Index: gcc/ChangeLog =================================================================== --- gcc/ChangeLog (revision 240100) +++ gcc/ChangeLog (working copy) @@ -1,3 +1,14 @@ +2016-09-12 Andrew Pinski + + * config/aarch64/aarch64-tuning-flags.def (SLOW_UNALIGNED_LDPW): + New tuning option. + * config/aarch64/aarch64.c (thunderx_tunings): Enable + AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW. + (aarch64_operands_ok_for_ldpstp): Return false if + AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW and the mode + was SImode and the alignment is less than 8 byte. + (aarch64_operands_adjust_ok_for_ldpstp): Likewise. + 2016-09-12 Marek Polacek * doc/extend.texi: Use lowercase "boolean". Index: gcc/config/aarch64/aarch64-tuning-flags.def =================================================================== --- gcc/config/aarch64/aarch64-tuning-flags.def (revision 240100) +++ gcc/config/aarch64/aarch64-tuning-flags.def (working copy) @@ -29,3 +29,8 @@ AARCH64_TUNE_ to give an enum name. */ AARCH64_EXTRA_TUNING_OPTION ("rename_fma_regs", RENAME_FMA_REGS) + +/* Don't create non-8 byte aligned load/store pair. That is if the +two load/stores are not at least 8 byte aligned don't create load/store +pairs. */ +AARCH64_EXTRA_TUNING_OPTION ("slow_unaligned_ldpw", SLOW_UNALIGNED_LDPW) Index: gcc/config/aarch64/aarch64.c =================================================================== --- gcc/config/aarch64/aarch64.c (revision 240100) +++ gcc/config/aarch64/aarch64.c (working copy) @@ -712,7 +712,7 @@ 0, /* max_case_values. */ 0, /* cache_line_size. */ tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ - (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ + (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) /* tune_flags. */ }; static const struct tune_params xgene1_tunings = @@ -13629,6 +13629,15 @@ if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2)) return false; + /* If we have SImode and slow unaligned ldp, + check the alignment to be at least 8 byte. */ + if (mode == SImode + && (aarch64_tune_params.extra_tuning_flags + & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) + && !optimize_size + && MEM_ALIGN (mem_1) < 8 * BITS_PER_UNIT) + return false; + /* Check if the addresses are in the form of [base+offset]. */ extract_base_offset_in_addr (mem_1, &base_1, &offset_1); if (base_1 == NULL_RTX || offset_1 == NULL_RTX) @@ -13788,6 +13797,15 @@ return false; } + /* If we have SImode and slow unaligned ldp, + check the alignment to be at least 8 byte. */ + if (mode == SImode + && (aarch64_tune_params.extra_tuning_flags + & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) + && !optimize_size + && MEM_ALIGN (mem_1) < 8 * BITS_PER_UNIT) + return false; + if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) rclass_1 = FP_REGS; else Index: gcc/testsuite/ChangeLog =================================================================== --- gcc/testsuite/ChangeLog (revision 240100) +++ gcc/testsuite/ChangeLog (working copy) @@ -1,3 +1,8 @@ +2016-09-12 Andrew Pinski + + * gcc.target/aarch64/thunderxloadpair.c: New testcase. + * gcc.target/aarch64/thunderxnoloadpair.c: New testcase. + 2016-09-12 Uros Bizjak * gcc.dg/compat/scalar-by-value-4_x.c: Also test passing of Index: gcc/testsuite/gcc.target/aarch64/thunderxloadpair.c =================================================================== --- gcc/testsuite/gcc.target/aarch64/thunderxloadpair.c (nonexistent) +++ gcc/testsuite/gcc.target/aarch64/thunderxloadpair.c (working copy) @@ -0,0 +1,20 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -mcpu=thunderx" } */ + +struct ldp +{ + long long c; + int a, b; +}; + + +int f(struct ldp *a) +{ + return a->a + a->b; +} + + +/* We know the alignement of a->a to be 8 byte aligned so it is profitable + to do ldp. */ +/* { dg-final { scan-assembler-times "ldp\tw\[0-9\]+, w\[0-9\]" 1 } } */ + Index: gcc/testsuite/gcc.target/aarch64/thunderxnoloadpair.c =================================================================== --- gcc/testsuite/gcc.target/aarch64/thunderxnoloadpair.c (nonexistent) +++ gcc/testsuite/gcc.target/aarch64/thunderxnoloadpair.c (working copy) @@ -0,0 +1,17 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -mcpu=thunderx" } */ + +struct noldp +{ + int a, b; +}; + + +int f(struct noldp *a) +{ + return a->a + a->b; +} + +/* We know the alignement of a->a to be 4 byte aligned so it is not profitable + to do ldp. */ +/* { dg-final { scan-assembler-not "ldp\tw\[0-9\]+, w\[0-9\]" } } */