+void mono_emit_stack_alloc (MBCGEN_TYPE* s, MBTREE_TYPE* tree)
+{
+#ifdef PLATFORM_WIN32
+ guint8* br[5];
+ int sreg;
+
+ /*
+ * Under Windows:
+ * If requested stack size is larger than one page,
+ * perform stack-touch operation
+ * (see comments in mono_emit_stack_alloc_const below).
+ */
+ x86_test_reg_imm (s->code, tree->left->reg1, ~0xFFF);
+ br[0] = s->code; x86_branch8 (s->code, X86_CC_Z, 0, FALSE);
+
+ sreg = tree->left->reg1;
+
+ br[2] = s->code; /* loop */
+ x86_alu_reg_imm (s->code, X86_SUB, X86_ESP, 0x1000);
+ x86_test_membase_reg (s->code, X86_ESP, 0, X86_ESP);
+ x86_alu_reg_imm (s->code, X86_SUB, sreg, 0x1000);
+ x86_alu_reg_imm (s->code, X86_CMP, sreg, 0x1000);
+ br[3] = s->code; x86_branch8 (s->code, X86_CC_AE, 0, FALSE);
+ x86_patch (br[3], br[2]);
+ x86_test_reg_reg (s->code, sreg, sreg);
+ br[4] = s->code; x86_branch8 (s->code, X86_CC_Z, 0, FALSE);
+ x86_alu_reg_reg (s->code, X86_SUB, X86_ESP, sreg);
+
+ br[1] = s->code; x86_jump8 (s->code, 0);
+
+ x86_patch (br[0], s->code);
+ x86_alu_reg_reg (s->code, X86_SUB, X86_ESP, tree->left->reg1);
+ x86_patch (br[1], s->code);
+ x86_patch (br[4], s->code);
+#else /* PLATFORM_WIN32 */
+ x86_alu_reg_reg (s->code, X86_SUB, X86_ESP, tree->left->reg1);
+#endif
+}
+
+void mono_emit_stack_alloc_const (MBCGEN_TYPE* s, MBTREE_TYPE* tree, int size)
+{
+#ifdef PLATFORM_WIN32
+ int i, npages;
+ guint8* br[2];
+
+ if (size > 0xFFE) {
+ /*
+ * Generate stack probe code.
+ * Under Windows, it is necessary to allocate one page at a time,
+ * "touching" stack after each successful sub-allocation. This is
+ * because of the way stack growth is implemented - there is a
+ * guard page before the lowest stack page that is currently commited.
+ * Stack normally grows sequentially so OS traps access to the
+ * guard page and commits more pages when needed.
+ */
+ npages = ((unsigned) size) >> 12;
+ if (npages > 4) {
+ if (tree->reg1 != X86_EAX && tree->left->reg1 != X86_EAX) {
+ x86_alu_reg_imm (s->code, X86_SUB, X86_ESP, 0x1000);
+ x86_test_membase_reg (s->code, X86_ESP, 0, X86_ESP);
+ x86_mov_membase_reg (s->code, X86_ESP, 0x1000 - 4, X86_EAX, 4); /* save EAX */
+ x86_mov_reg_imm (s->code, X86_EAX, npages - 1);
+ } else {
+ x86_mov_reg_imm (s->code, X86_EAX, npages);
+ }
+ br[0] = s->code; /* loop */
+ x86_alu_reg_imm (s->code, X86_SUB, X86_ESP, 0x1000);
+ x86_test_membase_reg (s->code, X86_ESP, 0, X86_ESP);
+ x86_dec_reg (s->code, X86_EAX);
+ br[1] = s->code; x86_branch8 (s->code, X86_CC_NZ, 0, TRUE);
+ x86_patch (br[1], br[0]);
+ if (tree->reg1 != X86_EAX && tree->left->reg1 != X86_EAX)
+ x86_mov_reg_membase (s->code, X86_EAX, X86_ESP, (npages * 0x1000) - 4, 4); /* restore EAX */
+ } else {
+ /* generate unrolled code for relatively small allocs */
+ for (i = npages; --i >= 0;) {
+ x86_alu_reg_imm (s->code, X86_SUB, X86_ESP, 0x1000);
+ x86_test_membase_reg (s->code, X86_ESP, 0, X86_ESP);
+ }
+ }
+ }
+
+ if (size & 0xFFF) x86_alu_reg_imm (s->code, X86_ADD, X86_ESP, -(size & 0xFFF));
+#else /* PLATFORM_WIN32 */
+ x86_alu_reg_imm (s->code, X86_SUB, X86_ESP, size);
+#endif
+}
+