http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45813
--- Comment #9 from Mikael Pettersson <mikpe at it dot uu.se> 2010-09-29
19:41:08 UTC ---
I can reproduce it with crosses to arm-elf and arm-linux-gnueabi. The
combination of -mthumb (1 not 2) and -fno-omit-frame-pointer is the trigger.
GCC 4.3.5 works, 4.4 and 4.5 generate broken code, 4.6 works.
gcc-4.3.5 -march=armv5te -mthumb -fno-omit-frame-pointer -Os generates:
ReadLE16U:
push {r7, lr}
ldrb r3, [r0]
sub sp, sp, #8
add r7, sp, #0
add r2, r7, #6
strb r3, [r2]
ldrb r3, [r0, #1]
mov sp, r7
add sp, sp, #8
strb r3, [r2, #1]
ldrh r0, [r2]
@ sp needed for prologue
pop {r7, pc}
The second byte is correctly stored before the short is read and returned.
gcc-4.4-20100928 and gcc-4.5-20100923 generate:
ReadLE16U:
push {r7, lr}
ldrb r2, [r0]
sub sp, sp, #8
add r7, sp, #0
add r3, r7, #6
mov sp, r7
strb r2, [r3]
add sp, sp, #8
ldrb r2, [r0, #1]
@ sp needed for prologue
ldrh r0, [r3]
pop {r7, pc}
The second byte is read but not written before the short is read.
gcc-4.6-20100925 generates:
ReadLE16U:
push {r0, r1, r7, lr}
ldrb r3, [r0]
add r7, sp, #0
mov sp, r7
strb r3, [r7, #6]
ldrb r3, [r0, #1]
@ sp needed for prologue
strb r3, [r7, #7]
ldrh r0, [r7, #6]
pop {r1, r2, r7, pc}
This is correct.
