From: Tamar Christina Date: Thu, 30 Aug 2018 10:40:07 +0000 (+0000) Subject: Limit movmem copies to TImode on AArch64 X-Git-Tag: basepoints/gcc-10~4462 X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=f7e1d19d58660dee3fb0bd41496610a343f93e45;p=thirdparty%2Fgcc.git Limit movmem copies to TImode on AArch64 On AArch64 we have integer modes larger than TImode, and while we can generate moves for these they're not as efficient. So instead make sure we limit the maximum we can copy to TImode. This means copying a 16 byte struct will issue 1 TImode copy, which will be done using a single STP as we expect but an CImode sized copy won't issue CImode operations. I am also moving the residual code inside the if since smallest_mode_for_int may trap if the mode doesn't exist. And the only time we know the mode to exist for sure is when the condition of the if is true. This also saves repeated calls to the iterator. gcc/ 2018-08-30 Tamar Christina * config/aarch64/aarch64.c (aarch64_expand_movmem): Set TImode max. gcc/testsuite/ 2018-08-30 Tamar Christina * gcc.target/aarch64/large_struct_copy_2.c: New. From-SVN: r263974 --- diff --git a/gcc/ChangeLog b/gcc/ChangeLog index e57a5921559a..f7fb162f9255 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,7 @@ +2018-08-30 Tamar Christina + + * config/aarch64/aarch64.c (aarch64_expand_movmem): Set TImode max. + 2018-08-30 Vlad Lazar PR middle-end/86995 diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index 0d7ca9998466..1de76e075471 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -15938,13 +15938,17 @@ aarch64_expand_movmem (rtx *operands) /* Convert n to bits to make the rest of the code simpler. */ n = n * BITS_PER_UNIT; + /* Maximum amount to copy in one go. The AArch64 back-end has integer modes + larger than TImode, but we should not use them for loads/stores here. */ + const int copy_limit = GET_MODE_BITSIZE (TImode); + while (n > 0) { /* Find the largest mode in which to do the copy in without over reading or writing. */ opt_scalar_int_mode mode_iter; FOR_EACH_MODE_IN_CLASS (mode_iter, MODE_INT) - if (GET_MODE_BITSIZE (mode_iter.require ()) <= n) + if (GET_MODE_BITSIZE (mode_iter.require ()) <= MIN (n, copy_limit)) cur_mode = mode_iter.require (); gcc_assert (cur_mode != BLKmode); @@ -15958,10 +15962,10 @@ aarch64_expand_movmem (rtx *operands) cheaper. i.e. less instructions to do so. For instance doing a 15 byte copy it's more efficient to do two overlapping 8 byte copies than 8 + 6 + 1. */ - next_mode = smallest_mode_for_size (n, MODE_INT); - int n_bits = GET_MODE_BITSIZE (next_mode).to_constant (); - if (n > 0 && n_bits > n && n_bits <= 8 * BITS_PER_UNIT) + if (n > 0 && n <= 8 * BITS_PER_UNIT) { + next_mode = smallest_mode_for_size (n, MODE_INT); + int n_bits = GET_MODE_BITSIZE (next_mode).to_constant (); src = aarch64_move_pointer (src, (n - n_bits) / BITS_PER_UNIT); dst = aarch64_move_pointer (dst, (n - n_bits) / BITS_PER_UNIT); n = n_bits; diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog index 317cb4795754..fd78dcf08920 100644 --- a/gcc/testsuite/ChangeLog +++ b/gcc/testsuite/ChangeLog @@ -1,3 +1,7 @@ +2018-08-30 Tamar Christina + + * gcc.target/aarch64/large_struct_copy_2.c: New. + 2018-08-29 Bernd Edlinger PR middle-end/87053 diff --git a/gcc/testsuite/gcc.target/aarch64/large_struct_copy_2.c b/gcc/testsuite/gcc.target/aarch64/large_struct_copy_2.c new file mode 100644 index 000000000000..565434244e82 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/large_struct_copy_2.c @@ -0,0 +1,26 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +typedef unsigned __attribute__((mode(DI))) uint64_t; + +struct S0 { + uint64_t f1; + uint64_t f2; + uint64_t f3; + uint64_t f4; + uint64_t f5; +} a; +struct S2 { + uint64_t f0; + uint64_t f2; + struct S0 f3; +}; + +void fn1 () { + struct S2 b = {0, 1, 7, 4073709551611, 4, 8, 7}; + a = b.f3; +} + +/* { dg-final { scan-assembler-times {ldp\s+x[0-9]+} 2 } } */ +/* { dg-final { scan-assembler-times {stp\s+x[0-9]+} 2 } } */ +/* { dg-final { scan-assembler-not {ld[1-3]} } } */