[Openmp-commits] [openmp] [openmp] Add initial ARM64EC support (PR #138769)

[Jacek Caban via Openmp-commits]

================
@@ -1301,6 +1302,106 @@ KMP_LABEL(kmp_no_args):
 // '
 #if (KMP_OS_LINUX || KMP_OS_DARWIN || KMP_OS_WINDOWS) && (KMP_ARCH_AARCH64 || KMP_ARCH_AARCH64_32)
 
+#ifdef __arm64ec__
+
+//------------------------------------------------------------------------
+// int
+// __kmp_invoke_microtask( void (*pkfn) (int gtid, int tid, ...),
+//                         int gtid, int tid,
+//                         int argc, void *p_argv[]) {
+//   (*pkfn)( & gtid, & tid, argv[0], ... );
+//   return 1;
+// }
+//
+// parameters:
+//	x0:	pkfn
+//	w1:	gtid
+//	w2:	tid
+//	w3:	argc
+//	x4:	p_argv
+//
+// locals:
+//	__gtid:	gtid parm pushed on stack so can pass &gtid to pkfn
+//	__tid:	tid parm pushed on stack so can pass &tid to pkfn
+//
+// reg temps:
+//	 x8:	used as temporary for stack placement calculation
+//	 w9:	used as temporary for number of pkfn parms
+//	x10:	used to traverse p_argv array
+//	x11:	used to hold pkfn address
+//	x12:	used as temporary for stack parameters
+//	x19:	used to preserve exit_frame_ptr, callee-save
+//
+// return:	w0	(always 1/TRUE)
+//
+
+__gtid = 4
+__tid = 8
+
+// -- Begin __kmp_invoke_microtask
+// mark_begin;
+	.section .text,"xr",discard,"#__kmp_invoke_microtask"
+	.globl "#__kmp_invoke_microtask"
+	ALIGN 2
+"#__kmp_invoke_microtask":
+	stp	x29, x30, [sp, #-16]!
+	mov	x29, sp
+
+	mov	w9, #1
+	add	w9, w9, w3, lsr #1
+	sub	sp, sp, w9, uxtw #4
+	mov	x8, sp
+
+	mov	x11, x0
+	str	w1, [x29, #-__gtid]
+	str	w2, [x29, #-__tid]
+	mov	w9, w3
+	mov	x10, x4
+
+	sub	x0, x29, #__gtid
+	sub	x1, x29, #__tid
+	mov     x4, sp
+	mov     w5, #0
+
+	cbz	w9, KMP_LABEL(kmp_1)
+	ldr	x2, [x10]
+
+	sub	w9, w9, #1
+	cbz	w9, KMP_LABEL(kmp_1)
+	ldr	x3, [x10, #8]!
----------------
cjacek wrote:

`__os_arm64x_check_icall` can’t resolve the target correctly if it’s ARM64EC. In that case, it leaves the target address unchanged and we call the function directly. As you noted, this means that if the target isn’t actually a varargs function, it will misinterpret the arguments.

If we want to support such cases, ARM64EC alone likely won’t suffice. We may need to build x86_64 assembly instead. If we do that and the target is ARM64EC, control will return from the emulator to the entry chunk, which should handle it correctly. The downside is that it's a bit tricky for the build system to handle. I’ll investigate this further.

https://github.com/llvm/llvm-project/pull/138769