From cba10bb3c8875f428821608f9dfb860a97a63aa1 Mon Sep 17 00:00:00 2001 From: Richard Henderson Date: Sat, 5 Aug 2023 01:55:23 +0000 Subject: [PATCH] tcg/ppc: Implement negsetcond_* In the general case we simply negate. However with isel we may load -1 instead of 1 with no extra effort. Consolidate EQ0 and NE0 logic. Replace the NE0 zero-extension with inversion+negation of EQ0, which is never worse and may eliminate one insn. Provide a special case for -EQ0. Reviewed-by: Daniel Henrique Barboza Signed-off-by: Richard Henderson --- tcg/ppc/tcg-target.c.inc | 127 ++++++++++++++++++++++++--------------- tcg/ppc/tcg-target.h | 4 +- 2 files changed, 82 insertions(+), 49 deletions(-) diff --git a/tcg/ppc/tcg-target.c.inc b/tcg/ppc/tcg-target.c.inc index 511e14b180..10448aa0e6 100644 --- a/tcg/ppc/tcg-target.c.inc +++ b/tcg/ppc/tcg-target.c.inc @@ -1548,8 +1548,20 @@ static void tcg_out_cmp(TCGContext *s, int cond, TCGArg arg1, TCGArg arg2, } static void tcg_out_setcond_eq0(TCGContext *s, TCGType type, - TCGReg dst, TCGReg src) + TCGReg dst, TCGReg src, bool neg) { + if (neg && (TCG_TARGET_REG_BITS == 32 || type == TCG_TYPE_I64)) { + /* + * X != 0 implies X + -1 generates a carry. + * RT = (~X + X) + CA + * = -1 + CA + * = CA ? 0 : -1 + */ + tcg_out32(s, ADDIC | TAI(TCG_REG_R0, src, -1)); + tcg_out32(s, SUBFE | TAB(dst, src, src)); + return; + } + if (type == TCG_TYPE_I32) { tcg_out32(s, CNTLZW | RS(src) | RA(dst)); tcg_out_shri32(s, dst, dst, 5); @@ -1557,18 +1569,28 @@ static void tcg_out_setcond_eq0(TCGContext *s, TCGType type, tcg_out32(s, CNTLZD | RS(src) | RA(dst)); tcg_out_shri64(s, dst, dst, 6); } + if (neg) { + tcg_out32(s, NEG | RT(dst) | RA(dst)); + } } -static void tcg_out_setcond_ne0(TCGContext *s, TCGReg dst, TCGReg src) +static void tcg_out_setcond_ne0(TCGContext *s, TCGType type, + TCGReg dst, TCGReg src, bool neg) { - /* X != 0 implies X + -1 generates a carry. Extra addition - trickery means: R = X-1 + ~X + C = X-1 + (-X+1) + C = C. */ - if (dst != src) { - tcg_out32(s, ADDIC | TAI(dst, src, -1)); - tcg_out32(s, SUBFE | TAB(dst, dst, src)); - } else { + if (!neg && (TCG_TARGET_REG_BITS == 32 || type == TCG_TYPE_I64)) { + /* + * X != 0 implies X + -1 generates a carry. Extra addition + * trickery means: R = X-1 + ~X + C = X-1 + (-X+1) + C = C. + */ tcg_out32(s, ADDIC | TAI(TCG_REG_R0, src, -1)); tcg_out32(s, SUBFE | TAB(dst, TCG_REG_R0, src)); + return; + } + tcg_out_setcond_eq0(s, type, dst, src, false); + if (neg) { + tcg_out32(s, ADDI | TAI(dst, dst, -1)); + } else { + tcg_out_xori32(s, dst, dst, 1); } } @@ -1590,9 +1612,10 @@ static TCGReg tcg_gen_setcond_xor(TCGContext *s, TCGReg arg1, TCGArg arg2, static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond, TCGArg arg0, TCGArg arg1, TCGArg arg2, - int const_arg2) + int const_arg2, bool neg) { - int crop, sh; + int sh; + bool inv; tcg_debug_assert(TCG_TARGET_REG_BITS == 64 || type == TCG_TYPE_I32); @@ -1605,14 +1628,10 @@ static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond, if (arg2 == 0) { switch (cond) { case TCG_COND_EQ: - tcg_out_setcond_eq0(s, type, arg0, arg1); + tcg_out_setcond_eq0(s, type, arg0, arg1, neg); return; case TCG_COND_NE: - if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) { - tcg_out_ext32u(s, TCG_REG_R0, arg1); - arg1 = TCG_REG_R0; - } - tcg_out_setcond_ne0(s, arg0, arg1); + tcg_out_setcond_ne0(s, type, arg0, arg1, neg); return; case TCG_COND_GE: tcg_out32(s, NOR | SAB(arg1, arg0, arg1)); @@ -1621,9 +1640,17 @@ static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond, case TCG_COND_LT: /* Extract the sign bit. */ if (type == TCG_TYPE_I32) { - tcg_out_shri32(s, arg0, arg1, 31); + if (neg) { + tcg_out_sari32(s, arg0, arg1, 31); + } else { + tcg_out_shri32(s, arg0, arg1, 31); + } } else { - tcg_out_shri64(s, arg0, arg1, 63); + if (neg) { + tcg_out_sari64(s, arg0, arg1, 63); + } else { + tcg_out_shri64(s, arg0, arg1, 63); + } } return; default: @@ -1641,7 +1668,7 @@ static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond, isel = tcg_to_isel[cond]; - tcg_out_movi(s, type, arg0, 1); + tcg_out_movi(s, type, arg0, neg ? -1 : 1); if (isel & 1) { /* arg0 = (bc ? 0 : 1) */ tab = TAB(arg0, 0, arg0); @@ -1655,51 +1682,47 @@ static void tcg_out_setcond(TCGContext *s, TCGType type, TCGCond cond, return; } + inv = false; switch (cond) { case TCG_COND_EQ: arg1 = tcg_gen_setcond_xor(s, arg1, arg2, const_arg2); - tcg_out_setcond_eq0(s, type, arg0, arg1); - return; + tcg_out_setcond_eq0(s, type, arg0, arg1, neg); + break; case TCG_COND_NE: arg1 = tcg_gen_setcond_xor(s, arg1, arg2, const_arg2); - /* Discard the high bits only once, rather than both inputs. */ - if (TCG_TARGET_REG_BITS == 64 && type == TCG_TYPE_I32) { - tcg_out_ext32u(s, TCG_REG_R0, arg1); - arg1 = TCG_REG_R0; - } - tcg_out_setcond_ne0(s, arg0, arg1); - return; + tcg_out_setcond_ne0(s, type, arg0, arg1, neg); + break; + case TCG_COND_LE: + case TCG_COND_LEU: + inv = true; + /* fall through */ case TCG_COND_GT: case TCG_COND_GTU: - sh = 30; - crop = 0; - goto crtest; - - case TCG_COND_LT: - case TCG_COND_LTU: - sh = 29; - crop = 0; + sh = 30; /* CR7 CR_GT */ goto crtest; case TCG_COND_GE: case TCG_COND_GEU: - sh = 31; - crop = CRNOR | BT(7, CR_EQ) | BA(7, CR_LT) | BB(7, CR_LT); + inv = true; + /* fall through */ + case TCG_COND_LT: + case TCG_COND_LTU: + sh = 29; /* CR7 CR_LT */ goto crtest; - case TCG_COND_LE: - case TCG_COND_LEU: - sh = 31; - crop = CRNOR | BT(7, CR_EQ) | BA(7, CR_GT) | BB(7, CR_GT); crtest: tcg_out_cmp(s, cond, arg1, arg2, const_arg2, 7, type); - if (crop) { - tcg_out32(s, crop); - } tcg_out32(s, MFOCRF | RT(TCG_REG_R0) | FXM(7)); tcg_out_rlw(s, RLWINM, arg0, TCG_REG_R0, sh, 31, 31); + if (neg && inv) { + tcg_out32(s, ADDI | TAI(arg0, arg0, -1)); + } else if (neg) { + tcg_out32(s, NEG | RT(arg0) | RA(arg0)); + } else if (inv) { + tcg_out_xori32(s, arg0, arg0, 1); + } break; default: @@ -2982,11 +3005,19 @@ static void tcg_out_op(TCGContext *s, TCGOpcode opc, case INDEX_op_setcond_i32: tcg_out_setcond(s, TCG_TYPE_I32, args[3], args[0], args[1], args[2], - const_args[2]); + const_args[2], false); break; case INDEX_op_setcond_i64: tcg_out_setcond(s, TCG_TYPE_I64, args[3], args[0], args[1], args[2], - const_args[2]); + const_args[2], false); + break; + case INDEX_op_negsetcond_i32: + tcg_out_setcond(s, TCG_TYPE_I32, args[3], args[0], args[1], args[2], + const_args[2], true); + break; + case INDEX_op_negsetcond_i64: + tcg_out_setcond(s, TCG_TYPE_I64, args[3], args[0], args[1], args[2], + const_args[2], true); break; case INDEX_op_setcond2_i32: tcg_out_setcond2(s, args, const_args); @@ -3724,6 +3755,7 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op) case INDEX_op_rotl_i32: case INDEX_op_rotr_i32: case INDEX_op_setcond_i32: + case INDEX_op_negsetcond_i32: case INDEX_op_and_i64: case INDEX_op_andc_i64: case INDEX_op_shl_i64: @@ -3732,6 +3764,7 @@ static TCGConstraintSetIndex tcg_target_op_def(TCGOpcode op) case INDEX_op_rotl_i64: case INDEX_op_rotr_i64: case INDEX_op_setcond_i64: + case INDEX_op_negsetcond_i64: return C_O1_I2(r, r, ri); case INDEX_op_mul_i32: diff --git a/tcg/ppc/tcg-target.h b/tcg/ppc/tcg-target.h index a2ca0b44ce..8bfb14998e 100644 --- a/tcg/ppc/tcg-target.h +++ b/tcg/ppc/tcg-target.h @@ -97,7 +97,7 @@ typedef enum { #define TCG_TARGET_HAS_sextract_i32 0 #define TCG_TARGET_HAS_extract2_i32 0 #define TCG_TARGET_HAS_movcond_i32 1 -#define TCG_TARGET_HAS_negsetcond_i32 0 +#define TCG_TARGET_HAS_negsetcond_i32 1 #define TCG_TARGET_HAS_mulu2_i32 0 #define TCG_TARGET_HAS_muls2_i32 0 #define TCG_TARGET_HAS_muluh_i32 1 @@ -135,7 +135,7 @@ typedef enum { #define TCG_TARGET_HAS_sextract_i64 0 #define TCG_TARGET_HAS_extract2_i64 0 #define TCG_TARGET_HAS_movcond_i64 1 -#define TCG_TARGET_HAS_negsetcond_i64 0 +#define TCG_TARGET_HAS_negsetcond_i64 1 #define TCG_TARGET_HAS_add2_i64 1 #define TCG_TARGET_HAS_sub2_i64 1 #define TCG_TARGET_HAS_mulu2_i64 0