| Message ID | CAGWvny=AKTV9MYdyga7F-nvSSVqZ_QCCJgxk5NAQLnu9WWiG=w@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
Index: config/rs6000/rs6000.c =================================================================== --- config/rs6000/rs6000.c (revision 218496) +++ config/rs6000/rs6000.c (working copy) @@ -17379,12 +17379,7 @@ rs6000_secondary_reload_inner (rtx reg, rtx mem, r case SYMBOL_REF: case CONST: case LABEL_REF: - if (TARGET_TOC) - emit_insn (gen_rtx_SET (VOIDmode, scratch, - create_TOC_reference (addr, scratch))); - else - rs6000_emit_move (scratch, addr, Pmode); - + rs6000_emit_move (scratch, addr, Pmode); new_addr = scratch; break;
The patch removes code from rs6000_secondary_reload_inner that explicitly called create_TOC_reference on SYMBOL_REF, CONST, and LABEL_REF involved in a secondary reload. find_reload was choosing an equivalence of a symbol_ref that referenced the raw data and secondary reload code was creating a direct reference to the data instead of a TOC reference, treating the symbol offset as a TOC offset, which was invalid. This generated code that performed a load from a wild address and garbage data. As Uli pointed out, rs6000_emit_move already knows how and when to create TOC references. Mike Meissner tested the patch with PPC32 and PPC64 Linux. I have bootstrapped and tested on powerpc-ibm-aix7.1.0.0. This fixes all of failures in the GCC VMX testsuite on AIX. Thanks, David * config/rs6000/rs6000.c (rs6000_secondary_reload_inner) [SYMBOL_REF]: Do not explicitly call create_TOC_reference for TARGET_TOC. Always use rs6000_emit_move.