[x86] PR rtl-optimization/101617: Use neg/sbb in ix86_expand_int_movcc.
Commit Message
This patch resolves PR rtl-optimization/101617 where we should generate
the exact same code for (X ? -1 : 1) as we do for ((X ? -1 : 0) | 1).
The cause of the current difference on x86_64 is actually in
ix86_expand_int_movcc that doesn't know that negl;sbbl can be used
to create a -1/0 result depending on whether the input is zero/nonzero.
So for Andrew Pinski's test case:
int f1(int i)
{
return i ? -1 : 1;
}
GCC currently generates:
f1: cmpl $1, %edi
sbbl %eax, %eax // x ? 0 : -1
andl $2, %eax // x ? 0 : 2
subl $1, %eax // x ? -1 : 1
ret
but with the attached patch, now generates:
f1: negl %edi
sbbl %eax, %eax // x ? -1 : 0
orl $1, %eax // x ? -1 : 1
ret
To implement this I needed to add two expanders to i386.md to generate
the required instructions (in both SImode and DImode) matching the
pre-existing define_insns of the same name.
This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check, both with and without --target_board=unix{-m32},
with no new failures. Ok for mainline?
2022-05-30 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR rtl-optimization/101617
* config/i386/i386-expand.cc (ix86_expand_int_movcc): Add a
special case (indicated by negate_cc_compare_p) to generate a
-1/0 mask using neg;sbb.
* config/i386/i386.md (x86_neg<mode>_ccc): New define_expand
to generate an *x86_neg<mode>_ccc instruction.
(x86_mov<mode>cc_0_m1_neg): Likewise, a new define_expand to
generate a *x86_mov<mode>cc_0_m1_neg instruction.
gcc/testsuite/ChangeLog
PR rtl-optimization/101617
* gcc.target/i386/pr101617.c: New test case.
Thanks in advance,
Roger
--
Comments
On Mon, May 30, 2022 at 7:50 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> This patch resolves PR rtl-optimization/101617 where we should generate
> the exact same code for (X ? -1 : 1) as we do for ((X ? -1 : 0) | 1).
> The cause of the current difference on x86_64 is actually in
> ix86_expand_int_movcc that doesn't know that negl;sbbl can be used
> to create a -1/0 result depending on whether the input is zero/nonzero.
>
> So for Andrew Pinski's test case:
>
> int f1(int i)
> {
> return i ? -1 : 1;
> }
>
> GCC currently generates:
>
> f1: cmpl $1, %edi
> sbbl %eax, %eax // x ? 0 : -1
> andl $2, %eax // x ? 0 : 2
> subl $1, %eax // x ? -1 : 1
> ret
>
> but with the attached patch, now generates:
>
> f1: negl %edi
> sbbl %eax, %eax // x ? -1 : 0
> orl $1, %eax // x ? -1 : 1
> ret
>
> To implement this I needed to add two expanders to i386.md to generate
> the required instructions (in both SImode and DImode) matching the
> pre-existing define_insns of the same name.
>
> This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> and make -k check, both with and without --target_board=unix{-m32},
> with no new failures. Ok for mainline?
>
>
> 2022-05-30 Roger Sayle <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
> PR rtl-optimization/101617
> * config/i386/i386-expand.cc (ix86_expand_int_movcc): Add a
> special case (indicated by negate_cc_compare_p) to generate a
> -1/0 mask using neg;sbb.
> * config/i386/i386.md (x86_neg<mode>_ccc): New define_expand
> to generate an *x86_neg<mode>_ccc instruction.
> (x86_mov<mode>cc_0_m1_neg): Likewise, a new define_expand to
> generate a *x86_mov<mode>cc_0_m1_neg instruction.
>
> gcc/testsuite/ChangeLog
> PR rtl-optimization/101617
> * gcc.target/i386/pr101617.c: New test case.
LGTM.
Thanks,
Uros.
@@ -3142,6 +3142,7 @@ ix86_expand_int_movcc (rtx operands[])
rtx compare_op;
machine_mode mode = GET_MODE (operands[0]);
bool sign_bit_compare_p = false;
+ bool negate_cc_compare_p = false;
rtx op0 = XEXP (operands[1], 0);
rtx op1 = XEXP (operands[1], 1);
rtx op2 = operands[2];
@@ -3188,16 +3189,48 @@ ix86_expand_int_movcc (rtx operands[])
HOST_WIDE_INT cf = INTVAL (op3);
HOST_WIDE_INT diff;
+ if ((mode == SImode
+ || (TARGET_64BIT && mode == DImode))
+ && (GET_MODE (op0) == SImode
+ || (TARGET_64BIT && GET_MODE (op0) == DImode)))
+ {
+ /* Special case x != 0 ? -1 : y. */
+ if (code == NE && op1 == const0_rtx && ct == -1)
+ {
+ negate_cc_compare_p = true;
+ std::swap (ct, cf);
+ code = EQ;
+ }
+ else if (code == EQ && op1 == const0_rtx && cf == -1)
+ negate_cc_compare_p = true;
+ }
+
diff = ct - cf;
/* Sign bit compares are better done using shifts than we do by using
sbb. */
if (sign_bit_compare_p
+ || negate_cc_compare_p
|| ix86_expand_carry_flag_compare (code, op0, op1, &compare_op))
{
/* Detect overlap between destination and compare sources. */
rtx tmp = out;
- if (!sign_bit_compare_p)
+ if (negate_cc_compare_p)
+ {
+ if (GET_MODE (op0) == DImode)
+ emit_insn (gen_x86_negdi_ccc (gen_reg_rtx (DImode), op0));
+ else
+ emit_insn (gen_x86_negsi_ccc (gen_reg_rtx (SImode),
+ gen_lowpart (SImode, op0)));
+
+ tmp = gen_reg_rtx (mode);
+ if (mode == DImode)
+ emit_insn (gen_x86_movdicc_0_m1_neg (tmp));
+ else
+ emit_insn (gen_x86_movsicc_0_m1_neg (gen_lowpart (SImode,
+ tmp)));
+ }
+ else if (!sign_bit_compare_p)
{
rtx flags;
bool fpcmp = false;
@@ -11189,6 +11189,14 @@
[(set_attr "type" "negnot")
(set_attr "mode" "<MODE>")])
+(define_expand "x86_neg<mode>_ccc"
+ [(parallel
+ [(set (reg:CCC FLAGS_REG)
+ (ne:CCC (match_operand:SWI48 1 "register_operand")
+ (const_int 0)))
+ (set (match_operand:SWI48 0 "register_operand")
+ (neg:SWI48 (match_dup 1)))])])
+
(define_insn "*negqi_ext<mode>_2"
[(set (zero_extract:SWI248
(match_operand:SWI248 0 "register_operand" "+Q")
@@ -20700,6 +20708,12 @@
(set_attr "mode" "<MODE>")
(set_attr "length_immediate" "0")])
+(define_expand "x86_mov<mode>cc_0_m1_neg"
+ [(parallel
+ [(set (match_operand:SWI48 0 "register_operand")
+ (neg:SWI48 (ltu:SWI48 (reg:CCC FLAGS_REG) (const_int 0))))
+ (clobber (reg:CC FLAGS_REG))])])
+
(define_split
[(set (match_operand:SWI48 0 "register_operand")
(neg:SWI48
new file mode 100644
@@ -0,0 +1,19 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+int f(int i)
+{
+ int t = i ? -1 : 0;
+ return t | 1;
+}
+
+int f1(int i)
+{
+ int t = i ? -1 : 1;
+ return t;
+}
+
+/* { dg-final { scan-assembler-times "negl" 2 } } */
+/* { dg-final { scan-assembler-times "sbbl" 2 } } */
+/* { dg-final { scan-assembler-times "orl" 2 } } */
+/* { dg-final { scan-assembler-not "cmpl" } } */
+