Hello,this patch expands __builtin_shuffle for V4DF mode in at most 3 insn. It is simple and works really well, often generates only 2 insn. It is not very generic, because other modes don't have an instruction equivalent to vshufpd. For V8SF (and likely V4DI and V8SI with AVX2, but I still need to do that), my patch "default case" in PR 52607 seems more interesting.
I tried calling this new function after expand_vec_perm_vperm2f128_vblend (instead of before as in the patch), but it generated more instructions for some permutations, and never less. That function is still useful for V8SF though.
I bootstrapped gcc on a non-avx platform, compiled a program that tests all 4096 shuffles with -mavx/-mavx2, and ran the result using Intel's emulator (SDE).
There are still a few V4DF permutations that don't generate an optimal sequence (3 insn instead of 2), but not that many I think. Of course, I am assuming a constant cost of 1 per insn, which is completely false, but seems like a sensible first approximation.
(note that I can't commit) 2012-04-17 Marc Glisse <marc.gli...@inria.fr> PR target/502607 * config/i386/i386.c (ix86_expand_vec_perm_const): Move code to ... (canonicalize_perm): ... new function. (expand_vec_perm_2vperm2f128_vshuf): New function. (ix86_expand_vec_perm_const_1): Call it. -- Marc Glisse
Index: config/i386/i386.c =================================================================== --- config/i386/i386.c (revision 186523) +++ config/i386/i386.c (working copy) @@ -32946,6 +32946,7 @@ bool testing_p; }; +static bool canonicalize_perm (struct expand_vec_perm_d *d); static bool expand_vec_perm_1 (struct expand_vec_perm_d *d); static bool expand_vec_perm_broadcast_1 (struct expand_vec_perm_d *d); @@ -37003,6 +37004,57 @@ return true; } +/* A subroutine of ix86_expand_vec_perm_builtin_1. Implement a V4DF + permutation using two vperm2f128, followed by a vshufpd insn blending + the two vectors together. */ + +static bool +expand_vec_perm_2vperm2f128_vshuf (struct expand_vec_perm_d *d) +{ + struct expand_vec_perm_d dfirst, dsecond, dthird; + bool ok; + + if (!TARGET_AVX || (d->vmode != V4DFmode)) + return false; + + if (d->testing_p) + return true; + + dfirst = *d; + dsecond = *d; + dthird = *d; + + dfirst.perm[0] = (d->perm[0] & ~1); + dfirst.perm[1] = (d->perm[0] & ~1) + 1; + dfirst.perm[2] = (d->perm[2] & ~1); + dfirst.perm[3] = (d->perm[2] & ~1) + 1; + dsecond.perm[0] = (d->perm[1] & ~1); + dsecond.perm[1] = (d->perm[1] & ~1) + 1; + dsecond.perm[2] = (d->perm[3] & ~1); + dsecond.perm[3] = (d->perm[3] & ~1) + 1; + dthird.perm[0] = (d->perm[0] % 2); + dthird.perm[1] = (d->perm[1] % 2) + 4; + dthird.perm[2] = (d->perm[2] % 2) + 2; + dthird.perm[3] = (d->perm[3] % 2) + 6; + + dfirst.target = gen_reg_rtx (dfirst.vmode); + dsecond.target = gen_reg_rtx (dsecond.vmode); + dthird.op0 = dfirst.target; + dthird.op1 = dsecond.target; + dthird.one_operand_p = false; + + canonicalize_perm (&dfirst); + canonicalize_perm (&dsecond); + + ok = expand_vec_perm_1 (&dfirst) + && expand_vec_perm_1 (&dsecond) + && expand_vec_perm_1 (&dthird); + + gcc_assert (ok); + + return true; +} + /* A subroutine of expand_vec_perm_even_odd_1. Implement the double-word permutation with two pshufb insns and an ior. We should have already failed all two instruction sequences. */ @@ -37652,6 +37704,9 @@ /* Try sequences of three instructions. */ + if (expand_vec_perm_2vperm2f128_vshuf (d)) + return true; + if (expand_vec_perm_pshufb2 (d)) return true; @@ -37689,12 +37744,56 @@ return false; } +/* If a permutation only uses one operand, make it clear. Returns true + if the permutation references both operands. */ + +static bool +canonicalize_perm (struct expand_vec_perm_d *d) +{ + int i, which, nelt = d->nelt; + + for (i = which = 0; i < nelt; ++i) + which |= (d->perm[i] < nelt ? 1 : 2); + + d->one_operand_p = true; + switch (which) + { + default: + gcc_unreachable(); + + case 3: + if (!rtx_equal_p (d->op0, d->op1)) + { + d->one_operand_p = false; + break; + } + /* The elements of PERM do not suggest that only the first operand + is used, but both operands are identical. Allow easier matching + of the permutation by folding the permutation into the single + input vector. */ + /* FALLTHRU */ + + case 2: + for (i = 0; i < nelt; ++i) + d->perm[i] &= nelt - 1; + d->op0 = d->op1; + break; + + case 1: + d->op1 = d->op0; + break; + } + + return (which == 3); +} + bool ix86_expand_vec_perm_const (rtx operands[4]) { struct expand_vec_perm_d d; unsigned char perm[MAX_VECT_LEN]; - int i, nelt, which; + int i, nelt; + bool two_args; rtx sel; d.target = operands[0]; @@ -37711,45 +37810,16 @@ gcc_assert (XVECLEN (sel, 0) == nelt); gcc_checking_assert (sizeof (d.perm) == sizeof (perm)); - for (i = which = 0; i < nelt; ++i) + for (i = 0; i < nelt; ++i) { rtx e = XVECEXP (sel, 0, i); int ei = INTVAL (e) & (2 * nelt - 1); - - which |= (ei < nelt ? 1 : 2); d.perm[i] = ei; perm[i] = ei; } - d.one_operand_p = true; - switch (which) - { - default: - gcc_unreachable(); + two_args = canonicalize_perm (&d); - case 3: - if (!rtx_equal_p (d.op0, d.op1)) - { - d.one_operand_p = false; - break; - } - /* The elements of PERM do not suggest that only the first operand - is used, but both operands are identical. Allow easier matching - of the permutation by folding the permutation into the single - input vector. */ - /* FALLTHRU */ - - case 2: - for (i = 0; i < nelt; ++i) - d.perm[i] &= nelt - 1; - d.op0 = d.op1; - break; - - case 1: - d.op1 = d.op0; - break; - } - if (ix86_expand_vec_perm_const_1 (&d)) return true; @@ -37757,7 +37827,7 @@ same, the above tried to expand with one_operand_p and flattened selector. If that didn't work, retry without one_operand_p; we succeeded with that during testing. */ - if (which == 3 && d.one_operand_p) + if (two_args && d.one_operand_p) { d.one_operand_p = false; memcpy (d.perm, perm, sizeof (perm));