[gcc r15-47] RISC-V: Refine the condition for add additional vars in RVV cost model
https://gcc.gnu.org/g:ca2f531cc5db4f1020d4329976610356033e0246
commit r15-47-gca2f531cc5db4f1020d4329976610356033e0246
Author: demin.han
Date: Tue Mar 26 16:52:12 2024 +0800
RISC-V: Refine the condition for add additional vars in RVV cost model
The adjacent_dr_p is sufficient and unnecessary condition for contiguous
access.
So unnecessary live-ranges are added and result in smaller LMUL.
This patch uses MEMORY_ACCESS_TYPE as condition and constrains segment
load/store.
Tested on RV64 and no regression.
PR target/114506
gcc/ChangeLog:
* config/riscv/riscv-vector-costs.cc
(non_contiguous_memory_access_p): Rename
(need_additional_vector_vars_p): Rename and refine condition
gcc/testsuite/ChangeLog:
* gcc.dg/vect/costmodel/riscv/rvv/pr114506.c: New test.
Signed-off-by: demin.han
Diff:
---
gcc/config/riscv/riscv-vector-costs.cc | 23 ++
.../gcc.dg/vect/costmodel/riscv/rvv/pr114506.c | 23 ++
2 files changed, 38 insertions(+), 8 deletions(-)
diff --git a/gcc/config/riscv/riscv-vector-costs.cc
b/gcc/config/riscv/riscv-vector-costs.cc
index d27bb68a7b9..4582b0db425 100644
--- a/gcc/config/riscv/riscv-vector-costs.cc
+++ b/gcc/config/riscv/riscv-vector-costs.cc
@@ -563,14 +563,24 @@ get_store_value (gimple *stmt)
return gimple_assign_rhs1 (stmt);
}
-/* Return true if it is non-contiguous load/store. */
+/* Return true if addtional vector vars needed. */
static bool
-non_contiguous_memory_access_p (stmt_vec_info stmt_info)
+need_additional_vector_vars_p (stmt_vec_info stmt_info)
{
enum stmt_vec_info_type type
= STMT_VINFO_TYPE (vect_stmt_to_vectorize (stmt_info));
- return ((type == load_vec_info_type || type == store_vec_info_type)
- && !adjacent_dr_p (STMT_VINFO_DATA_REF (stmt_info)));
+ if (type == load_vec_info_type || type == store_vec_info_type)
+{
+ if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
+ && STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info) == VMAT_GATHER_SCATTER)
+ return true;
+
+ machine_mode mode = TYPE_MODE (STMT_VINFO_VECTYPE (stmt_info));
+ int lmul = riscv_get_v_regno_alignment (mode);
+ if (DR_GROUP_SIZE (stmt_info) * lmul > RVV_M8)
+ return true;
+}
+ return false;
}
/* Return the LMUL of the current analysis. */
@@ -739,10 +749,7 @@ update_local_live_ranges (
stmt_vec_info stmt_info = vinfo->lookup_stmt (gsi_stmt (si));
enum stmt_vec_info_type type
= STMT_VINFO_TYPE (vect_stmt_to_vectorize (stmt_info));
- if (non_contiguous_memory_access_p (stmt_info)
- /* LOAD_LANES/STORE_LANES doesn't need a perm indice. */
- && STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info)
- != VMAT_LOAD_STORE_LANES)
+ if (need_additional_vector_vars_p (stmt_info))
{
/* For non-adjacent load/store STMT, we will potentially
convert it into:
diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114506.c
b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114506.c
new file mode 100644
index 000..a88d24b2d2d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114506.c
@@ -0,0 +1,23 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -ftree-vectorize
-mrvv-max-lmul=dynamic -fdump-tree-vect-details" } */
+
+float a[32000], b[32000], c[32000], d[32000];
+float aa[256][256], bb[256][256], cc[256][256];
+
+void
+s2275 ()
+{
+ for (int i = 0; i < 256; i++)
+{
+ for (int j = 0; j < 256; j++)
+ {
+ aa[j][i] = aa[j][i] + bb[j][i] * cc[j][i];
+ }
+ a[i] = b[i] + c[i] * d[i];
+}
+}
+
+/* { dg-final { scan-assembler-times {e32,m8} 1 } } */
+/* { dg-final { scan-assembler-not {e32,m4} } } */
+/* { dg-final { scan-assembler-not {csrr} } } */
+/* { dg-final { scan-tree-dump-not "Preferring smaller LMUL loop because it
has unexpected spills" "vect" } } */
[gcc r14-9826] RISC-V: Minor fix for max_point
https://gcc.gnu.org/g:aa2ab7b79a87c25d113752401a6026c6823dfe57 commit r14-9826-gaa2ab7b79a87c25d113752401a6026c6823dfe57 Author: demin.han Date: Mon Apr 1 16:20:46 2024 +0800 RISC-V: Minor fix for max_point The program points start from 1, so max_point should be equal to length(). Tested on RV64 and no regression. gcc/ChangeLog: * config/riscv/riscv-vector-costs.cc: Use length() Signed-off-by: demin.han Diff: --- gcc/config/riscv/riscv-vector-costs.cc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/gcc/config/riscv/riscv-vector-costs.cc b/gcc/config/riscv/riscv-vector-costs.cc index f462c272a6e..5ceb313c118 100644 --- a/gcc/config/riscv/riscv-vector-costs.cc +++ b/gcc/config/riscv/riscv-vector-costs.cc @@ -752,7 +752,7 @@ update_local_live_ranges ( We will be likely using one more vector variable. */ unsigned int max_point - = (*program_points_per_bb.get (bb)).length () - 1; + = (*program_points_per_bb.get (bb)).length (); auto *live_ranges = live_ranges_per_bb.get (bb); bool existed_p = false; tree var = type == load_vec_info_type
[gcc r14-9380] RISC-V: Fix ICE in riscv vector costs
https://gcc.gnu.org/g:42187c6968af9907de1f9b7563d87de739857684
commit r14-9380-g42187c6968af9907de1f9b7563d87de739857684
Author: demin.han
Date: Wed Mar 6 17:34:34 2024 +0800
RISC-V: Fix ICE in riscv vector costs
The following code can result in ICE:
-march=rv64gcv --param riscv-autovec-lmul=dynamic -O3
char *jpeg_difference7_input_buf;
void jpeg_difference7(int *diff_buf) {
unsigned width;
int samp, Rb;
while (--width) {
Rb = samp = *jpeg_difference7_input_buf;
*diff_buf++ = -(int)(samp + (long)Rb >> 1);
}
}
One biggest_mode update missed in one branch and trigger assertion fail.
gcc_assert (biggest_size >= mode_size);
Tested On RV64 and no regression.
PR target/114264
gcc/ChangeLog:
* config/riscv/riscv-vector-costs.cc: Fix ICE
gcc/testsuite/ChangeLog:
* gcc.dg/vect/costmodel/riscv/rvv/pr114264.c: New test.
Signed-off-by: demin.han
Diff:
---
gcc/config/riscv/riscv-vector-costs.cc | 2 ++
gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114264.c | 15 +++
2 files changed, 17 insertions(+)
diff --git a/gcc/config/riscv/riscv-vector-costs.cc
b/gcc/config/riscv/riscv-vector-costs.cc
index adf9c197df5..5ac8655b4d8 100644
--- a/gcc/config/riscv/riscv-vector-costs.cc
+++ b/gcc/config/riscv/riscv-vector-costs.cc
@@ -414,6 +414,8 @@ compute_local_live_ranges (
auto *r = get_live_range (live_ranges, arg);
gcc_assert (r);
(*r).second = MAX (point, (*r).second);
+ biggest_mode = get_biggest_mode (
+ biggest_mode, TYPE_MODE (TREE_TYPE (arg)));
}
}
else
diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114264.c
b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114264.c
new file mode 100644
index 000..7853f292af7
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/costmodel/riscv/rvv/pr114264.c
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -ftree-vectorize
--param=riscv-autovec-lmul=dynamic" } */
+
+char *jpeg_difference7_input_buf;
+void
+jpeg_difference7 (int *diff_buf)
+{
+ unsigned width;
+ int samp, Rb;
+ while (--width)
+{
+ Rb = samp = *jpeg_difference7_input_buf;
+ *diff_buf++ = -(int) (samp + (long) Rb >> 1);
+}
+}
[gcc r14-9349] RISC-V: Refactor expand_vec_cmp [NFC]
https://gcc.gnu.org/g:1cd8254ebad7b73993d2acee80a7caf37c21878a
commit r14-9349-g1cd8254ebad7b73993d2acee80a7caf37c21878a
Author: demin.han
Date: Mon Feb 26 14:50:15 2024 +0800
RISC-V: Refactor expand_vec_cmp [NFC]
There are two expand_vec_cmp functions.
They have same structure and similar code.
We can use default arguments instead of overloading.
Tested on RV32 and RV64.
gcc/ChangeLog:
* config/riscv/riscv-protos.h (expand_vec_cmp): Change proto
* config/riscv/riscv-v.cc (expand_vec_cmp): Use default arguments
(expand_vec_cmp_float): Adapt arguments
Signed-off-by: demin.han
Diff:
---
gcc/config/riscv/riscv-protos.h | 2 +-
gcc/config/riscv/riscv-v.cc | 44 +
2 files changed, 15 insertions(+), 31 deletions(-)
diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
index 80efdf2b7e5..b8735593805 100644
--- a/gcc/config/riscv/riscv-protos.h
+++ b/gcc/config/riscv/riscv-protos.h
@@ -603,7 +603,7 @@ bool simm5_p (rtx);
bool neg_simm5_p (rtx);
#ifdef RTX_CODE
bool has_vi_variant_p (rtx_code, rtx);
-void expand_vec_cmp (rtx, rtx_code, rtx, rtx);
+void expand_vec_cmp (rtx, rtx_code, rtx, rtx, rtx = nullptr, rtx = nullptr);
bool expand_vec_cmp_float (rtx, rtx_code, rtx, rtx, bool);
void expand_cond_len_unop (unsigned, rtx *);
void expand_cond_len_binop (unsigned, rtx *);
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index 2d32db06dd1..967f4e38287 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -2775,7 +2775,8 @@ vectorize_related_mode (machine_mode vector_mode,
scalar_mode element_mode,
/* Expand an RVV comparison. */
void
-expand_vec_cmp (rtx target, rtx_code code, rtx op0, rtx op1)
+expand_vec_cmp (rtx target, rtx_code code, rtx op0, rtx op1, rtx mask,
+ rtx maskoff)
{
machine_mode mask_mode = GET_MODE (target);
machine_mode data_mode = GET_MODE (op0);
@@ -2785,8 +2786,8 @@ expand_vec_cmp (rtx target, rtx_code code, rtx op0, rtx
op1)
{
rtx lt = gen_reg_rtx (mask_mode);
rtx gt = gen_reg_rtx (mask_mode);
- expand_vec_cmp (lt, LT, op0, op1);
- expand_vec_cmp (gt, GT, op0, op1);
+ expand_vec_cmp (lt, LT, op0, op1, mask, maskoff);
+ expand_vec_cmp (gt, GT, op0, op1, mask, maskoff);
icode = code_for_pred (IOR, mask_mode);
rtx ops[] = {target, lt, gt};
emit_vlmax_insn (icode, BINARY_MASK_OP, ops);
@@ -2794,33 +2795,16 @@ expand_vec_cmp (rtx target, rtx_code code, rtx op0, rtx
op1)
}
rtx cmp = gen_rtx_fmt_ee (code, mask_mode, op0, op1);
- rtx ops[] = {target, cmp, op0, op1};
- emit_vlmax_insn (icode, COMPARE_OP, ops);
-}
-
-void
-expand_vec_cmp (rtx target, rtx_code code, rtx mask, rtx maskoff, rtx op0,
- rtx op1)
-{
- machine_mode mask_mode = GET_MODE (target);
- machine_mode data_mode = GET_MODE (op0);
- insn_code icode = get_cmp_insn_code (code, data_mode);
-
- if (code == LTGT)
+ if (!mask && !maskoff)
{
- rtx lt = gen_reg_rtx (mask_mode);
- rtx gt = gen_reg_rtx (mask_mode);
- expand_vec_cmp (lt, LT, mask, maskoff, op0, op1);
- expand_vec_cmp (gt, GT, mask, maskoff, op0, op1);
- icode = code_for_pred (IOR, mask_mode);
- rtx ops[] = {target, lt, gt};
- emit_vlmax_insn (icode, BINARY_MASK_OP, ops);
- return;
+ rtx ops[] = {target, cmp, op0, op1};
+ emit_vlmax_insn (icode, COMPARE_OP, ops);
+}
+ else
+{
+ rtx ops[] = {target, mask, maskoff, cmp, op0, op1};
+ emit_vlmax_insn (icode, COMPARE_OP_MU, ops);
}
-
- rtx cmp = gen_rtx_fmt_ee (code, mask_mode, op0, op1);
- rtx ops[] = {target, mask, maskoff, cmp, op0, op1};
- emit_vlmax_insn (icode, COMPARE_OP_MU, ops);
}
/* Expand an RVV floating-point comparison:
@@ -2898,7 +2882,7 @@ expand_vec_cmp_float (rtx target, rtx_code code, rtx op0,
rtx op1,
else
{
/* vmfeq.vvv0, vb, vb, v0.t */
- expand_vec_cmp (eq0, EQ, eq0, eq0, op1, op1);
+ expand_vec_cmp (eq0, EQ, op1, op1, eq0, eq0);
}
break;
default:
@@ -2916,7 +2900,7 @@ expand_vec_cmp_float (rtx target, rtx_code code, rtx op0,
rtx op1,
if (code == ORDERED)
emit_move_insn (target, eq0);
else
-expand_vec_cmp (eq0, code, eq0, eq0, op0, op1);
+expand_vec_cmp (eq0, code, op0, op1, eq0, eq0);
if (can_invert_p)
{
