>> Ah. So then just feed it cond_fn? I mean, we don't have >> LEN_FMA, the only LEN-but-not-MASK ifns are those used by >> power/s390, LEN_LOAD and LEN_STORE? Yes, that's why I feed cond_fn with get_len_internal_fn (cond_fn) >> Yes, but all of this depends on what the original ifn is, no? Yes. >> reduc_idx < 0 means this stmt isn't part of a reduction. So yes, >> you can vectorize FMA as COND_LEN_FMA with dummy mask and len if you >> don't have FMA expanders? Could you give me an example that reduction >= 0 when vectorizing FMA into COND_LEN_FMA? Actually, I failed to produce such circumstance in this patch: https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625697.html I only fully tested vectorizing COND_* into COND_LEN_* but I failed to produce the case that: FMA ---> COND_LEN_FMA. juzhe.zhong@rivai.ai From: Richard Biener Date: 2023-07-31 18:45 To: juzhe.zhong@rivai.ai CC: gcc-patches; richard.sandiford Subject: Re: Re: [PATCH V2] VECT: Support CALL vectorization for COND_LEN_* On Mon, 31 Jul 2023, juzhe.zhong@rivai.ai wrote: > Hi, Richard. Thanks a lot for the comment > > >> In your code above > >> you either use cond_len_fn or get_len_internal_fn (cond_fn) but > >> isn't that the very same?! So how come you in one case add two > >> and in the other add four args? > > cond_len_fn is not the same as get_len_internal_fn (cond_fn) when vectorizing FMA into COND_LEN_FMA. > > since "internal_fn cond_len_fn = get_len_internal_fn (ifn);" > > and the iterators: > > +#define FOR_EACH_LEN_FN_PAIR(T) \ > > + T (MASK_LOAD, MASK_LEN_LOAD) \ > > + T (MASK_STORE, MASK_LEN_STORE) \ > > + T (MASK_GATHER_LOAD, MASK_LEN_GATHER_LOAD) \ > > + T (MASK_SCATTER_STORE, MASK_LEN_SCATTER_STORE) \ > > + T (COND_ADD, COND_LEN_ADD) \ > > + T (COND_SUB, COND_LEN_SUB) \ > > + T (COND_MUL, COND_LEN_MUL) \ > > + T (COND_DIV, COND_LEN_DIV) \ > > + T (COND_MOD, COND_LEN_MOD) \ > > + T (COND_RDIV, COND_LEN_RDIV) \ > > + T (COND_FMIN, COND_LEN_FMIN) \ > > + T (COND_FMAX, COND_LEN_FMAX) \ > > + T (COND_MIN, COND_LEN_MIN) \ > > + T (COND_MAX, COND_LEN_MAX) \ > > + T (COND_AND, COND_LEN_AND) \ > > + T (COND_IOR, COND_LEN_IOR) \ > > + T (COND_XOR, COND_LEN_XOR) \ > > + T (COND_SHL, COND_LEN_SHL) \ > > + T (COND_SHR, COND_LEN_SHR) \ > > + T (COND_NEG, COND_LEN_NEG) \ > > + T (COND_FMA, COND_LEN_FMA) \ > > + T (COND_FMS, COND_LEN_FMS) \ > > + T (COND_FNMA, COND_LEN_FNMA) \ > > + T (COND_FNMS, COND_LEN_FNMS) > > So, cond_len_fn will be IFN_LAST when ifn = FMA. Ah. So then just feed it cond_fn? I mean, we don't have LEN_FMA, the only LEN-but-not-MASK ifns are those used by power/s390, LEN_LOAD and LEN_STORE? > Maybe is it reasonable that I add 4 more iterators here? > > + T (FMA, COND_LEN_FMA) \ > > + T (FMS, COND_LEN_FMS) \ > > + T (FNMA, COND_LEN_FNMA) \ > > + T (FNMS, COND_LEN_FNMS) > > So that we won't need to have "get_len_internal_fn (cond_fn)" No, as said we don't have LEN_FMA. > When vectorizing COND_ADD into COND_LEN_ADD we already have "mask" and "else" value. > So we only need to add 2 arguments. > > But when vectorizing FMA into COND_LEN_FMA, we need to add 4 arguments (mask,else,len,bias). Yes, but all of this depends on what the original ifn is, no? > >>as said, > >>no idea why we special case reduc_idx >= 0 at the moment > > I also want to vectorize FMA into COND_LEN_FMA even reduc_idx < 0. > Could I relax this condition for COND_LEN_* since it will improve RVV codegen a lot. reduc_idx < 0 means this stmt isn't part of a reduction. So yes, you can vectorize FMA as COND_LEN_FMA with dummy mask and len if you don't have FMA expanders? Richard. > Thanks. > > > juzhe.zhong@rivai.ai > > From: Richard Biener > Date: 2023-07-31 17:26 > To: Juzhe-Zhong > CC: gcc-patches; richard.sandiford > Subject: Re: [PATCH V2] VECT: Support CALL vectorization for COND_LEN_* > On Fri, 28 Jul 2023, Juzhe-Zhong wrote: > > > Hi, Richard and Richi. > > > > Base on the suggestions from Richard: > > https://gcc.gnu.org/pipermail/gcc-patches/2023-July/625396.html > > > > This patch choose (1) approach that Richard provided, meaning: > > > > RVV implements cond_* optabs as expanders. RVV therefore supports > > both IFN_COND_ADD and IFN_COND_LEN_ADD. No dummy length arguments > > are needed at the gimple level. > > > > Such approach can make codes much cleaner and reasonable. > > > > Consider this following case: > > void foo (float * __restrict a, float * __restrict b, int * __restrict cond, int n) > > { > > for (int i = 0; i < n; i++) > > if (cond[i]) > > a[i] = b[i] + a[i]; > > } > > > > > > Output of RISC-V (32-bits) gcc (trunk) (Compiler #3) > > :5:21: missed: couldn't vectorize loop > > :5:21: missed: not vectorized: control flow in loop. > > > > ARM SVE: > > > > ... > > mask__27.10_51 = vect__4.9_49 != { 0, ... }; > > ... > > vec_mask_and_55 = loop_mask_49 & mask__27.10_51; > > ... > > vect__9.17_62 = .COND_ADD (vec_mask_and_55, vect__6.13_56, vect__8.16_60, vect__6.13_56); > > > > For RVV, we want IR as follows: > > > > ... > > _68 = .SELECT_VL (ivtmp_66, POLY_INT_CST [4, 4]); > > ... > > mask__27.10_51 = vect__4.9_49 != { 0, ... }; > > ... > > vect__9.17_60 = .COND_LEN_ADD (mask__27.10_51, vect__6.13_55, vect__8.16_59, vect__6.13_55, _68, 0); > > ... > > > > Both len and mask of COND_LEN_ADD are real not dummy. > > > > This patch has been fully tested in RISC-V port with supporting both COND_* and COND_LEN_*. > > > > And also, Bootstrap and Regression on X86 passed. > > > > OK for trunk? > > > > gcc/ChangeLog: > > > > * internal-fn.cc (FOR_EACH_LEN_FN_PAIR): New macro. > > (get_len_internal_fn): New function. > > (CASE): Ditto. > > * internal-fn.h (get_len_internal_fn): Ditto. > > * tree-vect-stmts.cc (vectorizable_call): Support CALL vectorization with COND_LEN_*. > > > > --- > > gcc/internal-fn.cc | 46 ++++++++++++++++++++++ > > gcc/internal-fn.h | 1 + > > gcc/tree-vect-stmts.cc | 87 +++++++++++++++++++++++++++++++++++++----- > > 3 files changed, 125 insertions(+), 9 deletions(-) > > > > diff --git a/gcc/internal-fn.cc b/gcc/internal-fn.cc > > index 8e294286388..379220bebc7 100644 > > --- a/gcc/internal-fn.cc > > +++ b/gcc/internal-fn.cc > > @@ -4443,6 +4443,52 @@ get_conditional_internal_fn (internal_fn fn) > > } > > } > > > > +/* Invoke T(IFN) for each internal function IFN that also has an > > + IFN_COND_LEN_* or IFN_MASK_LEN_* form. */ > > +#define FOR_EACH_LEN_FN_PAIR(T) \ > > + T (MASK_LOAD, MASK_LEN_LOAD) \ > > + T (MASK_STORE, MASK_LEN_STORE) \ > > + T (MASK_GATHER_LOAD, MASK_LEN_GATHER_LOAD) \ > > + T (MASK_SCATTER_STORE, MASK_LEN_SCATTER_STORE) \ > > + T (COND_ADD, COND_LEN_ADD) \ > > + T (COND_SUB, COND_LEN_SUB) \ > > + T (COND_MUL, COND_LEN_MUL) \ > > + T (COND_DIV, COND_LEN_DIV) \ > > + T (COND_MOD, COND_LEN_MOD) \ > > + T (COND_RDIV, COND_LEN_RDIV) \ > > + T (COND_FMIN, COND_LEN_FMIN) \ > > + T (COND_FMAX, COND_LEN_FMAX) \ > > + T (COND_MIN, COND_LEN_MIN) \ > > + T (COND_MAX, COND_LEN_MAX) \ > > + T (COND_AND, COND_LEN_AND) \ > > + T (COND_IOR, COND_LEN_IOR) \ > > + T (COND_XOR, COND_LEN_XOR) \ > > + T (COND_SHL, COND_LEN_SHL) \ > > + T (COND_SHR, COND_LEN_SHR) \ > > + T (COND_NEG, COND_LEN_NEG) \ > > + T (COND_FMA, COND_LEN_FMA) \ > > + T (COND_FMS, COND_LEN_FMS) \ > > + T (COND_FNMA, COND_LEN_FNMA) \ > > + T (COND_FNMS, COND_LEN_FNMS) > > + > > +/* If there exists an internal function like IFN that operates on vectors, > > + but with additional length and bias parameters, return the internal_fn > > + for that function, otherwise return IFN_LAST. */ > > +internal_fn > > +get_len_internal_fn (internal_fn fn) > > +{ > > + switch (fn) > > + { > > +#define CASE(NAME, LEN_NAME) \ > > + case IFN_##NAME: \ > > + return IFN_##LEN_NAME; > > + FOR_EACH_LEN_FN_PAIR (CASE) > > +#undef CASE > > + default: > > + return IFN_LAST; > > + } > > +} > > + > > /* If IFN implements the conditional form of an unconditional internal > > function, return that unconditional function, otherwise return IFN_LAST. */ > > > > diff --git a/gcc/internal-fn.h b/gcc/internal-fn.h > > index a5c3f4765ff..410c1b623d6 100644 > > --- a/gcc/internal-fn.h > > +++ b/gcc/internal-fn.h > > @@ -224,6 +224,7 @@ extern bool set_edom_supported_p (void); > > > > extern internal_fn get_conditional_internal_fn (tree_code); > > extern internal_fn get_conditional_internal_fn (internal_fn); > > +extern internal_fn get_len_internal_fn (internal_fn); > > extern internal_fn get_conditional_len_internal_fn (tree_code); > > extern tree_code conditional_internal_fn_code (internal_fn); > > extern internal_fn get_unconditional_internal_fn (internal_fn); > > diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc > > index 6a4e8fce126..ae5b0b09c08 100644 > > --- a/gcc/tree-vect-stmts.cc > > +++ b/gcc/tree-vect-stmts.cc > > @@ -3540,7 +3540,10 @@ vectorizable_call (vec_info *vinfo, > > > > int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info); > > internal_fn cond_fn = get_conditional_internal_fn (ifn); > > + internal_fn cond_len_fn = get_len_internal_fn (ifn); > > + int len_opno = internal_fn_len_index (cond_len_fn); > > vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL); > > + vec_loop_lens *lens = (loop_vinfo ? &LOOP_VINFO_LENS (loop_vinfo) : NULL); > > if (!vec_stmt) /* transformation not required. */ > > { > > if (slp_node) > > @@ -3586,8 +3589,14 @@ vectorizable_call (vec_info *vinfo, > > Above for reduc_idx >= 0 there's a check whether cond_fn is supported, > don't you need to amend that with a check for cond_len_fn? > > > tree scalar_mask = NULL_TREE; > > if (mask_opno >= 0) > > scalar_mask = gimple_call_arg (stmt_info->stmt, mask_opno); > > - vect_record_loop_mask (loop_vinfo, masks, nvectors, > > - vectype_out, scalar_mask); > > + if (cond_len_fn != IFN_LAST > > + && direct_internal_fn_supported_p (cond_len_fn, vectype_out, > > + OPTIMIZE_FOR_SPEED)) > > + vect_record_loop_len (loop_vinfo, lens, nvectors, vectype_out, > > + 1); > > + else > > + vect_record_loop_mask (loop_vinfo, masks, nvectors, vectype_out, > > + scalar_mask); > > } > > } > > return true; > > @@ -3603,8 +3612,24 @@ vectorizable_call (vec_info *vinfo, > > vec_dest = vect_create_destination_var (scalar_dest, vectype_out); > > > > bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo); > > + bool len_loop_p = loop_vinfo && LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo); > > unsigned int vect_nargs = nargs; > > - if (masked_loop_p && reduc_idx >= 0) > > + if (len_loop_p) > > + { > > + if (len_opno >= 0) > > + { > > + ifn = cond_len_fn; > > + /* COND_* -> COND_LEN_* takes 2 extra arguments:LEN,BIAS. */ > > + vect_nargs += 2; > > + } > > + else if (reduc_idx >= 0) > > + { > > + /* FMA -> COND_LEN_FMA takes 4 extra arguments:MASK,ELSE,LEN,BIAS. */ > > + ifn = get_len_internal_fn (cond_fn); > > + vect_nargs += 4; > > I'm a bit confused (but also by the existing mask code), whether > vect_nargs needs adjustment depends on the IFN in the IL we analyze. > If if-conversion recognizes a .COND_ADD then we need to add nothing > for masking (that is, ifn == cond_fn already). In your code above > you either use cond_len_fn or get_len_internal_fn (cond_fn) but > isn't that the very same?! So how come you in one case add two > and in the other add four args? > > Please make sure to place gcc_unreachable () in each arm and check > you have test coverage. I believe that the else arm is unreachable > but when you vectorize .FMA you will need to add 4 and when you > vectorize .COND_FMA you will need to add two arguments (as said, > no idea why we special case reduc_idx >= 0 at the moment). > > Otherwise the patch looks OK to me. > > Thanks, > Richard. > > > + } > > + } > > + else if (masked_loop_p && reduc_idx >= 0) > > { > > ifn = cond_fn; > > vect_nargs += 2; > > @@ -3629,7 +3654,18 @@ vectorizable_call (vec_info *vinfo, > > FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0) > > { > > int varg = 0; > > - if (masked_loop_p && reduc_idx >= 0) > > + if (len_loop_p && reduc_idx >= 0) > > + { > > + /* Always true for SLP. */ > > + gcc_assert (ncopies == 1); > > + /* For COND_LEN_* operations used by reduction of > > + CALL vectorization, the LEN argument is the real > > + loop len produced by SELECT_VL or MIN wheras the > > + MASK argument here is the dummy mask. */ > > + vargs[varg++] > > + = build_minus_one_cst (truth_type_for (vectype_out)); > > + } > > + else if (masked_loop_p && reduc_idx >= 0) > > { > > unsigned int vec_num = vec_oprnds0.length (); > > /* Always true for SLP. */ > > @@ -3644,7 +3680,7 @@ vectorizable_call (vec_info *vinfo, > > vec vec_oprndsk = vec_defs[k]; > > vargs[varg++] = vec_oprndsk[i]; > > } > > - if (masked_loop_p && reduc_idx >= 0) > > + if ((masked_loop_p || len_loop_p) && reduc_idx >= 0) > > vargs[varg++] = vargs[reduc_idx + 1]; > > gimple *new_stmt; > > if (modifier == NARROW) > > @@ -3671,7 +3707,21 @@ vectorizable_call (vec_info *vinfo, > > } > > else > > { > > - if (mask_opno >= 0 && masked_loop_p) > > + if (len_opno >= 0 && len_loop_p) > > + { > > + unsigned int vec_num = vec_oprnds0.length (); > > + /* Always true for SLP. */ > > + gcc_assert (ncopies == 1); > > + tree len > > + = vect_get_loop_len (loop_vinfo, gsi, lens, vec_num, > > + vectype_out, i, 1); > > + signed char biasval > > + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); > > + tree bias = build_int_cst (intQI_type_node, biasval); > > + vargs[len_opno] = len; > > + vargs[len_opno + 1] = bias; > > + } > > + else if (mask_opno >= 0 && masked_loop_p) > > { > > unsigned int vec_num = vec_oprnds0.length (); > > /* Always true for SLP. */ > > @@ -3701,7 +3751,16 @@ vectorizable_call (vec_info *vinfo, > > } > > > > int varg = 0; > > - if (masked_loop_p && reduc_idx >= 0) > > + if (len_loop_p && reduc_idx >= 0) > > + { > > + /* For COND_LEN_* operations used by reduction of > > + CALL vectorization, the LEN argument is the real > > + loop len produced by SELECT_VL or MIN wheras the > > + MASK argument here is the dummy mask. */ > > + vargs[varg++] > > + = build_minus_one_cst (truth_type_for (vectype_out)); > > + } > > + else if (masked_loop_p && reduc_idx >= 0) > > vargs[varg++] = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, > > vectype_out, j); > > for (i = 0; i < nargs; i++) > > @@ -3716,10 +3775,20 @@ vectorizable_call (vec_info *vinfo, > > } > > vargs[varg++] = vec_defs[i][j]; > > } > > - if (masked_loop_p && reduc_idx >= 0) > > + if ((masked_loop_p || len_loop_p) && reduc_idx >= 0) > > vargs[varg++] = vargs[reduc_idx + 1]; > > > > - if (mask_opno >= 0 && masked_loop_p) > > + if (len_opno >= 0 && len_loop_p) > > + { > > + tree len = vect_get_loop_len (loop_vinfo, gsi, lens, ncopies, > > + vectype_out, j, 1); > > + signed char biasval > > + = LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS (loop_vinfo); > > + tree bias = build_int_cst (intQI_type_node, biasval); > > + vargs[len_opno] = len; > > + vargs[len_opno + 1] = bias; > > + } > > + else if (mask_opno >= 0 && masked_loop_p) > > { > > tree mask = vect_get_loop_mask (loop_vinfo, gsi, masks, ncopies, > > vectype_out, j); > > > > -- Richard Biener SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg, Germany; GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)