A slight oversight. and<mode>3_imm is predicated by rs6000_gen_cell_microcode so we had better not generate it otherwise; and in that case, we need to force the constant to a register.
Bootstrap and testsuite run are fine (of course, since this isn't tested at all); comparing testsuite runs between -mcpu=cell and default shows a few thousand fails (mainly ICEs in dfp), and no differences between with and without this patch. But this PR was a failure in Linux kernel build; I checked it did fail to build without this patch, and works with (I built cell_defconfig). Okay for trunk? Segher 2014-11-27 Segher Boessenkool <seg...@kernel.crashing.org> gcc/ PR target/64093 * config/rs6000/rs6000.md (and<mode>3): Don't generate and<mode>3_imm unless rs6000_gen_cell_microcode is true. --- gcc/config/rs6000/rs6000.md | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md index c8c3a68..f3b5aae 100644 --- a/gcc/config/rs6000/rs6000.md +++ b/gcc/config/rs6000/rs6000.md @@ -2646,8 +2646,13 @@ (define_expand "and<mode>3" if (logical_const_operand (operands[2], <MODE>mode) && !any_mask_operand (operands[2], <MODE>mode)) { - emit_insn (gen_and<mode>3_imm (operands[0], operands[1], operands[2])); - DONE; + if (rs6000_gen_cell_microcode) + { + emit_insn (gen_and<mode>3_imm (operands[0], operands[1], operands[2])); + DONE; + } + else + operands[2] = force_reg (<MODE>mode, operands[2]); } if ((<MODE>mode == DImode && !and64_2_operand (operands[2], <MODE>mode)) -- 1.8.1.4