC headers, nor
the documentation, seems to add any extra definitions on how to
construct the hint.
This s documented at:
<https://learn.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics?view=msvc-180>
---
clang/include/clang/Basic/BuiltinsAArch64.td | 1 +
clang/lib/CodeGen/TargetBuiltins/ARM.cpp | 20 +++++++++++++++++++
clang/lib/Headers/intrin.h | 1 +
clang/lib/Sema/SemaARM.cpp | 3 +++
.../test/CodeGen/arm64-microsoft-intrinsics.c | 10 ++++++++++
5 files changed, 35 insertions(+)
diff --git a/clang/include/clang/Basic/BuiltinsAArch64.td b/clang/include/clang/Basic/BuiltinsAArch64.td
index 1a04f86f602fc..90824e955c072 100644
--- a/clang/include/clang/Basic/BuiltinsAArch64.td
+++ b/clang/include/clang/Basic/BuiltinsAArch64.td
@@ -403,6 +403,7 @@ let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES",
let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES", Header = "intrin.h" in {
def __prefetch : AArch64NoPrefixTargetLibBuiltin<"void (void const *)">;
+ def __prefetch2 : AArch64NoPrefixTargetLibBuiltin<"void (void const *, unsigned char)">;
}
let Attributes = [NoThrow, RequireDeclaration], Languages = "ALL_MS_LANGUAGES", Header = "intrin.h" in {
diff --git a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
index 751d9f8613de1..b2e090489c7d9 100644
--- a/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
+++ b/clang/lib/CodeGen/TargetBuiltins/ARM.cpp
@@ -5293,6 +5293,26 @@ Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
return Builder.CreateCall(F, {Address, RW, Locality, Data});
}
+ if (BuiltinID == AArch64::BI__prefetch2) {
+ Value *Address = EmitScalarExpr(E->getArg(0));
+ llvm::APSInt PrfOp = E->getArg(1)->EvaluateKnownConstInt(CGM.getContext());
+ // Decode 5-bit PRFM encoding: bits[4:3]=type, bits[2:1]=target,
+ // bit[0]=policy
+ // type: PLD=0(load), PLI=1(instr), PST=2(store)
+ // target: L1=0, L2=1, L3=2
+ // policy: KEEP=0, STRM=1
+ uint64_t Op = PrfOp.getZExtValue();
+ uint64_t Type = (Op >> 3) & 0x3;
+ uint64_t Target = (Op >> 1) & 0x3;
+ uint64_t Policy = Op & 0x1;
+ Value *RW = ConstantInt::get(Int32Ty, Type == 2 ? 1 : 0);
+ Value *Local = ConstantInt::get(Int32Ty, Target);
+ Value *Retain = ConstantInt::get(Int32Ty, Policy == 0 ? 1 : 0);
+ Value *IsData = ConstantInt::get(Int32Ty, Type == 1 ? 0 : 1);
+ Function *F = CGM.getIntrinsic(Intrinsic::aarch64_prefetch);
+ return Builder.CreateCall(F, {Address, RW, Local, Retain, IsData});
+ }
+
if (BuiltinID == AArch64::BI__hlt) {
Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hlt);
Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0))});
diff --git a/clang/lib/Headers/intrin.h b/clang/lib/Headers/intrin.h
index 039846bafacf5..d8b09748402a7 100644
--- a/clang/lib/Headers/intrin.h
+++ b/clang/lib/Headers/intrin.h
@@ -447,6 +447,7 @@ unsigned int _CountTrailingZeros64(unsigned __int64);
unsigned int __hlt(unsigned int, ...);
void __cdecl __prefetch(const void *);
+void __cdecl __prefetch2(const void *, unsigned char);
#endif
diff --git a/clang/lib/Sema/SemaARM.cpp b/clang/lib/Sema/SemaARM.cpp
index d3bfd75223220..c45580902db94 100644
--- a/clang/lib/Sema/SemaARM.cpp
+++ b/clang/lib/Sema/SemaARM.cpp
@@ -1173,6 +1173,9 @@ bool SemaARM::CheckAArch64BuiltinFunctionCall(const TargetInfo &TI,
BuiltinID == AArch64::BI__getRegFp || BuiltinID == AArch64::BI__setRegFp)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 31);
+ if (BuiltinID == AArch64::BI__prefetch2)
+ return SemaRef.BuiltinConstantArgRange(TheCall, 1, 0, 31);
+
if (BuiltinID == AArch64::BI__break)
return SemaRef.BuiltinConstantArgRange(TheCall, 0, 0, 0xffff);
diff --git a/clang/test/CodeGen/arm64-microsoft-intrinsics.c b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
index 4d4ea29fc125a..4289ab6758dfc 100644
--- a/clang/test/CodeGen/arm64-microsoft-intrinsics.c
+++ b/clang/test/CodeGen/arm64-microsoft-intrinsics.c
@@ -647,6 +647,16 @@ void check__prefetch(void *arg1) {
// CHECK-MSCOMPAT: call void @llvm.prefetch.p0(ptr %[[VAR0]], i32 0, i32 3, i32 1)
// CHECK-MSCOMPAT: ret void
+void check__prefetch2(void *arg1) {
+ __prefetch2(arg1, 0x00);
+ __prefetch2(arg1, 0x13);
+}
+
+// CHECK-MSCOMPAT-LABEL: define{{.*}}void @check__prefetch2(ptr{{.*}}%arg1){{.*}}{
+// CHECK-MSCOMPAT: call void @llvm.aarch64.prefetch(ptr %{{.*}}, i32 0, i32 0, i32 1, i32 1)
+// CHECK-MSCOMPAT: call void @llvm.aarch64.prefetch(ptr %{{.*}}, i32 1, i32 1, i32 0, i32 1)
+// CHECK-LINUX: error: call to undeclared function '__prefetch2'
+
// CHECK-MSCOMPAT: ![[MD2]] = !{!"x18"}
// CHECK-MSCOMPAT: ![[MD3]] = !{!"sp"}
More information about the cfe-commits
mailing list