This simple patch improves the implementation of 128-bit (TImode)
rotations on x86_64 (a missed optimization opportunity spotted
during the recent V1TImode improvements).
Currently, the function:
unsigned __int128 rotrti3(unsigned __int128 x, unsigned int i) {
return (x >> i) | (x << (128-i));
}
produces:
rotrti3:
movq %rsi, %r8
movq %rdi, %r9
movl %edx, %ecx
movq %rdi, %rsi
movq %r9, %rax
movq %r8, %rdx
movq %r8, %rdi
shrdq %r8, %rax
shrq %cl, %rdx
xorl %r8d, %r8d
testb $64, %cl
cmovne %rdx, %rax
cmovne %r8, %rdx
negl %ecx
andl $127, %ecx
shldq %r9, %rdi
salq %cl, %rsi
xorl %r9d, %r9d
testb $64, %cl
cmovne %rsi, %rdi
cmovne %r9, %rsi
orq %rdi, %rdx
orq %rsi, %rax
ret
with this patch, GCC will now generate the much nicer:
rotrti3:
movl %edx, %ecx
movq %rdi, %rdx
shrdq %rsi, %rdx
shrdq %rdi, %rsi
andl $64, %ecx
movq %rdx, %rax
cmove %rsi, %rdx
cmovne %rsi, %rax
ret
Even I wasn't expecting the optimizer's choice of the final three
instructions; a thing of beauty. For rotations larger than 64,
the lowpart and the highpart (%rax and %rdx) are transposed, and
it would be nice to have a conditional swap/exchange. The inspired
solution the compiler comes up with is to store/duplicate the same
value in both %rax/%rdx, and then use complementary conditional moves
to either update the lowpart or highpart, which cleverly avoids the
potential decode-stage pipeline stall (on some microarchitectures)
from having multiple instructions conditional on the same condition.
See X86_TUNE_ONE_IF_CONV_INSN, and notice there are two such stalls
in the original expansion of rot[rl]ti3.
2021-11-02 Roger Sayle <roger@nextmovesoftware.com>
Uroš Bizjak <ubizjak@gmail.com>
* config/i386/i386.md (<any_rotate>ti3): Provide expansion for
rotations by non-constant amounts.
emit_insn (gen_ix86_<insn>ti3_doubleword
(operands[0], operands[1], operands[2]));
else
- FAIL;
-
+ {
+ rtx amount = force_reg (QImode, operands[2]);
+ rtx src_lo = gen_lowpart (DImode, operands[1]);
+ rtx src_hi = gen_highpart (DImode, operands[1]);
+ rtx tmp_lo = gen_reg_rtx (DImode);
+ rtx tmp_hi = gen_reg_rtx (DImode);
+ emit_move_insn (tmp_lo, src_lo);
+ emit_move_insn (tmp_hi, src_hi);
+ rtx (*shiftd) (rtx, rtx, rtx)
+ = (<CODE> == ROTATE) ? gen_x86_64_shld : gen_x86_64_shrd;
+ emit_insn (shiftd (tmp_lo, src_hi, amount));
+ emit_insn (shiftd (tmp_hi, src_lo, amount));
+ rtx dst_lo = gen_lowpart (DImode, operands[0]);
+ rtx dst_hi = gen_highpart (DImode, operands[0]);
+ emit_move_insn (dst_lo, tmp_lo);
+ emit_move_insn (dst_hi, tmp_hi);
+ emit_insn (gen_x86_shiftdi_adj_1 (dst_lo, dst_hi, amount, tmp_lo));
+ }
DONE;
})