xtensa: Optimize several boolean evaluations of EQ/NE against constant zero
Checks
Commit Message
An idiomatic implementation of boolean evaluation of whether a register is
zero or not in Xtensa is to assign 0 and 1 to the temporary and destination,
and then issue the MOV[EQ/NE]Z machine instruction
(See 8.3.2 Instruction Idioms, Xtensa ISA refman., p.599):
;; A2 = (A3 != 0) ? 1 : 0;
movi.n a9, 1
movi.n a2, 0
movnez a2, a9, a3 ;; if (A3 != 0) A2 = A9;
As you can see in the above idiom, if the source and destination are the
same register, a move instruction from the source to another temporary
register must be prepended:
;; A2 = (A2 == 0) ? 1 : 0;
mov.n a10, a2
movi.n a9, 1
movi.n a2, 0
moveqz a2, a9, a10 ;; if (A10 == 0) A2 = A9;
Fortunately, we can reduce the number of instructions and temporary
registers with a few tweaks:
;; A2 = (A3 != 0) ? 1 : 0;
movi.n a2, 1
moveqz a2, a3, a3 ;; if (A3 == 0) A2 = A3;
;; A2 = (A2 != 0) ? 1 : 0;
movi.n a9, 1
movnez a2, a9, a2 ;; if (A2 != 0) A2 = A9;
;; A2 = (A3 == 0) ? 1 : 0;
movi.n a2, -1
moveqz a2, a3, a3 ;; if (A3 == 0) A2 = A3;
addi.n a2, a2, 1
;; A2 = (A2 == 0) ? 1 : 0;
movi.n a9, -1
movnez a2, a9, a2 ;; if (A2 != 0) A2 = A9;
addi.n a2, a2, 1
Additionally, if TARGET_NSA is configured, the fact that it returns 32 iff
the source of the NSAU machine instruction is 0, otherwise less than, can be
used in boolean evaluation of EQ comparison.
;; A2 = (A3 == 0) ? 1 : 0;
nsau a2, a3 ;; Source and destination can be the same register
srli a2, a2, 5
Furthermore, this patch also saves one instruction when determining whether
the ANDing with mask values in which 1s are lined up from the upper or lower
bit end (for example, 0xFFE00000 or 0x003FFFFF) is 0 or not.
gcc/ChangeLog:
* config/xtensa/xtensa.cc (xtensa_expand_scc):
Revert the changes from the last patch, as the work in the RTL
expansion pass is too far to determine the physical registers.
* config/xtensa/xtensa.md (*eqne_INT_MIN): Ditto.
(eq_zero_NSA, eqne_zero, *eqne_zero_masked_bits): New patterns.
---
gcc/config/xtensa/xtensa.cc | 35 +----------
gcc/config/xtensa/xtensa.md | 112 ++++++++++++++++++++++++++++++++++++
2 files changed, 113 insertions(+), 34 deletions(-)
Comments
On Fri, Sep 8, 2023 at 1:49 AM Takayuki 'January June' Suwa
<jjsuwa_sys3175@yahoo.co.jp> wrote:
>
> An idiomatic implementation of boolean evaluation of whether a register is
> zero or not in Xtensa is to assign 0 and 1 to the temporary and destination,
> and then issue the MOV[EQ/NE]Z machine instruction
> (See 8.3.2 Instruction Idioms, Xtensa ISA refman., p.599):
>
> ;; A2 = (A3 != 0) ? 1 : 0;
> movi.n a9, 1
> movi.n a2, 0
> movnez a2, a9, a3 ;; if (A3 != 0) A2 = A9;
>
> As you can see in the above idiom, if the source and destination are the
> same register, a move instruction from the source to another temporary
> register must be prepended:
>
> ;; A2 = (A2 == 0) ? 1 : 0;
> mov.n a10, a2
> movi.n a9, 1
> movi.n a2, 0
> moveqz a2, a9, a10 ;; if (A10 == 0) A2 = A9;
>
> Fortunately, we can reduce the number of instructions and temporary
> registers with a few tweaks:
>
> ;; A2 = (A3 != 0) ? 1 : 0;
> movi.n a2, 1
> moveqz a2, a3, a3 ;; if (A3 == 0) A2 = A3;
>
> ;; A2 = (A2 != 0) ? 1 : 0;
> movi.n a9, 1
> movnez a2, a9, a2 ;; if (A2 != 0) A2 = A9;
>
> ;; A2 = (A3 == 0) ? 1 : 0;
> movi.n a2, -1
> moveqz a2, a3, a3 ;; if (A3 == 0) A2 = A3;
> addi.n a2, a2, 1
>
> ;; A2 = (A2 == 0) ? 1 : 0;
> movi.n a9, -1
> movnez a2, a9, a2 ;; if (A2 != 0) A2 = A9;
> addi.n a2, a2, 1
>
> Additionally, if TARGET_NSA is configured, the fact that it returns 32 iff
> the source of the NSAU machine instruction is 0, otherwise less than, can be
> used in boolean evaluation of EQ comparison.
>
> ;; A2 = (A3 == 0) ? 1 : 0;
> nsau a2, a3 ;; Source and destination can be the same register
> srli a2, a2, 5
>
> Furthermore, this patch also saves one instruction when determining whether
> the ANDing with mask values in which 1s are lined up from the upper or lower
> bit end (for example, 0xFFE00000 or 0x003FFFFF) is 0 or not.
>
> gcc/ChangeLog:
>
> * config/xtensa/xtensa.cc (xtensa_expand_scc):
> Revert the changes from the last patch, as the work in the RTL
> expansion pass is too far to determine the physical registers.
> * config/xtensa/xtensa.md (*eqne_INT_MIN): Ditto.
> (eq_zero_NSA, eqne_zero, *eqne_zero_masked_bits): New patterns.
> ---
> gcc/config/xtensa/xtensa.cc | 35 +----------
> gcc/config/xtensa/xtensa.md | 112 ++++++++++++++++++++++++++++++++++++
> 2 files changed, 113 insertions(+), 34 deletions(-)
Regtested for target=xtensa-linux-uclibc, no new regressions.
Committed to master.
@@ -994,41 +994,8 @@ xtensa_expand_scc (rtx operands[4], machine_mode cmp_mode)
rtx cmp;
rtx one_tmp, zero_tmp;
rtx (*gen_fn) (rtx, rtx, rtx, rtx, rtx);
- enum rtx_code code = GET_CODE (operands[1]);
- if (cmp_mode == SImode && CONST_INT_P (operands[3])
- && (code == EQ || code == NE))
- switch (INTVAL (operands[3]))
- {
- case 0:
- if (TARGET_MINMAX)
- {
- one_tmp = force_reg (SImode, const1_rtx);
- emit_insn (gen_uminsi3 (dest, operands[2], one_tmp));
- if (code == EQ)
- emit_insn (gen_xorsi3 (dest, dest, one_tmp));
- return 1;
- }
- break;
- case -2147483648:
- if (TARGET_ABS)
- {
- emit_insn (gen_abssi2 (dest, operands[2]));
- if (code == EQ)
- emit_insn (gen_lshrsi3 (dest, dest, GEN_INT (31)));
- else
- {
- emit_insn (gen_ashrsi3 (dest, dest, GEN_INT (31)));
- emit_insn (gen_addsi3 (dest, dest, const1_rtx));
- }
- return 1;
- }
- break;
- default:
- break;
- }
-
- if (! (cmp = gen_conditional_move (code, cmp_mode,
+ if (! (cmp = gen_conditional_move (GET_CODE (operands[1]), cmp_mode,
operands[2], operands[3])))
return 0;
@@ -3188,6 +3188,118 @@
(const_int 5)
(const_int 6)))])
+(define_insn_and_split "eq_zero_NSA"
+ [(set (match_operand:SI 0 "register_operand" "=a")
+ (eq:SI (match_operand:SI 1 "register_operand" "r")
+ (const_int 0)))]
+ "TARGET_NSA"
+ "#"
+ "&& 1"
+ [(set (match_dup 0)
+ (clz:SI (match_dup 1)))
+ (set (match_dup 0)
+ (lshiftrt:SI (match_dup 0)
+ (const_int 5)))]
+ ""
+ [(set_attr "type" "move")
+ (set_attr "mode" "SI")
+ (set_attr "length" "6")])
+
+(define_insn_and_split "eqne_zero"
+ [(set (match_operand:SI 0 "register_operand" "=a,&a")
+ (match_operator:SI 2 "boolean_operator"
+ [(match_operand:SI 1 "register_operand" "0,r")
+ (const_int 0)]))
+ (clobber (match_scratch:SI 3 "=&a,X"))]
+ ""
+ "#"
+ "&& reload_completed"
+ [(const_int 0)]
+{
+ enum rtx_code code = GET_CODE (operands[2]);
+ int same_p = REGNO (operands[0]) == REGNO (operands[1]);
+ emit_move_insn (same_p ? operands[3] : operands[0],
+ code == EQ ? constm1_rtx : const1_rtx);
+ emit_insn (gen_movsicc_internal0 (operands[0], operands[1],
+ same_p ? operands[3] : operands[1],
+ operands[0],
+ gen_rtx_fmt_ee (same_p ? NE : EQ,
+ VOIDmode,
+ operands[1],
+ const0_rtx)));
+ if (code == EQ)
+ emit_insn (gen_addsi3 (operands[0], operands[0], const1_rtx));
+ DONE;
+}
+ [(set_attr "type" "move")
+ (set_attr "mode" "SI")
+ (set (attr "length")
+ (if_then_else (match_test "GET_CODE (operands[2]) == EQ")
+ (if_then_else (match_test "TARGET_DENSITY")
+ (const_int 7)
+ (const_int 9))
+ (if_then_else (match_test "TARGET_DENSITY")
+ (const_int 5)
+ (const_int 6))))])
+
+(define_insn_and_split "*eqne_zero_masked_bits"
+ [(set (match_operand:SI 0 "register_operand" "=a")
+ (match_operator 3 "boolean_operator"
+ [(and:SI (match_operand:SI 1 "register_operand" "r")
+ (match_operand:SI 2 "const_int_operand" "i"))
+ (const_int 0)]))]
+ "IN_RANGE (exact_log2 (INTVAL (operands[2]) + 1), 17, 31)
+ || IN_RANGE (exact_log2 (-INTVAL (operands[2])), 1, 30)"
+ "#"
+ "&& 1"
+ [(const_int 0)]
+{
+ HOST_WIDE_INT mask = INTVAL (operands[2]);
+ int n;
+ enum rtx_code code = GET_CODE (operands[3]);
+ if (IN_RANGE (n = exact_log2 (mask + 1), 17, 31))
+ emit_insn (gen_ashlsi3 (operands[0], operands[1], GEN_INT (32 - n)));
+ else
+ emit_insn (gen_lshrsi3 (operands[0], operands[1],
+ GEN_INT (floor_log2 (-mask))));
+ if (TARGET_NSA && code == EQ)
+ emit_insn (gen_eq_zero_NSA (operands[0], operands[0]));
+ else
+ emit_insn (gen_eqne_zero (operands[0], operands[0],
+ gen_rtx_fmt_ee (code, VOIDmode,
+ operands[0], const0_rtx)));
+ DONE;
+})
+
+(define_insn_and_split "*eqne_INT_MIN"
+ [(set (match_operand:SI 0 "register_operand" "=a")
+ (match_operator:SI 2 "boolean_operator"
+ [(match_operand:SI 1 "register_operand" "r")
+ (const_int -2147483648)]))]
+ "TARGET_ABS"
+ "#"
+ "&& 1"
+ [(const_int 0)]
+{
+ emit_insn (gen_abssi2 (operands[0], operands[1]));
+ if (GET_CODE (operands[2]) == EQ)
+ emit_insn (gen_lshrsi3 (operands[0], operands[0], GEN_INT (31)));
+ else
+ {
+ emit_insn (gen_ashrsi3 (operands[0], operands[0], GEN_INT (31)));
+ emit_insn (gen_addsi3 (operands[0], operands[0], const1_rtx));
+ }
+ DONE;
+}
+ [(set_attr "type" "move")
+ (set_attr "mode" "SI")
+ (set (attr "length")
+ (if_then_else (match_test "GET_CODE (operands[2]) == EQ")
+ (const_int 6)
+ (if_then_else (match_test "TARGET_DENSITY")
+ (const_int 8)
+ (const_int 9))))])
+
(define_peephole2
[(set (match_operand:SI 0 "register_operand")
(match_operand:SI 6 "reload_operand"))