This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
combine corrupts insns + dumps with insn cost problems
- From: Rask Ingemann Lambertsen <rask at sygehus dot dk>
- To: gcc at gcc dot gnu dot org
- Date: Wed, 27 Jun 2007 19:32:56 +0200
- Subject: combine corrupts insns + dumps with insn cost problems
I'm seeing this on my 16-bit ix86 port. Something isn't right:
insn_cost 5: 12
insn_cost 6: 8
insn_cost 7: 4
...
rejecting combination of insns 5 and 6
original costs 12 + 8 = 24
replacement cost 28
Now, 12 + 8 = 20, not 24. The cost obviously includes insn 7 also. What's
happening is that combine is trying to combine insns 5, 6 but needs a CCmode
change in insn 7 because we have plain CCmode but SELECT_CC_MODE chooses
CCZ_Cmode for the combined insn 5+6. The original insns:
(insn 5 2 6 2 .../gcc.c-torture/execute/20000801-4.c:12 (set (reg/f:HI 22)
(const:HI (plus:HI (symbol_ref/f:HI ("*.LC0") [flags 0x2] <string_cst 0xb7ea5ee0>)
(const_int 1 [0x1])))) 9 {*movhi} (nil))
(insn 6 5 7 2 .../gcc/testsuite/gcc.c-torture/execute/20000801-4.c:12 (set (reg:CC 13 cc)
(compare:CC (mem/s:QI (reg/f:HI 22) [0 S1 A8])
(const_int 0 [0x0]))) 436 {cmpqi_cc} (expr_list:REG_DEAD (reg/f:HI 22)
(nil)))
(insn 7 6 10 2 .../gcc/testsuite/gcc.c-torture/execute/20000801-4.c:12 (parallel [
(set (reg:HI 24)
(eq:HI (reg:CC 13 cc)
(const_int 0 [0x0])))
(clobber (scratch:QI))
(clobber (reg:CC 13 cc))
]) 454 {*seqhi_cc} (expr_list:REG_DEAD (reg:CC 13 cc)
(expr_list:REG_UNUSED (reg:CC 13 cc)
(nil))))
The replacements:
(set (reg:CCZ_C 13 cc)
(compare:CCZ_C (mem/s:QI (const:HI (plus:HI (symbol_ref/f:HI ("*.LC0") [flags 0x2] <string_cst 0xb7ea5ee0>)
(const_int 1 [0x1]))) [0 S1 A8])
(const_int 0 [0x0])))
(set (reg:HI 24)
(eq:HI (reg:CCZ_C 13 cc)
(const_int 0 [0x0])))
As noted, combine rejects the replacement. But the structure of insn 7
has now been corrupted:
(insn 5 2 6 2 .../gcc.c-torture/execute/20000801-4.c:12 (set (reg/f:HI 22)
(const:HI (plus:HI (symbol_ref/f:HI ("*.LC0") [flags 0x2] <string_cst 0xb7ea5ee0>)
(const_int 1 [0x1])))) 9 {*movhi} (nil))
(insn 6 5 7 2 .../gcc.c-torture/execute/20000801-4.c:12 (set (reg:CC 13 cc)
(compare:CC (mem/s:QI (reg/f:HI 22) [0 S1 A8])
(const_int 0 [0x0]))) 436 {cmpqi_cc} (expr_list:REG_DEAD (reg/f:HI 22)
(nil)))
(insn 7 6 10 2 .../gcc.c-torture/execute/20000801-4.c:12 (parallel [
(set (reg:HI 24)
(eq:HI (reg:CC 13 cc)
(const_int 0 [0x0])))
(clobber (reg:CC 13 cc))
]) 454 {*seqhi_cc} (expr_list:REG_DEAD (reg:CC 13 cc)
(expr_list:REG_UNUSED (reg:CC 13 cc)
(nil))))
The clobber of the scratch register has disappeared!
A possible clue as to what sets up the failure is that the second
replacement insn (to replace insn 7)
(set (reg:HI 24)
(eq:HI (reg:CCZ_C 13 cc)
(const_int 0 [0x0])))
needs to have a clobber added. It really looks like this:
(set (reg:HI 24)
(eq:HI (reg:CCZ_C 13 cc)
(const_int 0 [0x0])))
(clobber (reg:CC 13 cc))
Combine knows how to add clobbers to make insns recognizable. I'm guessing
it accidentally clobbers the original insn in doing so. Where would I look?
--
Rask Ingemann Lambertsen