This patch fixes the following FAILs in "full coverage" testing:
Running target riscv-sim/-march=rv64gcv_zvl256b/-mabi=lp64d/-mcmodel=medlow/--param=riscv-autovec-lmul=m8/--param=riscv-autovec-preference=fixed-vlmax
FAIL: gcc.dg/vect/vect-strided-mult-char-ls.c -flto -ffat-lto-objects execution test
FAIL: gcc.dg/vect/vect-strided-mult-char-ls.c execution test
FAIL: gcc.dg/vect/vect-strided-u8-i2.c -flto -ffat-lto-objects execution test
FAIL: gcc.dg/vect/vect-strided-u8-i2.c execution test
The root cause is vmerge optimization on this following IR:
_45 = VEC_PERM_EXPR <vect__3.13_47, vect__4.14_46, { 0, 257, 2, 259, 4, 261, 6, 263, 8, 265, 10, 267, 12, 269, 14, 271, 16, 273, 18, 275, 20, 277, 22, 279, 24, 281, 26, 283, 28, 285, 30, 287, 32, 289, 34, 291, 36, 293, 38, 295, 40, 297, 42, 299, 44, 301, 46, 303, 48, 305, 50, 307, 52, 309, 54, 311, 56, 313, 58, 315, 60, 317, 62, 319, 64, 321, 66, 323, 68, 325, 70, 327, 72, 329, 74, 331, 76, 333, 78, 335, 80, 337, 82, 339, 84, 341, 86, 343, 88, 345, 90, 347, 92, 349, 94, 351, 96, 353, 98, 355, 100, 357, 102, 359, 104, 361, 106, 363, 108, 365, 110, 367, 112, 369, 114, 371, 116, 373, 118, 375, 120, 377, 122, 379, 124, 381, 126, 383, 128, 385, 130, 387, 132, 389, 134, 391, 136, 393, 138, 395, 140, 397, 142, 399, 144, 401, 146, 403, 148, 405, 150, 407, 152, 409, 154, 411, 156, 413, 158, 415, 160, 417, 162, 419, 164, 421, 166, 423, 168, 425, 170, 427, 172, 429, 174, 431, 176, 433, 178, 435, 180, 437, 182, 439, 184, 441, 186, 443, 188, 445, 190, 447, 192, 449, 194, 451, 196, 453, 198, 455, 200, 457, 202, 459, 204, 461, 206, 463, 208, 465, 210, 467, 212, 469, 214, 471, 216, 473, 218, 475, 220, 477, 222, 479, 224, 481, 226, 483, 228, 485, 230, 487, 232, 489, 234, 491, 236, 493, 238, 495, 240, 497, 242, 499, 244, 501, 246, 503, 248, 505, 250, 507, 252, 509, 254, 511 }>;
It's obvious we have many index > 255 in shuffle indice. Here we use vmerge optimizaiton which is available but incorrect codgen cause run fail.
The bug codegen:
vsetvli zero,a4,e8,m8,ta,ma
vmsltu.vi v0,v0,0 -> it should be 256 instead of 0, but since it is EEW8 vector, 256 is not a available value that 8bit register can hold it.
vmerge.vvm v8,v8,v16,v0
After this patch:
vmv.v.x v0,a6
vmerge.vvm v8,v8,v16,v0
gcc/ChangeLog:
* config/riscv/riscv-v.cc (shuffle_merge_patterns): Fix bug.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/bug-1.c: New test.
---
gcc/config/riscv/riscv-v.cc | 59 ++++++++++++++++---
.../gcc.target/riscv/rvv/autovec/bug-1.c | 39 ++++++++++++
2 files changed, 90 insertions(+), 8 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/bug-1.c
@@ -2987,20 +2987,63 @@ shuffle_merge_patterns (struct expand_vec_perm_d *d)
&& !d->perm.series_p (i, n_patterns, vec_len + i, n_patterns))
return false;
+ /* We need to use precomputed mask for such situation and such mask
+ can only be computed in compile-time known size modes. */
+ bool indices_fit_selector_p
+ = GET_MODE_BITSIZE (GET_MODE_INNER (vmode)) > 8 || known_lt (vec_len, 256);
+ if (!indices_fit_selector_p && !vec_len.is_constant ())
+ return false;
+
if (d->testing_p)
return true;
machine_mode mask_mode = get_mask_mode (vmode);
rtx mask = gen_reg_rtx (mask_mode);
- rtx sel = vec_perm_indices_to_rtx (sel_mode, d->perm);
-
- /* MASK = SELECTOR < NUNTIS ? 1 : 0. */
- rtx x = gen_int_mode (vec_len, GET_MODE_INNER (sel_mode));
- insn_code icode = code_for_pred_cmp_scalar (sel_mode);
- rtx cmp = gen_rtx_fmt_ee (LTU, mask_mode, sel, x);
- rtx ops[] = {mask, cmp, sel, x};
- emit_vlmax_insn (icode, COMPARE_OP, ops);
+ if (indices_fit_selector_p)
+ {
+ /* MASK = SELECTOR < NUNTIS ? 1 : 0. */
+ rtx sel = vec_perm_indices_to_rtx (sel_mode, d->perm);
+ rtx x = gen_int_mode (vec_len, GET_MODE_INNER (sel_mode));
+ insn_code icode = code_for_pred_cmp_scalar (sel_mode);
+ rtx cmp = gen_rtx_fmt_ee (LTU, mask_mode, sel, x);
+ rtx ops[] = {mask, cmp, sel, x};
+ emit_vlmax_insn (icode, COMPARE_OP, ops);
+ }
+ else
+ {
+ /* For EEW8 and NUNITS may be larger than 255, we can't use vmsltu
+ directly to generate the selector mask, instead, we can only use
+ precomputed mask.
+
+ E.g. selector = <0, 257, 2, 259> for EEW8 vector with NUNITS = 256, we
+ don't have a QImode scalar register to hold larger than 255.
+ We also cannot hold that in a vector QImode register if LMUL = 8, and,
+ since there is no larger HI mode vector we cannot create a larger
+ selector.
+
+ As the mask is a simple {0, 1, ...} pattern and the length is known we
+ can store it in a scalar register and broadcast it to a mask register.
+ */
+ gcc_assert (vec_len.is_constant ());
+ int size = CEIL (GET_MODE_NUNITS (mask_mode).to_constant (), 8);
+ machine_mode mode = get_vector_mode (QImode, size).require ();
+ rtx tmp = gen_reg_rtx (mode);
+ rvv_builder v (mode, 1, size);
+ for (int i = 0; i < vec_len.to_constant () / 8; i++)
+ {
+ uint8_t value = 0;
+ for (int j = 0; j < 8; j++)
+ {
+ int index = i * 8 + j;
+ if (known_lt (d->perm[index], 256))
+ value |= 1 << j;
+ }
+ v.quick_push (gen_int_mode (value, QImode));
+ }
+ emit_move_insn (tmp, v.build ());
+ emit_move_insn (mask, gen_lowpart (mask_mode, tmp));
+ }
/* TARGET = MASK ? OP0 : OP1. */
/* swap op0 and op1 since the order is opposite to pred_merge. */
new file mode 100644
@@ -0,0 +1,39 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv_zvl256b -mabi=lp64d --param=riscv-autovec-lmul=m8 --param=riscv-autovec-preference=fixed-vlmax -fno-vect-cost-model -O3 -fdump-tree-optimized" } */
+
+#include <assert.h>
+
+#define N 64
+
+typedef struct
+{
+ unsigned char a;
+ unsigned char b;
+} s;
+
+int
+main1 (s *arr)
+{
+ s *ptr = arr;
+ s res[N];
+ int i;
+
+ for (i = 0; i < N; i++)
+ {
+ res[i].a = ptr->b - ptr->a;
+ res[i].b = ptr->b + ptr->a;
+ ptr++;
+ }
+ /* check results: */
+#pragma GCC novector
+ for (i = 0; i < N; i++)
+ {
+ if (res[i].a != arr[i].b - arr[i].a || res[i].b != arr[i].a + arr[i].b)
+ assert (0);
+ }
+
+ return 0;
+}
+
+/* { dg-final { scan-tree-dump-times "\.VEC_PERM" 3 "optimized" } } */
+/* { dg-final { scan-assembler-not {vmsltu\.vi} } } */