[compiler-rt] Add initial ARM64EC builtins support (#139279)

Use the aarch64 variants of assembly functions.

Co-authored-by: Billy Laws <[email protected]>

Author: cjacek

compiler-rt/cmake/Modules/AddCompilerRT.cmake

@@ -123,7 +123,7 @@ macro(set_output_name output name arch)
   else()
     if(ANDROID AND ${arch} STREQUAL "i386")
       set(${output} "${name}-i686${COMPILER_RT_OS_SUFFIX}")
-    elseif("${arch}" MATCHES "^arm")
+    elseif(NOT "${arch}" MATCHES "^arm64" AND "${arch}" MATCHES "^arm")
       if(COMPILER_RT_DEFAULT_TARGET_ONLY)
         set(triple "${COMPILER_RT_DEFAULT_TARGET_TRIPLE}")
       else()

compiler-rt/cmake/builtin-config-ix.cmake

@@ -59,7 +59,7 @@ else()
 endif()
 
 set(AMDGPU amdgcn)
-set(ARM64 aarch64)
+set(ARM64 aarch64 arm64ec)
 set(ARM32 arm armhf armv4t armv5te armv6 armv6m armv7m armv7em armv7 armv7s armv7k armv8m.base armv8m.main armv8.1m.main)
 set(AVR avr)
 set(HEXAGON hexagon)

compiler-rt/lib/builtins/CMakeLists.txt

@@ -668,6 +668,7 @@ set(armv7k_SOURCES ${arm_SOURCES})
 set(arm64_SOURCES ${aarch64_SOURCES})
 set(arm64e_SOURCES ${aarch64_SOURCES})
 set(arm64_32_SOURCES ${aarch64_SOURCES})
+set(arm64ec_SOURCES ${aarch64_SOURCES})
 
 # macho_embedded archs
 set(armv6m_SOURCES ${thumb1_SOURCES})

compiler-rt/lib/builtins/aarch64/chkstk.S

@@ -15,12 +15,18 @@
 //      bl      __chkstk
 //      sub     sp, sp, x15, lsl #4
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__arm64ec__)
+
+#ifdef __arm64ec__
+#define CHKSTK_FUNC __chkstk_arm64ec
+#else
+#define CHKSTK_FUNC __chkstk
+#endif
 
 #define PAGE_SIZE 4096
 
         .p2align 2
-DEFINE_COMPILERRT_FUNCTION(__chkstk)
+DEFINE_COMPILERRT_FUNCTION(CHKSTK_FUNC)
         lsl    x16, x15, #4
         mov    x17, sp
 1:
@@ -30,6 +36,6 @@ DEFINE_COMPILERRT_FUNCTION(__chkstk)
         b.gt   1b
 
         ret
-END_COMPILERRT_FUNCTION(__chkstk)
+END_COMPILERRT_FUNCTION(CHKSTK_FUNC)
 
-#endif // __aarch64__
+#endif // defined(__aarch64__) || defined(__arm64ec__)

compiler-rt/lib/builtins/aarch64/lse.S

@@ -20,7 +20,7 @@
 // Routines may modify temporary registers tmp0, tmp1, tmp2,
 // return value x0 and the flags only.
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__arm64ec__)
 
 #ifdef HAS_ASM_LSE
 .arch armv8-a+lse
@@ -267,4 +267,4 @@ NO_EXEC_STACK_DIRECTIVE
 // GNU property note for BTI and PAC
 GNU_PROPERTY_BTI_PAC
 
-#endif // __aarch64__
+#endif // defined(__aarch64__) || defined(__arm64ec__)

compiler-rt/lib/builtins/aarch64/sme-libc-mem-routines.S

@@ -235,7 +235,7 @@ END_COMPILERRT_FUNCTION(__arm_sc_memcpy)
 DEFINE_COMPILERRT_FUNCTION_ALIAS(__arm_sc_memmove, __arm_sc_memcpy)
 
 // This version uses FP registers. Use this only on targets with them
-#if defined(__aarch64__) && __ARM_FP != 0
+#if (defined(__aarch64__) && __ARM_FP != 0) || defined(__arm64ec__)
 //
 //  __arm_sc_memset
 //

compiler-rt/lib/builtins/clear_cache.c

@@ -59,13 +59,14 @@ uintptr_t GetCurrentProcess(void);
 // specified range.
 
 void __clear_cache(void *start, void *end) {
-#if __i386__ || __x86_64__ || defined(_M_IX86) || defined(_M_X64)
+#if defined(_WIN32) &&                                                         \
+    (defined(__arm__) || defined(__aarch64__) || defined(__arm64ec__))
+  FlushInstructionCache(GetCurrentProcess(), start, end - start);
+#elif __i386__ || __x86_64__ || defined(_M_IX86) || defined(_M_X64)
 // Intel processors have a unified instruction and data cache
 // so there is nothing to do
 #elif defined(__s390__)
 // no-op
-#elif defined(_WIN32) && (defined(__arm__) || defined(__aarch64__))
-  FlushInstructionCache(GetCurrentProcess(), start, end - start);
 #elif defined(__arm__) && !defined(__APPLE__)
 #if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)
   struct arm_sync_icache_args arg;

compiler-rt/lib/builtins/cpu_model/aarch64.c

@@ -14,7 +14,8 @@
 
 #include "aarch64.h"
 
-#if !defined(__aarch64__) && !defined(__arm64__) && !defined(_M_ARM64)
+#if !defined(__aarch64__) && !defined(__arm64__) && !defined(_M_ARM64) &&      \
+    !defined(__arm64ec__) && !defined(_M_ARM64EC)
 #error This file is intended only for aarch64-based targets
 #endif
 

compiler-rt/lib/builtins/cpu_model/aarch64.h

@@ -8,7 +8,8 @@
 
 #include "cpu_model.h"
 
-#if !defined(__aarch64__) && !defined(__arm64__) && !defined(_M_ARM64)
+#if !defined(__aarch64__) && !defined(__arm64__) && !defined(_M_ARM64) &&      \
+    !defined(__arm64ec__) && !defined(_M_ARM64EC)
 #error This file is intended only for aarch64-based targets
 #endif
 

compiler-rt/lib/builtins/fp_compare_impl.inc

@@ -12,7 +12,7 @@
 // functions. We need to ensure that the return value is sign-extended in the
 // same way as GCC expects (since otherwise GCC-generated __builtin_isinf
 // returns true for finite 128-bit floating-point numbers).
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(__arm64ec__)
 // AArch64 GCC overrides libgcc_cmp_return to use int instead of long.
 typedef int CMP_RESULT;
 #elif __SIZEOF_POINTER__ == 8 && __SIZEOF_LONG__ == 4

compiler-rt/lib/builtins/fp_lib.h

@@ -359,7 +359,7 @@ static __inline fp_t __compiler_rt_scalbn(fp_t x, int y) {
   return __compiler_rt_scalbnX(x, y);
 }
 static __inline fp_t __compiler_rt_fmax(fp_t x, fp_t y) {
-#if defined(__aarch64__)
+#if defined(__aarch64__) || defined(__arm64ec__)
   // Use __builtin_fmax which turns into an fmaxnm instruction on AArch64.
   return __builtin_fmax(x, y);
 #else

compiler-rt/lib/builtins/udivmodti4.c

@@ -83,7 +83,7 @@ static inline du_int udiv128by64to64default(du_int u1, du_int u0, du_int v,
 
 static inline du_int udiv128by64to64(du_int u1, du_int u0, du_int v,
                                      du_int *r) {
-#if defined(__x86_64__)
+#if defined(__x86_64__) && !defined(__arm64ec__)
   du_int result;
   __asm__("divq %[v]"
           : "=a"(result), "=d"(*r)

compiler-rt/test/builtins/Unit/enable_execute_stack_test.c

@@ -10,9 +10,22 @@ extern void __enable_execute_stack(void* addr);
 
 typedef int (*pfunc)(void);
 
+#ifdef __arm64ec__
+// On ARM64EC, we need the x86_64 version of this function, but the compiler
+// would normally generate the AArch64 variant, so we hardcode it here.
+static char func1[] = {
+    0xb8, 0x01, 0x00, 0x00, 0x00, // movl    $0x1, %eax
+    0xc3                          // retq
+};
+static char func2[] = {
+    0xb8, 0x02, 0x00, 0x00, 0x00, // movl    $0x2, %eax
+    0xc3                          // retq
+};
+#else
 // Make these static to avoid ILT jumps for incremental linking on Windows.
 static int func1() { return 1; }
 static int func2() { return 2; }
+#endif
 
 void *__attribute__((noinline))
 memcpy_f(void *dst, const void *src, size_t n) {

compiler-rt/test/builtins/Unit/fixunstfdi_test.c

@@ -4,7 +4,7 @@
 
 #include <stdio.h>
 
-#if _ARCH_PPC || __aarch64__
+#if _ARCH_PPC || __aarch64__ || __arm64ec__
 
 #include "int_lib.h"
 
@@ -35,7 +35,7 @@ char assumption_3[sizeof(long double)*CHAR_BIT == 128] = {0};
 
 int main()
 {
-#if _ARCH_PPC || __aarch64__
+#if _ARCH_PPC || __aarch64__ || __arm64ec__
     if (test__fixunstfdi(0.0, 0))
         return 1;
 

compiler-rt/test/builtins/Unit/multc3_test.c

@@ -4,7 +4,7 @@
 
 #include <stdio.h>
 
-#if _ARCH_PPC || __aarch64__
+#if _ARCH_PPC || __aarch64__ || __arm64ec__
 
 #include "int_lib.h"
 #include <math.h>
@@ -348,7 +348,7 @@ long double x[][2] =
 
 int main()
 {
-#if _ARCH_PPC || __aarch64__
+#if _ARCH_PPC || __aarch64__ || __arm64ec__
     const unsigned N = sizeof(x) / sizeof(x[0]);
     unsigned i, j;
     for (i = 0; i < N; ++i)