[armel, avmv6] wrong assembler with -O[13s], not -O2

[Matthias Klose]

Binary package hint: gcc-4.4

maverick:

$ gcc-4.4 -c -march=armv6 -Os inlining.i
inlining.c: In function 'rewriteLock':
inlining.c:850: error: insn does not satisfy its constraints:
(insn 62 20 21 3 inlining.c:846 (set (reg:SI 3 r3)
        (reg/f:SI 13 sp)) 174 {*thumb1_movsi_insn_osize} (nil))
inlining.c:850: internal compiler error: in reload_cse_simplify_operands, at postreload.c:397
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-4.4/README.Bugs> for instructions.

natty, with both gcc-4.4 and gcc-4.5:
$ gcc-4.5 -c -march=armv6 -Os inlining.i
/tmp/ccl1qfKd.s: Assembler messages:
/tmp/ccl1qfKd.s:794: Error: invalid immediate: 983040 is out of range
/tmp/ccl1qfKd.s:794: Error: value of 983040 too large for field of 2 bytes at 1310
/tmp/ccl1qfKd.s:1088: Error: invalid immediate: 983040 is out of range
/tmp/ccl1qfKd.s:1088: Error: value of 983040 too large for field of 2 bytes at 1800

works with -O2, fails with -O1, -Os, -O3.

[Dr. David Alan Gilbert (davidgil-uk)]

Still fails on 4.5-2011.03-0 release when built with -mthumb

/tmp/cc491lp9.s: Assembler messages:
/tmp/cc491lp9.s:793: Error: invalid immediate: 983040 is out of range
/tmp/cc491lp9.s:793: Error: value of 00000000000f0000 too large for field of 2 bytes at 000000000000051e
/tmp/cc491lp9.s:1087: Error: invalid immediate: 983040 is out of range
/tmp/cc491lp9.s:1087: Error: value of 00000000000f0000 too large for field of 2 bytes at 0000000000000708

those seem to come from a line in the input newCodeBlock() and newDuplicateBlock function with inline asm, of

mov r7, #0xf0000

For Thumb1 that just doesn't fit; compiling with -march=armv6t2 works.

Dave

[Michael Hope (michaelh1)]

From what you're saying Dave, this is an inline assembler bug which is invalid for Thumb-1.

Matthias, if this needs to compile under Thumb-1 then could you tag it as arm-porting-queue and notify Steve L?

[Dr. David Alan Gilbert (davidgil-uk)]

Correct, my reading is that the inline assembler is invalid on Thumb-1.

Dave

[Mans Rullgard (mansr)]

The problem is indeed inline asm which is invalid for Thumb-1.

This code comes from JamVM, and the inline asm performs a syscall to invalidate the I-cache after generating some code. It would be better off using the gcc __builtin__clear_cache() function for this, or least changing the code to let gcc generate the constants.