[PATCH/AARCH64 v2] Enable software prefetching (-fprefetch-loop-arrays) for ThunderX 88xxx

[PATCH/AARCH64 v2] Enable software prefetching (-fprefetch-loop-arrays) for ThunderX 88xxx

[Andrew Pinski]

[-- Attachment #1: Type: text/plain, Size: 714 bytes --]

Here is the updated patch based on the new infrastructure which is now included.

OK?  Bootstrapped and tested on aarch64-linux-gnu with no regressions
and tested again on SPEC CPU 2006 on THunderX T88 with the speed up
mentioned before.

Thanks,
Andrew Pinski

ChangeLog:
* config/aarch64/aarch64-cores.def (thunderxt88p1): Use thunderxt88 tunings.
(thunderxt88): Likewise.

* config/aarch64/aarch64.c (thunderxt88_prefetch_tune): New variable.
(thunderx_prefetch_tune): New variable.
(thunderx2t99_prefetch_tune): Update for the correct values.
(thunderxt88_tunings): New variable.
(thunderx_tunings): Use thunderx_prefetch_tune instead of generic_prefetch_tune.
(thunderx2t99_tunings): Use AUTOPREFETCHER_WEAK.

[-- Attachment #2: prefetch.diff.txt --]
[-- Type: text/plain, Size: 4143 bytes --]

Index: gcc/config/aarch64/aarch64-cores.def
===================================================================
--- gcc/config/aarch64/aarch64-cores.def	(revision 249422)
+++ gcc/config/aarch64/aarch64-cores.def	(working copy)
@@ -56,8 +56,8 @@ AARCH64_CORE("cortex-a73",  cortexa73, c
 AARCH64_CORE("thunderx",      thunderx,      thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO, thunderx,  0x43, 0x0a0, -1)
 /* Do not swap around "thunderxt88p1" and "thunderxt88",
    this order is required to handle variant correctly. */
-AARCH64_CORE("thunderxt88p1", thunderxt88p1, thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO,	thunderx,  0x43, 0x0a1, 0)
-AARCH64_CORE("thunderxt88",   thunderxt88,   thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO, thunderx,  0x43, 0x0a1, -1)
+AARCH64_CORE("thunderxt88p1", thunderxt88p1, thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO,	thunderxt88,  0x43, 0x0a1, 0)
+AARCH64_CORE("thunderxt88",   thunderxt88,   thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO, thunderxt88,  0x43, 0x0a1, -1)
 AARCH64_CORE("thunderxt81",   thunderxt81,   thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO, thunderx,  0x43, 0x0a2, -1)
 AARCH64_CORE("thunderxt83",   thunderxt83,   thunderx,  8A,  AARCH64_FL_FOR_ARCH8 | AARCH64_FL_CRC | AARCH64_FL_CRYPTO, thunderx,  0x43, 0x0a3, -1)
 
Index: gcc/config/aarch64/aarch64.c
===================================================================
--- gcc/config/aarch64/aarch64.c	(revision 249422)
+++ gcc/config/aarch64/aarch64.c	(working copy)
@@ -554,12 +554,30 @@ static const cpu_prefetch_tune qdf24xx_p
   3			/* default_opt_level  */
 };
 
+static const cpu_prefetch_tune thunderxt88_prefetch_tune =
+{
+  8,			/* num_slots  */
+  32,			/* l1_cache_size  */
+  128,			/* l1_cache_line_size  */
+  16*1024,		/* l2_cache_size  */
+  3			/* default_opt_level  */
+};
+
+static const cpu_prefetch_tune thunderx_prefetch_tune =
+{
+  8,			/* num_slots  */
+  32,			/* l1_cache_size  */
+  128,			/* l1_cache_line_size  */
+  -1,			/* l2_cache_size  */
+  -1			/* default_opt_level  */
+};
+
 static const cpu_prefetch_tune thunderx2t99_prefetch_tune =
 {
-  0,			/* num_slots  */
-  -1,			/* l1_cache_size  */
+  8,			/* num_slots  */
+  32,			/* l1_cache_size  */
   64,			/* l1_cache_line_size  */
-  -1,			/* l2_cache_size  */
+  256,			/* l2_cache_size  */
   -1			/* default_opt_level  */
 };
 
@@ -745,6 +763,31 @@ static const struct tune_params exynosm1
   &exynosm1_prefetch_tune
 };
 
+static const struct tune_params thunderxt88_tunings =
+{
+  &thunderx_extra_costs,
+  &generic_addrcost_table,
+  &thunderx_regmove_cost,
+  &thunderx_vector_cost,
+  &generic_branch_cost,
+  &generic_approx_modes,
+  6, /* memmov_cost  */
+  2, /* issue_rate  */
+  AARCH64_FUSE_CMP_BRANCH, /* fusible_ops  */
+  8,	/* function_align.  */
+  8,	/* jump_align.  */
+  8,	/* loop_align.  */
+  2,	/* int_reassoc_width.  */
+  4,	/* fp_reassoc_width.  */
+  1,	/* vec_reassoc_width.  */
+  2,	/* min_div_recip_mul_sf.  */
+  2,	/* min_div_recip_mul_df.  */
+  0,	/* max_case_values.  */
+  tune_params::AUTOPREFETCHER_OFF,	/* autoprefetcher_model.  */
+  (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW),	/* tune_flags.  */
+  &thunderxt88_prefetch_tune
+};
+
 static const struct tune_params thunderx_tunings =
 {
   &thunderx_extra_costs,
@@ -767,7 +810,7 @@ static const struct tune_params thunderx
   0,	/* max_case_values.  */
   tune_params::AUTOPREFETCHER_OFF,	/* autoprefetcher_model.  */
   (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW),	/* tune_flags.  */
-  &generic_prefetch_tune
+  &thunderx_prefetch_tune
 };
 
 static const struct tune_params xgene1_tunings =
@@ -841,7 +884,7 @@ static const struct tune_params thunderx
   2,	/* min_div_recip_mul_sf.  */
   2,	/* min_div_recip_mul_df.  */
   0,	/* max_case_values.  */
-  tune_params::AUTOPREFETCHER_OFF,	/* autoprefetcher_model.  */
+  tune_params::AUTOPREFETCHER_WEAK,	/* autoprefetcher_model.  */
   (AARCH64_EXTRA_TUNE_NONE),	/* tune_flags.  */
   &thunderx2t99_prefetch_tune
 };

Re: [PATCH/AARCH64 v2] Enable software prefetching (-fprefetch-loop-arrays) for ThunderX 88xxx

[James Greenhalgh]

On Tue, Jun 20, 2017 at 11:13:24AM -0700, Andrew Pinski wrote:
> Here is the updated patch based on the new infrastructure which is now included.
> 
> OK?  Bootstrapped and tested on aarch64-linux-gnu with no regressions
> and tested again on SPEC CPU 2006 on THunderX T88 with the speed up
> mentioned before.

OK.

Thanks,
James

> 
> Thanks,
> Andrew Pinski
> 
> ChangeLog:
> * config/aarch64/aarch64-cores.def (thunderxt88p1): Use thunderxt88 tunings.
> (thunderxt88): Likewise.
> 
> * config/aarch64/aarch64.c (thunderxt88_prefetch_tune): New variable.
> (thunderx_prefetch_tune): New variable.
> (thunderx2t99_prefetch_tune): Update for the correct values.
> (thunderxt88_tunings): New variable.
> (thunderx_tunings): Use thunderx_prefetch_tune instead of generic_prefetch_tune.
> (thunderx2t99_tunings): Use AUTOPREFETCHER_WEAK.