Unnecessary jump lengthening for 127 bytes
Brought to you by:
japheth
Menu
▾
▴
In v2.02pre dated 12/28/09 short jumps with a displacement of 127 bytes are lengthened unnecessarily to the opposite condition and a near jump. Did not occur in the previous build.
There is a bug concerning backpatching in v1.96-v2.01. I started to fix it before xmas, but then took a vacation,. so it's somewhat "half-fixed" currently.
However, "half-fixed" just means that JWasm may need more passes than necessary. It doesn't mean that "sub-optimal" code may be generated. So please, if possible, upload your test-case which reveals this behavior!
japheth
Retested with build dated 12/29/09 and it still happens. This happens with displacements 7Dh-7Fh. Generates opposite condition short jump and near jump for 8086/286, conditional rel16 jump (0F 8x) for 386+
This is a regression - the issue does not occur in v2.02pre build from 12/22/09
> This is a regression -
This may be true, but is more or less irrelevant because the backpatching issue has to be fixed, and that's why I have to know under what circumstances this "regression" occurs. Just a simple forward reference does NOT trigger it - a valid test case is needed!
japheth
code segment para 'code'
assume cs:code,ds:code
org 100h
entry: jmp skip1
test ax,ax
jnz skip2
db 126 dup(90h)
skip1: ret
skip2: ret
code ends
end entry
The jmp skip1 causes the issue, removing this jmp makes the problem go away. Note JWasm v2.02 from 12/22 does 3 passes on this code while the 12/29 build only does 2 passes.
Listing (.8086)
code segment para 'code'
assume cs:code,ds:code
org 100h
00000100 E98500 entry: jmp skip1
00000103 85C0 test ax,ax
00000105 7403E97F00 jnz skip2
0000010A 9090909090909090 db 126 dup(90h)
00000188 C3 skip1: ret
00000189 C3 skip2: ret
0000018A code ends
end entry
Listing (.386)
code segment para 'code'
assume cs:code,ds:code
.386
org 100h
00000100 E98400 entry: jmp skip1
00000103 85C0 test ax,ax
00000105 0F857F00 jnz skip2
00000109 9090909090909090 db 126 dup(90h)
00000187 C3 skip1: ret
00000188 C3 skip2: ret
00000189 code ends
end entry
The change which causes the lenghening has been reverted because a full fix would significantly delay JWasm v2.02. The fix is now "scheduled" for v2.03.
japheth