Fix breakages in Linux-only build modes

- compile.com now polyfills -march=native which gcc/clang removed
- Guarantee zero Windows code is linked into non-Windows binaries
- MODE=tinylinux binaries are now back to being as tiny as ~4kb
- Improve the runtime's stack allocation / alignment hack
- GitHub Actions now tests Linux modes for assurance

.github/workflows/build.yml

@@ -17,7 +17,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        mode: ["", tiny, rel]
+        mode: ["", tiny, rel, tinylinux, optlinux]
     steps:
       - uses: actions/checkout@v3
 

ape/aarch64.lds

@@ -275,6 +275,7 @@ SECTIONS {
 }
 
 ape_stack_memsz = DEFINED(ape_stack_memsz) ? ape_stack_memsz : APE_STACKSIZE;
+ape_stack_memsz2 = ape_stack_memsz * 2;
 _tls_size = _tbss_end - _tdata_start;
 _tdata_size = _tdata_end - _tdata_start;
 _tbss_size = _tbss_end - _tbss_start;

ape/ape.S

@@ -794,7 +794,7 @@ ape_loader_end:
 	.stub	ape_stack_vaddr,quad		// is mmap()'d with MAP_FIXED
 	.stub	ape_stack_paddr,quad		// ignored
 	.stub	ape_stack_filesz,quad		// ignored
-	.stub	ape_stack_memsz,quad		// is mmap(size) argument
+	.stub	ape_stack_memsz2,quad		// is mmap(size) argument
 	.stub	ape_stack_align,quad		// must be 16+
 
 #if SupportsOpenbsd() || SupportsNetbsd()
@@ -1718,7 +1718,7 @@ ape_grub_entry:
 │ αcτµαlly pδrταblε εxεcµταblε § cosmopolitan libc runtime runtime         ─╬─│┼
 ╚────────────────────────────────────────────────────────────────────────────│*/
 kernel:	movabs	$ape_stack_vaddr,%rsp
-	add	$ape_stack_memsz,%rsp
+	add	$ape_stack_memsz2,%rsp
 	movl	$0,0x7b000			// unmap null 2mb
 #if USE_SYMBOL_HACK
 	.byte	0x0f,0x1f,0207			// nop rdi binbase

ape/ape.lds

@@ -317,7 +317,6 @@ SECTIONS {
     __privileged_start = .;
     *(.privileged)
     __privileged_end = .;
-    . += . > 0 ? CODE_GRANULE : 0;
 
     KEEP(*(.ape.pad.text))
     . = ALIGN(CONSTANT(COMMONPAGESIZE));
@@ -344,9 +343,11 @@ SECTIONS {
 
 /*BEGIN: read-only data that's only needed for initialization */
 
+#if SupportsWindows()
     /* Windows DLL Import Directory */
     KEEP(*(.idata.ro));
     KEEP(*(SORT_BY_NAME(.idata.ro.*)))
+#endif
 
     . = ALIGN(__SIZEOF_POINTER__);
     __init_array_start = .;
@@ -405,7 +406,9 @@ SECTIONS {
 
   .data . : {
 /*BEGIN: Read/Write Data */
+#if SupportsWindows()
     KEEP(*(SORT_BY_NAME(.piro.data.sort.iat.*)))
+#endif
 /*BEGIN: NT FORK COPYING */
     KEEP(*(.dataprologue))
     *(.data .data.*)
@@ -517,6 +520,11 @@ SECTIONS {
   }
 
   /DISCARD/ : {
+#if !SupportsWindows()
+    *(.idata.ro);
+    *(.idata.ro.*)
+    *(.piro.data.sort.iat.*)
+#endif
     *(__patchable_function_entries)
     *(__mcount_loc)
     *(.discard)
@@ -569,6 +577,7 @@ ape_stack_vaddr = DEFINED(ape_stack_vaddr) ? ape_stack_vaddr : 0x700000000000;
 ape_stack_paddr = ape_ram_paddr + ape_ram_filesz;
 ape_stack_filesz = 0;
 ape_stack_memsz = DEFINED(ape_stack_memsz) ? ape_stack_memsz : APE_STACKSIZE;
+ape_stack_memsz2 = ape_stack_memsz * 2;
 ape_stack_align = 16;
 
 ape_note_offset = ape_cod_offset + (ape_note - ape_cod_vaddr);

build/config.mk

@@ -76,12 +76,13 @@ endif
 ifeq ($(MODE), opt)
 ENABLE_FTRACE = 1
 CONFIG_OFLAGS ?= -g
-CONFIG_CPPFLAGS += -DNDEBUG -DSYSDEBUG -msse2avx -Wa,-msse2avx
+CONFIG_CPPFLAGS += -DNDEBUG -DSYSDEBUG
 CONFIG_CCFLAGS += $(BACKTRACES) -O3 -fmerge-all-constants
-TARGET_ARCH ?= -march=skylake
+TARGET_ARCH ?= -march=native
 endif
 
 # Optimized Linux Mode
+# The Fastest Mode of All
 #
 #   - `make MODE=optlinux`
 #   - Turns on red zone
@@ -91,8 +92,9 @@ endif
 #   - Turns off support for other operating systems
 #
 ifeq ($(MODE), optlinux)
+ENABLE_FTRACE = 1
 CONFIG_OFLAGS ?= -g
-CONFIG_CPPFLAGS += -DNDEBUG -msse2avx -Wa,-msse2avx -DSUPPORT_VECTOR=1
+CONFIG_CPPFLAGS += -DNDEBUG -DSYSDEBUG -DSUPPORT_VECTOR=1
 CONFIG_CCFLAGS += -O3 -fmerge-all-constants
 CONFIG_COPTS += -mred-zone
 TARGET_ARCH ?= -march=native
@@ -114,7 +116,7 @@ endif
 #   - CHECK_xx() won't leak strings into binary
 #
 ifeq ($(MODE), rel)
-CONFIG_CPPFLAGS += -DNDEBUG
+CONFIG_CPPFLAGS += -DNDEBUG -DDWARFLESS
 CONFIG_CCFLAGS += $(BACKTRACES) -O2
 TARGET_ARCH ?= -msse3
 PYFLAGS += -O1

libc/calls/clock_nanosleep.c

@@ -55,8 +55,10 @@ static errno_t sys_clock_nanosleep(int clock, int flags,
     rc = sys_clock_nanosleep_xnu(clock, flags, req, rem);
   } else if (IsOpenbsd()) {
     rc = sys_clock_nanosleep_openbsd(clock, flags, req, rem);
-  } else {
+  } else if (IsWindows()) {
     rc = sys_clock_nanosleep_nt(clock, flags, req, rem);
+  } else {
+    rc = enosys();
   }
   if (rc == -1) {
     rc = errno;

libc/calls/getpriority.c

@@ -25,6 +25,7 @@
 #include "libc/intrin/describeflags.internal.h"
 #include "libc/intrin/strace.internal.h"
 #include "libc/limits.h"
+#include "libc/sysv/errfuns.h"
 
 /**
  * Returns nice value of thing.
@@ -72,8 +73,10 @@ int getpriority(int which, unsigned who) {
       errno = rc;
       rc = -1;
     }
-  } else {
+  } else if (IsWindows()) {
     rc = sys_getpriority_nt(which, who);
+  } else {
+    rc = enosys();
   }
 #else
   rc = sys_getpriority(which, who);

libc/crt/crt.S

@@ -71,14 +71,15 @@ _start:
 //	align stack to GetStackSize() so GetStackAddr() is fast
 	.weak	ape_stack_memsz
 	mov	$ape_stack_memsz,%r9d
-	mov	$ape_stack_align,%r8d
-	cmp	$_HOSTLINUX,%cl
-	cmove	%r9d,%r8d
+	mov	$16,%r8d
+	test	%r9d,%r9d
+	cmovnz	%r9,%r8
 	neg	%r8
 	and	%r8,%rsp
 	xor	%ebp,%ebp
 //	bofram	9f
 
+#if SupportsWindows()
 //	make win32 imps noop
 	.weak	ape_idata_iat
 	.weak	ape_idata_iatend
@@ -88,6 +89,7 @@ _start:
 	sub	%rdi,%rcx
 	shr	$3,%ecx
 	rep stosq
+#endif
 
 //	scan through environment varis
 //	find start of auxiliary values

libc/errno.h

@@ -9,7 +9,7 @@ COSMOPOLITAN_C_START_
  */
 
 #if defined(__GNUC__) && defined(__x86_64__) && defined(__MNO_RED_ZONE__) && \
-    !defined(__STRICT_ANSI__)
+    !defined(__STRICT_ANSI__) && !defined(__cplusplus)
 #define errno                                                           \
   (*({                                                                  \
     errno_t *_ep;                                                       \

libc/integral/normalize.inc

@@ -80,7 +80,7 @@
 #define FRAMESIZE     0x10000
 #define PAGESIZE      0x1000 /* i386+ */
 #else
-#define APE_STACKSIZE 8388608 /* default 8mb stack */
+#define APE_STACKSIZE 4194304 /* default 4mb stack */
 #endif
 #define APE_PAGESIZE  0x10000 /* i386+ */
 #define APE_GUARDSIZE 0x4000  /* b/c apple m1 */

libc/intrin/strerror_r.c

@@ -16,6 +16,7 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
+#include "libc/dce.h"
 #include "libc/nt/runtime.h"
 #include "libc/str/str.h"
 
@@ -26,5 +27,6 @@
  * @return 0 on success, or error code
  */
 int strerror_r(int err, char *buf, size_t size) {
-  return strerror_wr(err, GetLastError(), buf, size);
+  int winerr = IsWindows() ? GetLastError() : 0;
+  return strerror_wr(err, winerr, buf, size);
 }

libc/runtime/enable_tls.c

@@ -222,7 +222,9 @@ textstartup void __enable_tls(void) {
 
 #ifdef __x86_64__
   // rewrite the executable tls opcodes in memory
-  __morph_tls();
+  if (IsWindows() || IsXnu()) {
+    __morph_tls();
+  }
 #endif
 
   // we are now allowed to use tls

libc/runtime/init.S

@@ -76,7 +76,7 @@ _init_check_rdi_rsi:
 	jne	1b
 3:	.endfn	_init_check_rdi_rsi
 #endif
-_woot:	leave
+	leave
 #elif defined(__aarch64__)
 	ldp	x29,x30,[sp],#16
 #endif

libc/runtime/morph_tls.c

@@ -45,77 +45,75 @@ privileged void __morph_tls(void) {
   // We check `_tls_content` which is generated by the linker script
   // since it lets us determine ahead of time if _Thread_local vars
   // have actually been linked into this program.
-  if (IsWindows() || IsXnu()) {
-    int n;
-    uint64_t w;
-    sigset_t mask;
-    unsigned m, dis;
-    unsigned char *p;
-    __morph_begin(&mask);
+  int n;
+  uint64_t w;
+  sigset_t mask;
+  unsigned m, dis;
+  unsigned char *p;
+  __morph_begin(&mask);
 
-    if (IsXnu()) {
-      // Apple is quite straightforward to patch. We basically
-      // just change the segment register, and the linear slot
-      // address 0x30 was promised to us, according to Go team
-      // https://github.com/golang/go/issues/23617
-      dis = 0x30;
-    } else {
-      // MSVC __declspec(thread) generates binary code for this
-      // %gs:0x1480 abi. So long as TlsAlloc() isn't called >64
-      // times we should be good.
-      dis = 0x1480 + __tls_index * 8;
-    }
-
-    // iterate over modifiable code looking for 9 byte instruction
-    // this would take 30 ms using xed to enable tls on python.com
-    for (p = _ereal; p + 9 <= __privileged_start; p += n) {
-
-      // use sse to zoom zoom to fs register prefixes
-      // that way it'll take 1 ms to morph python.com
-      while (p + 9 + 16 <= __privileged_start) {
-        if ((m = __builtin_ia32_pmovmskb128(
-                 *(xmm_t *)p == (xmm_t){0144, 0144, 0144, 0144, 0144, 0144,
-                                        0144, 0144, 0144, 0144, 0144, 0144,
-                                        0144, 0144, 0144, 0144}))) {
-          m = __builtin_ctzll(m);
-          p += m;
-          break;
-        } else {
-          p += 16;
-        }
-      }
-
-      // we're checking for the following expression:
-      //   0144 == p[0] &&           // %fs
-      //   0110 == (p[1] & 0373) &&  // rex.w (and ignore rex.r)
-      //   (0213 == p[2] ||          // mov reg/mem → reg (word-sized)
-      //   0003 == p[2]) &&          // add reg/mem → reg (word-sized)
-      //   0004 == (p[3] & 0307) &&  // mod/rm (4,reg,0) means sib → reg
-      //   0045 == p[4] &&           // sib (5,4,0) → (rbp,rsp,0) → disp32
-      //   0000 == p[5] &&           // displacement (von Neumann endian)
-      //   0000 == p[6] &&           // displacement
-      //   0000 == p[7] &&           // displacement
-      //   0000 == p[8]              // displacement
-      w = READ64LE(p) & READ64LE("\377\373\377\307\377\377\377\377");
-      if ((w == READ64LE("\144\110\213\004\045\000\000\000") ||
-           w == READ64LE("\144\110\003\004\045\000\000\000")) &&
-          !p[8]) {
-
-        // now change the code
-        p[0] = 0145;                       // change %fs to %gs
-        p[5] = (dis & 0x000000ff) >> 000;  // displacement
-        p[6] = (dis & 0x0000ff00) >> 010;  // displacement
-        p[7] = (dis & 0x00ff0000) >> 020;  // displacement
-        p[8] = (dis & 0xff000000) >> 030;  // displacement
-
-        // advance to the next instruction
-        n = 9;
-      } else {
-        n = 1;
-      }
-    }
-
-    __morph_end(&mask);
+  if (IsXnu()) {
+    // Apple is quite straightforward to patch. We basically
+    // just change the segment register, and the linear slot
+    // address 0x30 was promised to us, according to Go team
+    // https://github.com/golang/go/issues/23617
+    dis = 0x30;
+  } else {
+    // MSVC __declspec(thread) generates binary code for this
+    // %gs:0x1480 abi. So long as TlsAlloc() isn't called >64
+    // times we should be good.
+    dis = 0x1480 + __tls_index * 8;
   }
+
+  // iterate over modifiable code looking for 9 byte instruction
+  // this would take 30 ms using xed to enable tls on python.com
+  for (p = _ereal; p + 9 <= __privileged_start; p += n) {
+
+    // use sse to zoom zoom to fs register prefixes
+    // that way it'll take 1 ms to morph python.com
+    while (p + 9 + 16 <= __privileged_start) {
+      if ((m = __builtin_ia32_pmovmskb128(
+               *(xmm_t *)p == (xmm_t){0144, 0144, 0144, 0144, 0144, 0144, 0144,
+                                      0144, 0144, 0144, 0144, 0144, 0144, 0144,
+                                      0144, 0144}))) {
+        m = __builtin_ctzll(m);
+        p += m;
+        break;
+      } else {
+        p += 16;
+      }
+    }
+
+    // we're checking for the following expression:
+    //   0144 == p[0] &&           // %fs
+    //   0110 == (p[1] & 0373) &&  // rex.w (and ignore rex.r)
+    //   (0213 == p[2] ||          // mov reg/mem → reg (word-sized)
+    //   0003 == p[2]) &&          // add reg/mem → reg (word-sized)
+    //   0004 == (p[3] & 0307) &&  // mod/rm (4,reg,0) means sib → reg
+    //   0045 == p[4] &&           // sib (5,4,0) → (rbp,rsp,0) → disp32
+    //   0000 == p[5] &&           // displacement (von Neumann endian)
+    //   0000 == p[6] &&           // displacement
+    //   0000 == p[7] &&           // displacement
+    //   0000 == p[8]              // displacement
+    w = READ64LE(p) & READ64LE("\377\373\377\307\377\377\377\377");
+    if ((w == READ64LE("\144\110\213\004\045\000\000\000") ||
+         w == READ64LE("\144\110\003\004\045\000\000\000")) &&
+        !p[8]) {
+
+      // now change the code
+      p[0] = 0145;                       // change %fs to %gs
+      p[5] = (dis & 0x000000ff) >> 000;  // displacement
+      p[6] = (dis & 0x0000ff00) >> 010;  // displacement
+      p[7] = (dis & 0x00ff0000) >> 020;  // displacement
+      p[8] = (dis & 0xff000000) >> 030;  // displacement
+
+      // advance to the next instruction
+      n = 9;
+    } else {
+      n = 1;
+    }
+  }
+
+  __morph_end(&mask);
 #endif
 }

libc/sock/epoll.c

@@ -1410,6 +1410,7 @@ err:
   return -1;
 }
 
+#if SupportsWindows()
 textwindows int sys_close_epoll_nt(int fd) {
   struct PortState *port_state;
   struct TsTreeNode *tree_node;
@@ -1427,6 +1428,7 @@ err:
   err_check_handle(g_fds.p[fd].handle);
   return -1;
 }
+#endif
 
 /**
  * Creates new epoll instance.

libc/testlib/testrunner.c

@@ -223,7 +223,9 @@ void testlib_runtestcases(testfn_t *start, testfn_t *end, testfn_t warmup) {
     if (_weaken(testlib_enable_tmp_setup_teardown)) SetupTmpDir();
     if (_weaken(SetUp)) _weaken(SetUp)();
     errno = 0;
-    SetLastError(0);
+    if (IsWindows()) {
+      SetLastError(0);
+    }
     if (!IsWindows()) sys_getpid();
     if (warmup) warmup();
     testlib_clearxmmregisters();

test/libc/calls/mkntpath_test.c

@@ -18,8 +18,10 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/internal.h"
 #include "libc/calls/syscall_support-nt.internal.h"
+#include "libc/dce.h"
 #include "libc/mem/gc.internal.h"
 #include "libc/testlib/testlib.h"
+#if SupportsWindows()
 
 char16_t p[PATH_MAX];
 
@@ -48,3 +50,5 @@ TEST(mkntpath, testRemoveDoubleSlash) {
   EXPECT_EQ(21, __mkntpath("C:\\Users\\jart\\\\.config", p));
   EXPECT_STREQ(u"C:\\Users\\jart\\.config", p);
 }
+
+#endif /* SupportsWindows() */

test/libc/thread/pthread_setname_np_test.c

@@ -100,13 +100,14 @@ TEST(pthread_setname_np, GetNameOfOtherThread) {
   ASSERT_EQ(0, pthread_create(&id, 0, GetNameOfOtherThreadWorker, 0));
   while (!atomic_load(&sync1)) pthread_yield();
   errno_t e = pthread_getname_np(id, me, sizeof(me));
-  if (IsLinux() && e == ENOENT) return;  // bah old kernel
-  if (IsLinux() && e == EACCES) return;  // meh landlock
+  if (IsLinux() && e == ENOENT) goto GiveUp;  // bah old kernel
+  if (IsLinux() && e == EACCES) goto GiveUp;  // meh landlock
   ASSERT_EQ(0, e);
   EXPECT_STREQ("justine", me);
   ASSERT_EQ(0, pthread_setname_np(id, "tunney"));
   ASSERT_EQ(0, pthread_getname_np(id, me, sizeof(me)));
   EXPECT_STREQ("tunney", me);
+GiveUp:
   atomic_store(&sync2, 1);
   ASSERT_EQ(0, pthread_join(id, 0));
 }

third_party/lua/lunix.c

@@ -205,7 +205,7 @@ int LuaUnixSysretErrno(lua_State *L, const char *call, int olderr) {
   struct UnixErrno *ep;
   int i, unixerr, winerr;
   unixerr = errno;
-  winerr = GetLastError();
+  winerr = IsWindows() ? GetLastError() : 0;
   if (!IsTiny() && !(0 < unixerr && unixerr < (!IsWindows() ? 4096 : 65536))) {
     WARNF("errno should not be %d", unixerr);
   }

third_party/python/Lib/ntpath.py

@@ -664,10 +664,7 @@ def commonpath(paths):
 try:
     # GetFinalPathNameByHandle is available starting with Windows 6.0.
     # Windows XP and non-Windows OS'es will mock _getfinalpathname.
-    if sys.getwindowsversion()[:2] >= (6, 0):
-        _getfinalpathname = posix._getfinalpathname
-    else:
-        raise ImportError
+    _getfinalpathname = posix._getfinalpathname
 except (AttributeError, ImportError, OSError):
     # On Windows XP and earlier, two files are the same if their absolute
     # pathnames are the same.

third_party/python/Modules/clinic/posixmodule.inc

@@ -946,6 +946,10 @@ os__getfinalpathname(PyObject *module, PyObject *arg)
     PyObject *return_value = NULL;
     PyObject *path;
 
+    if (!IsWindows()) {
+        Py_RETURN_NONE;
+    }
+
     if (!PyArg_Parse(arg, "U:_getfinalpathname", &path)) {
         goto exit;
     }
@@ -1037,6 +1041,9 @@ os__getvolumepathname_impl(PyObject *module, PyObject *path);
 static PyObject *
 os__getvolumepathname(PyObject *module, PyObject **args, Py_ssize_t nargs, PyObject *kwnames)
 {
+    if (!IsWindows()) {
+        Py_RETURN_NONE;
+    }
     PyObject *return_value = NULL;
     static const char * const _keywords[] = {"path", NULL};
     static _PyArg_Parser _parser = {"U:_getvolumepathname", _keywords, 0};

third_party/python/Python/errors.c

@@ -4,6 +4,7 @@
 │ Python 3                                                                     │
 │ https://docs.python.org/3/license.html                                       │
 ╚─────────────────────────────────────────────────────────────────────────────*/
+#include "libc/dce.h"
 #include "libc/errno.h"
 #include "libc/mem/mem.h"
 #include "libc/nt/enum/formatmessageflags.h"
@@ -11,6 +12,7 @@
 #include "libc/nt/memory.h"
 #include "libc/nt/process.h"
 #include "libc/nt/runtime.h"
+#include "libc/runtime/runtime.h"
 #include "libc/x/x.h"
 #include "third_party/python/Include/abstract.h"
 #include "third_party/python/Include/dictobject.h"
@@ -589,6 +591,10 @@ PyObject *PyErr_SetExcFromWindowsErrWithFilenameObjects(
     PyObject *message;
     PyObject *args, *v;
     uint32_t err = (uint32_t)ierr;
+    if (!IsWindows()) {
+        PyErr_SetString(PyExc_SystemError, "this is not windows");
+        return 0;
+    }
     if (err==0) err = GetLastError();
     len = FormatMessage(
         /* Error API error */

third_party/python/Python/sysmodule.c

@@ -1032,6 +1032,11 @@ sys_getwindowsversion(PyObject *self)
     void *verblock;
     uint32_t verblock_size;
 
+    if (!IsWindows()) {
+        PyErr_SetString(PyExc_SystemError, "this is not windows");
+        return 0;
+    }
+
     ver.dwOSVersionInfoSize = sizeof(ver);
     if (!GetVersionEx(&ver))
         return PyErr_SetFromWindowsErr(0);

third_party/sqlite3/zipfile.c

@@ -1420,8 +1420,8 @@ static int zipfileGetMode(
 ** Both (const char*) arguments point to nul-terminated strings. Argument
 ** nB is the value of strlen(zB). This function returns 0 if the strings are
 ** identical, ignoring any trailing '/' character in either path.  */
-static int zipfileComparePath(const char *zA, const char *zB, int nB){
-  int nA = (int)strlen(zA);
+static int zipfileComparePath(const char *zA, const char *zB, size_t nB){
+  size_t nA = strlen(zA);
   if( nA>0 && zA[nA-1]=='/' ) nA--;
   if( nB>0 && zB[nB-1]=='/' ) nB--;
   if( nA==nB && memcmp(zA, zB, nA)==0 ) return 0;

third_party/xxhash/xxh_x86dispatch.c

@@ -82,10 +82,12 @@ extern "C" {
 #  define XXH_X86DISPATCH_ALLOW_AVX
 #endif
 
+#if 0 /* [jart] be quiet */
 #if defined(__AVX__) && !defined(XXH_X86DISPATCH_ALLOW_AVX)
 #  error "Error: if xxh_x86dispatch.c is compiled with AVX enabled, the resulting binary will crash on sse2-only cpus !! " \
          "If you nonetheless want to do that, please enable the XXH_X86DISPATCH_ALLOW_AVX build variable"
 #endif
+#endif
 
 #ifdef __has_include
 #  define XXH_HAS_INCLUDE(header) __has_include(header)

tool/build/assimilate.c

@@ -232,7 +232,7 @@ void Assimilate(void) {
     kprintf("%s: fstat() failed: %m\n", prog);
     exit(14);
   }
-  if (st.st_size < 8192) {
+  if (st.st_size < 4096) {
     kprintf("%s: ape binaries must be at least 4096 bytes\n", prog);
     exit(15);
   }

tool/build/compile.c

@@ -48,6 +48,7 @@
 #include "libc/mem/mem.h"
 #include "libc/nexgen32e/kcpuids.h"
 #include "libc/nexgen32e/x86feature.h"
+#include "libc/nexgen32e/x86info.h"
 #include "libc/runtime/runtime.h"
 #include "libc/stdio/append.h"
 #include "libc/stdio/stdio.h"
@@ -911,10 +912,10 @@ int main(int argc, char *argv[]) {
   }
 
   s = basename(strdup(cmd));
-  if (strstr(s, "gcc")) {
+  if (strstr(s, "gcc") || strstr(s, "g++")) {
     iscc = true;
     isgcc = true;
-  } else if (strstr(s, "clang")) {
+  } else if (strstr(s, "clang") || strstr(s, "clang++")) {
     iscc = true;
     isclang = true;
   } else if (strstr(s, "ld.bfd")) {
@@ -1027,6 +1028,113 @@ int main(int argc, char *argv[]) {
       if (isgcc) {
         AddArg(argv[i]);
       }
+
+#ifdef __x86_64__
+    } else if (!strcmp(argv[i], "-march=native")) {
+      struct X86ProcessorModel *model;
+      if (X86_HAVE(ABM)) AddArg("-mabm");
+      if (X86_HAVE(XOP)) AddArg("-mxop");
+      if (X86_HAVE(SSE4A)) AddArg("-msse4a");
+      if (X86_HAVE(SSE3)) AddArg("-msse3");
+      if (X86_HAVE(SSSE3)) AddArg("-mssse3");
+      if (X86_HAVE(SSE4_1)) AddArg("-msse4.1");
+      if (X86_HAVE(SSE4_2)) AddArg("-msse4.2");
+      if (X86_HAVE(AVX)) AddArg("-mavx");
+      if (X86_HAVE(AVX2)) {
+        AddArg("-mavx2");
+        if (isgcc) {
+          AddArg("-msse2avx");
+          AddArg("-Wa,-msse2avx");
+        }
+      }
+      if (X86_HAVE(AVX512F)) AddArg("-mavx512f");
+      if (X86_HAVE(AVX512PF)) AddArg("-mavx512pf");
+      if (X86_HAVE(AVX512ER)) AddArg("-mavx512er");
+      if (X86_HAVE(AVX512CD)) AddArg("-mavx512cd");
+      if (X86_HAVE(AVX512VL)) AddArg("-mavx512vl");
+      if (X86_HAVE(AVX512BW)) AddArg("-mavx512bw");
+      if (X86_HAVE(AVX512DQ)) AddArg("-mavx512dq");
+      if (X86_HAVE(AVX512IFMA)) AddArg("-mavx512ifma");
+      if (X86_HAVE(AVX512VBMI)) AddArg("-mavx512vbmi");
+      if (X86_HAVE(SHA)) AddArg("-msha");
+      if (X86_HAVE(AES)) AddArg("-maes");
+      if (X86_HAVE(VAES)) AddArg("-mvaes");
+      if (X86_HAVE(PCLMUL)) AddArg("-mpclmul");
+      if (X86_HAVE(FSGSBASE)) AddArg("-mfsgsbase");
+      if (X86_HAVE(F16C)) AddArg("-mf16c");
+      if (X86_HAVE(FMA)) AddArg("-mfma");
+      if (X86_HAVE(POPCNT)) AddArg("-mpopcnt");
+      if (X86_HAVE(BMI)) AddArg("-mbmi");
+      if (X86_HAVE(BMI2)) AddArg("-mbmi2");
+      if (X86_HAVE(ADX)) AddArg("-madx");
+      if (X86_HAVE(FXSR)) AddArg("-mfxsr");
+      if ((model = getx86processormodel(kX86ProcessorModelKey))) {
+        switch (model->march) {
+          case X86_MARCH_CORE2:
+            AddArg("-march=core2");
+            break;
+          case X86_MARCH_NEHALEM:
+            AddArg("-march=nehalem");
+            break;
+          case X86_MARCH_WESTMERE:
+            AddArg("-march=westmere");
+            break;
+          case X86_MARCH_SANDYBRIDGE:
+            AddArg("-march=sandybridge");
+            break;
+          case X86_MARCH_IVYBRIDGE:
+            AddArg("-march=ivybridge");
+            break;
+          case X86_MARCH_HASWELL:
+            AddArg("-march=haswell");
+            break;
+          case X86_MARCH_BROADWELL:
+            AddArg("-march=broadwell");
+            break;
+          case X86_MARCH_SKYLAKE:
+          case X86_MARCH_KABYLAKE:
+            AddArg("-march=skylake");
+            break;
+          case X86_MARCH_CANNONLAKE:
+            AddArg("-march=cannonlake");
+            break;
+          case X86_MARCH_ICELAKE:
+            if (model->grade >= X86_GRADE_SERVER) {
+              AddArg("-march=icelake-server");
+            } else {
+              AddArg("-march=icelake-client");
+            }
+            break;
+          case X86_MARCH_TIGERLAKE:
+            AddArg("-march=tigerlake");
+            break;
+          case X86_MARCH_BONNELL:
+          case X86_MARCH_SALTWELL:
+            AddArg("-march=bonnell");
+            break;
+          case X86_MARCH_SILVERMONT:
+          case X86_MARCH_AIRMONT:
+            AddArg("-march=silvermont");
+            break;
+          case X86_MARCH_GOLDMONT:
+            AddArg("-march=goldmont");
+            break;
+          case X86_MARCH_GOLDMONTPLUS:
+            AddArg("-march=goldmont-plus");
+            break;
+          case X86_MARCH_TREMONT:
+            AddArg("-march=tremont");
+            break;
+          case X86_MARCH_KNIGHTSLANDING:
+            AddArg("-march=knl");
+            break;
+          case X86_MARCH_KNIGHTSMILL:
+            AddArg("-march=knm");
+            break;
+        }
+      }
+#endif /* __x86_64__ */
+
     } else if (!strcmp(argv[i], "-fsanitize=address")) {
       if (isgcc && ccversion >= 6) wantasan = true;
     } else if (!strcmp(argv[i], "-fsanitize=undefined")) {

tool/emacs/cosmo-stuff.el

@@ -25,7 +25,7 @@
 (require 'ld-script)
 (require 'make-mode)
 
-(setq cosmo-dbg-mode "zero")
+(setq cosmo-dbg-mode "dbg")
 (setq cosmo-default-mode "")
 (setq c-doc-comment-style 'javadown)
 

tool/viz/fixconsole.c

@@ -22,6 +22,8 @@
 #include "libc/nt/enum/consolemodeflags.h"
 #include "libc/nt/runtime.h"
 #include "libc/runtime/runtime.h"
+#include "libc/stdio/stdio.h"
+#if defined(__x86_64__) && SupportsWindows()
 
 int main(int argc, char *argv[]) {
   if (!IsWindows()) {
@@ -40,3 +42,11 @@ int main(int argc, char *argv[]) {
                      kNtEnableVirtualTerminalProcessing);
   _Exit(0);
 }
+
+#else
+int main(int argc, char *argv[]) {
+  fprintf(stderr,
+          "fixconsole not supported on this cpu arch or build config\n");
+  return 1;
+}
+#endif /* __x86_64__ && SupportsWindows() */

tool/viz/printpeb.c

@@ -41,7 +41,7 @@
 #include "libc/time/time.h"
 #include "tool/decode/lib/flagger.h"
 #include "tool/decode/lib/idname.h"
-#ifdef __x86_64__
+#if defined(__x86_64__) && SupportsWindows()
 
 char *GetString(struct NtUnicodeString *s) {
   static char buf[1024];
@@ -538,4 +538,9 @@ int main(int argc, char *argv[]) {
   return 0;
 }
 
-#endif /* __x86_64__ */
+#else
+int main(int argc, char *argv[]) {
+  fprintf(stderr, "printpeb not supported on this cpu arch or build config\n");
+  return 1;
+}
+#endif /* __x86_64__ && SupportsWindows() */

tool/viz/virtualquery.c

@@ -26,6 +26,7 @@
 #include "libc/runtime/runtime.h"
 #include "libc/stdio/stdio.h"
 #include "libc/str/str.h"
+#if defined(__x86_64__) && SupportsWindows()
 
 /**
  * @fileoverview WIN32 Virtual Memory Layout Dump Utility
@@ -73,3 +74,11 @@ int main(int argc, char *argv[]) {
            (DescribeNtPageFlags)(b[4], mi.Protect));
   }
 }
+
+#else
+int main(int argc, char *argv[]) {
+  fprintf(stderr,
+          "virtualquery not supported on this cpu arch or build config\n");
+  return 1;
+}
+#endif /* __x86_64__ && SupportsWindows() */