Win64 LLP64 ABI support with cross-compiler tooling

tools/lkl/Makefile.autoconf

@@ -106,10 +106,12 @@ endef
 
 define nt_host
   $(call set_autoconf_var,NT,y)
+  $(call set_kernel_config,INIT_STACK_ALL_ZERO,n)
   KOPT = "KALLSYMS_EXTRA_PASS=1"
   KOPT += "HOSTCFLAGS=-Wno-char-subscripts"
   KOPT += "HOSTLDFLAGS=-s"
   LDLIBS += -lws2_32
+  LDFLAGS += -Wl,--image-base,0x10000
   EXESUF := .exe
   SOSUF := .dll
   CFLAGS += -Iinclude/mingw32

tools/lkl/bin/x86_64-w64-mingw32-cc

@@ -0,0 +1 @@
+x86_64-w64-mingw32-gcc

tools/lkl/bin/x86_64-w64-mingw32-gcc

@@ -0,0 +1,22 @@
+#!/bin/bash
+# Shim: routes kernel code to cygwin cross-gcc, user-space to mingw-gcc.
+# Cygwin-gcc: LP64 (long=64-bit), direct PE/COFF output, no Wine needed.
+# Mingw-gcc: LLP64 (long=32-bit), standard Win64 libraries.
+# The LKL API header uses __lkl_long_t (always 64-bit) to bridge the difference.
+REAL_MINGW_GCC="/usr/bin/x86_64-w64-mingw32-gcc"
+CYGWIN_GCC="/usr/bin/x86_64-pc-cygwin-gcc"
+LKL_BIN_DIR="$(cd "$(dirname "$0")" && pwd)"
+
+# Kernel code → cygwin-gcc (LP64)
+for arg in "$@"; do
+    if [[ "$arg" == "-D__KERNEL__" ]]; then
+        exec "$CYGWIN_GCC" "$@"
+    fi
+done
+
+# User-space: use mingw-gcc with patched ld via specs
+SPECS="$LKL_BIN_DIR/.lkl-linker.specs"
+if [[ ! -f "$SPECS" ]]; then
+    printf '*linker:\n%s/x86_64-w64-mingw32-ld\n' "$LKL_BIN_DIR" > "$SPECS"
+fi
+exec "$REAL_MINGW_GCC" -specs="$SPECS" "$@"

tools/lkl/lib/jmp_buf.c

@@ -1,6 +1,92 @@
-#include <setjmp.h>
 #include <lkl_host.h>
 
+#if defined(__x86_64__) && defined(_WIN64)
+/*
+ * Custom non-SEH setjmp/longjmp for x86_64 Windows.
+ *
+ * Windows x64 longjmp uses RtlUnwindEx which walks the stack frame-by-frame.
+ * LKL uses longjmp for cooperative context switching between different stacks
+ * (each kernel thread has its own malloc'd stack). RtlUnwindEx cannot handle
+ * cross-stack jumps, causing stack overflow during unwinding.
+ * We bypass SEH entirely by doing a raw register save/restore.
+ *
+ * jmp_buf layout (16 bytes aligned):
+ *   [0]  rbx
+ *   [1]  rbp
+ *   [2]  r12
+ *   [3]  r13
+ *   [4]  r14
+ *   [5]  r15
+ *   [6]  rsp (after return from setjmp)
+ *   [7]  rip (return address)
+ *   [8]  rdi
+ *   [9]  rsi
+ */
+
+static inline int lkl_setjmp(void *buf)
+{
+	int ret;
+	__asm__ __volatile__ (
+		"movq %%rbx,   0(%[b])\n\t"
+		"movq %%rbp,   8(%[b])\n\t"
+		"movq %%r12,  16(%[b])\n\t"
+		"movq %%r13,  24(%[b])\n\t"
+		"movq %%r14,  32(%[b])\n\t"
+		"movq %%r15,  40(%[b])\n\t"
+		"leaq 8(%%rsp), %%rax\n\t"
+		"movq %%rax,  48(%[b])\n\t"
+		"movq (%%rsp), %%rax\n\t"
+		"movq %%rax,  56(%[b])\n\t"
+		"movq %%rdi,  64(%[b])\n\t"
+		"movq %%rsi,  72(%[b])\n\t"
+		"xorl %[ret], %[ret]\n\t"
+		: [ret] "=a" (ret)
+		: [b] "r" (buf)
+		: "memory"
+	);
+	return ret;
+}
+
+__attribute__((noreturn))
+static void lkl_longjmp(void *buf, int val)
+{
+	__asm__ __volatile__ (
+		"movq  0(%[b]), %%rbx\n\t"
+		"movq  8(%[b]), %%rbp\n\t"
+		"movq 16(%[b]), %%r12\n\t"
+		"movq 24(%[b]), %%r13\n\t"
+		"movq 32(%[b]), %%r14\n\t"
+		"movq 40(%[b]), %%r15\n\t"
+		"movq 48(%[b]), %%rsp\n\t"
+		"movq 64(%[b]), %%rdi\n\t"
+		"movq 72(%[b]), %%rsi\n\t"
+		"movl %[v], %%eax\n\t"
+		"testl %%eax, %%eax\n\t"
+		"jnz 1f\n\t"
+		"incl %%eax\n\t"
+		"1:\n\t"
+		"jmpq *56(%[b])\n\t"
+		:
+		: [b] "r" (buf), [v] "r" (val)
+		: "memory"
+	);
+	__builtin_unreachable();
+}
+
+void jmp_buf_set(struct lkl_jmp_buf *jmpb, void (*f)(void))
+{
+	if (!lkl_setjmp(jmpb->buf))
+		f();
+}
+
+void jmp_buf_longjmp(struct lkl_jmp_buf *jmpb, int val)
+{
+	lkl_longjmp(jmpb->buf, val);
+}
+
+#else
+#include <setjmp.h>
+
 void jmp_buf_set(struct lkl_jmp_buf *jmpb, void (*f)(void))
 {
 	if (!setjmp(*((jmp_buf *)jmpb->buf)))
@@ -11,3 +97,4 @@ void jmp_buf_longjmp(struct lkl_jmp_buf *jmpb, int val)
 {
 	longjmp(*((jmp_buf *)jmpb->buf), val);
 }
+#endif