diff --git a/Makefile b/Makefile
index d59f1fcd7..22bd54c65 100644
--- a/Makefile
+++ b/Makefile
@@ -60,7 +60,6 @@
 #   build/config.mk
 
 SHELL      = build/bootstrap/cocmd.com
-HOSTS     ?= freebsd openbsd netbsd rhel7 rhel5 xnu win10
 MAKEFLAGS += --no-builtin-rules
 
 .SUFFIXES:
diff --git a/build/definitions.mk b/build/definitions.mk
index 630b52abd..fdc598244 100644
--- a/build/definitions.mk
+++ b/build/definitions.mk
@@ -82,8 +82,10 @@ endif
 ifneq ($(findstring aarch64,$(MODE)),)
 ARCH = aarch64
 VM = o/third_party/qemu/qemu-aarch64
+HOSTS ?= pi silicon
 else
 ARCH = x86_64
+HOSTS ?= freebsd openbsd netbsd rhel7 rhel5 xnu win10
 endif
 
 ifneq ("$(wildcard o/third_party/gcc/bin/x86_64-pc-linux-gnu-*)","")
diff --git a/build/online.mk b/build/online.mk
index 0c405ca88..82b94e7a3 100644
--- a/build/online.mk
+++ b/build/online.mk
@@ -23,9 +23,9 @@
 #   - tool/build/runitd.c
 
 .PRECIOUS: o/$(MODE)/%.com.ok
-o/$(MODE)/%.com.ok: .PLEDGE = stdio rpath wpath cpath proc fattr inet
+o/$(MODE)/%.com.ok: private .PLEDGE = stdio rpath wpath cpath proc fattr inet
 o/$(MODE)/%.com.ok:				\
-		o/$(MODE)/tool/build/runit.com	\
-		o/$(MODE)/tool/build/runitd.com	\
+		o//tool/build/runit.com		\
+		o//tool/build/runitd.com	\
 		o/$(MODE)/%.com
 	@$(COMPILE) -wATEST -tT$@ $^ $(HOSTS)
diff --git a/examples/crashreport.c b/examples/crashreport.c
index 326356f74..b8daf4488 100644
--- a/examples/crashreport.c
+++ b/examples/crashreport.c
@@ -9,6 +9,7 @@
 #endif
 #include "libc/intrin/kprintf.h"
 #include "libc/log/log.h"
+#include "libc/math.h"
 #include "libc/runtime/symbols.internal.h"
 
 /**
@@ -25,13 +26,17 @@
  */
 
 noubsan int main(int argc, char *argv[]) {
-
   kprintf("----------------\n");
   kprintf(" THIS IS A TEST \n");
   kprintf("SIMULATING CRASH\n");
   kprintf("----------------\n");
 
-  volatile int64_t x;
   ShowCrashReports();
+
+  volatile double a = 0;
+  volatile double b = 23;
+  volatile double c = exp(b) / a;
+
+  volatile int64_t x;
   return 1 / (x = 0);
 }
diff --git a/libc/calls/struct/aarch64.h b/libc/calls/struct/aarch64.h
new file mode 100644
index 000000000..71a7d11a2
--- /dev/null
+++ b/libc/calls/struct/aarch64.h
@@ -0,0 +1,43 @@
+#ifndef COSMOPOLITAN_LIBC_CALLS_STRUCT_AARCH64_H_
+#define COSMOPOLITAN_LIBC_CALLS_STRUCT_AARCH64_H_
+#if !(__ASSEMBLER__ + __LINKER__ + 0)
+COSMOPOLITAN_C_START_
+
+#define FPSIMD_MAGIC 0x46508001
+#define ESR_MAGIC    0x45535201
+#define EXTRA_MAGIC  0x45585401
+#define SVE_MAGIC    0x53564501
+
+struct _aarch64_ctx {
+  uint32_t magic;
+  uint32_t size;
+};
+
+struct fpsimd_context {
+  struct _aarch64_ctx head;
+  uint32_t fpsr;
+  uint32_t fpcr;
+  uint128_t vregs[32];
+};
+
+struct esr_context {
+  struct _aarch64_ctx head;
+  uint64_t esr;
+};
+
+struct extra_context {
+  struct _aarch64_ctx head;
+  uint64_t datap;
+  uint32_t size;
+  uint32_t __reserved[3];
+};
+
+struct sve_context {
+  struct _aarch64_ctx head;
+  uint16_t vl;
+  uint16_t __reserved[3];
+};
+
+COSMOPOLITAN_C_END_
+#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */
+#endif /* COSMOPOLITAN_LIBC_CALLS_STRUCT_AARCH64_H_ */
diff --git a/libc/intrin/g_fds.c b/libc/intrin/g_fds.c
index f1ef9df32..90961e857 100644
--- a/libc/intrin/g_fds.c
+++ b/libc/intrin/g_fds.c
@@ -49,8 +49,6 @@ static textwindows dontinline void SetupWinStd(struct Fds *fds, int i, int x) {
 textstartup void InitializeFileDescriptors(void) {
   struct Fds *fds;
   __fds_lock_obj._type = PTHREAD_MUTEX_RECURSIVE;
-  pthread_atfork(_weaken(__fds_lock), _weaken(__fds_unlock),
-                 _weaken(__fds_funlock));
   fds = VEIL("r", &g_fds);
   fds->n = 4;
   atomic_store_explicit(&fds->f, 3, memory_order_relaxed);
diff --git a/libc/log/oncrash_arm64.c b/libc/log/oncrash_arm64.c
index 96d885143..e732e4950 100644
--- a/libc/log/oncrash_arm64.c
+++ b/libc/log/oncrash_arm64.c
@@ -19,6 +19,7 @@
 #include "ape/sections.internal.h"
 #include "libc/assert.h"
 #include "libc/calls/calls.h"
+#include "libc/calls/struct/aarch64.h"
 #include "libc/calls/struct/rusage.internal.h"
 #include "libc/calls/struct/siginfo.h"
 #include "libc/calls/struct/sigset.h"
@@ -44,6 +45,7 @@ STATIC_YOINK("strerror_wr");  // for kprintf %m
 STATIC_YOINK("strsignal_r");  // for kprintf %G
 
 #define RESET   "\e[0m"
+#define BOLD    "\e[1m"
 #define STRONG  "\e[30;101m"
 #define RED     "\e[31;1m"
 #define GREEN   "\e[32;1m"
@@ -203,6 +205,7 @@ relegated void __oncrash_arm64(int sig, struct siginfo *si, void *arg) {
       const char *addr2line;
       struct StackFrame *fp;
       struct SymbolTable *st;
+      struct fpsimd_context *vc;
       st = GetSymbolTable();
       debugbin = FindDebugBinary();
       addr2line = GetAddr2linePath();
@@ -217,13 +220,30 @@ relegated void __oncrash_arm64(int sig, struct siginfo *si, void *arg) {
         for (j = 0; j < 4; ++j) {
           int r = 8 * j + i;
           if (j) Append(b, " ");
-          Append(b, "%s%016lx%s %d%s", ColorRegister(r),
+          Append(b, "%s%016lx%s x%d%s", ColorRegister(r),
                  ctx->uc_mcontext.regs[r], reset, r,
                  r == 8 || r == 9 ? " " : "");
         }
         Append(b, "\n");
       }
 
+      // PRINT VECTORS
+      vc = (struct fpsimd_context *)ctx->uc_mcontext.__reserved;
+      if (vc->head.magic == FPSIMD_MAGIC) {
+        int n = 16;
+        while (n && !vc->vregs[n - 1] && !vc->vregs[n - 2]) n -= 2;
+        for (i = 0; i * 2 < n; ++i) {
+          Append(b, " ");
+          for (j = 0; j < 2; ++j) {
+            int r = j + 2 * i;
+            if (j) Append(b, " ");
+            Append(b, "%016lx ..%s %016lx v%d%s", (long)(vc->vregs[r] >> 64),
+                   !j ? "" : ".", (long)vc->vregs[r], r, r < 10 ? " " : "");
+          }
+          Append(b, "\n");
+        }
+      }
+
       // PRINT CURRENT LOCATION
       //
       // We can get the address of the currently executing function by
diff --git a/libc/nexgen32e/checkstackalign.S b/libc/nexgen32e/checkstackalign.S
index 60bd5cce8..767cc43c3 100644
--- a/libc/nexgen32e/checkstackalign.S
+++ b/libc/nexgen32e/checkstackalign.S
@@ -26,22 +26,20 @@ CheckStackIsAligned:
 
 	push	%rbp
 	mov	%rsp,%rbp
-
-//	allocate sixteen bytes
-	push	%rax
-	push	%rax
-
-//	put a value in it
-	xorps	%xmm0,%xmm0
-	movaps	%xmm0,-16(%rbp)
-
-	leave
+	test	$15,%rsp
+	jz	1f
+	int3				// misaligned stack trap
+1:	pop	%rbp
 	ret
 
 #elif defined(__aarch64__)
 
-//	TODO: support me
-	mov	x0,#1
+	stp	x29,x30,[sp,#-16]!
+	mov	x29,sp
+	and	x0,x29,#15
+	cbz	x0,1f
+	brk	#666			// misaligned stack trap
+1:	ldp	x29,x30,[sp],#16
 	ret
 
 #else
diff --git a/libc/runtime/clone.c b/libc/runtime/clone.c
index 4de63de01..716dcb983 100644
--- a/libc/runtime/clone.c
+++ b/libc/runtime/clone.c
@@ -313,7 +313,7 @@ static errno_t CloneOpenbsd(int (*func)(void *, int), char *stk, size_t stksz,
   sp &= -alignof(struct __tfork);
   tf = (struct __tfork *)sp;
   sp -= sizeof(struct CloneArgs);
-  sp &= -MAX(16, alignof(struct CloneArgs));
+  sp &= -alignof(struct CloneArgs);
   wt = (struct CloneArgs *)sp;
   wt->ctid = flags & CLONE_CHILD_SETTID ? ctid : &wt->tid;
   wt->ztid = flags & CLONE_CHILD_CLEARTID ? ctid : &wt->tid;
@@ -442,10 +442,10 @@ static int CloneNetbsd(int (*func)(void *, int), char *stk, size_t stksz,
 
 static void *SiliconThreadMain(void *arg) {
   register struct CloneArgs *wt asm("x21") = arg;
-  asm volatile("ldr\tx28,%0" : /* no outputs */ : "m"(wt->tls));
   *wt->ctid = wt->this;
-  register long x0 asm("x0") = (long)wt->arg;
-  register long x1 asm("x1") = (long)wt->tid;
+  register void *x0 asm("x0") = wt->arg;
+  register int x1 asm("w1") = wt->this;
+  register void *x28 asm("x28") = wt->tls;
   asm volatile("mov\tx19,x29\n\t"  // save frame pointer
                "mov\tx20,sp\n\t"   // save stack pointer
                "mov\tx29,#0\n\t"   // reset backtrace
@@ -454,7 +454,7 @@ static void *SiliconThreadMain(void *arg) {
                "mov\tx29,x19\n\t"  // restore frame pointer
                "mov\tsp,x20"       // restore stack pointer
                : "+r"(x0)
-               : "r"(x1), "r"(wt->func), "r"(wt)
+               : "r"(x1), "r"(wt->func), "r"(wt), "r"(x28)
                : "x19", "x20", "memory");
   *wt->ztid = 0;
   return 0;
diff --git a/libc/runtime/cosmo2.c b/libc/runtime/cosmo2.c
index 85e9cf53d..ef4dce934 100644
--- a/libc/runtime/cosmo2.c
+++ b/libc/runtime/cosmo2.c
@@ -19,6 +19,7 @@
 #include "libc/calls/calls.h"
 #include "libc/calls/syscall-sysv.internal.h"
 #include "libc/dce.h"
+#include "libc/intrin/kprintf.h"
 #include "libc/intrin/strace.internal.h"
 #include "libc/macros.internal.h"
 #include "libc/nexgen32e/rdtsc.h"
@@ -44,12 +45,12 @@ typedef int init_f(int argc, char **argv, char **envp, unsigned long *auxv);
 
 extern long syscon_start[];
 extern long syscon_end[];
-extern long syscon_linux[];
-extern long syscon_xnu[];
-extern long syscon_freebsd[];
-extern long syscon_openbsd[];
-extern long syscon_netbsd[];
-extern long syscon_windows[];
+extern char syscon_linux[];
+extern char syscon_xnu[];
+extern char syscon_freebsd[];
+extern char syscon_openbsd[];
+extern char syscon_netbsd[];
+extern char syscon_windows[];
 extern init_f __strace_init;
 extern init_f *__preinit_array_start[] __attribute__((__weak__));
 extern init_f *__preinit_array_end[] __attribute__((__weak__));
@@ -61,13 +62,24 @@ extern char ape_stack_prot[] __attribute__((__weak__));
 extern pthread_mutex_t __mmi_lock_obj;
 extern int hostos asm("__hostos");
 
+static const char *DecodeMagnum(const char *p, long *r) {
+  int k = 0;
+  unsigned long c, x = 0;
+  do {
+    c = *p++;
+    x |= (c & 127) << k;
+    k += 7;
+  } while (c & 128);
+  return *r = x, p;
+}
+
 textstartup void cosmo(long *sp, struct Syslib *m1) {
   int argc;
+  long *mp;
   init_f **fp;
   uintptr_t *pp;
-  long *mp, *magnums;
-  char **argv, **envp;
   unsigned long *auxv;
+  char **argv, **envp, *magnums;
 
   // get startup timestamp as early as possible
   // its used by --strace and also kprintf() %T
@@ -93,7 +105,7 @@ textstartup void cosmo(long *sp, struct Syslib *m1) {
 
   // setup system magic numbers
   for (mp = syscon_start; mp < syscon_end; ++mp) {
-    *mp = *magnums++;
+    magnums = DecodeMagnum(magnums, mp);
   }
 
   // check system call abi compatibility
@@ -131,6 +143,7 @@ textstartup void cosmo(long *sp, struct Syslib *m1) {
 #endif
 #endif
 
+  // initialize file system
   InitializeFileDescriptors();
 
   // set helpful globals
diff --git a/libc/runtime/set_tls.c b/libc/runtime/set_tls.c
index cb74e559f..e0a0d50c1 100644
--- a/libc/runtime/set_tls.c
+++ b/libc/runtime/set_tls.c
@@ -61,9 +61,7 @@ textstartup void __set_tls(struct CosmoTib *tib) {
                    "d"((uint32_t)(val >> 32)));
   }
 #else
-  asm volatile("mov\tx28,%0" : /* no outputs */ : "r"(tib));
-  if (!IsXnu()) {
-    asm volatile("msr\ttpidr_el0,%0" : /* no outputs */ : "r"(tib));
-  }
+  register long x28 asm("x28") = (long)tib;
+  asm volatile("" : "+r"(x28));
 #endif
 }
diff --git a/libc/runtime/vfork.S b/libc/runtime/vfork.S
index 342673c23..aa1cff693 100644
--- a/libc/runtime/vfork.S
+++ b/libc/runtime/vfork.S
@@ -67,13 +67,13 @@ vfork:
 #endif
 	pop	%rbp
 #endif
-	mov	%fs:0,%r9		# get thread information block
+	mov	%fs:0,%r9		// get thread information block
 #if SupportsWindows()
 	testb	IsWindows()
-	jnz	6f			# and we're lucky to have that
+	jnz	6f			// and we're lucky to have that
 #endif
 #ifdef __SANITIZE_ADDRESS__
-	jmp	5f			# TODO: asan and vfork don't mix?
+	jmp	5f			// TODO: asan and vfork don't mix?
 #endif
 #if SupportsXnu()
 	testb	IsXnu()
@@ -81,10 +81,10 @@ vfork:
 #endif
 #if SupportsOpenbsd()
 	testb	IsOpenbsd()
-	jnz	5f			# fake vfork plus msyscall issues
+	jnz	5f			// fake vfork plus msyscall issues
 #endif
-	mov	0x3c(%r9),%r8d		# avoid question of @vforksafe errno
-	pop	%rsi			# saves return address in a register
+	mov	0x3c(%r9),%r8d		// avoid question of @vforksafe errno
+	pop	%rsi			// saves return address in a register
 	mov	__NR_vfork(%rip),%eax
 #if SupportsBsd()
 	clc
@@ -95,10 +95,10 @@ vfork:
 	neg	%rax
 0:
 #endif
-	push	%rsi			# note it happens twice in same page
+	push	%rsi			// note it happens twice in same page
 	cmp	$-4095,%eax
 	jae	systemfive_error
-	mov	%r8d,0x3c(%r9)		# restore errno
+	mov	%r8d,0x3c(%r9)		// restore errno
 1:	test	%eax,%eax
 	jnz	.Lpar
 .Lchi:	orb	$TIB_FLAG_VFORKED,0x40(%r9)
@@ -121,7 +121,7 @@ vfork:
 	mov	%rsp,%rbp
 	push	%r9
 	push	%r9
-	xor	%edi,%edi		# dwCreationFlags
+	xor	%edi,%edi		// dwCreationFlags
 	call	sys_fork_nt
 	pop	%r9
 	pop	%r9
@@ -131,11 +131,31 @@ vfork:
 
 #elif defined(__aarch64__)
 
-	mov	x8,#220			// __NR_clone
+	adrp	x0,__hostos
+	ldr	w0,[x0,#:lo12:__hostos]
+	tbz	x0,3,1f			// bit 3 is xnu
+	b	fork			// which doesn't support vfork()
+1:	mov	x8,#220			// __NR_clone
 	mov	x0,#0x4111		// SIGCHLD | CLONE_VM | CLONE_VFORK
 	mov	x1,#0
 	svc	0
-	.hidden	_sysret
+
+//	if (!rc) {
+//	  __get_tls()->tib_flags |= TIB_FLAG_VFORKED;
+//	} else {
+//	  __get_tls()->tib_flags &= ~TIB_FLAG_VFORKED;
+//	}
+	sub	x1,x28,#1152
+	ldr	x2,[x1,0x40]
+	cbnz	x0,2f
+	orr	x2,x2,#TIB_FLAG_VFORKED
+1:	str	x2,[x1,0x40]
+	b	3f
+2:	and	x2,x2,#~TIB_FLAG_VFORKED
+	b	1b
+
+//	if (rc < 0) errno = -rc, rc = -1;
+3:	.hidden	_sysret
 	b	_sysret
 
 #else
diff --git a/libc/sysv/consts.sh b/libc/sysv/consts.sh
index 50f759358..663894bca 100755
--- a/libc/sysv/consts.sh
+++ b/libc/sysv/consts.sh
@@ -601,8 +601,8 @@ syscon	clock	CLOCK_MONOTONIC_PRECISE			1			1			1			6			11			3			3			1			#
 syscon	clock	CLOCK_MONOTONIC_FAST			1			1			1			6			12			3			3			1			#
 syscon	clock	CLOCK_MONOTONIC_COARSE			6			6			1			6			12			3			3			1			# Linux 2.6.32+; bsd consensus; not available on RHEL5
 syscon	clock	CLOCK_MONOTONIC_RAW			4			4			127			4			127			127			127			127			# actually monotonic; not subject to NTP adjustments; Linux 2.6.28+; XNU/NT/FreeBSD/OpenBSD faked; not available on RHEL5
-syscon	clock	CLOCK_PROCESS_CPUTIME_ID		2			2			127			127			15			2			0x40000000		127			#
-syscon	clock	CLOCK_THREAD_CPUTIME_ID			3			3			127			127			14			4			0x20000000		127			#
+syscon	clock	CLOCK_PROCESS_CPUTIME_ID		2			2			127			12			15			2			0x40000000		127			#
+syscon	clock	CLOCK_THREAD_CPUTIME_ID			3			3			127			16			14			4			0x20000000		127			#
 syscon	clock	CLOCK_PROF				127			127			127			127			2			127			2			127			#
 syscon	clock	CLOCK_BOOTTIME				7			7			127			127			127			6			127			127			#
 syscon	clock	CLOCK_REALTIME_ALARM			8			8			127			127			127			127			127			127			#
diff --git a/libc/sysv/consts/CLOCK_PROCESS_CPUTIME_ID.S b/libc/sysv/consts/CLOCK_PROCESS_CPUTIME_ID.S
index e6f376dac..2b8c354db 100644
--- a/libc/sysv/consts/CLOCK_PROCESS_CPUTIME_ID.S
+++ b/libc/sysv/consts/CLOCK_PROCESS_CPUTIME_ID.S
@@ -1,2 +1,2 @@
 #include "libc/sysv/consts/syscon.internal.h"
-.syscon clock,CLOCK_PROCESS_CPUTIME_ID,2,2,127,127,15,2,0x40000000,127
+.syscon clock,CLOCK_PROCESS_CPUTIME_ID,2,2,127,12,15,2,0x40000000,127
diff --git a/libc/sysv/consts/CLOCK_THREAD_CPUTIME_ID.S b/libc/sysv/consts/CLOCK_THREAD_CPUTIME_ID.S
index 13c493389..7d5893688 100644
--- a/libc/sysv/consts/CLOCK_THREAD_CPUTIME_ID.S
+++ b/libc/sysv/consts/CLOCK_THREAD_CPUTIME_ID.S
@@ -1,2 +1,2 @@
 #include "libc/sysv/consts/syscon.internal.h"
-.syscon clock,CLOCK_THREAD_CPUTIME_ID,3,3,127,127,14,4,0x20000000,127
+.syscon clock,CLOCK_THREAD_CPUTIME_ID,3,3,127,16,14,4,0x20000000,127
diff --git a/libc/sysv/consts/syscon.internal.h b/libc/sysv/consts/syscon.internal.h
index f14368282..3f4677d6a 100644
--- a/libc/sysv/consts/syscon.internal.h
+++ b/libc/sysv/consts/syscon.internal.h
@@ -21,10 +21,7 @@
 // clang-format off
 
 #ifdef __x86_64__
-#define ENCODE .uleb128
 .yoink	_init_systemfive
-#else
-#define ENCODE .balign 8 ; .quad
 #endif
 
 .macro	.syscon	group:req name:req linux:req linux_aarch64:req xnu:req xnu_aarch64:req freebsd:req openbsd:req netbsd:req windows:req
@@ -38,9 +35,9 @@
 #if SupportsLinux() || SupportsMetal()
 	.section .sort.rodata.syscon.linux.2.\group\().\name,"a",@progbits
 #ifdef __aarch64__
-	ENCODE	\linux_aarch64
+	.uleb128	\linux_aarch64
 #else
-	ENCODE	\linux
+	.uleb128	\linux
 #endif
 	.previous
 #endif
@@ -48,34 +45,34 @@
 #if SupportsXnu()
 	.section .sort.rodata.syscon.xnu.2.\group\().\name,"a",@progbits
 #ifdef __aarch64__
-	ENCODE	\xnu_aarch64
+	.uleb128	\xnu_aarch64
 #else
-	ENCODE	\xnu
+	.uleb128	\xnu
 #endif
 	.previous
 #endif
 
 #if SupportsFreebsd()
 	.section .sort.rodata.syscon.freebsd.2.\group\().\name,"a",@progbits
-	ENCODE	\freebsd
+	.uleb128	\freebsd
 	.previous
 #endif
 
 #if SupportsOpenbsd()
 	.section .sort.rodata.syscon.openbsd.2.\group\().\name,"a",@progbits
-	ENCODE	\openbsd
+	.uleb128	\openbsd
 	.previous
 #endif
 
 #if SupportsNetbsd()
 	.section .sort.rodata.syscon.netbsd.2.\group\().\name,"a",@progbits
-	ENCODE	\netbsd
+	.uleb128	\netbsd
 	.previous
 #endif
 
 #if SupportsWindows()
 	.section .sort.rodata.syscon.windows.2.\group\().\name,"a",@progbits
-	ENCODE	\windows
+	.uleb128	\windows
 	.previous
 #endif
 
diff --git a/libc/sysv/syscon.S b/libc/sysv/syscon.S
index e1ab0704b..42ef038c2 100644
--- a/libc/sysv/syscon.S
+++ b/libc/sysv/syscon.S
@@ -19,12 +19,6 @@
 #include "libc/dce.h"
 #include "libc/macros.internal.h"
 
-#ifdef __x86_64__
-#define ROALIGN 1
-#else
-#define ROALIGN 8
-#endif
-
 //	Sections for varint encoded magic numbers.
 //
 //	These sections are all ordered by (group_name, constant_name).
@@ -46,7 +40,6 @@ syscon_end:
 	.globl	syscon_end
 #if SupportsLinux() || SupportsMetal()
 	.section .sort.rodata.syscon.linux.1,"a",@progbits
-	.balign	ROALIGN
 syscon_linux:/*
 	...decentralized leb128...
 	*/.previous
@@ -55,7 +48,6 @@ syscon_linux:/*
 #endif
 #if SupportsXnu()
 	.section .sort.rodata.syscon.xnu.1,"a",@progbits
-	.balign	ROALIGN
 syscon_xnu:/*
 	...decentralized leb128...
 	*/.previous
@@ -64,7 +56,6 @@ syscon_xnu:/*
 #endif
 #if SupportsFreebsd()
 	.section .sort.rodata.syscon.freebsd.1,"a",@progbits
-	.balign	ROALIGN
 syscon_freebsd:/*
 	...decentralized leb128...
 	*/.previous
@@ -73,7 +64,6 @@ syscon_freebsd:/*
 #endif
 #if SupportsOpenbsd()
 	.section .sort.rodata.syscon.openbsd.1,"a",@progbits
-	.balign	ROALIGN
 syscon_openbsd:/*
 	...decentralized leb128...
 	*/.previous
@@ -82,7 +72,6 @@ syscon_openbsd:/*
 #endif
 #if SupportsNetbsd()
 	.section .sort.rodata.syscon.netbsd.1,"a",@progbits
-	.balign	ROALIGN
 syscon_netbsd:/*
 	...decentralized leb128...
 	*/.previous
@@ -91,7 +80,6 @@ syscon_netbsd:/*
 #endif
 #if SupportsWindows()
 	.section .sort.rodata.syscon.windows.1,"a",@progbits
-	.balign	ROALIGN
 syscon_windows:/*
 	...decentralized leb128...
 	*/.previous
diff --git a/libc/thread/pthread_atfork.c b/libc/thread/pthread_atfork.c
index 361c45593..7b94ff7c2 100644
--- a/libc/thread/pthread_atfork.c
+++ b/libc/thread/pthread_atfork.c
@@ -17,6 +17,7 @@
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/assert.h"
+#include "libc/calls/state.internal.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
 #include "libc/intrin/kmalloc.h"
@@ -58,13 +59,15 @@ static void _pthread_onfork(int i) {
 void _pthread_onfork_prepare(void) {
   _pthread_onfork(0);
   pthread_spin_lock(&_pthread_lock);
-  __kmalloc_lock();
+  __fds_lock();
   __mmi_lock();
+  __kmalloc_lock();
 }
 
 void _pthread_onfork_parent(void) {
-  __mmi_unlock();
   __kmalloc_unlock();
+  __mmi_unlock();
+  __fds_unlock();
   pthread_spin_unlock(&_pthread_lock);
   _pthread_onfork(1);
 }
@@ -82,11 +85,12 @@ void _pthread_onfork_child(void) {
   atomic_store_explicit(&pt->cancelled, false, memory_order_relaxed);
 
   // wipe core runtime locks
+  __kmalloc_unlock();
   pthread_mutexattr_init(&attr);
   pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
   pthread_mutex_init(&__mmi_lock_obj, &attr);
+  pthread_mutex_init(&__fds_lock_obj, &attr);
   pthread_spin_init(&_pthread_lock, 0);
-  __kmalloc_unlock();
 
   // call user-supplied forked child callbacks
   _pthread_onfork(2);
diff --git a/test/libc/thread/pthread_kill_test.c b/test/libc/thread/pthread_kill_test.c
index cfb84c325..34257f16f 100644
--- a/test/libc/thread/pthread_kill_test.c
+++ b/test/libc/thread/pthread_kill_test.c
@@ -18,7 +18,9 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/calls/calls.h"
 #include "libc/calls/struct/sigaction.h"
+#include "libc/dce.h"
 #include "libc/errno.h"
+#include "libc/stdio/stdio.h"
 #include "libc/sysv/consts/sig.h"
 #include "libc/testlib/testlib.h"
 #include "libc/thread/thread.h"
@@ -26,6 +28,13 @@
 int fds[2];
 _Thread_local sig_atomic_t gotsig;
 
+void SetUp(void) {
+  if (IsXnuSilicon()) {
+    fprintf(stderr, "TODO(jart): Get pthread_kill() working on XNU silicon\n");
+    exit(0);
+  }
+}
+
 void OnSig(int sig) {
   gotsig = 1;
 }
diff --git a/tool/build/runitd.c b/tool/build/runitd.c
index 4c6fd9d87..18af831ee 100644
--- a/tool/build/runitd.c
+++ b/tool/build/runitd.c
@@ -432,11 +432,20 @@ void HandleClient(void) {
     sigaction(SIGQUIT, &(struct sigaction){0}, 0);
     sigprocmask(SIG_SETMASK, &savemask, 0);
     int i = 0;
-    char *args[4] = {0};
+    const char *exe;
+    char *args[8] = {0};
+    if (!IsXnuSilicon()) {
+      exe = g_exepath;
+    } else {
+      exe = "ape-m1.com";
+      args[i++] = exe;
+      args[i++] = "-";
+      args[i++] = g_exepath;
+    }
     args[i++] = g_exepath;
     if (use_strace) args[i++] = "--strace";
     if (use_ftrace) args[i++] = "--ftrace";
-    execv(g_exepath, args);
+    execvp(exe, args);
     _Exit(127);
   }
   close(pipefds[1]);