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

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();