[gcc(refs/users/meissner/heads/work255-sha)] PR target/117251: Improve vector fusion #1
Michael Meissner via Gcc-cvs <[email protected]> Mon, 3 Aug 2026 19:53:31 +0000 (GMT)
| Newsgroups | gmane.comp.gcc.cvs |
|---|---|
| Message-ID | <[email protected]> |
https://gcc.gnu.org/g:576fec5571ee9be4cb101a869eb8b57279be1d65 commit 576fec5571ee9be4cb101a869eb8b57279be1d65 Author: Michael Meissner <[email protected]> Date: Mon Aug 3 15:49:27 2026 -0400 PR target/117251: Improve vector fusion #1 See the following post for a complete explanation of what the patches for PR target/117251: * https://gcc.gnu.org/pipermail/gcc-patches/2025-June/686474.html This is patch #1 of 45 to generate the 'XXEVAL' instruction on power10 and power11 instead of using the Altivec 'VAND' instruction feeding into 'VAND'. The 'XXEVAL' instruction can use all 64 vector registers, instead of the 32 registers that traditional Altivec vector instructions use. By allowing all of the vector registers to be used, it reduces the amount of spilling that a large benchmark generated. Currently the following code: vector int a, b, c, d; a = (c & d) & b; Generates: vand t,c,d vand a,t,b Now in addition with this patch, if the arguments or result is allocated to a traditional FPR register, the GCC compiler will now generate the following code instead of adding vector move instructions: xxeval a,b,c,1 Since fusion using 2 Altivec instructions is slightly faster than using the 'XXEVAL' instruction we prefer to generate the Altivec instructions if we can. In addition, because 'XXEVAL' is a prefixed instruction, it possibly might generate an extra NOP instruction to align the 'XXEVAL' instruction. I have tested these patches on both big endian and little endian PowerPC servers, with no regressions. Can I check these patches into the trunk? I have committed all of the patches in my backlog (dense math registers, other -mcpu=future instructions, random bug fixes, support for _Float16 and __bfloat16, and optimizations for vector logical operations on power10/power11) into the IBM vendor branch: vendors/ibm/gcc-17-future 2026-08-03 Michael Meissner <[email protected]> gcc/ PR target/117251 * config/rs6000/fusion.md: Regenerate. * config/rs6000/genfusion.pl (gen_logical_addsubf): Add support to generate vector/vector and/and fusion if XXEVAL is supported. * config/rs6000/predicates.md (vector_fusion_operand): New predicate. * config/rs6000/rs6000.md (isa attribute): Add p10p. (enabled attribute): Add support for XXEVAL support. * config/rs6000/rs6000.h (TARGET_XXEVAL): New macro. Diff: --- gcc/config/rs6000/fusion.md | 15 ++++++----- gcc/config/rs6000/genfusion.pl | 58 ++++++++++++++++++++++++++++++++++++++--- gcc/config/rs6000/predicates.md | 12 +++++++++ gcc/config/rs6000/rs6000.h | 4 +++ gcc/config/rs6000/rs6000.md | 6 ++++- 5 files changed, 84 insertions(+), 11 deletions(-) diff --git a/gcc/config/rs6000/fusion.md b/gcc/config/rs6000/fusion.md index 416017149d6f..68c41114d131 100644 --- a/gcc/config/rs6000/fusion.md +++ b/gcc/config/rs6000/fusion.md @@ -1871,20 +1871,23 @@ ;; logical-logical fusion pattern generated by gen_logical_addsubf ;; vector vand -> vand (define_insn "*fuse_vand_vand" - [(set (match_operand:VM 3 "altivec_register_operand" "=&0,&1,&v,v") - (and:VM (and:VM (match_operand:VM 0 "altivec_register_operand" "v,v,v,v") - (match_operand:VM 1 "altivec_register_operand" "%v,v,v,v")) - (match_operand:VM 2 "altivec_register_operand" "v,v,v,v"))) - (clobber (match_scratch:VM 4 "=X,X,X,&v"))] + [(set (match_operand:VM 3 "vector_fusion_operand" "=&0,&1,&v,wa,v") + (and:VM (and:VM (match_operand:VM 0 "vector_fusion_operand" "v,v,v,wa,v") + (match_operand:VM 1 "vector_fusion_operand" "%v,v,v,wa,v")) + (match_operand:VM 2 "vector_fusion_operand" "v,v,v,wa,v"))) + (clobber (match_scratch:VM 4 "=X,X,X,X,&v"))] "(TARGET_P10_FUSION)" "@ vand %3,%1,%0\;vand %3,%3,%2 vand %3,%1,%0\;vand %3,%3,%2 vand %3,%1,%0\;vand %3,%3,%2 + xxeval %x3,%x2,%x1,%x0,1 vand %4,%1,%0\;vand %3,%4,%2" [(set_attr "type" "fused_vector") (set_attr "cost" "6") - (set_attr "length" "8")]) + (set_attr "length" "8") + (set_attr "prefixed" "*,*,*,yes,*") + (set_attr "isa" "*,*,*,p10p,*")]) ;; logical-logical fusion pattern generated by gen_logical_addsubf ;; vector vandc -> vand diff --git a/gcc/config/rs6000/genfusion.pl b/gcc/config/rs6000/genfusion.pl index 04e6ab15da6e..3cebfd6d9549 100755 --- a/gcc/config/rs6000/genfusion.pl +++ b/gcc/config/rs6000/genfusion.pl @@ -211,25 +211,33 @@ sub gen_logical_addsubf $inner_comp, $inner_inv, $inner_rtl, $inner_op, $both_commute, $c4, $bc, $inner_arg0, $inner_arg1, $inner_exp, $outer_arg2, $outer_exp, $ftype, $insn, $is_subf, $is_rsubf, $outer_32, $outer_42,$outer_name, - $fuse_type); - KIND: foreach $kind ('scalar','vector') { + $fuse_type, $xxeval, $c5, $vect_pred, $vect_inner_arg0, $vect_inner_arg1, + $vect_inner_exp, $vect_outer_arg2, $vect_outer_exp); + + my %xxeval_fusions = ( + "vand_vand" => 1, + ); + + KIND: foreach $kind ('scalar','vector') { @outer_ops = @logicals; if ( $kind eq 'vector' ) { $vchr = "v"; $mode = "VM"; $pred = "altivec_register_operand"; + $vect_pred = "vector_fusion_operand"; $constraint = "v"; $fuse_type = "fused_vector"; } else { $vchr = ""; $mode = "GPR"; - $pred = "gpc_reg_operand"; + $vect_pred = $pred = "gpc_reg_operand"; $constraint = "r"; $fuse_type = "fused_arith_logical"; push (@outer_ops, @addsub); push (@outer_ops, ( "rsubf" )); } $c4 = "${constraint},${constraint},${constraint},${constraint}"; + $c5 = "${constraint},${constraint},${constraint},wa,${constraint}"; OUTER: foreach $outer ( @outer_ops ) { $outer_name = "${vchr}${outer}"; $is_subf = ( $outer eq "subf" ); @@ -263,23 +271,33 @@ sub gen_logical_addsubf $bc = ""; if ( $both_commute ) { $bc = "%"; } $inner_arg0 = "(match_operand:${mode} 0 \"${pred}\" \"${c4}\")"; $inner_arg1 = "(match_operand:${mode} 1 \"${pred}\" \"${bc}${c4}\")"; + $vect_inner_arg0 = "(match_operand:${mode} 0 \"${vect_pred}\" \"${c5}\")"; + $vect_inner_arg1 = "(match_operand:${mode} 1 \"${vect_pred}\" \"${bc}${c5}\")"; if ( ($inner_comp & 1) == 1 ) { $inner_arg0 = "(not:${mode} $inner_arg0)"; + $vect_inner_arg0 = "(not:${mode} $vect_inner_arg0)"; } if ( ($inner_comp & 2) == 2 ) { $inner_arg1 = "(not:${mode} $inner_arg1)"; + $vect_inner_arg1 = "(not:${mode} $vect_inner_arg1)"; } $inner_exp = "(${inner_rtl}:${mode} ${inner_arg0} ${inner_arg1})"; + $vect_inner_exp = "(${inner_rtl}:${mode} ${vect_inner_arg0} + ${vect_inner_arg1})"; if ( $inner_inv == 1 ) { $inner_exp = "(not:${mode} $inner_exp)"; + $vect_inner_exp = "(not:${mode} $vect_inner_exp)"; } $outer_arg2 = "(match_operand:${mode} 2 \"${pred}\" \"${c4}\")"; + $vect_outer_arg2 = "(match_operand:${mode} 2 \"${vect_pred}\" \"${c5}\")"; if ( ($outer_comp & 1) == 1 ) { $outer_arg2 = "(not:${mode} $outer_arg2)"; + $vect_outer_arg2 = "(not:${mode} $vect_outer_arg2)"; } if ( ($outer_comp & 2) == 2 ) { $inner_exp = "(not:${mode} $inner_exp)"; + $vect_inner_exp = "(not:${mode} $vect_inner_exp)"; } if ( $is_subf ) { $outer_32 = "%2,%3"; @@ -291,15 +309,23 @@ sub gen_logical_addsubf if ( $is_rsubf == 1 ) { $outer_exp = "(${outer_rtl}:${mode} ${outer_arg2} ${inner_exp})"; + $vect_outer_exp = "(${outer_rtl}:${mode} ${vect_outer_arg2} + ${vect_inner_exp})"; } else { $outer_exp = "(${outer_rtl}:${mode} ${inner_exp} ${outer_arg2})"; + $vect_outer_exp = "(${outer_rtl}:${mode} ${vect_inner_exp} + ${vect_outer_arg2})"; } if ( $outer_inv == 1 ) { $outer_exp = "(not:${mode} $outer_exp)"; + $vect_outer_exp = "(not:${mode} $vect_outer_exp)"; } - $insn = <<"EOF"; + # See if we can use xxeval on vector fusion + $xxeval = $xxeval_fusions{"${inner_op}_${outer_name}"}; + if (!$xxeval) { + $insn = <<"EOF"; ;; $ftype fusion pattern generated by gen_logical_addsubf ;; $kind $inner_op -> $outer_name @@ -318,6 +344,30 @@ sub gen_logical_addsubf (set_attr "length" "8")]) EOF + } else { + $insn = <<"EOF"; + +;; $ftype fusion pattern generated by gen_logical_addsubf +;; $kind $inner_op -> $outer_name +(define_insn "*fuse_${inner_op}_${outer_name}" + [(set (match_operand:${mode} 3 "${vect_pred}" "=&0,&1,&${constraint},wa,${constraint}") + ${vect_outer_exp}) + (clobber (match_scratch:${mode} 4 "=X,X,X,X,&${constraint}"))] + "(TARGET_P10_FUSION)" + "@ + ${inner_op} %3,%1,%0\\;${outer_op} %3,${outer_32} + ${inner_op} %3,%1,%0\\;${outer_op} %3,${outer_32} + ${inner_op} %3,%1,%0\\;${outer_op} %3,${outer_32} + xxeval %x3,%x2,%x1,%x0,${xxeval} + ${inner_op} %4,%1,%0\\;${outer_op} %3,${outer_42}" + [(set_attr "type" "$fuse_type") + (set_attr "cost" "6") + (set_attr "length" "8") + (set_attr "prefixed" "*,*,*,yes,*") + (set_attr "isa" "*,*,*,p10p,*")]) +EOF + } + print $insn; } } diff --git a/gcc/config/rs6000/predicates.md b/gcc/config/rs6000/predicates.md index a6ec04f5922d..467a9b5e9ece 100644 --- a/gcc/config/rs6000/predicates.md +++ b/gcc/config/rs6000/predicates.md @@ -119,6 +119,18 @@ return VSX_REGNO_P (REGNO (op)); }) +;; Return 1 if op is a register that can be used for vector fusion. If XXEVAL +;; is supported, return true for all VSX registers, otherwise the fusion is +;; limited to Altivec registers since the machine only fuses Altivec +;; operations. +(define_predicate "vector_fusion_operand" + (match_operand 0 "register_operand") +{ + return (TARGET_XXEVAL + ? vsx_register_operand (op, mode) + : altivec_register_operand (op, mode)); +}) + ;; Return 1 if op is a vector register that operates on floating point vectors ;; (either altivec or VSX). (define_predicate "vfloat_operand" diff --git a/gcc/config/rs6000/rs6000.h b/gcc/config/rs6000/rs6000.h index bcd4c39d24dc..3eb27be05a56 100644 --- a/gcc/config/rs6000/rs6000.h +++ b/gcc/config/rs6000/rs6000.h @@ -500,6 +500,10 @@ extern int rs6000_vector_align[]; #define TARGET_MINMAX (TARGET_HARD_FLOAT && TARGET_PPC_GFXOPT \ && (TARGET_P9_MINMAX || !flag_trapping_math)) +/* Enable XXEVAL support if we support prefixed instructions and at least + power10. */ +#define TARGET_XXEVAL (TARGET_POWER10 && TARGET_PREFIXED) + /* In switching from using target_flags to using rs6000_isa_flags, the options machinery creates OPTION_MASK_<xxx> instead of MASK_<xxx>. The MASK_<xxxx> options that have not yet been replaced by their OPTION_MASK_<xxx> diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md index e06dcec65889..1d2fcc3c32f8 100644 --- a/gcc/config/rs6000/rs6000.md +++ b/gcc/config/rs6000/rs6000.md @@ -371,7 +371,7 @@ (const (symbol_ref "(enum attr_cpu) rs6000_tune"))) ;; The ISA we implement. -(define_attr "isa" "any,p5,p6,p7,p7v,p8,p8v,p9,p9v,p9kf,p9tf,p10,future,mma,dmf" +(define_attr "isa" "any,p5,p6,p7,p7v,p8,p8v,p9,p9v,p9kf,p9tf,p10,p10p,future,mma,dmf" (const_string "any")) ;; Is this alternative enabled for the current CPU/ISA/etc.? @@ -424,6 +424,10 @@ (match_test "TARGET_POWER10")) (const_int 1) + (and (eq_attr "isa" "p10p") + (match_test "TARGET_POWER10 && TARGET_PREFIXED")) + (const_int 1) + (and (eq_attr "isa" "future") (match_test "TARGET_FUTURE")) (const_int 1)