On Tue, Aug 25, 2026 at 7:27 AM feedable via Sourceware Forge
<[email protected]> wrote:
>
> From: feedable <[email protected]>
>
> ix86_split_long_move is using parts of the the operands array that are not
> set by the recog machionery as storage space, which makes the logic confusing
> and also interferes with making the operand count dynamic.
>
> Use local arrays instead.
One question, does the match_dup in the split output template but not
in input pattern also be considered improper storage of operands
array?
.i.e for below pattern in i386.md, input pattern only has 2 operands,
but the output template has match_dup 4 and match_dup 5.
Does genconfig count operand numbers for split's output template? Does
dynamic operand count handle that?
(define_insn_and_split "*cmp<dwi>_doubleword"
[(set (reg:CCZ FLAGS_REG)
(compare:CCZ (match_operand:<DWI> 0 "nonimmediate_operand")
(match_operand:<DWI> 1 "general_operand")))]
"ix86_pre_reload_split ()"
"#"
"&& 1"
[(parallel [(set (reg:CCZ FLAGS_REG)
(compare:CCZ (ior:DWIH (match_dup 4) (match_dup 5))
(const_int 0)))
(set (match_dup 4) (ior:DWIH (match_dup 4) (match_dup 5)))])]
split_double_mode (<DWI>mode, &operands[0], 2, &operands[0], &operands[2]);
operands[4] = gen_reg_rtx (<MODE>mode);
/* Special case comparisons against -1. */
if (operands[1] == constm1_rtx && operands[3] == constm1_rtx)
{
emit_insn (gen_and<mode>3 (operands[4], operands[0], operands[2]));
emit_insn (gen_cmp_1 (<MODE>mode, operands[4], constm1_rtx));
DONE;
}
>
> gcc/ChangeLog:
>
> * config/i386/i386-expand.cc (ix86_split_to_parts): Use automatic
> storage for temporaries.
> (ix86_split_long_move): Likewise.
> * config/i386/i386.md: Likewise.
> ---
> gcc/config/i386/i386-expand.cc | 26 ++++----
> gcc/config/i386/i386.md | 113 ++++++++++++++++-----------------
> 2 files changed, 69 insertions(+), 70 deletions(-)
>
> diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
> index 34c9599f928e6..8f42297263d3c 100644
> --- a/gcc/config/i386/i386-expand.cc
> +++ b/gcc/config/i386/i386-expand.cc
> @@ -6420,8 +6420,7 @@ ix86_split_to_parts (rtx operand, rtx *parts,
> machine_mode mode)
>
> /* Emit insns to perform a move or push of DI, DF, XF, and TF values.
> Return false when normal moves are needed; true when all required
> - insns have been emitted. Operands 2-4 contain the input values
> - int the correct order; operands 5-7 contain the output values. */
> + insns have been emitted. */
>
> void
> ix86_split_long_move (rtx operands[])
> @@ -6602,6 +6601,7 @@ ix86_split_long_move (rtx operands[])
> return;
> }
>
> + rtx move_src[4], move_dst[4];
> /* Choose correct order to not overwrite the source before it is copied.
> */
> if ((REG_P (part[0][0])
> && REG_P (part[1][1])
> @@ -6615,31 +6615,31 @@ ix86_split_long_move (rtx operands[])
> {
> for (i = 0, j = nparts - 1; i < nparts; i++, j--)
> {
> - operands[2 + i] = part[0][j];
> - operands[6 + i] = part[1][j];
> + move_src[i] = part[0][j];
> + move_dst[i] = part[1][j];
> }
> }
> else
> {
> for (i = 0; i < nparts; i++)
> {
> - operands[2 + i] = part[0][i];
> - operands[6 + i] = part[1][i];
> + move_src[i] = part[0][i];
> + move_dst[i] = part[1][i];
> }
> }
>
> /* Attempt to locally unCSE nonzero constants. */
> for (j = 0; j < nparts - 1; j++)
> - if (CONST_INT_P (operands[6 + j])
> - && operands[6 + j] != const0_rtx
> - && REG_P (operands[2 + j]))
> + if (CONST_INT_P (move_dst[j])
> + && move_dst[j] != const0_rtx
> + && REG_P (move_src[j]))
> for (i = j; i < nparts - 1; i++)
> - if (CONST_INT_P (operands[7 + i])
> - && INTVAL (operands[7 + i]) == INTVAL (operands[6 + j]))
> - operands[7 + i] = operands[2 + j];
> + if (CONST_INT_P (move_dst[1 + i])
> + && INTVAL (move_dst[1 + i]) == INTVAL (move_dst[j]))
> + move_dst[1 + i] = move_src[j];
>
> for (i = 0; i < nparts; i++)
> - emit_move_insn (operands[2 + i], operands[6 + i]);
> + emit_move_insn (move_src[i], move_dst[i]);
>
> return;
> }
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index 4e18797743f48..63f744daab63e 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -3779,8 +3779,8 @@
> "&& reload_completed"
> [(const_int 0)]
> {
> - operands[4] = gen_lowpart (DImode, operands[1]);
> - split_double_concat (TImode, operands[0], operands[4], operands[2]);
> + rtx temp = gen_lowpart (DImode, operands[1]);
> + split_double_concat (TImode, operands[0], temp, operands[2]);
> DONE;
> })
>
> @@ -3801,8 +3801,8 @@
> "&& reload_completed"
> [(const_int 0)]
> {
> - operands[4] = gen_highpart (DImode, operands[1]);
> - split_double_concat (TImode, operands[0], operands[2], operands[4]);
> + rtx temp = gen_highpart (DImode, operands[1]);
> + split_double_concat (TImode, operands[0], operands[2], temp);
> DONE;
> })
>
> @@ -3821,8 +3821,8 @@
> "&& reload_completed"
> [(const_int 0)]
> {
> - operands[4] = gen_highpart (SImode, operands[1]);
> - split_double_concat (DImode, operands[0], operands[2], operands[4]);
> + rtx temp = gen_highpart (SImode, operands[1]);
> + split_double_concat (DImode, operands[0], operands[2], temp);
> DONE;
> })
>
> @@ -5117,10 +5117,11 @@
> [(const_int 0)]
> {
> rtx bits = GEN_INT (<MODE_SIZE> * BITS_PER_UNIT - 1);
> + rtx lowpart, highpart;
>
> - split_double_mode (<DWI>mode, &operands[0], 1, &operands[3], &operands[4]);
> + split_double_mode (<DWI>mode, &operands[0], 1, &lowpart, &highpart);
>
> - emit_move_insn (operands[3], operands[1]);
> + emit_move_insn (lowpart, operands[1]);
>
> /* Generate a cltd if possible and doing so it profitable. */
> if ((optimize_function_for_size_p (cfun) || TARGET_USE_CLTD)
> @@ -5134,7 +5135,7 @@
> emit_move_insn (operands[2], operands[1]);
> emit_insn (gen_ashr<mode>3_cvt (operands[2], operands[2], bits));
> }
> - emit_move_insn (operands[4], operands[2]);
> + emit_move_insn (highpart, operands[2]);
> DONE;
> })
>
> @@ -5191,29 +5192,30 @@
> [(const_int 0)]
> {
> rtx bits = GEN_INT (<MODE_SIZE> * BITS_PER_UNIT - 1);
> + rtx lowpart, highpart;
>
> - split_double_mode (<DWI>mode, &operands[0], 1, &operands[3], &operands[4]);
> + split_double_mode (<DWI>mode, &operands[0], 1, &lowpart, &highpart);
>
> - if (REGNO (operands[3]) != REGNO (operands[1]))
> - emit_move_insn (operands[3], operands[1]);
> + if (REGNO (lowpart) != REGNO (operands[1]))
> + emit_move_insn (lowpart, operands[1]);
>
> rtx src = operands[1];
> - if (REGNO (operands[3]) == AX_REG)
> - src = operands[3];
> + if (REGNO (lowpart) == AX_REG)
> + src = lowpart;
>
> /* Generate a cltd if possible and doing so it profitable. */
> if ((optimize_function_for_size_p (cfun) || TARGET_USE_CLTD)
> && REGNO (src) == AX_REG
> - && REGNO (operands[4]) == DX_REG)
> + && REGNO (highpart) == DX_REG)
> {
> - emit_insn (gen_ashr<mode>3_cvt (operands[4], src, bits));
> + emit_insn (gen_ashr<mode>3_cvt (highpart, src, bits));
> DONE;
> }
>
> - if (REGNO (operands[4]) != REGNO (operands[1]))
> - emit_move_insn (operands[4], operands[1]);
> + if (REGNO (highpart) != REGNO (operands[1]))
> + emit_move_insn (highpart, operands[1]);
>
> - emit_insn (gen_ashr<mode>3_cvt (operands[4], operands[4], bits));
> + emit_insn (gen_ashr<mode>3_cvt (highpart, highpart, bits));
> DONE;
> })
>
> @@ -6317,16 +6319,16 @@
> || TARGET_AVX512VL)"
> [(const_int 0)]
> {
> - operands[3] = lowpart_subreg (<ssevecmode>mode, operands[0], <MODE>mode);
> - operands[4] = lowpart_subreg (V4SImode, operands[0], <MODE>mode);
> + rtx lowpart_sse = lowpart_subreg (<ssevecmode>mode, operands[0],
> <MODE>mode);
> + rtx lowpart_v4si = lowpart_subreg (V4SImode, operands[0], <MODE>mode);
>
> - emit_insn (gen_sse2_loadld (operands[4],
> + emit_insn (gen_sse2_loadld (lowpart_v4si,
> CONST0_RTX (V4SImode), operands[1]));
>
> if (<ssevecmode>mode == V4SFmode)
> - emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
> + emit_insn (gen_floatv4siv4sf2 (lowpart_sse, lowpart_v4si));
> else
> - emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
> + emit_insn (gen_sse2_cvtdq2pd (lowpart_sse, lowpart_v4si));
> DONE;
> })
>
> @@ -13330,32 +13332,32 @@
> [(const_int:DWIH 0)]
> {
> bool emit_insn_deleted_note_p = false;
> + rtx lowpart[3], highpart[3];
> + split_double_mode (<DWI>mode, &operands[0], 3, lowpart, highpart);
>
> - split_double_mode (<DWI>mode, &operands[0], 3, &operands[0], &operands[3]);
> -
> - if (operands[2] == const0_rtx)
> - emit_move_insn (operands[0], const0_rtx);
> - else if (operands[2] == constm1_rtx)
> + if (lowpart[2] == const0_rtx)
> + emit_move_insn (lowpart[0], const0_rtx);
> + else if (lowpart[2] == constm1_rtx)
> {
> - if (!rtx_equal_p (operands[0], operands[1]))
> - emit_move_insn (operands[0], operands[1]);
> + if (!rtx_equal_p (lowpart[0], lowpart[1]))
> + emit_move_insn (lowpart[0], lowpart[1]);
> else
> emit_insn_deleted_note_p = true;
> }
> else
> - ix86_expand_binary_operator (AND, <MODE>mode, &operands[0],
> TARGET_APX_NDD);
> + ix86_expand_binary_operator (AND, <MODE>mode, &lowpart[0],
> TARGET_APX_NDD);
>
> - if (operands[5] == const0_rtx)
> - emit_move_insn (operands[3], const0_rtx);
> - else if (operands[5] == constm1_rtx)
> + if (highpart[2] == const0_rtx)
> + emit_move_insn (highpart[0], const0_rtx);
> + else if (highpart[2] == constm1_rtx)
> {
> - if (!rtx_equal_p (operands[3], operands[4]))
> - emit_move_insn (operands[3], operands[4]);
> + if (!rtx_equal_p (highpart[0], highpart[1]))
> + emit_move_insn (highpart[0], highpart[1]);
> else if (emit_insn_deleted_note_p)
> emit_note (NOTE_INSN_DELETED);
> }
> else
> - ix86_expand_binary_operator (AND, <MODE>mode, &operands[3],
> TARGET_APX_NDD);
> + ix86_expand_binary_operator (AND, <MODE>mode, &highpart[0],
> TARGET_APX_NDD);
>
> DONE;
> }
> @@ -14388,45 +14390,42 @@
> and operands[0] == operands[1], which requires a NOTE_INSN_DELETED. */
> bool emit_insn_deleted_note_p = false;
>
> - split_double_mode (<DWI>mode, &operands[0], 3, &operands[0], &operands[3]);
> + rtx lowpart[3], highpart[3];
> + split_double_mode (<DWI>mode, &operands[0], 3, lowpart, highpart);
>
> - if (operands[2] == const0_rtx)
> + if (lowpart[2] == const0_rtx)
> {
> - if (!rtx_equal_p (operands[0], operands[1]))
> - emit_move_insn (operands[0], operands[1]);
> + if (!rtx_equal_p (lowpart[0], lowpart[1]))
> + emit_move_insn (lowpart[0], lowpart[1]);
> else
> emit_insn_deleted_note_p = true;
> }
> - else if (operands[2] == constm1_rtx)
> + else if (lowpart[2] == constm1_rtx)
> {
> if (<CODE> == IOR)
> - emit_move_insn (operands[0], constm1_rtx);
> + emit_move_insn (lowpart[0], constm1_rtx);
> else
> - ix86_expand_unary_operator (NOT, <MODE>mode, &operands[0],
> - TARGET_APX_NDD);
> + ix86_expand_unary_operator (NOT, <MODE>mode, lowpart, TARGET_APX_NDD);
> }
> else
> - ix86_expand_binary_operator (<CODE>, <MODE>mode, &operands[0],
> - TARGET_APX_NDD);
> + ix86_expand_binary_operator (<CODE>, <MODE>mode, lowpart,
> TARGET_APX_NDD);
>
> - if (operands[5] == const0_rtx)
> + if (highpart[2] == const0_rtx)
> {
> - if (!rtx_equal_p (operands[3], operands[4]))
> - emit_move_insn (operands[3], operands[4]);
> + if (!rtx_equal_p (highpart[0], highpart[1]))
> + emit_move_insn (highpart[0], highpart[1]);
> else if (emit_insn_deleted_note_p)
> emit_note (NOTE_INSN_DELETED);
> }
> - else if (operands[5] == constm1_rtx)
> + else if (highpart[2] == constm1_rtx)
> {
> if (<CODE> == IOR)
> - emit_move_insn (operands[3], constm1_rtx);
> + emit_move_insn (highpart[0], constm1_rtx);
> else
> - ix86_expand_unary_operator (NOT, <MODE>mode, &operands[3],
> - TARGET_APX_NDD);
> + ix86_expand_unary_operator (NOT, <MODE>mode, highpart,
> TARGET_APX_NDD);
> }
> else
> - ix86_expand_binary_operator (<CODE>, <MODE>mode, &operands[3],
> - TARGET_APX_NDD);
> + ix86_expand_binary_operator (<CODE>, <MODE>mode, highpart,
> TARGET_APX_NDD);
>
> DONE;
> }
> --
> 2.54.0
>
--
BR,
Hongtao
