On 8/12/22 16:48, Robin Dapp wrote: > Hi, > > similar to other backends this patch implements vec_set via > vec_merge and vec_duplicate instead of an unspec. This opens up > more possibilites to combine instructions. > > Bootstrapped and regtested. No regressions. > > Is it OK? > > Regards > Robin > > gcc/ChangeLog: > > * config/s390/s390.md: Implement vec_set with vec_merge and > vec_duplicate. > * config/s390/vector.md: Likewise. > * config/s390/vx-builtins.md: Likewise. > * config/s390/s390.cc (s390_expand_vec_init): Emit new pattern. > (print_operand_address): New output modifier. > (print_operand): New output modifier.
The way you handle the element selector doesn't look right to me. It appears to be an index if it is a CONST_INT and a bitmask otherwise. I don't think it is legal to change operand semantics like this depending on the operand type. This would break e.g. if LRA would decide to load the immediate index in a register. Couldn't you make the shift part of the RTX instead and have the parameter always as an index? Bye, Andreas > --- > > diff --git a/gcc/config/s390/s390.cc b/gcc/config/s390/s390.cc > index c86b26933d7a..ff89fb83360a 100644 > --- a/gcc/config/s390/s390.cc > +++ b/gcc/config/s390/s390.cc > @@ -7073,11 +7073,10 @@ s390_expand_vec_init (rtx target, rtx vals) > if (!general_operand (elem, GET_MODE (elem))) > elem = force_reg (inner_mode, elem); > > - emit_insn (gen_rtx_SET (target, > - gen_rtx_UNSPEC (mode, > - gen_rtvec (3, elem, > - GEN_INT (i), target), > - UNSPEC_VEC_SET))); > + emit_insn > + (gen_rtx_SET > + (target, gen_rtx_VEC_MERGE > + (mode, gen_rtx_VEC_DUPLICATE (mode, elem), target, GEN_INT (1 << > i)))); > } > } > > @@ -8057,6 +8056,8 @@ print_operand_address (FILE *file, rtx addr) > 'S': print S-type memory reference (base+displacement). > 'Y': print address style operand without index (e.g. shift count or > setmem > operand). > + 'P': print address-style operand without index but with the offset as > + if it were specified by a 'p' format flag. > > 'b': print integer X as if it's an unsigned byte. > 'c': print integer X as if it's an signed byte. > @@ -8068,6 +8069,7 @@ print_operand_address (FILE *file, rtx addr) > 'k': print the first nonzero SImode part of X. > 'm': print the first SImode part unequal to -1 of X. > 'o': print integer X as if it's an unsigned 32bit word. > + 'p': print N such that 2^N == X (X must be a power of 2 and const int). > 's': "start" of contiguous bitmask X in either DImode or vector > inner mode. > 't': CONST_INT: "start" of contiguous bitmask X in SImode. > CONST_VECTOR: Generate a bitmask for vgbm instruction. > @@ -8237,6 +8239,16 @@ print_operand (FILE *file, rtx x, int code) > print_shift_count_operand (file, x); > return; > > + case 'P': > + if (CONST_INT_P (x)) > + { > + ival = exact_log2 (INTVAL (x)); > + fprintf (file, HOST_WIDE_INT_PRINT_DEC, ival); > + } > + else > + print_shift_count_operand (file, x); > + return; > + > case 'K': > /* Append @PLT to both local and non-local symbols in order to > support > Linux Kernel livepatching: patches contain individual functions and > @@ -8321,6 +8333,9 @@ print_operand (FILE *file, rtx x, int code) > case 'o': > ival &= 0xffffffff; > break; > + case 'p': > + ival = exact_log2 (INTVAL (x)); > + break; > case 'e': case 'f': > case 's': case 't': > { > diff --git a/gcc/config/s390/s390.md b/gcc/config/s390/s390.md > index f37d8fd33a15..a82db4c624fa 100644 > --- a/gcc/config/s390/s390.md > +++ b/gcc/config/s390/s390.md > @@ -183,7 +183,6 @@ (define_c_enum "unspec" [ > UNSPEC_VEC_GFMSUM_128 > UNSPEC_VEC_GFMSUM_ACCUM > UNSPEC_VEC_GFMSUM_ACCUM_128 > - UNSPEC_VEC_SET > > UNSPEC_VEC_VSUMG > UNSPEC_VEC_VSUMQ > diff --git a/gcc/config/s390/vector.md b/gcc/config/s390/vector.md > index c50451a8326c..bde3a39db3d4 100644 > --- a/gcc/config/s390/vector.md > +++ b/gcc/config/s390/vector.md > @@ -467,12 +467,17 @@ (define_insn "mov<mode>" > ; vec_set is supposed to *modify* an existing vector so operand 0 is > ; duplicated as input operand. > (define_expand "vec_set<mode>" > - [(set (match_operand:V 0 "register_operand" "") > - (unspec:V [(match_operand:<non_vec> 1 "general_operand" "") > - (match_operand:SI 2 "nonmemory_operand" "") > - (match_dup 0)] > - UNSPEC_VEC_SET))] > - "TARGET_VX") > + [(set (match_operand:V 0 "register_operand" "") > + (vec_merge:V > + (vec_duplicate:V > + (match_operand:<non_vec> 1 "general_operand" "")) > + (match_dup 0) > + (match_operand:SI 2 "nonmemory_operand")))] > + "" > +{ > + if (CONST_INT_P (operands[2])) > + operands[2] = GEN_INT (1 << INTVAL (operands[2])); > +}) > > ; FIXME: Support also vector mode operands for 1 > ; FIXME: A target memory operand seems to be useful otherwise we end > @@ -480,28 +485,31 @@ (define_expand "vec_set<mode>" > ; that itself? > ; vlvgb, vlvgh, vlvgf, vlvgg, vleb, vleh, vlef, vleg, vleib, vleih, > vleif, vleig > (define_insn "*vec_set<mode>" > - [(set (match_operand:V 0 "register_operand" "=v,v,v") > - (unspec:V [(match_operand:<non_vec> 1 "general_operand" "d,R,K") > - (match_operand:SI 2 "nonmemory_operand" "an,I,I") > - (match_operand:V 3 "register_operand" "0,0,0")] > - UNSPEC_VEC_SET))] > + [(set (match_operand:V 0 "register_operand" "=v,v,v") > + (vec_merge:V > + (vec_duplicate:V > + (match_operand:<non_vec> 1 "general_operand" "d,R,K")) > + (match_operand:V 3 "register_operand" "0,0,0") > + (match_operand:SI 2 "nonmemory_operand" "an,I,I")))] > "TARGET_VX > && (!CONST_INT_P (operands[2]) > - || UINTVAL (operands[2]) < GET_MODE_NUNITS (<V:MODE>mode))" > + || exact_log2 (UINTVAL (operands[2])) < (GET_MODE_NUNITS > (<V:MODE>mode)))" > "@ > - vlvg<bhfgq>\t%v0,%1,%Y2 > - vle<bhfgq>\t%v0,%1,%2 > - vlei<bhfgq>\t%v0,%1,%2" > + vlvg<bhfgq>\t%v0,%1,%P2 > + vle<bhfgq>\t%v0,%1,%p2 > + vlei<bhfgq>\t%v0,%1,%p2" > [(set_attr "op_type" "VRS,VRX,VRI")]) > > ; vlvgb, vlvgh, vlvgf, vlvgg > (define_insn "*vec_set<mode>_plus" > - [(set (match_operand:V 0 "register_operand" "=v") > - (unspec:V [(match_operand:<non_vec> 1 "general_operand" "d") > - (plus:SI (match_operand:SI 2 "register_operand" "a") > - (match_operand:SI 4 "const_int_operand" "n")) > - (match_operand:V 3 "register_operand" "0")] > - UNSPEC_VEC_SET))] > + [(set (match_operand:V 0 "register_operand" "=v") > + (vec_merge:V > + (vec_duplicate:V > + (match_operand:<non_vec> 1 "general_operand" "d")) > + (match_operand:V 3 "register_operand" "0") > + (plus:SI > + (match_operand:SI 2 "register_operand" "a") > + (match_operand:SI 4 "const_int_operand" "n"))))] > "TARGET_VX" > "vlvg<bhfgq>\t%v0,%1,%Y4(%2)" > [(set_attr "op_type" "VRS")]) > @@ -575,7 +583,7 @@ (define_insn "*vec_splat<mode>" > (match_operand:V_128_NOSINGLE 1 "register_operand" "v") > (parallel > [(match_operand:QI 2 "const_mask_operand" "C")]))))] > - "TARGET_VX && UINTVAL (operands[2]) < GET_MODE_NUNITS (<MODE>mode)" > + "TARGET_VX && exact_log2 (UINTVAL (operands[2])) < GET_MODE_NUNITS > (<MODE>mode)" > "vrep<bhfgq>\t%v0,%v1,%2" > [(set_attr "op_type" "VRI")]) > > @@ -678,13 +686,18 @@ (define_split > (vec_duplicate:V_128_NOSINGLE (match_operand:<non_vec> 1 > "register_operand" "")))] > "TARGET_VX && GENERAL_REG_P (operands[1])" > [(set (match_dup 0) > - (unspec:V_128_NOSINGLE [(match_dup 1) (match_dup 2) (match_dup 0)] > UNSPEC_VEC_SET)) > + (vec_merge:V_128_NOSINGLE > + (vec_duplicate:V_128_NOSINGLE > + (match_dup 1)) > + (match_dup 0) > + (match_dup 2))) > (set (match_dup 0) > (vec_duplicate:V_128_NOSINGLE > (vec_select:<non_vec> > - (match_dup 0) (parallel [(match_dup 2)]))))] > + (match_dup 0) (parallel [(match_dup 3)]))))] > { > - operands[2] = GEN_INT (GET_MODE_NUNITS (<MODE>mode) - 1); > + operands[2] = GEN_INT (1 << (GET_MODE_NUNITS (<MODE>mode) - 1)); > + operands[3] = GEN_INT (GET_MODE_NUNITS (<MODE>mode) - 1); > }) > > (define_predicate "vcond_comparison_operator" > @@ -1136,9 +1149,12 @@ (define_expand "popcountv8hi2_vx" > (set (match_dup 3) (match_dup 2)) > ; Generate the shift count operand in a VR (8->byte 7) > (set (match_dup 4) (match_dup 5)) > - (set (match_dup 4) (unspec:V16QI [(const_int 8) > - (const_int 7) > - (match_dup 4)] UNSPEC_VEC_SET)) > + (set (match_dup 4) > + (vec_merge:V16QI > + (vec_duplicate:V16QI > + (const_int 8)) > + (match_dup 4) > + (const_int 128))) > ; Vector shift right logical by one byte > (set (match_dup 3) > (unspec:V16QI [(match_dup 3) (match_dup 4)] UNSPEC_VEC_SRLB)) > @@ -1339,10 +1355,11 @@ (define_insn "*vec_slb<mode>" > ; this means it is a left shift on BE targets! > (define_expand "vec_shr_<mode>" > [(set (match_dup 3) > - (unspec:V16QI [(match_operand:SI 2 "const_shift_by_byte_operand" "") > - (const_int 7) > - (match_dup 3)] > - UNSPEC_VEC_SET)) > + (vec_merge:V16QI > + (vec_duplicate:V16QI > + (match_operand:SI 2 "const_shift_by_byte_operand" "")) > + (match_dup 3) > + (const_int 128))) > (set (match_operand:V_128 0 "register_operand" "") > (unspec:V_128 [(match_operand:V_128 1 "register_operand" "") > (match_dup 3)] > diff --git a/gcc/config/s390/vx-builtins.md b/gcc/config/s390/vx-builtins.md > index 01dfde438af5..2bdd694112a1 100644 > --- a/gcc/config/s390/vx-builtins.md > +++ b/gcc/config/s390/vx-builtins.md > @@ -150,22 +150,25 @@ (define_expand "vec_splats<mode>" > (vec_duplicate:VEC_HW (match_operand:<non_vec> 1 "general_operand" > "")))] > "TARGET_VX") > > + > (define_expand "vec_insert<mode>" > - [(set (match_operand:VEC_HW 0 "register_operand" "") > - (unspec:VEC_HW [(match_operand:<non_vec> 2 "register_operand" "") > - (match_operand:SI 3 "nonmemory_operand" "") > - (match_operand:VEC_HW 1 "register_operand" "")] > - UNSPEC_VEC_SET))] > + [(set (match_operand:VEC_HW 0 "register_operand" "") > + (vec_merge:VEC_HW > + (vec_duplicate:VEC_HW > + (match_operand:<non_vec> 2 "register_operand" "")) > + (match_operand:VEC_HW 1 "register_operand" "") > + (match_operand:SI 3 "nonmemory_operand" "")))] > "TARGET_VX" > "") > > ; This is vec_set + modulo arithmetic on the element selector (op 2) > (define_expand "vec_promote<mode>" > - [(set (match_operand:VEC_HW 0 "register_operand" "") > - (unspec:VEC_HW [(match_operand:<non_vec> 1 "register_operand" "") > - (match_operand:SI 2 "nonmemory_operand" "") > - (match_dup 0)] > - UNSPEC_VEC_SET))] > + [(set (match_operand:VEC_HW 0 "register_operand" "") > + (vec_merge:VEC_HW > + (vec_duplicate:VEC_HW > + (match_operand:<non_vec> 1 "register_operand" "")) > + (match_dup 0) > + (match_operand:SI 2 "nonmemory_operand" "")))] > "TARGET_VX" > "") > > @@ -457,11 +460,11 @@ (define_insn "vec_scatter_element<V_HW_2:mode>_SI" > [(set (mem:<non_vec> > (plus:SI (subreg:SI > (vec_select:<non_vec_int> > - (match_operand:V_HW_2 1 "register_operand" "v") > - (parallel [(match_operand:QI 3 "const_mask_operand" > "C")])) 4) > - (match_operand:SI 2 "address_operand" "ZQ"))) > + (match_operand:<TOINTVEC> 1 "register_operand" "v") > + (parallel [(match_operand:QI 3 "const_mask_operand" > "C")])) 4) > + (match_operand:SI 2 "address_operand" "ZQ"))) > (vec_select:<non_vec> > - (match_operand:V_HW_2 0 "register_operand" "v") > + (match_operand:V_HW_2 0 "register_operand" "v") > (parallel [(match_dup 3)])))] > "TARGET_VX && !TARGET_64BIT && UINTVAL (operands[3]) < > GET_MODE_NUNITS (<V_HW_2:MODE>mode)" > "vsce<V_HW_2:bhfgq>\t%v0,%O2(%v1,%R2),%3" > @@ -1858,14 +1861,18 @@ (define_expand "vec_ld2f" > ; necessary since all elements of the vector will be set anyway. > ; This is just to make it explicit to the data flow framework. > (set (match_dup 2) (match_dup 3)) > - (set (match_dup 2) (unspec:V4SF [(match_operand:SF 1 > "memory_operand" "") > - (const_int 0) > - (match_dup 2)] > - UNSPEC_VEC_SET)) > - (set (match_dup 2) (unspec:V4SF [(match_dup 4) > - (const_int 2) > - (match_dup 2)] > - UNSPEC_VEC_SET)) > + (set (match_dup 2) > + (vec_merge:V4SF > + (vec_duplicate:V4SF > + (match_operand:SF 1 "memory_operand" "")) > + (match_dup 2) > + (const_int 1))) > + (set (match_dup 2) > + (vec_merge:V4SF > + (vec_duplicate:V4SF > + (match_dup 4)) > + (match_dup 2) > + (const_int 4))) > (set (match_operand:V2DF 0 "register_operand" "") > (unspec:V2DF [(match_dup 2)] UNSPEC_VEC_VFLL))] > "TARGET_VX" > @@ -2303,26 +2310,29 @@ (define_insn_and_split "*eltswap<mode>_emu" > ; vlebrh, vlebrf, vlebrg > (define_insn "*vec_set_bswap_elem<mode>" > [(set (match_operand:V_HW_HSD 0 > "register_operand" "=v") > - (unspec:V_HW_HSD [(bswap:<non_vec> (match_operand:<non_vec> 1 > "memory_operand" "R")) > - (match_operand:SI 2 > "const_int_operand" "C") > - (match_operand:V_HW_HSD 3 > "register_operand" "0")] > - UNSPEC_VEC_SET))] > - "TARGET_VXE2 && UINTVAL (operands[2]) < GET_MODE_NUNITS > (<V_HW_HSD:MODE>mode)" > - "vlebr<bhfgq>\t%v0,%1,%2" > + (vec_merge:V_HW_HSD > + (vec_duplicate:V_HW_HSD > + (bswap:<non_vec> > + (match_operand:<non_vec> 1 "memory_operand" "R"))) > + (match_operand:V_HW_HSD 3 "register_operand" "0") > + (match_operand:SI 2 "const_int_operand" "C")))] > + "TARGET_VXE2 && exact_log2 (UINTVAL (operands[2])) < GET_MODE_NUNITS > (<V_HW_HSD:MODE>mode)" > + "vlebr<bhfgq>\t%v0,%1,%p2" > [(set_attr "op_type" "VRX")]) > > ; vec_revb (vec_insert (*a, vec_revb (b), 1)) set-element-bswap-1.c > ; vlebrh, vlebrf, vlebrg > (define_insn "*vec_set_bswap_vec<mode>" > [(set (match_operand:V_HW_HSD 0 > "register_operand" "=v") > - (bswap:V_HW_HSD > - (unspec:V_HW_HSD [(match_operand:<non_vec> 1 > "memory_operand" "R") > - (match_operand:SI 2 > "const_int_operand" "C") > - (bswap:V_HW_HSD (match_operand:V_HW_HSD 3 > "register_operand" > "0"))] > - UNSPEC_VEC_SET))) > - (use (match_operand:V16QI 4 > "permute_pattern_operand" "X"))] > - "TARGET_VXE2 && UINTVAL (operands[2]) < GET_MODE_NUNITS > (<V_HW_HSD:MODE>mode)" > - "vlebr<bhfgq>\t%v0,%1,%2" > + (bswap:V_HW_HSD > + (vec_merge:V_HW_HSD > + (vec_duplicate:V_HW_HSD > + (match_operand:<non_vec> 1 "memory_operand" "R")) > + (match_operand:V_HW_HSD 3 "register_operand" "0") > + (match_operand:SI 2 "const_int_operand" "C")))) > + (use (match_operand:V16QI 4 "permute_pattern_operand" "X"))] > + "TARGET_VXE2 && exact_log2 (UINTVAL (operands[2])) < GET_MODE_NUNITS > (<V_HW_HSD:MODE>mode)" > + "vlebr<bhfgq>\t%v0,%1,%p2" > [(set_attr "op_type" "VRX")]) > > ; *a = vec_extract (vec_revb (b), 1); get-element-bswap-3.c