Hojun-Cho/ww
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

w6l+selfhost: BSS optimisation — trim trailing .data zeros from filesz

Browse Source 
Scan the consolidated .data buffer (post-relocation) for trailing zero
bytes; set the R+W PT_LOAD's p_filesz to exclude them while p_memsz
covers the full region. The loader zero-fills the gap, so behaviour is
unchanged. Saves up to a page per binary on programs whose globals are
zero-init.

Mirrored in selfhost/cmd/w6l/out.ww so test 992's byte-identity diff
still holds. Dynamic-link path is untouched — it still errors on any
mutable global; that's the next feature.
This commit is contained in:
Hojun-Cho
2026-05-12 13:44:27 +09:00
parent 548547a1d0
commit 2480f4c272
4 changed files with 142 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
selfhost/cmd/w6l/out.ww
View File

@@ -63,6 +63,21 @@ export fn emitelf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
dataoff = (rxend + PAGE_SZ - 1u64) & ~(PAGE_SZ - 1u64);
};
// BSS optimisation: trailing zero bytes in .data can be left
// out of the file. The loader zero-fills the gap between
// p_filesz and p_memsz. Scan after l_relocate has applied any
// DATAR patches — anything still zero at the tail genuinely is
// zero-init. Matches cmd/w6l/out.c byte-for-byte.
let bsslen: u64 = 0u64;
if (hasdata) {
for (bsslen < l.datalen) {
let b: u8 = l.data[l.datalen - 1u64 - bsslen];
if (b != 0u8) { break; };
bsslen += 1u64;
};
};
let datafilelen: u64 = l.datalen - bsslen;
// One contiguous header buffer covering [0..0x1000), then .text.
let hdr: *u8 = os.alloc(TEXT_OFF): *u8; // zero-initialised by mmap
@@ -106,7 +121,7 @@ export fn emitelf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
wru64(hdr, 128u64, dataoff); // p_offset
wru64(hdr, 136u64, base + dataoff); // p_vaddr
wru64(hdr, 144u64, base + dataoff); // p_paddr
wru64(hdr, 152u64, l.datalen); // p_filesz
wru64(hdr, 152u64, datafilelen); // p_filesz
wru64(hdr, 160u64, l.datalen); // p_memsz
wru64(hdr, 168u64, PAGE_SZ); // p_align
};
@@ -128,8 +143,10 @@ export fn emitelf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
};
if (n2 != l.textlen: i64) { return -1; };
};
if (hasdata) {
// Pad to the page-aligned data offset, then write .data.
if (hasdata && datafilelen > 0u64) {
// Pad to the page-aligned data offset, then write only
// the non-zero prefix of .data. The rest is BSS — the
// loader zero-fills from p_filesz to p_memsz.
let here: u64 = TEXT_OFF + l.textlen;
let zero: u8 = 0u8;
for (here < dataoff) {
@@ -140,13 +157,13 @@ export fn emitelf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
};
here += 1u64;
};
let r4: (i64 | os.oserror) = os.writeall(fd, l.data, l.datalen);
let r4: (i64 | os.oserror) = os.writeall(fd, l.data, datafilelen);
let n4: i64 = 0i64;
match (r4) {
case let v: i64 => n4 = v;
case let e: os.oserror => return -1;
};
if (n4 != l.datalen: i64) { return -1; };
if (n4 != datafilelen: i64) { return -1; };
};
return 0;
};