Update in V2:
1. Also disable vect unroll for Znver1.
2. double the count for {AVX512,AVX256}_SPLIT_REGS for corresponding vector
length.
Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
Since it regressed SPEC performance(Refer to PR121994), I guess
it's related to register pressure and can be tuned by adjusting
reduc_lat_mult_thr. I don't have Zen2 machine, so for simplity, I'll
just disable unroll in vectorizer for Zen2.
Also adjust count number for {AVX256,AVX512}_SPLIT_REGS.
gcc/ChangeLog:
PR target/121994
* config/i386/x86-tune-costs.h (znver2_cost): Set
vect_unroll_limit to 1.
(znver1_cost): Ditto.
* config/i386/i386.cc (ix86_vector_costs::add_stmt_cost):
Adjust count number for {AVX256,AVX512}_SPLIT_REGS.
---
gcc/config/i386/i386.cc | 18 +++++++++++++-----
gcc/config/i386/x86-tune-costs.h | 4 ++--
2 files changed, 15 insertions(+), 7 deletions(-)
diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index 5ef7c315091..6eb26cd7b82 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -26144,6 +26144,14 @@ ix86_vector_costs::add_stmt_cost (int count,
vect_cost_for_stmt kind,
/* Record number of load/store/gather/scatter in vectorized body. */
if (where == vect_body && !m_costing_for_scalar)
{
+ int scale = 1;
+ if (vectype
+ && ((GET_MODE_SIZE (TYPE_MODE (vectype)) == 64
+ && TARGET_AVX512_SPLIT_REGS)
+ || (GET_MODE_SIZE (TYPE_MODE (vectype)) == 32
+ && TARGET_AVX256_SPLIT_REGS)))
+ scale = 2;
+
switch (kind)
{
/* Emulated gather/scatter or any scalarization. */
@@ -26166,7 +26174,7 @@ ix86_vector_costs::add_stmt_cost (int count,
vect_cost_for_stmt kind,
/* Handle __builtin_fma. */
if (gimple_call_combined_fn (stmt_info->stmt) == CFN_FMA)
{
- m_num_reduc[X86_REDUC_FMA] += count;
+ m_num_reduc[X86_REDUC_FMA] += count * scale;
break;
}
@@ -26203,12 +26211,12 @@ ix86_vector_costs::add_stmt_cost (int count,
vect_cost_for_stmt kind,
&& (def = SSA_NAME_DEF_STMT (rhs1), true)
&& is_gimple_assign (def)
&& gimple_assign_rhs_code (def) == MULT_EXPR)
- m_num_reduc[X86_REDUC_FMA] += count;
+ m_num_reduc[X86_REDUC_FMA] += count * scale;
else if (TREE_CODE (rhs2) == SSA_NAME
&& (def = SSA_NAME_DEF_STMT (rhs2), true)
&& is_gimple_assign (def)
&& gimple_assign_rhs_code (def) == MULT_EXPR)
- m_num_reduc[X86_REDUC_FMA] += count;
+ m_num_reduc[X86_REDUC_FMA] += count * scale;
break;
/* Vectorizer lane_reducing_op_p supports DOT_PROX_EXPR,
@@ -26237,7 +26245,7 @@ ix86_vector_costs::add_stmt_cost (int count,
vect_cost_for_stmt kind,
? TARGET_AVX10_2
: (TARGET_AVXVNNIINT8 || TARGET_AVX10_2));
}
- m_num_reduc[X86_REDUC_DOT_PROD] += count;
+ m_num_reduc[X86_REDUC_DOT_PROD] += count * scale;
/* Dislike to do unroll and partial sum for
emulated DOT_PROD_EXPR. */
@@ -26246,7 +26254,7 @@ ix86_vector_costs::add_stmt_cost (int count,
vect_cost_for_stmt kind,
break;
case SAD_EXPR:
- m_num_reduc[X86_REDUC_SAD] += count;
+ m_num_reduc[X86_REDUC_SAD] += count * scale;
break;
default:
diff --git a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h
index 1649ea2fe3e..c7a0f6805ca 100644
--- a/gcc/config/i386/x86-tune-costs.h
+++ b/gcc/config/i386/x86-tune-costs.h
@@ -1744,7 +1744,7 @@ struct processor_costs znver1_cost = {
FMA/DOT_PROD_EXPR/SAD_EXPR,
it's used to determine unroll
factor in the vectorizer. */
- 4, /* Limit how much the autovectorizer
+ 1, /* Limit how much the autovectorizer
may unroll a loop. */
znver1_memcpy,
znver1_memset,
@@ -1918,7 +1918,7 @@ struct processor_costs znver2_cost = {
FMA/DOT_PROD_EXPR/SAD_EXPR,
it's used to determine unroll
factor in the vectorizer. */
- 4, /* Limit how much the autovectorizer
+ 1, /* Limit how much the autovectorizer
may unroll a loop. */
znver2_memcpy,
znver2_memset,
--
2.34.1
