Add arm64 alloca() support.

Author: kbkpbot

lib/Makefile

@@ -38,7 +38,7 @@ XFLAGS += -I$(TOP)
 I386_O = libtcc1.o alloca.o alloca-bt.o $(COMMON_O)
 X86_64_O = libtcc1.o alloca.o alloca-bt.o $(COMMON_O)
 ARM_O = libtcc1.o armeabi.o alloca.o armflush.o $(COMMON_O)
-ARM64_O = lib-arm64.o $(COMMON_O)
+ARM64_O = lib-arm64.o alloca.o $(COMMON_O)
 RISCV64_O = lib-arm64.o $(COMMON_O)
 COMMON_O = stdatomic.o atomic.o builtin.o
 WIN_O = crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o

lib/alloca.S

@@ -70,16 +70,76 @@ p3:
 /* ---------------------------------------------- */
 #elif defined __arm__
 
-	.text
-	.align	2
-	.global	alloca
-	.type	alloca, %function
+    .text
+    .align
+    .global alloca
+    .type   alloca, %function
 alloca:
-	rsb	sp, r0, sp
-	bic	sp, sp, #7
-	mov	r0, sp
-	mov	pc, lr
-	.size	alloca, .-alloca
+    rsb sp, r0, sp
+    bic sp, sp, #7
+    mov r0, sp
+    mov pc, lr
+    .size   alloca, .-alloca
 
 /* ---------------------------------------------- */
+#elif defined __aarch64__ || defined __arm64__
+
+    .text
+    .align  2
+    .global alloca
+    .type   alloca, %function
+alloca:
+#ifdef __TINYC__
+    .int 0x91003c00
+    .int 0x927cec00
+#ifdef _WIN32
+    .int 0xb4000160
+    .int 0xd2820001
+    .int 0xeb01001f
+    .int 0x540000c3
+    .int 0xcb2163e2
+    .int 0xf940005f
+    .int 0xcb2163ff
+    .int 0xcb010000
+    .int 0x17fffffa
+#endif
+    .int 0xb4000040
+    .int 0xcb2063ff
+    .int 0x910003e0
+    .int 0xd65f03c0
+#else
+    add x0, x0, #15     // Round up to 16-byte boundary
+    and x0, x0, #-16    // Ensure 16-byte alignment
+    cbz x0, p100        // If size is 0, skip to return
+#ifdef _WIN32
+    // Windows requires page-wise allocation with stack probing
+    mov x1, #4096       // Page size = 4096 bytes
+
+p101:
+    cmp x0, x1          // Compare remaining size with page size
+    b.lo    p102        // If less than page, jump to remainder
+
+    // Probe first, then allocate
+    sub x2, sp, x1      // Calculate guard page address (sp - 4096)
+    ldr xzr, [x2]       // Touch guard page FIRST
+    sub sp, sp, x1      // THEN allocate the page
+
+    sub x0, x0, x1      // Decrement remaining size
+    b   p101            // Continue loop
+
+p102:
+    // Allocate remaining bytes (less than one page)
+    cbz x0, p100        // If no remaining bytes, skip
+    sub sp, sp, x0      // Allocate remaining space
+#else
+    // Non-Windows: simple one-time allocation
+    sub sp, sp, x0      // Allocate space on stack
+#endif
+
+p100:
+    mov x0, sp          // Return allocated address
+    ret                 // Return to caller
+#endif
+    .size   alloca, .-alloca
+/* ---------------------------------------------- */
 #endif