Hello, On 15 Aug 20:35, Uros Bizjak wrote: > On Fri, Aug 15, 2014 at 1:56 PM, Kirill Yukhin <kirill.yuk...@gmail.com> > wrote: > Again, please split insn pattern to avoid: > > + "TARGET_SSE2 > + && <mask_mode512bit_condition> > + && ((<MODE>mode != V16HImode && <MODE>mode != V8HImode) > + || TARGET_AVX512BW > + || !<mask_applied>)" > > insn constraints. The insn constraint should use baseline TARGET_* and > mode iterator should use TARGET_* that results in "baseline TARGET_ && > iterator TARGET_" for certain mode. If these are properly used, then > there is no need to use <MODE>mode checks in the insn constraint. I've refactored the pattern. But it should be mentioned, that it still uses <MODE>mode checkes implicitly. Problem is that masking allowed either for ZMM or (XMM|YMM)+TARGET_AVX512VL, supposing that TARGET_AVX512F is on for both cases. That is what "mask_mode512bit_condition" check: (define_subst_attr "mask_mode512bit_condition" "mask" "1" "(<MODE_SIZE> == 64 || TARGET_AVX512VL)")
So, this pattern: (define_insn "<shift_insn><mode>3<mask_name>" [(set (match_operand:VI2_AVX2_AVX512BW 0 "register_operand" "=x,v") (any_lshift:VI2_AVX2_AVX512BW (match_operand:VI2_AVX2_AVX512BW 1 "register_operand" "0,v") (match_operand:SI 2 "nonmemory_operand" "xN,vN")))] "TARGET_SSE2 && <mask_mode512bit_condition> && <mask_avx512bw_condition>" is expanded to (16hi, mask, ashift): (define_insn ("ashlv16hi3_mask") [ (set (match_operand:V16HI 0 ("register_operand") ("=x,v")) (vec_merge:V16HI (ashift:V16HI (match_operand:V16HI 1 ("register_operand") ("0,v")) (match_operand:SI 2 ("nonmemory_operand") ("xN,vN"))) (match_operand:V16HI 3 ("vector_move_operand") ("0C,0C")) (match_operand:HI 4 ("register_operand") ("Yk,Yk")))) ] ("(TARGET_AVX512F) && ((TARGET_SSE2 && (32 == 64 || TARGET_AVX512VL) && TARGET_AVX512BW) && (TARGET_AVX2))") (TARGET_AVX512F) came from `mask' subst, (32 == 64 || TARGET_AVX512VL) is what I've described above, (TARGET_AVX512BW) came from coresponding subst attr and TARGET_AVX2 belongs to mode iterator. Bootstrapped and avx512-regtested. gcc/ * config/i386/sse.md (define_mode_iterator VI248_AVX2): Delete. (define_mode_iterator VI2_AVX2_AVX512BW): New. (define_mode_iterator VI48_AVX2): Ditto. (define_insn ""<shift_insn><mode>3<mask_name>"): Add masking. Split into two similar patterns, which use different mode iterators: VI2_AVX2_AVX512BW and VI48_AVX2. * config/i386/subst.md (define_subst_attr "mask_avx512bw_condition"): New. -- Thanks, K diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md index 6a5faee..5b1e5c1 100644 --- a/gcc/config/i386/sse.md +++ b/gcc/config/i386/sse.md @@ -347,9 +347,11 @@ (V16HI "TARGET_AVX2") V8HI (V8SI "TARGET_AVX2") V4SI]) -(define_mode_iterator VI248_AVX2 - [(V16HI "TARGET_AVX2") V8HI - (V8SI "TARGET_AVX2") V4SI +(define_mode_iterator VI2_AVX2_AVX512BW + [(V32HI "TARGET_AVX512BW") (V16HI "TARGET_AVX2") V8HI]) + +(define_mode_iterator VI48_AVX2 + [(V8SI "TARGET_AVX2") V4SI (V4DI "TARGET_AVX2") V2DI]) (define_mode_iterator VI248_AVX2_8_AVX512F @@ -8585,15 +8587,34 @@ (const_string "0"))) (set_attr "mode" "<sseinsnmode>")]) -(define_insn "<shift_insn><mode>3" - [(set (match_operand:VI248_AVX2 0 "register_operand" "=x,x") - (any_lshift:VI248_AVX2 - (match_operand:VI248_AVX2 1 "register_operand" "0,x") - (match_operand:SI 2 "nonmemory_operand" "xN,xN")))] - "TARGET_SSE2" +(define_insn "<shift_insn><mode>3<mask_name>" + [(set (match_operand:VI2_AVX2_AVX512BW 0 "register_operand" "=x,v") + (any_lshift:VI2_AVX2_AVX512BW + (match_operand:VI2_AVX2_AVX512BW 1 "register_operand" "0,v") + (match_operand:SI 2 "nonmemory_operand" "xN,vN")))] + "TARGET_SSE2 && <mask_mode512bit_condition> && <mask_avx512bw_condition>" + "@ + p<vshift><ssemodesuffix>\t{%2, %0|%0, %2} + vp<vshift><ssemodesuffix>\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}" + [(set_attr "isa" "noavx,avx") + (set_attr "type" "sseishft") + (set (attr "length_immediate") + (if_then_else (match_operand 2 "const_int_operand") + (const_string "1") + (const_string "0"))) + (set_attr "prefix_data16" "1,*") + (set_attr "prefix" "orig,vex") + (set_attr "mode" "<sseinsnmode>")]) + +(define_insn "<shift_insn><mode>3<mask_name>" + [(set (match_operand:VI48_AVX2 0 "register_operand" "=x,v") + (any_lshift:VI48_AVX2 + (match_operand:VI48_AVX2 1 "register_operand" "0,v") + (match_operand:SI 2 "nonmemory_operand" "xN,vN")))] + "TARGET_SSE2 && <mask_mode512bit_condition>" "@ p<vshift><ssemodesuffix>\t{%2, %0|%0, %2} - vp<vshift><ssemodesuffix>\t{%2, %1, %0|%0, %1, %2}" + vp<vshift><ssemodesuffix>\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}" [(set_attr "isa" "noavx,avx") (set_attr "type" "sseishft") (set (attr "length_immediate") diff --git a/gcc/config/i386/subst.md b/gcc/config/i386/subst.md index edb0965..b05cb17 100644 --- a/gcc/config/i386/subst.md +++ b/gcc/config/i386/subst.md @@ -56,6 +56,7 @@ (define_subst_attr "mask_operand_arg34" "mask" "" ", operands[3], operands[4]") (define_subst_attr "mask_mode512bit_condition" "mask" "1" "(<MODE_SIZE> == 64 || TARGET_AVX512VL)") (define_subst_attr "mask_avx512vl_condition" "mask" "1" "TARGET_AVX512VL") +(define_subst_attr "mask_avx512bw_condition" "mask" "1" "TARGET_AVX512BW") (define_subst_attr "store_mask_constraint" "mask" "vm" "v") (define_subst_attr "store_mask_predicate" "mask" "nonimmediate_operand" "register_operand") (define_subst_attr "mask_prefix" "mask" "vex" "evex")