------- Comment #3 from jakub at gcc dot gnu dot org 2009-03-13 17:54 ------- I can reproduce even with r144843. Comparing -da -fdump-unnumbered -fdump-noaddr dumps between redhat/gcc-4_4-branch and power7 patched redhat/gcc-4_4-branch doesn't show any significant differences still in sched1 (just different line numbers in *.md files and different D.NNNN numbers of artificial variables), but in the *.ira dump I see very high costs of link and control register classes: - a0(r308,l0) costs: BASE_REGS:0,0 GENERAL_REGS:0,0 FLOAT_REGS:0,0 LINK_REGS:3030,3030 CTR_REGS:3030,3030 SPECIAL_REGS:3030,3030 MEM:202 0 + a0(r308,l0) costs: BASE_REGS:0,0 GENERAL_REGS:0,0 FLOAT_REGS:0,0 LINK_REGS:1002424,1002424 CTR_REGS:1002424,1002424 SPECIAL_REGS:10024 24,1002424 MEM:2020 ... - a36(r209,l0) costs: BASE_REGS:1010,1010 GENERAL_REGS:1010,1010 FLOAT_REGS:202,202 LINK_REGS:3232,3232 CTR_REGS:3232,3232 SPECIAL_REGS: 3232,3232 MEM:2222 + a36(r209,l0) costs: BASE_REGS:1010,1010 GENERAL_REGS:1010,1010 FLOAT_REGS:202,202 MEM:2222 etc. This is the default arch compilation, so ideally power7 patches should make no difference on the generated code.
-- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39457