diff --git a/src/stub/src/amd64-linux.elf-main2.c b/src/stub/src/amd64-linux.elf-main2.c
index 18636d89..84f8e4a4 100644
--- a/src/stub/src/amd64-linux.elf-main2.c
+++ b/src/stub/src/amd64-linux.elf-main2.c
@@ -502,18 +502,14 @@ do_xmap(
     ElfW(Addr) reloc = 0;
     if (xi) { // compressed main program:
         // C_BASE space reservation, C_TEXT compressed data and stub
-        ElfW(Addr)  ehdr0 = *p_reloc;  // the 'hi' copy!
+        ElfW(Addr)  ehdr0 = *p_reloc;
         ElfW(Phdr) *phdr0 = (ElfW(Phdr) *)(1+ (ElfW(Ehdr) *)ehdr0);  // cheats .e_phoff
-        // Clear the 'lo' space reservation for use by PT_LOADs
-        ehdr0 -= phdr0[1].p_vaddr;  // the 'lo' copy
-        if (ET_EXEC == ehdr->e_type) {
-            ehdr0 = phdr0[0].p_vaddr;
+        v_brk = ehdr0 + phdr0->p_vaddr + phdr0->p_memsz;
+        if (ET_DYN == ehdr->e_type) {
+            reloc = ehdr0 - phdr0[1].p_vaddr;
         }
-        else {
-            reloc = ehdr0;
-        }
-        v_brk = phdr0->p_memsz + ehdr0;
-        munmap((void *)ehdr0, phdr0->p_memsz);
+        // paranoia: prevent "hangover" from VMA for C_BASE
+        munmap((void *)(reloc + phdr0->p_vaddr), phdr0->p_memsz);
     }
     else { // PT_INTERP
         DPRINTF("INTERP\\n", 0);
diff --git a/src/stub/src/arm.v4a-linux.elf-entry.S b/src/stub/src/arm.v4a-linux.elf-entry.S
index 8c87a830..65762a79 100644
--- a/src/stub/src/arm.v4a-linux.elf-entry.S
+++ b/src/stub/src/arm.v4a-linux.elf-entry.S
@@ -237,6 +237,13 @@ mfd  .req old_sp
         call f_expand
         add sp,sp,#2*NBPW  @ P_02, P_01  remove 5th param
 
+
+// faster than clearcache + msync; also avoids msync() bug
+        ldr arg3,[sp,#F_LENU]
+        ldr arg2,[sp,#F_ADRU]
+        mov arg1,mfd
+        do_sys __NR_write
+
         mov r5,#0
         mov r4,mfd
         mov r3,#MAP_PRIVATE|MAP_FIXED
diff --git a/src/stub/src/i386-linux.elf-main2.c b/src/stub/src/i386-linux.elf-main2.c
index a761a0b9..957ce409 100644
--- a/src/stub/src/i386-linux.elf-main2.c
+++ b/src/stub/src/i386-linux.elf-main2.c
@@ -603,18 +603,14 @@ do_xmap(
     ElfW(Addr) reloc = 0;
     if (xi) { // compressed main program:
         // C_BASE space reservation, C_TEXT compressed data and stub
-        ElfW(Addr)  ehdr0 = *p_reloc;  // the 'hi' copy!
+        ElfW(Addr)  ehdr0 = *p_reloc;
         ElfW(Phdr) *phdr0 = (ElfW(Phdr) *)(1+ (ElfW(Ehdr) *)ehdr0);  // cheats .e_phoff
-        // Clear the 'lo' space reservation for use by PT_LOADs
-        ehdr0 -= phdr0[1].p_vaddr;  // the 'lo' copy
-        if (ET_EXEC == ehdr->e_type) {
-            ehdr0 = phdr0[0].p_vaddr;
+        v_brk = ehdr0 + phdr0->p_vaddr + phdr0->p_memsz;
+        if (ET_DYN == ehdr->e_type) {
+            reloc = ehdr0 - phdr0[1].p_vaddr;
         }
-        else {
-            reloc = ehdr0;
-        }
-        v_brk = phdr0->p_memsz + ehdr0;
-        munmap((void *)ehdr0, phdr0->p_memsz);
+        // paranoia: prevent "hangover" from VMA for C_BASE
+        munmap((void *)(reloc + phdr0->p_vaddr), phdr0->p_memsz);
     }
     else { // PT_INTERP
         DPRINTF("INTERP ehdr=%%p  av=%%p\\n", ehdr, av);
@@ -662,8 +658,8 @@ do_xmap(
         DPRINTF("  mlen=%%p\\n", mlen);
 #endif
 
-        DPRINTF("mmap addr=%%p  mlen=%%p  offset=%%p  frag=%%p  prot=%%x\\n",
-            addr, mlen, phdr->p_offset - frag, frag, prot);
+        DPRINTF("mmap addr=%%p  mlen=%%p  phdr=%%p  offset=%%p  frag=%%p  prot=%%x\\n",
+            addr, mlen, phdr, phdr->p_offset - frag, frag, prot);
         int mfd = 0;
         if (xi && phdr->p_flags & PF_X) { // SELinux
             // Cannot set PROT_EXEC except via mmap() into a region (Linux "vma")