Hi,
This is Zhongyao's slightly adjusted patch backported to GCC 15.
Obviously the CI won't be able to apply it but I have regtested
it on rv64gcv_zvl512b.
Going to push soon.
--
Regards
Robin
This patch removes the "hi/lo" optimization path in interleaved
stepped constant vector synthesis. This optimization was found
to be not really better than the fallback merge version. Once the
overflow issue is fixed, the extra masking makes the code generation
"even worse". So remove it instead of fixing.
Obsolete tests slp-interleave-[1-5].c are also removed as they were
specifically designed to verify this now-removed path.
My local test shows no regression.
PR target/125215
gcc/ChangeLog:
* config/riscv/riscv-v.cc
(expand_const_vector_interleaved_stepped_npatterns):
Remove hi/lo optimization and always use the merge fallback.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr125215.c: New test.
* gcc.target/riscv/rvv/autovec/slp-interleave-1.c: Remove.
* gcc.target/riscv/rvv/autovec/slp-interleave-2.c: Remove.
* gcc.target/riscv/rvv/autovec/slp-interleave-3.c: Remove.
* gcc.target/riscv/rvv/autovec/slp-interleave-4.c: Remove.
* gcc.target/riscv/rvv/autovec/slp-interleave-5.c: Remove.
---
gcc/config/riscv/riscv-v.cc | 164 ++++--------------
.../gcc.target/riscv/rvv/autovec/pr125215.c | 24 +++
.../riscv/rvv/autovec/slp-interleave-1.c | 17 --
.../riscv/rvv/autovec/slp-interleave-2.c | 18 --
.../riscv/rvv/autovec/slp-interleave-3.c | 19 --
.../riscv/rvv/autovec/slp-interleave-4.c | 19 --
.../riscv/rvv/autovec/slp-interleave-5.c | 35 ----
7 files changed, 57 insertions(+), 239 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/pr125215.c
delete mode 100644
gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-1.c
delete mode 100644
gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-2.c
delete mode 100644
gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-3.c
delete mode 100644
gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-4.c
delete mode 100644
gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-5.c
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index d7a69a4847f..242dfee1a1f 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -1480,140 +1480,42 @@ expand_const_vector (rtx target, rtx src)
= rtx_to_poly_int64 (builder.elt (builder.npatterns () + 1))
- rtx_to_poly_int64 (base2);
- /* For { 1, 0, 2, 0, ... , n - 1, 0 }, we can use larger EEW
- integer vector mode to generate such vector efficiently.
-
- E.g. EEW = 16, { 2, 0, 4, 0, ... }
-
- can be interpreted into:
-
- EEW = 32, { 2, 4, ... }.
-
- Both the series1 and series2 may overflow before taking the IOR
- to generate the final result. However, only series1 matters
- because the series2 will shift before IOR, thus the overflow
- bits will never pollute the final result.
-
- For now we forbid the negative steps and overflow, and they
- will fall back to the default merge way to generate the
- const_vector. */
-
- unsigned int new_smode_bitsize = builder.inner_bits_size () * 2;
- scalar_int_mode new_smode;
- machine_mode new_mode;
- poly_uint64 new_nunits
- = exact_div (GET_MODE_NUNITS (builder.mode ()), 2);
-
- poly_int64 base1_poly = rtx_to_poly_int64 (base1);
- bool overflow_smode_p = false;
-
- if (!step1.is_constant ())
- overflow_smode_p = true;
- else
- {
- int elem_count = XVECLEN (src, 0);
- uint64_t step1_val = step1.to_constant ();
- uint64_t base1_val = base1_poly.to_constant ();
- uint64_t elem_val = base1_val + (elem_count - 1) * step1_val;
-
- if ((elem_val >> builder.inner_bits_size ()) != 0)
- overflow_smode_p = true;
- }
-
- if (known_ge (step1, 0) && known_ge (step2, 0)
- && int_mode_for_size (new_smode_bitsize, 0).exists (&new_smode)
- && get_vector_mode (new_smode, new_nunits).exists (&new_mode)
- && !overflow_smode_p)
+ rtx vid = gen_reg_rtx (mode);
+ expand_vec_series (vid, const0_rtx, const1_rtx);
+ /* Transform into { 0, 0, 1, 1, 2, 2, ... }. */
+ rtx shifted_vid;
+ if (lra_in_progress)
{
- rtx tmp1 = gen_reg_rtx (new_mode);
- base1 = gen_int_mode (base1_poly, new_smode);
- expand_vec_series (tmp1, base1, gen_int_mode (step1, new_smode));
-
- if (rtx_equal_p (base2, const0_rtx) && known_eq (step2, 0))
- /* { 1, 0, 2, 0, ... }. */
- emit_move_insn (result, gen_lowpart (mode, tmp1));
- else if (known_eq (step2, 0))
- {
- /* { 1, 1, 2, 1, ... }. */
- rtx scalar = expand_simple_binop (
- Xmode, ASHIFT,
- gen_int_mode (rtx_to_poly_int64 (base2), Xmode),
- gen_int_mode (builder.inner_bits_size (), Xmode),
- NULL_RTX, false, OPTAB_DIRECT);
- scalar = simplify_gen_subreg (new_smode, scalar, Xmode, 0);
- rtx tmp2 = gen_reg_rtx (new_mode);
- rtx ior_ops[] = {tmp2, tmp1, scalar};
- emit_vlmax_insn (code_for_pred_scalar (IOR, new_mode),
- BINARY_OP, ior_ops);
- emit_move_insn (result, gen_lowpart (mode, tmp2));
- }
- else
- {
- /* { 1, 3, 2, 6, ... }. */
- rtx tmp2 = gen_reg_rtx (new_mode);
- base2 = gen_int_mode (rtx_to_poly_int64 (base2), new_smode);
- expand_vec_series (tmp2, base2,
- gen_int_mode (step2, new_smode));
- rtx shifted_tmp2;
- rtx shift = gen_int_mode (builder.inner_bits_size (), Xmode);
- if (lra_in_progress)
- {
- shifted_tmp2 = gen_reg_rtx (new_mode);
- rtx shift_ops[] = {shifted_tmp2, tmp2, shift};
- emit_vlmax_insn (code_for_pred_scalar
- (ASHIFT, new_mode), BINARY_OP,
- shift_ops);
- }
- else
- shifted_tmp2 = expand_simple_binop (new_mode, ASHIFT, tmp2,
- shift, NULL_RTX, false,
- OPTAB_DIRECT);
- rtx tmp3 = gen_reg_rtx (new_mode);
- rtx ior_ops[] = {tmp3, tmp1, shifted_tmp2};
- emit_vlmax_insn (code_for_pred (IOR, new_mode), BINARY_OP,
- ior_ops);
- emit_move_insn (result, gen_lowpart (mode, tmp3));
- }
+ shifted_vid = gen_reg_rtx (mode);
+ rtx shift = gen_int_mode (1, Xmode);
+ rtx shift_ops[] = {shifted_vid, vid, shift};
+ emit_vlmax_insn (code_for_pred_scalar
+ (LSHIFTRT, mode), BINARY_OP,
+ shift_ops);
}
else
- {
- rtx vid = gen_reg_rtx (mode);
- expand_vec_series (vid, const0_rtx, const1_rtx);
- /* Transform into { 0, 0, 1, 1, 2, 2, ... }. */
- rtx shifted_vid;
- if (lra_in_progress)
- {
- shifted_vid = gen_reg_rtx (mode);
- rtx shift = gen_int_mode (1, Xmode);
- rtx shift_ops[] = {shifted_vid, vid, shift};
- emit_vlmax_insn (code_for_pred_scalar
- (LSHIFTRT, mode), BINARY_OP,
- shift_ops);
- }
- else
- shifted_vid = expand_simple_binop (mode, LSHIFTRT, vid,
- const1_rtx, NULL_RTX,
- false, OPTAB_DIRECT);
- rtx tmp1 = gen_reg_rtx (mode);
- rtx tmp2 = gen_reg_rtx (mode);
- expand_vec_series (tmp1, base1,
- gen_int_mode (step1, builder.inner_mode ()),
- shifted_vid);
- expand_vec_series (tmp2, base2,
- gen_int_mode (step2, builder.inner_mode ()),
- shifted_vid);
-
- /* Transform into { 0, 1, 0, 1, 0, 1, ... }. */
- rtx and_vid = gen_reg_rtx (mode);
- rtx and_ops[] = {and_vid, vid, const1_rtx};
- emit_vlmax_insn (code_for_pred_scalar (AND, mode), BINARY_OP,
- and_ops);
- rtx mask = gen_reg_rtx (builder.mask_mode ());
- expand_vec_cmp (mask, EQ, and_vid, CONST1_RTX (mode));
-
- rtx ops[] = {result, tmp1, tmp2, mask};
- emit_vlmax_insn (code_for_pred_merge (mode), MERGE_OP, ops);
- }
+ shifted_vid = expand_simple_binop (mode, LSHIFTRT, vid,
+ const1_rtx, NULL_RTX,
+ false, OPTAB_DIRECT);
+ rtx tmp1 = gen_reg_rtx (mode);
+ rtx tmp2 = gen_reg_rtx (mode);
+ expand_vec_series (tmp1, base1,
+ gen_int_mode (step1, builder.inner_mode ()),
+ shifted_vid);
+ expand_vec_series (tmp2, base2,
+ gen_int_mode (step2, builder.inner_mode ()),
+ shifted_vid);
+
+ /* Transform into { 0, 1, 0, 1, 0, 1, ... }. */
+ rtx and_vid = gen_reg_rtx (mode);
+ rtx and_ops[] = {and_vid, vid, const1_rtx};
+ emit_vlmax_insn (code_for_pred_scalar (AND, mode), BINARY_OP,
+ and_ops);
+ rtx mask = gen_reg_rtx (builder.mask_mode ());
+ expand_vec_cmp (mask, EQ, and_vid, CONST1_RTX (mode));
+
+ rtx ops[] = {result, tmp1, tmp2, mask};
+ emit_vlmax_insn (code_for_pred_merge (mode), MERGE_OP, ops);
}
else
/* TODO: We will enable more variable-length vector in the future. */
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr125215.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr125215.c
new file mode 100644
index 00000000000..7c280701a38
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr125215.c
@@ -0,0 +1,24 @@
+/* { dg-do run { target { riscv_v } } } */
+/* { dg-options "-O3 -march=rv64gcv -mabi=lp64d -fno-vect-cost-model" } */
+
+#include <stdint-gcc.h>
+
+__attribute__((noipa))
+void foo(uint8_t *d)
+{
+ d[0] = 128; d[1] = 135; d[2] = 130; d[3] = 149;
+ d[4] = 132; d[5] = 163; d[6] = 134; d[7] = 177;
+ d[8] = 136; d[9] = 191; d[10] = 138; d[11] = 205;
+ d[12] = 140; d[13] = 219; d[14] = 142; d[15] = 233;
+}
+
+int main()
+{
+ uint8_t d[16];
+ uint8_t e[] = {128, 135, 130, 149, 132, 163, 134, 177, 136, 191, 138, 205,
140, 219, 142, 233};
+ foo(d);
+ for (int i = 0; i < 16; i++)
+ if (d[i] != e[i])
+ __builtin_abort ();
+ return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-1.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-1.c
deleted file mode 100644
index 296fd4a45c0..00000000000
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-1.c
+++ /dev/null
@@ -1,17 +0,0 @@
-/* { dg-do compile } */
-/* { dg-options "-march=rv64gcv_zvl1024b -mabi=lp64d -fno-vect-cost-model
-mrvv-max-lmul=m8 -O3 -fdump-tree-optimized-details" } */
-
-struct S { int a, b; } s[8];
-
-void
-foo ()
-{
- int i;
- for (i = 0; i < 8; i++)
- {
- s[i].b = 0;
- s[i].a = i;
- }
-}
-
-/* { dg-final { scan-tree-dump-times "\{ 0, 0, 1, 0, 2, 0, ... \}" 1
"optimized" } } */
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-2.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-2.c
deleted file mode 100644
index d992c4c087c..00000000000
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-2.c
+++ /dev/null
@@ -1,18 +0,0 @@
-/* { dg-do compile } */
-/* { dg-options "-march=rv64gcv_zvl1024b -mabi=lp64d -fno-vect-cost-model
-mrvv-max-lmul=m8 -O3 -fdump-tree-optimized-details" } */
-
-struct S { int a, b; } s[8];
-
-void
-foo ()
-{
- int i;
- for (i = 0; i < 8; i++)
- {
- s[i].b = 1;
- s[i].a = i;
- }
-}
-
-/* { dg-final { scan-tree-dump-times "\{ 0, 1, 1, 1, 2, 1, ... \}" 1
"optimized" } } */
-/* { dg-final { scan-assembler-times {slli\t[a-x0-9]+,\s*[a-x0-9]+,\s*32} 1 }
} */
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-3.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-3.c
deleted file mode 100644
index 96bd8ec34f4..00000000000
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-3.c
+++ /dev/null
@@ -1,19 +0,0 @@
-/* { dg-do compile } */
-/* { dg-options "-march=rv64gcv_zvl1024b -mabi=lp64d -fno-vect-cost-model
-mrvv-max-lmul=m8 -O3 -fdump-tree-optimized-details" } */
-
-struct S { int a, b; } s[8];
-
-void
-foo ()
-{
- int i;
- for (i = 0; i < 8; i++)
- {
- s[i].b = i*3 + 100;
- s[i].a = i + 200;
- }
-}
-
-/* { dg-final { scan-tree-dump-times "\{ 200, 100, 201, 103, 202, 106, ... \}"
1 "optimized" } } */
-/* { dg-final { scan-assembler-times {vsll\.vx} 1 } } */
-/* { dg-final { scan-assembler-times {vor\.vv} 1 } } */
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-4.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-4.c
deleted file mode 100644
index 64483c58ea2..00000000000
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-4.c
+++ /dev/null
@@ -1,19 +0,0 @@
-/* { dg-do compile } */
-/* { dg-options "-march=rv64gc_zve32f_zvl1024b -mabi=lp64d
-fno-vect-cost-model -mrvv-max-lmul=m8 -O3 -fdump-tree-optimized-details" } */
-
-struct S { int a, b; } s[8];
-
-void
-foo ()
-{
- int i;
- for (i = 0; i < 8; i++)
- {
- s[i].b = i*3 + 100;
- s[i].a = i + 200;
- }
-}
-
-/* { dg-final { scan-tree-dump-times "\{ 200, 100, 201, 103, 202, 106, ... \}"
1 "optimized" } } */
-/* { dg-final { scan-assembler-times {vand\.vi} 1 } } */
-/* { dg-final { scan-assembler-times {vmseq\.vi} 1 } } */
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-5.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-5.c
deleted file mode 100644
index 32cfe8a8688..00000000000
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/slp-interleave-5.c
+++ /dev/null
@@ -1,35 +0,0 @@
-/* { dg-do compile } */
-/* { dg-options "-march=rv64gcv_zvl1024b -mabi=lp64d -O3
-fdump-tree-optimized-details" } */
-
-struct S { int a, b; } s[8];
-
-void
-foo ()
-{
- int i;
- for (i = 0; i < 8; i++)
- {
- s[i].b = 1;
- s[i].a = i+1;
- }
-}
-
-/* { dg-final { scan-tree-dump-times "\{ 1, 1, 2, 1, 3, 1, 4, 1 \}" 1
"optimized" } } */
-/* { dg-final { scan-assembler {vid\.v} } } */
-/* { dg-final { scan-assembler {vadd\.v} } } */
-/* { dg-final { scan-assembler {vor\.v} } } */
-
-void
-foo2 ()
-{
- int i;
- for (i = 0; i < 8; i++)
- {
- s[i].b = 0;
- s[i].a = i+1;
- }
-}
-
-/* { dg-final { scan-tree-dump-times "\{ 1, 0, 2, 0, 3, 0, 4, 0 \}" 1
"optimized" } } */
-/* { dg-final { scan-assembler {vid\.v} } } */
-/* { dg-final { scan-assembler {vadd\.v} } } */
--
2.54.0
