------- Comment #20 from pthaugen at gcc dot gnu dot org 2009-07-08 21:53
-------
Created an attachment (id=18165)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18165&action=view)
Reduced testcase
The attatched testcase exhibits the problem with the load-hit-store. It's
resulting from choosing a bad register class (GENERAL_REGS) for a pseudo that
should get assigned to FLOAT_REGS. Since there is no FPR -> GPR move for
-mcpu=power6 the copy must go through memory. I compiled the testcase with
-m64 -O3 -mcpu=power6 using trunk revision 149376. The pseudo in question is
361.
Following are the 3 insns referencing reg 361 in the sched1 dump (before ira):
(insn 51 238 241 8 thin6d_reduced.f:178 (set (reg:DF 361 [ prephitmp.35 ])
(reg:DF 358 [ prephitmp.35 ])) 351 {*movdf_hardfloat64} (nil))
...
(insn 47 46 231 9 thin6d_reduced.f:178 (set (reg:DF 361 [ prephitmp.35 ])
(reg:DF 179 [ prephitmp.35 ])) 351 {*movdf_hardfloat64} (nil))
...
(insn 196 194 198 11 thin6d_reduced.f:169 (set (mem/c/i:DF (plus:DI (reg/f:DI
477)
(const_int 56 [0x38])) [2 crkve+0 S8 A64])
(reg:DF 361 [ prephitmp.35 ])) 351 {*movdf_hardfloat64}
(expr_list:REG_DEAD (reg:DF 361 [ prephitmp.35 ])
(nil)))
And from the ira dump:
Pass1 cost computation:
a71 (r361,l1) best GENERAL_REGS, cover GENERAL_REGS
a3 (r361,l0) best GENERAL_REGS, cover GENERAL_REGS
a3(r361,l0) costs: BASE_REGS:0,0 GENERAL_REGS:0,0 FLOAT_REGS:0,0
LINK_REGS:156,1836 CTR_REGS:156,1836 SPECIAL_REGS:156,1836 MEM:156
a71(r361,l1) costs: BASE_REGS:0,0 GENERAL_REGS:0,0 FLOAT_REGS:0,0
LINK_REGS:1680,1680 CTR_REGS:1680,1680 SPECIAL_REGS:1680,1680 MEM:1120
Pass 2 cost computation:
r361: preferred GENERAL_REGS, alternative NO_REGS
a3(r361,l0) costs: BASE_REGS:0,2240 GENERAL_REGS:0,2240 FLOAT_REGS:312,2552
LINK_REGS:234,4154 CTR_REGS:234,4154 SPECIAL_REGS:234,4154 MEM:156
a71(r361,l1) costs: BASE_REGS:2240,2240 GENERAL_REGS:2240,2240
FLOAT_REGS:2240,2240 LINK_REGS:3920,3920 CTR_REGS:3920,3920
SPECIAL_REGS:3920,3920 MEM:3360
Not sure what's causing the FLOAT cost to be higher than the GENERAL cost yet.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39976
