From: Robin Dapp <rd...@ventanamicro.com>
This patch adds a second variant to implement the extract/slide1up
pattern. In order to do a permutation like
<3, 4, 5, 6> from vectors <0, 1, 2, 3> and <4, 5, 6, 7>
we currently extract <3> from the first vector and re-insert it into the
second vector. Unless register-file crossing latency is essentially
zero it should be preferable to first slide the second vector up by
one, then slide down the first vector by (nunits - 1).
gcc/ChangeLog:
* config/riscv/riscv-protos.h (riscv_register_move_cost):
Export.
* config/riscv/riscv-v.cc (shuffle_extract_and_slide1up_patterns):
Rename...
(shuffle_off_by_one_patterns): ... to this and add slideup/slidedown
variant.
(expand_vec_perm_const_1): Call renamed function.
* config/riscv/riscv.cc (riscv_secondary_memory_needed): Remove
static.
(riscv_register_move_cost): Add VR<->GR/FR handling.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/pr112599-2.c: Adjust test
expectation.
---
gcc/config/riscv/riscv-protos.h | 1 +
gcc/config/riscv/riscv-v.cc | 52 ++++++++++++++-----
gcc/config/riscv/riscv.cc | 18 ++++++-
.../gcc.target/riscv/rvv/autovec/pr112599-2.c | 2 +-
4 files changed, 57 insertions(+), 16 deletions(-)
diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
index 500b357f6eb..ecb4e64cdf8 100644
--- a/gcc/config/riscv/riscv-protos.h
+++ b/gcc/config/riscv/riscv-protos.h
@@ -139,6 +139,7 @@ extern void riscv_expand_ussub (rtx, rtx, rtx);
extern void riscv_expand_sssub (rtx, rtx, rtx);
extern void riscv_expand_ustrunc (rtx, rtx);
extern void riscv_expand_sstrunc (rtx, rtx);
+extern int riscv_register_move_cost (machine_mode, reg_class_t, reg_class_t);
#ifdef RTX_CODE
extern void riscv_expand_int_scc (rtx, enum rtx_code, rtx, rtx, bool
*invert_ptr = 0);
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index ded8d52d9f9..12251b7f90b 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -3796,11 +3796,13 @@ shuffle_bswap_pattern (struct expand_vec_perm_d *d)
return true;
}
-/* Recognize the pattern that can be shuffled by vec_extract and slide1up
- approach. */
+/* Recognize patterns like [3 4 5 6] where we combine the last element
+ of the first vector and the first n - 1 elements of the second vector.
+ This can be implemented by slides or by extracting and re-inserting
+ (slide1up) the first vector's last element. */
static bool
-shuffle_extract_and_slide1up_patterns (struct expand_vec_perm_d *d)
+shuffle_off_by_one_patterns (struct expand_vec_perm_d *d)
{
poly_int64 nunits = GET_MODE_NUNITS (d->vmode);
@@ -3818,17 +3820,39 @@ shuffle_extract_and_slide1up_patterns (struct
expand_vec_perm_d *d)
if (d->testing_p)
return true;
- /* Extract the last element of the first vector. */
- scalar_mode smode = GET_MODE_INNER (d->vmode);
- rtx tmp = gen_reg_rtx (smode);
- emit_vec_extract (tmp, d->op0, gen_int_mode (nunits - 1, Pmode));
+ int scalar_cost = riscv_register_move_cost (d->vmode, V_REGS, GR_REGS)
+ + riscv_register_move_cost (d->vmode, GR_REGS, V_REGS) + 2;
+ int slide_cost = 2;
+
+ if (slide_cost < scalar_cost)
+ {
+ /* This variant should always be preferable because we just need two
+ slides. The extract-variant also requires two slides but additionally
+ pays the latency for register-file crossing. */
+ rtx tmp = gen_reg_rtx (d->vmode);
+ rtx ops[] = {tmp, d->op1, gen_int_mode (1, Pmode)};
+ insn_code icode = code_for_pred_slide (UNSPEC_VSLIDEUP, d->vmode);
+ emit_vlmax_insn (icode, BINARY_OP, ops);
+
+ rtx ops2[] = {d->target, tmp, d->op0, gen_int_mode (nunits - 1, Pmode)};
+ icode = code_for_pred_slide (UNSPEC_VSLIDEDOWN, d->vmode);
+ emit_nonvlmax_insn (icode, BINARY_OP_TUMA, ops2, gen_int_mode (1,
Pmode));
+ }
+ else
+ {
+ /* Extract the last element of the first vector. */
+ scalar_mode smode = GET_MODE_INNER (d->vmode);
+ rtx tmp = gen_reg_rtx (smode);
+ emit_vec_extract (tmp, d->op0, gen_int_mode (nunits - 1, Pmode));
+
+ /* Insert the scalar into element 0. */
+ unsigned int unspec
+ = FLOAT_MODE_P (d->vmode) ? UNSPEC_VFSLIDE1UP : UNSPEC_VSLIDE1UP;
+ insn_code icode = code_for_pred_slide (unspec, d->vmode);
+ rtx ops[] = {d->target, d->op1, tmp};
+ emit_vlmax_insn (icode, BINARY_OP, ops);
+ }
- /* Insert the scalar into element 0. */
- unsigned int unspec
- = FLOAT_MODE_P (d->vmode) ? UNSPEC_VFSLIDE1UP : UNSPEC_VSLIDE1UP;
- insn_code icode = code_for_pred_slide (unspec, d->vmode);
- rtx ops[] = {d->target, d->op1, tmp};
- emit_vlmax_insn (icode, BINARY_OP, ops);
return true;
}
@@ -3960,7 +3984,7 @@ expand_vec_perm_const_1 (struct expand_vec_perm_d *d)
return true;
if (shuffle_bswap_pattern (d))
return true;
- if (shuffle_extract_and_slide1up_patterns (d))
+ if (shuffle_off_by_one_patterns (d))
return true;
if (shuffle_series_patterns (d))
return true;
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 93702f71ec9..9a6d05bd3da 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -9468,7 +9468,7 @@ riscv_secondary_memory_needed (machine_mode mode,
reg_class_t class1,
/* Implement TARGET_REGISTER_MOVE_COST. */
-static int
+int
riscv_register_move_cost (machine_mode mode,
reg_class_t from, reg_class_t to)
{
@@ -9476,6 +9476,22 @@ riscv_register_move_cost (machine_mode mode,
(from == GR_REGS && to == FP_REGS))
return tune_param->fmv_cost;
+ if (from == V_REGS)
+ {
+ if (to == GR_REGS)
+ return get_vector_costs ()->regmove->VR2GR;
+ else if (to == FP_REGS)
+ return get_vector_costs ()->regmove->VR2FR;
+ }
+
+ if (to == V_REGS)
+ {
+ if (from == GR_REGS)
+ return get_vector_costs ()->regmove->GR2VR;
+ else if (from == FP_REGS)
+ return get_vector_costs ()->regmove->FR2VR;
+ }
+
return riscv_secondary_memory_needed (mode, from, to) ? 8 : 2;
}
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
index fd87565b054..79d87196bf7 100644
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
@@ -48,4 +48,4 @@ foo(void)
}
/* { dg-final { scan-assembler-not {vrgather} } } */
-/* { dg-final { scan-assembler-times {vslide1up\.vx} 1 } } */
+/* { dg-final { scan-assembler {vslide} } } */
--
2.47.0
