From 2e20b2ecd9ff57360b75ba583e2f4bc9da750849 Mon Sep 17 00:00:00 2001 From: Michael Meissner Date: Wed, 15 Oct 2025 11:25:38 -0400 Subject: [PATCH] Delete _Float16 -Ofast vectorization support. 2025-10-15 Michael Meissner gcc/ * config/rs6000/float16.cc (bfloat16_operation_as_v4sf): Delete fp16 operation cases only used by _Float16 vectorization. (float16_vectorization): Delete function. * config/rs6000/float16.md (VFP16): Delete. (FLOAT16_UNARY_OP): Likewise. (FLOAT16_BINARY_OP): Likewise. (float16_names): Likewise. (v8hf3, FLOAT16_UNARY_OP iterator): Likewise. (v8hf3, FLOAT16_BINARY_OP iterator): Likewise. (neg_v8hf3): Likewise. (abs_v8hf): Likewise. (fmav8hf3): Likewise. (fmsv8hf3): Likewise. (nfmav8hf3): Likewise. (nfmsv8hf3): Likewise. * config/rs6000/rs6000-protos.h (FP16_UNARY): Likewise. (FP16_ABS_BINAR): Likewise. (FP16_NEG_BINARY): Likewise. (float16_vectorization): Likewise. --- gcc/config/rs6000/float16.cc | 157 ----------------------------- gcc/config/rs6000/float16.md | 159 ------------------------------ gcc/config/rs6000/rs6000-protos.h | 5 - 3 files changed, 321 deletions(-) diff --git a/gcc/config/rs6000/float16.cc b/gcc/config/rs6000/float16.cc index b887d400312a..7539125a61fc 100644 --- a/gcc/config/rs6000/float16.cc +++ b/gcc/config/rs6000/float16.cc @@ -88,9 +88,6 @@ bfloat16_operation_as_v4sf (enum rtx_code icode, n_opts = 3; break; - case FP16_UNARY: - case FP16_ABS_BINARY: - case FP16_NEG_BINARY: default: gcc_unreachable (); } @@ -177,9 +174,6 @@ bfloat16_operation_as_v4sf (enum rtx_code icode, } break; - case FP16_UNARY: - case FP16_ABS_BINARY: - case FP16_NEG_BINARY: default: gcc_unreachable (); } @@ -197,154 +191,3 @@ bfloat16_operation_as_v4sf (enum rtx_code icode, else gcc_unreachable (); } - - -/* Expand a _Float16 vector operation: - - ICODE: Operation to perform. - RESULT: Result of the operation. - OP1: Input operand1. - OP2: Input operand2. - OP3: Input operand3 or NULL_RTX. - SUBTYPE: Describe the operation. */ - -void -float16_vectorization (enum rtx_code icode, - rtx result, - rtx op1, - rtx op2, - rtx op3, - enum fp16_operation subtype) -{ - gcc_assert (can_create_pseudo_p ()); - - enum rtx_code unary_op = UNKNOWN; - rtx op_orig[3] = { op1, op2, op3 }; - rtx op_hi[3]; - rtx op_lo[3]; - rtx result_hi; - rtx result_lo; - size_t n_opts; - - switch (subtype) - { - case FP16_UNARY: - n_opts = 1; - break; - - case FP16_BINARY: - n_opts = 2; - break; - - case FP16_ABS_BINARY: - unary_op = ABS; - n_opts = 2; - break; - - case FP16_NEG_BINARY: - unary_op = NEG; - n_opts = 2; - break; - - case FP16_FMA: - case FP16_FMS: - case FP16_NFMA: - case FP16_NFMS: - n_opts = 3; - break; - - default: - gcc_unreachable (); - } - - /* Allocate 2 temporaries for the results and the input operands. */ - result_hi = gen_reg_rtx (V4SFmode); - result_lo = gen_reg_rtx (V4SFmode); - - for (size_t i = 0; i < n_opts; i++) - { - gcc_assert (op_orig[i] != NULL_RTX); - op_hi[i] = gen_reg_rtx (V4SFmode); /* high register. */ - op_lo[i] = gen_reg_rtx (V4SFmode); /* low register. */ - - emit_insn (gen_vec_unpacks_hi_v8hf (op_hi[i], op_orig[i])); - emit_insn (gen_vec_unpacks_lo_v8hf (op_lo[i], op_orig[i])); - } - - /* Do 2 sets of V4SFmode operations. */ - switch (subtype) - { - case FP16_UNARY: - emit_insn (gen_rtx_SET (result_hi, - gen_rtx_fmt_e (icode, V4SFmode, op_hi[0]))); - - emit_insn (gen_rtx_SET (result_lo, - gen_rtx_fmt_e (icode, V4SFmode, op_lo[0]))); - break; - - case FP16_BINARY: - case FP16_ABS_BINARY: - case FP16_NEG_BINARY: - emit_insn (gen_rtx_SET (result_hi, - gen_rtx_fmt_ee (icode, V4SFmode, - op_hi[0], - op_hi[1]))); - - emit_insn (gen_rtx_SET (result_lo, - gen_rtx_fmt_ee (icode, V4SFmode, - op_lo[0], - op_lo[1]))); - break; - - case FP16_FMA: - case FP16_FMS: - case FP16_NFMA: - case FP16_NFMS: - { - rtx op1_hi = op_hi[0]; - rtx op2_hi = op_hi[1]; - rtx op3_hi = op_hi[2]; - - rtx op1_lo = op_lo[0]; - rtx op2_lo = op_lo[1]; - rtx op3_lo = op_lo[2]; - - if (subtype == FP16_FMS || subtype == FP16_NFMS) - { - op3_hi = gen_rtx_NEG (V4SFmode, op3_hi); - op3_lo = gen_rtx_NEG (V4SFmode, op3_lo); - } - - rtx op_fma_hi = gen_rtx_FMA (V4SFmode, op1_hi, op2_hi, op3_hi); - rtx op_fma_lo = gen_rtx_FMA (V4SFmode, op1_lo, op2_lo, op3_lo); - - if (subtype == FP16_NFMA || subtype == FP16_NFMS) - { - op_fma_hi = gen_rtx_NEG (V4SFmode, op_fma_hi); - op_fma_lo = gen_rtx_NEG (V4SFmode, op_fma_lo); - } - - emit_insn (gen_rtx_SET (result_hi, op_fma_hi)); - emit_insn (gen_rtx_SET (result_lo, op_fma_lo)); - } - break; - - default: - gcc_unreachable (); - } - - /* Add any unary operator modifications. */ - if (unary_op != UNKNOWN) - { - emit_insn (gen_rtx_SET (result_hi, - gen_rtx_fmt_e (unary_op, V4SFmode, result_hi))); - - emit_insn (gen_rtx_SET (result_lo, - gen_rtx_fmt_e (unary_op, V4SFmode, result_lo))); - } - - /* Combine the 2 V4SFmode operations into one V8HFmode vector. */ - emit_insn (gen_vec_pack_trunc_v4sf_v8hf (result, result_hi, result_lo)); - return; -} - diff --git a/gcc/config/rs6000/float16.md b/gcc/config/rs6000/float16.md index eaba1784362b..50205db44c2e 100644 --- a/gcc/config/rs6000/float16.md +++ b/gcc/config/rs6000/float16.md @@ -25,9 +25,6 @@ (define_mode_iterator FP16 [(BF "TARGET_BFLOAT16") (HF "TARGET_FLOAT16")]) -(define_mode_iterator VFP16 [(V8BF "TARGET_BFLOAT16") - (V8HF "TARGET_FLOAT16")]) - ;; Mode iterator for 16-bit floating point modes on machines with ;; hardware support both as a scalar and as a vector. (define_mode_iterator FP16_HW [(BF "TARGET_BFLOAT16_HW") @@ -62,22 +59,6 @@ (V8BF "V4BF") (V8HF "V4HF")]) -;; Unary operators for float16 vectorization. -(define_code_iterator FLOAT16_UNARY_OP [abs neg]) - -;; Binary operators for float16 vectorization. -(define_code_iterator FLOAT16_BINARY_OP [plus minus mult smax smin]) - -;; Standard names for the unary/binary/ternary operators -(define_code_attr float16_names [(abs "abs") - (fma "fma") - (plus "add") - (minus "sub") - (mult "mul") - (neg "neg") - (smax "smax") - (smin "smin")]) - ;; UNSPEC constants (define_c_enum "unspec" [UNSPEC_FP16_SHIFT_LEFT_32BIT @@ -711,146 +692,6 @@ DONE; }) - -;; Add vectorization support for _Float16. Unfortunately, since there -;; can only be one vec_pack_trunc_v4sf, we choose to support automatic -;; vectorization for BFmode. The following insns define vectorization -;; for HFmode. - -;; Unary operators being vectorized. -(define_insn_and_split "v8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (FLOAT16_UNARY_OP:V8HF - (match_operand:V8HF 1 "vsx_register_operand")))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (, operands[0], operands[1], NULL_RTX, NULL_RTX, - FP16_UNARY); - DONE; -}) - -;; Binary operators being vectorized. -(define_insn_and_split "v8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (FLOAT16_BINARY_OP:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand")))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (, operands[0], operands[1], operands[2], - NULL_RTX, FP16_BINARY); - DONE; -}) - -;; Negative of binary operators being vectorized. -(define_insn_and_split "*neg_v8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (neg:V8HF - (FLOAT16_BINARY_OP:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand"))))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (, operands[0], operands[1], operands[2], - NULL_RTX, FP16_NEG_BINARY); - DONE; -}) - -;; Absolute value of binary operators being vectorized. -(define_insn_and_split "*abs_v8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (abs:V8HF - (FLOAT16_BINARY_OP:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand"))))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (, operands[0], operands[1], operands[2], - NULL_RTX, FP16_ABS_BINARY); - DONE; -}) - -;; FMA operations being vectorized. -(define_insn_and_split "fmav8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (fma:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand") - (match_operand:V8HF 3 "vsx_register_operand")))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (FMA, operands[0], operands[1], operands[2], - operands[3], FP16_FMA); - DONE; -}) - -(define_insn_and_split "*fmsv8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (fma:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand") - (neg:V8HF - (match_operand:V8HF 3 "vsx_register_operand"))))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (FMA, operands[0], operands[1], operands[2], - operands[3], FP16_FMS); - DONE; -}) - -(define_insn_and_split "*nfmav8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (neg:V8HF - (fma:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand") - (match_operand:V8HF 3 "vsx_register_operand"))))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (FMA, operands[0], operands[1], operands[2], - operands[3], FP16_NFMA); - DONE; -}) - -(define_insn_and_split "*nfmsv8hf3" - [(set (match_operand:V8HF 0 "vsx_register_operand") - (neg:V8HF - (fma:V8HF - (match_operand:V8HF 1 "vsx_register_operand") - (match_operand:V8HF 2 "vsx_register_operand") - (neg:V8HF - (match_operand:V8HF 3 "vsx_register_operand")))))] - "TARGET_FLOAT16_HW && can_create_pseudo_p ()" - "#" - "&& 1" - [(pc)] -{ - float16_vectorization (FMA, operands[0], operands[1], operands[2], - operands[3], FP16_NFMS); - DONE; -}) - ;; If we do multiple __bfloat16 operations, between the first and ;; second operation, GCC will want to convert the first operation from diff --git a/gcc/config/rs6000/rs6000-protos.h b/gcc/config/rs6000/rs6000-protos.h index f677506b4aa0..be76cdb507ba 100644 --- a/gcc/config/rs6000/rs6000-protos.h +++ b/gcc/config/rs6000/rs6000-protos.h @@ -262,10 +262,7 @@ extern unsigned constant_generates_xxspltidp (vec_const_128bit_type *); /* From float16.cc. */ /* Optimize bfloat16 and float16 operations. */ enum fp16_operation { - FP16_UNARY, /* Bfloat16/float16 unary op. */ FP16_BINARY, /* Bfloat16/float16 binary op. */ - FP16_ABS_BINARY, /* abs (binary op). */ - FP16_NEG_BINARY, /* - binary op. */ FP16_FMA, /* (a * b) + c. */ FP16_FMS, /* (a * b) - c. */ FP16_NFMA, /* - ((a * b) + c). */ @@ -274,8 +271,6 @@ enum fp16_operation { extern void bfloat16_operation_as_v4sf (enum rtx_code, rtx, rtx, rtx, rtx, enum fp16_operation); -extern void float16_vectorization (enum rtx_code, rtx, rtx, rtx, rtx, - enum fp16_operation); #endif /* RTX_CODE */ #ifdef TREE_CODE