From 17a597978e7c7aa4311fe1b8fb186a6232ccfe4d Mon Sep 17 00:00:00 2001
From: Amand Tihon <amand.tihon@alrj.org>
Date: Sat, 30 Jul 2011 20:38:45 +0200
Subject: [PATCH] First support for .gnu.hash/DT_GNU_HASH.

* Not optimized yet.
* Strange bug when using cmove that I'll have to investigate.
* Thanks go to kernel_error and flure for the motivation :)
---
 runtime/bold_ibh-x86_64.asm |  68 ++++++++++++++++++++++++------------
 runtime/bold_ibh-x86_64.o   | Bin 1504 -> 1776 bytes
 2 files changed, 46 insertions(+), 22 deletions(-)

diff --git a/runtime/bold_ibh-x86_64.asm b/runtime/bold_ibh-x86_64.asm
index a9a3732..f309ce7 100644
--- a/runtime/bold_ibh-x86_64.asm
+++ b/runtime/bold_ibh-x86_64.asm
@@ -65,33 +65,55 @@ _bold__ibh_start:
       mov rdx, [r14 + 16]                 ; link_map->l_ld
 
       ; Find the interesting entries in the DYNAMIC table.
-      .dynamic_loop:
-
-        push byte DT_HASH                 ; DT_HASH == 4
-        pop rax
-        cmp [rdx], eax
-        cmove r9, [rdx+8]                 ; r9 : pointer to the hash table
-
-        inc eax                           ; DT_STRTAB == 5
-        cmp [rdx], eax
-        cmove r10, [rdx+8]                ; r10 : pointer to strtab
 
-        inc eax                           ; DT_SYMTAB == 6
-        cmp [rdx], eax
-        cmove r11, [rdx+8]                ; r11 : pointer to symtab
+      .dynamic_loop:
+        mov rsi, [rdx+8]
+
+        cmp [rdx], dword 4                ; DT_HASH == 4
+        jne .no_hash                      ; Why the heck does it segfault when
+          mov ecx, [rsi+4]                ; I use cmove ecx, [rsi+4] instead
+        .no_hash:                         ; of this jne construction ????
+
+        cmp [rdx], dword 5                ; DT_STRTAB == 5
+        cmove r10, rsi                    ; r10 : pointer to strtab
+
+        cmp [rdx], dword 6                ; DT_SYMTAB == 6
+        cmove r11, rsi                    ; r11 : pointer to symtab
+
+        cmp [rdx], dword 0x6ffffef5       ; DT_GNU_HASH
+        jne .not_gnu_hash
+          mov ecx, [rsi]                  ; ecx is nbuckets
+          mov r9d, [rsi + 8]              ; maskwords (64 bits entries)
+          lea r9, [rsi + 16 + r9 * 8]     ; rdx points to the first bucket
+
+          xor eax, eax
+          .walk_buckets:                  ; Find the last chain
+            cmp eax, [r9]
+            cmovl eax, [r9]
+            add r9, 4                     ; next bucket
+            loop .walk_buckets            ; eax is offset to the last chain
+          ; Now, rdx points to the beginning of the first chain
+
+          mov ecx, [rsi+4]                ; symndx
+          xchg eax, ecx
+          shl rax, 2                      ; symndx * 4
+          sub r9, rax                     ; hash values - symndx * 4
+          .walk_chain:
+            bt dword [r9 + rcx*4], 0      ;
+            inc ecx                       ; Increment ecx for each entry found
+            jc .walk_chain                ; in the chain.
+          dec ecx                         ; fixup.
+        .not_gnu_hash:
 
         ; Next dynamic entry
-        lea rdx, [rdx + 16]               ; add rdx, 16
-        xor eax, eax
-        cmp [rdx], eax
-        ;cmp [rdx], dword 0               ; Gut feeling: would harm compression
+        add rdx, 16
+        cmp [rdx], dword 0
         jnz short .dynamic_loop
 
       ; All DYNAMIC entries have been read.
-      mov ecx, [r9 + 4]                   ; nchain, number of exported symbols
 
       ; Iterate over the symbols in the library (symtab entries).
-      .symbolloop:
+      .symtabloop:
         ; Find the symbol name in strtab
         mov esi, [r11]                    ; st_name, offset in strtab
         add rsi, r10                      ; pointer to symbol name
@@ -110,11 +132,11 @@ _bold__ibh_start:
         cmp edx, ebx                      ; Compare with stored hash
         je short .found
         lea r11, [r11 + 24]               ; Next symtab entry
-      loop .symbolloop
+      loop .symtabloop
 
       ; Symbol was not found in this library
       mov r14, [r14 + 24]                 ; Next link_map entry
-      jmp short .libloop
+      jmp .libloop
     .found:
     mov rax, [r11 + 8]                    ; st_value, offset of the symbol
     add rax, [r14]                        ; add link_map->l_addr
@@ -123,7 +145,9 @@ _bold__ibh_start:
 
     pop rcx
     pop rsi
-    loop .symbol_loop
+    ;loop .symbol_loop
+    dec ecx
+    jnz .symbol_loop
 ; }}}
 
   ; When all is resolved, call main()
diff --git a/runtime/bold_ibh-x86_64.o b/runtime/bold_ibh-x86_64.o
index b72a08bb7c6cbca470093a90230f6a1a6a230ec0..786f97b7a3c79659f48e8ad4550379aeee04d924 100644
GIT binary patch
delta 607
zcmaFB{egGF0>+aQ7n(Cpn0V7(^Z*kB82n&_&@4cTZ?Z0<Hs=8l6A0E#PGmG^Dqx#D
zi_x#%qq~Ts*@~r<x!aGW*^1SN-{o4f6&sMc*J$<i-~aql!)`v8Za)scUY~OUh6fz2
z86Ek37(JU0vOM|>)H2cI-~%R4&4Z5oyHq+E{vRy*_Wz_u^C5v|D~8e~j@`ySjMoe=
z86L>Kq-c0(P0NAOZ<e=8_<eh$B_2KU?Jkpesb2a2|9{VJXAVzhHjmYjasN;9w<ZCF
zUVt1rJ<BG7pB2PnnC!=NO40<(L=Z2*Oh$(A$%f40_4x=1FiRO60E`UPXyT4g@fI}k
zSg3d(ns^0Nd<L5MB&hgmi1_4m=9!!;fnov-j0}e-I!a7dWO)#ummi;zSe(HSpOl}I
z5+9$Llo4NCl2}xtSDu)Y9iLR1oSj-yj6*g#BQY}%Q!X#RBtAW_6m0I~_bj5!C5cIs
z*;ykPnI|7)6`x$e$|K2-<Wxy0n+r$_OrFast|N#fyC2F1B^n_h{sRnq5P&HG0Hghx
ALjV8(

delta 334
zcmeys`+$4G0>-|H3(Xk=Cf>9cZD3*mgBgqvngvL4P1a@9=2QSNfxvchBBMD|0_)^i
zjDBfZED@GWKKw3T9RCl189^{c6o*G|kbogjwDgu^w<n8Zx3Lf7HN#7W2eL0I8Xj8H
za-j5^<*gDP-`;46M~{5F%OqZ}^Xzu!@MLE5SREPnXzmM;Wz(~4BKTP+_c5K43<B|y
zz#A}|ks)SsA+xw-8j=i%sR(vIBSYC_Mwa=U7C@c=10zGnL`R9q2U#Al6<6ja<>yS6
xV2j{mV1{Z1`J8KVEt|L`H&g^fNkZ9NK$>UrS~hWBUZ@C&YMETfCe8s9004UDMZf?6

-- 
2.39.5