------- Comment #7 from hjl dot tools at gmail dot com 2008-11-28 22:28 ------- (In reply to comment #6) > I think, H.J., that is one more latent bug (i already saw several of them) in > reload inheritance optimization triggered by IRA which allocates dx for p69 > and > p87 in subsequent insns > > 47:p65<-p69 > 151:p87<-mem[...]. >
With my patch, I got (insn 47 46 151 3 /tmp/foo.f90:53 (set (mem/c:SI (plus:SI (reg/f:SI 6 bp) (const_int -48 [0xffffffffffffffd0])) [0 %sfp+-24 S4 A32]) (reg:SI 1 dx [orig:69 S.10 ] [69])) 47 {*movsi_1} (nil)) (insn 151 47 48 3 /tmp/foo.f90:6 (set (reg:SI 1 dx [87]) (reg:SI 2 cx [80])) 47 {*movsi_1} (expr_list:REG_EQUIV (mem/u/c:SI (const:SI (unspec:SI [ (symbol_ref:SI ("__threadprivate2_MOD_foo") [flags 0x60] <var_decl 0x7fad44900780 foo>) ] 8)) [0 S4 A8]) (nil))) (insn 48 151 167 3 /tmp/foo.f90:6 (set (reg:SI 1 dx) (mem/s/j:SI (plus:SI (plus:SI (unspec:SI [ (const_int 0 [0x0]) ] 18) (reg:SI 1 dx [87])) (const_int 24 [0x18])) [0 <variable>.stride+0 S4 A32])) 47 {*movsi_1} (nil)) (insn 167 48 168 3 /tmp/foo.f90:6 (set (mem/c:SI (plus:SI (reg/f:SI 6 bp) (const_int -56 [0xffffffffffffffc8])) [0 %sfp+-32 S4 A32]) (reg:SI 1 dx)) 47 {*movsi_1} (nil)) (insn 168 167 49 3 /tmp/foo.f90:6 (set (reg:SI 1 dx [orig:62 D.2019 ] [62]) (mem/c:SI (plus:SI (reg/f:SI 6 bp) (const_int -56 [0xffffffffffffffc8])) [0 %sfp+-32 S4 A32])) 47 {*movsi_1} (nil)) (insn 49 168 50 3 /tmp/foo.f90:6 (parallel [ (set (reg:SI 1 dx [orig:62 D.2019 ] [62]) (mult:SI (reg:SI 1 dx [orig:62 D.2019 ] [62]) (mem/c:SI (plus:SI (reg/f:SI 6 bp) (const_int -48 [0xffffffffffffffd0])) [0 %sfp+-24 S4 A32]))) (clobber (reg:CC 17 flags)) ]) 333 {*mulsi3_1} (nil)) DX (r69) is dead at insn 47. DX (r87) at insn 151 is a different register. It looks OK to me. > I am not sure that your patch is right but you can commit it for a review to > get some comments. > I will submit it for comments after I finish tests on Linux/ia32, Linux/ia64 and Linux/Intel64. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38272