Embed APE loader inside APE

ape/ape.S

@@ -540,11 +540,32 @@ apesh:	.ascii	"'\n#'\"\n"			# sixth edition shebang
 //	modify the binary to follow the local system's convention.
 //	There isn't a one-size-fits-all approach for this, thus we
 //	present two choices.
+	.ascii	"o=\"$(command -v \"$0\")\"\n"
+//	Try to use a system-wide APE loader.
+	.ascii	"type ape-loader >/dev/null 2>&1 && "
+	.ascii	"exec ape-loader \"$0\" \"$@\"\n"
+#ifdef APE_LOADER
+//	There is no system-wide APE loader, but there is one
+//	embedded inside the APE. So if the system is not MacOs,
+//	extract the loader into a temp folder, and use it to
+//	load the APE without modifying it.
+	.ascii	"if [ ! -d /Applications ]; then\n"
+	.ascii	  "t=\"${TMPDIR:-/tmp}/ape-loader.$$\"\n"
+	.ascii	  "dd if=\"$o\" of=\"$t\" skip=\""
+	.shstub	    ape_loader_dd_skip,2
+	.ascii	    "\" count=\""
+	.shstub	    ape_loader_dd_count,2
+	.ascii	    "\" bs=\""
+	.shstub	    ape_loader_dd_bs,2
+	.ascii	    "\" 2>/dev/null\n"
+	.ascii	  "chmod 700 \"$t\"\n"
+	.ascii	  "exec \"$t\" \"$0\" \"$@\"\n"
+	.ascii	"fi\n"
+#endif
 #ifndef APE_NO_MODIFY_SELF
 //	The default behavior is: to overwrite the header in place.
 //	We prefer this because it's a tiny constant one time cost.
 //	We simply printf a 64-byte header and call execve() again.
-	.ascii	"o=\"$(command -v \"$0\")\"\n"
 #ifdef APE_BUILDSAFE
 //	This code is intended for binaries in build/bootstrap/. It
 //	causes them to not self-modify since they're checked-in to
@@ -645,6 +666,11 @@ apesh:	.ascii	"'\n#'\"\n"			# sixth edition shebang
 	.ascii	"fi\n"
 	.ascii	"exit $R\n"
 	.endobj	apesh
+#ifdef APE_LOADER
+        .section .ape.loader,"a",@progbits
+        .incbin APE_LOADER
+        .previous
+#endif /* APE_LOADER */
 #endif /* SupportsWindows() || SupportsMetal() || SupportsXnu() */
 
 #if SupportsSystemv() || SupportsMetal()

ape/ape.lds

@@ -205,6 +205,13 @@ SECTIONS {
     KEEP(*(.head2))
     KEEP(*(.text.head))
     
+    /* Loader */
+    . = ALIGN(64);
+    HIDDEN(ape_loader = .);
+    KEEP(*(.ape.loader))
+    . = ALIGN(64);
+    HIDDEN(ape_loader_end = .);
+
     /* Executable & Linkable Format */
     . = ALIGN(__SIZEOF_POINTER__);
     HIDDEN(ape_phdrs = .);
@@ -478,6 +485,17 @@ HIDDEN(ape_stack_filesz = 0);
 HIDDEN(ape_stack_memsz = STACKSIZE);
 HIDDEN(ape_stack_align = 16);
 
+HIDDEN(ape_loader_offset = ape_rom_offset + ape_loader - ape_rom_vaddr);
+HIDDEN(ape_loader_vaddr = ape_loader);
+HIDDEN(ape_loader_paddr = ape_rom_paddr + ape_loader_offset);
+HIDDEN(ape_loader_filesz = ape_loader_end - ape_loader);
+HIDDEN(ape_loader_memsz = ape_loader_filesz);
+HIDDEN(ape_loader_align = 64);
+HIDDEN(ape_loader_rva = RVA(ape_loader_vaddr));
+SHSTUB2(ape_loader_dd_bs, ape_loader_align);
+SHSTUB2(ape_loader_dd_skip, ape_loader_rva / ape_loader_align);
+SHSTUB2(ape_loader_dd_count, ape_loader_filesz / ape_loader_align);
+
 HIDDEN(ape_note_offset = ape_rom_offset + (ape_note - ape_rom_vaddr));
 HIDDEN(ape_note_vaddr = ape_note);
 HIDDEN(ape_note_paddr = ape_rom_paddr + ape_note_offset);
@@ -589,6 +607,13 @@ CHURN(ape_elf_shnum);
 CHURN(ape_elf_shoff);
 CHURN(ape_elf_shstrndx);
 CHURN(ape_macho_end);
+CHURN(ape_loader_offset);
+CHURN(ape_loader_vaddr);
+CHURN(ape_loader_paddr);
+CHURN(ape_loader_filesz);
+CHURN(ape_loader_memsz);
+CHURN(ape_loader_align);
+CHURN(ape_loader_rva);
 CHURN(ape_note);
 CHURN(ape_note_align);
 CHURN(ape_note_end);

ape/ape.mk

@@ -51,8 +51,8 @@ o/ape/idata.inc:			\
 $(APE_OBJS):	$(BUILD_FILES)		\
 		ape/ape.mk
 
-o/$(MODE)/ape/ape-no-modify-self.o: ape/ape.S
+o/$(MODE)/ape/ape-no-modify-self.o: ape/ape.S o/$(MODE)/examples/loader.elf
-	@$(COMPILE) -AOBJECTIFY.S $(OBJECTIFY.S) $(OUTPUT_OPTION) -DAPE_NO_MODIFY_SELF $<
+	@$(COMPILE) -AOBJECTIFY.S $(OBJECTIFY.S) $(OUTPUT_OPTION) -DAPE_LOADER="\"o/$(MODE)/examples/loader.elf\"" -DAPE_NO_MODIFY_SELF $<
 
 o/$(MODE)/ape/ape-buildsafe.o: ape/ape.S
 	@$(COMPILE) -AOBJECTIFY.S $(OBJECTIFY.S) $(OUTPUT_OPTION) -DAPE_BUILDSAFE $<