diff options
author | Richard Henderson <rth@twiddle.net> | 2013-08-14 19:32:56 -0700 |
---|---|---|
committer | Richard Henderson <rth@redhat.com> | 2014-04-16 12:12:58 -0400 |
commit | dfeb5fe7700d5e29a276f571aee7f6fc4267ee96 (patch) | |
tree | 5cda104066bb7bb484c32774a6637abaf786b143 /tcg/aarch64 | |
parent | 929f8b55509fb92320e085504a8492b1908b8110 (diff) |
tcg-aarch64: Use MOVN in tcg_out_movi
When profitable, initialize the register with MOVN instead of MOVZ,
before setting the remaining lanes with MOVK.
Reviewed-by: Claudio Fontana <claudio.fontana@huawei.com>
Signed-off-by: Richard Henderson <rth@twiddle.net>
Diffstat (limited to 'tcg/aarch64')
-rw-r--r-- | tcg/aarch64/tcg-target.c | 63 |
1 files changed, 50 insertions, 13 deletions
diff --git a/tcg/aarch64/tcg-target.c b/tcg/aarch64/tcg-target.c index 5e6d10b46e..1d7612c6e6 100644 --- a/tcg/aarch64/tcg-target.c +++ b/tcg/aarch64/tcg-target.c @@ -531,24 +531,61 @@ static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg rd, tcg_target_long value) { AArch64Insn insn; - - if (type == TCG_TYPE_I32) { + int i, wantinv, shift; + tcg_target_long svalue = value; + tcg_target_long ivalue = ~value; + tcg_target_long imask; + + /* For 32-bit values, discard potential garbage in value. For 64-bit + values within [2**31, 2**32-1], we can create smaller sequences by + interpreting this as a negative 32-bit number, while ensuring that + the high 32 bits are cleared by setting SF=0. */ + if (type == TCG_TYPE_I32 || (value & ~0xffffffffull) == 0) { + svalue = (int32_t)value; value = (uint32_t)value; + ivalue = (uint32_t)ivalue; + type = TCG_TYPE_I32; + } + + /* Would it take fewer insns to begin with MOVN? For the value and its + inverse, count the number of 16-bit lanes that are 0. */ + for (i = wantinv = imask = 0; i < 64; i += 16) { + tcg_target_long mask = 0xffffull << i; + if ((value & mask) == 0) { + wantinv -= 1; + } + if ((ivalue & mask) == 0) { + wantinv += 1; + imask |= mask; + } } - /* count trailing zeros in 16 bit steps, mapping 64 to 0. Emit the - first MOVZ with the half-word immediate skipping the zeros, with a shift - (LSL) equal to this number. Then all next instructions use MOVKs. - Zero the processed half-word in the value, continue until empty. - We build the final result 16bits at a time with up to 4 instructions, - but do not emit instructions for 16bit zero holes. */ + /* If we had more 0xffff than 0x0000, invert VALUE and use MOVN. */ insn = I3405_MOVZ; - do { - unsigned shift = ctz64(value) & (63 & -16); - tcg_out_insn_3405(s, insn, shift >= 32, rd, value >> shift, shift); + if (wantinv > 0) { + value = ivalue; + insn = I3405_MOVN; + } + + /* Find the lowest lane that is not 0x0000. */ + shift = ctz64(value) & (63 & -16); + tcg_out_insn_3405(s, insn, type, rd, value >> shift, shift); + + if (wantinv > 0) { + /* Re-invert the value, so MOVK sees non-inverted bits. */ + value = ~value; + /* Clear out all the 0xffff lanes. */ + value ^= imask; + } + /* Clear out the lane that we just set. */ + value &= ~(0xffffUL << shift); + + /* Iterate until all lanes have been set, and thus cleared from VALUE. */ + while (value) { + shift = ctz64(value) & (63 & -16); + tcg_out_insn(s, 3405, MOVK, type, rd, value >> shift, shift); value &= ~(0xffffUL << shift); - insn = I3405_MOVK; - } while (value); + } } static inline void tcg_out_ldst_r(TCGContext *s, |