OK. How about this following situation:
for (i ... i < 64.. i++) {
a[i] = a[i] + b[i]
}
When TARGET_MAX_LMUL is 1 or 2. Can auto-vectorize use LMUL 8 in VLS mode to
vectorize the case above ?
juzhe.zh...@rivai.ai
From: Kito Cheng
Date: 2025-08-14 17:39
To: 钟居哲
CC: kito.cheng; gcc-patches; palmer; jeffreyalaw; rdapp; Li, Pan2; vineetg;
patrick
Subject: Re: [PATCH] RISC-V: Allow VLS types using up to LMUL 8
Yeah, I guess you definitely got something wrong at that timing, I
also got some trouble when I relaxed the constraint...and then I
finally found the right way to enable that without changing behavior
of the auto vectorizer: only constraint the VLS mode in
autovectorize_vector_modes.
On Thu, Aug 14, 2025 at 5:29 PM 钟居哲 <juzhe.zh...@rivai.ai> wrote:
>
> As I remembered, I have ever trided to use larger LMUL in VLS mode than VLA
> modes.
> But it may cause ICE. I don't remember what scenario now... That's why
> originally I use MAX_LMUL to control both VLA modes and VLS modes.
>
> ________________________________
> juzhe.zh...@rivai.ai
>
>
> From: Kito Cheng
> Date: 2025-08-14 15:17
> To: gcc-patches; kito.cheng; palmer; jeffreyalaw; rdapp; juzhe.zhong;
> pan2.li; vineetg; patrick
> CC: Kito Cheng
> Subject: [PATCH] RISC-V: Allow VLS types using up to LMUL 8
> We used to apply -mrvv-max-lmul= to limit VLS code gen, auto vectorizer,
> and builtin string function expansion. But I think the VLS code gen part
> doesn't
> need this limit, since it only happens when the user explicitly writes vector
> types.
>
> For example, int32x8_t under -mrvv-max-lmul=m1 with VLEN=128 would be split
> into
> two int32x4_t, which generate more instructions and runs slower.
>
> In this patch, I changed -mrvv-max-lmul= to only affect auto vectorization and
> builtin string function expansion. Actually, the option's help text already
> says it only controls the LMUL used by auto-vectorization, so I believe this
> change is makes sense :)
>
> gcc/ChangeLog:
>
> * config/riscv/riscv-protos.h (vls_mode_valid_p): New argument
> allow_up_to_lmul_8.
> * config/riscv/riscv-v.cc (autovectorize_vector_modes): Set
> allow_up_to_lmul_8 to false.
> (vls_mode_valid_p): Add new argument allow_up_to_lmul_8, and use
> it to determine whether to allow LMUL 8.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target.riscv/rvv/vls-type-rvv-max-lmul.c: New test.
> ---
> gcc/config/riscv/riscv-protos.h | 2 +-
> gcc/config/riscv/riscv-v.cc | 31 ++++++++++---------
> .../riscv/rvv/vls-type-rvv-max-lmul.c | 12 +++++++
> 3 files changed, 29 insertions(+), 16 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/vls-type-rvv-max-lmul.c
>
> diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
> index 539321ff95b..045ee09b23f 100644
> --- a/gcc/config/riscv/riscv-protos.h
> +++ b/gcc/config/riscv/riscv-protos.h
> @@ -763,7 +763,7 @@ opt_machine_mode vectorize_related_mode (machine_mode,
> scalar_mode,
> unsigned int autovectorize_vector_modes (vec<machine_mode> *, bool);
> bool cmp_lmul_le_one (machine_mode);
> bool cmp_lmul_gt_one (machine_mode);
> -bool vls_mode_valid_p (machine_mode);
> +bool vls_mode_valid_p (machine_mode, bool allow_up_to_lmul_8 = true);
> bool vlmax_avl_type_p (rtx_insn *);
> bool has_vl_op (rtx_insn *);
> bool tail_agnostic_p (rtx_insn *);
> diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
> index c9c83282732..3484f6442e7 100644
> --- a/gcc/config/riscv/riscv-v.cc
> +++ b/gcc/config/riscv/riscv-v.cc
> @@ -2910,7 +2910,7 @@ autovectorize_vector_modes (vector_modes *modes, bool)
> machine_mode mode;
> while (size > 0 && get_vector_mode (QImode, size).exists (&mode))
> {
> - if (vls_mode_valid_p (mode))
> + if (vls_mode_valid_p (mode, /* allow_up_to_lmul_8 */ false))
> modes->safe_push (mode);
> i++;
> @@ -5027,26 +5027,27 @@ cmp_lmul_gt_one (machine_mode mode)
> Then we can have the condition for VLS mode in fixed-vlmax, aka:
> PRECISION (VLSmode) < VLEN / (64 / PRECISION(VLS_inner_mode)). */
> bool
> -vls_mode_valid_p (machine_mode vls_mode)
> +vls_mode_valid_p (machine_mode vls_mode, bool allow_up_to_lmul_8)
> {
> if (!TARGET_VECTOR || TARGET_XTHEADVECTOR)
> return false;
> if (rvv_vector_bits == RVV_VECTOR_BITS_SCALABLE)
> {
> - if (GET_MODE_CLASS (vls_mode) != MODE_VECTOR_BOOL
> - && !ordered_p (TARGET_MAX_LMUL * BITS_PER_RISCV_VECTOR,
> - GET_MODE_PRECISION (vls_mode)))
> - /* We enable VLS modes which are aligned with TARGET_MAX_LMUL and
> - BITS_PER_RISCV_VECTOR.
> -
> - e.g. When TARGET_MAX_LMUL = 1 and BITS_PER_RISCV_VECTOR = (128,128).
> - We enable VLS modes have fixed size <= 128bit. Since ordered_p is
> - false between VLA modes with size = (128, 128) bits and VLS mode
> - with size = 128 bits, we will end up with multiple ICEs in
> - middle-end generic codes. */
> - return false;
> - return true;
> + if (GET_MODE_CLASS (vls_mode) != MODE_VECTOR_BOOL)
> + return true;
> + if (allow_up_to_lmul_8)
> + return true;
> + /* We enable VLS modes which are aligned with TARGET_MAX_LMUL and
> + BITS_PER_RISCV_VECTOR.
> +
> + e.g. When TARGET_MAX_LMUL = 1 and BITS_PER_RISCV_VECTOR = (128,128).
> + We enable VLS modes have fixed size <= 128bit. Since ordered_p is
> + false between VLA modes with size = (128, 128) bits and VLS mode
> + with size = 128 bits, we will end up with multiple ICEs in
> + middle-end generic codes. */
> + return !ordered_p (TARGET_MAX_LMUL * BITS_PER_RISCV_VECTOR,
> + GET_MODE_PRECISION (vls_mode));
> }
> if (rvv_vector_bits == RVV_VECTOR_BITS_ZVL)
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/vls-type-rvv-max-lmul.c
> b/gcc/testsuite/gcc.target/riscv/rvv/vls-type-rvv-max-lmul.c
> new file mode 100644
> index 00000000000..5d52f7798d5
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/vls-type-rvv-max-lmul.c
> @@ -0,0 +1,12 @@
> +/* { dg-do compile } */
> +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -mrvv-max-lmul=m1
> -fdump-tree-optimized" } */
> +
> +typedef long long int64x8_t __attribute__((vector_size(64)));
> +
> +int64x8_t foo(int64x8_t a, int64x8_t b)
> +{
> + return a + b;
> +}
> +/* Make sure we can us up to LMUL 4 to process int64x8_t at once rather than
> + break that into 4 LMUL 1 operations. */
> +/* { dg-final { scan-assembler {vsetivli\s+zero,8,e64,m4,t[au],m[au]} } } */
> --
> 2.34.1
>
>
