[x86] Pre-reload splitter to transform and;cmp into not;test.
Commit Message
A common idiom for testing if a specific set of bits is set in a value
is to use "(X & Y) == Y", which on x86 results in an AND followed by a
CMP. A slightly improved implementation is to instead use (~X & Y)==0,
that uses a NOT and a TEST (or ANDN where available); still two "fast"
instructions, but typically shorter especially if Y is an immediate
constant. Because the above transformation would require more gimple
statements in SSA, and may only be a win on targets with flags registers,
it isn't performed by the middle-end, instead leaving this choice to
the backend.
As an example, here's the change in code generation for pr91400-1.c
[which now requires a tweak to its dg-final clauses].
Before:
movl __cpu_model+12(%rip), %eax
andl $68, %eax // 3 bytes
cmpl $68, %eax // 3 bytes
sete %al
ret
After:
movl __cpu_model+12(%rip), %eax
notl %eax // 2 bytes
testb $68, %al // 2 bytes
sete %al
ret
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?
2022-05-26 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/i386.md (*test<mode>_not): New define_insn_and_split
to split a combined "and;cmp" sequence into "not;test".
gcc/testsuite/ChangeLog
* gcc.target/i386/pr91400-1.c: Update for improved code generation.
* gcc.target/i386/pr91400-2.c: Likewise.
* gcc.target/i386/testnot-1.c: New test case.
* gcc.target/i386/testnot-2.c: Likewise.
Thanks in advance,
Roger
--
Comments
On Thu, May 26, 2022 at 8:41 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> A common idiom for testing if a specific set of bits is set in a value
> is to use "(X & Y) == Y", which on x86 results in an AND followed by a
> CMP. A slightly improved implementation is to instead use (~X & Y)==0,
> that uses a NOT and a TEST (or ANDN where available); still two "fast"
> instructions, but typically shorter especially if Y is an immediate
> constant. Because the above transformation would require more gimple
> statements in SSA, and may only be a win on targets with flags registers,
> it isn't performed by the middle-end, instead leaving this choice to
> the backend.
>
> As an example, here's the change in code generation for pr91400-1.c
> [which now requires a tweak to its dg-final clauses].
>
> Before:
> movl __cpu_model+12(%rip), %eax
> andl $68, %eax // 3 bytes
> cmpl $68, %eax // 3 bytes
> sete %al
> ret
>
> After:
> movl __cpu_model+12(%rip), %eax
> notl %eax // 2 bytes
> testb $68, %al // 2 bytes
> sete %al
> ret
>
>
> 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?
>
>
> 2022-05-26 Roger Sayle <roger@nextmovesoftware.com>
>
> gcc/ChangeLog
> * config/i386/i386.md (*test<mode>_not): New define_insn_and_split
> to split a combined "and;cmp" sequence into "not;test".
>
> gcc/testsuite/ChangeLog
> * gcc.target/i386/pr91400-1.c: Update for improved code generation.
> * gcc.target/i386/pr91400-2.c: Likewise.
> * gcc.target/i386/testnot-1.c: New test case.
> * gcc.target/i386/testnot-2.c: Likewise.
OK.
Thanks,
Uros.
@@ -9716,6 +9716,27 @@
operands[2] = gen_rtx_AND (mode, val, immed_wide_int_const (mask, mode));
})
+;; Split and;cmp (as optimized by combine) into not;test
+;; Except when TARGET_BMI provides andn (*andn_<mode>_ccno).
+(define_insn_and_split "*test<mode>_not"
+ [(set (reg:CCZ FLAGS_REG)
+ (compare:CCZ
+ (and:SWI
+ (not:SWI (match_operand:SWI 0 "register_operand"))
+ (match_operand:SWI 1 "<nonmemory_szext_operand>"))
+ (const_int 0)))]
+ "ix86_pre_reload_split ()
+ && (!TARGET_BMI || !REG_P (operands[1]))"
+ "#"
+ "&& 1"
+ [(set (match_dup 2) (not:SWI (match_dup 0)))
+ (set (reg:CCZ FLAGS_REG)
+ (compare:CCZ (and:SWI (match_dup 2) (match_dup 1))
+ (const_int 0)))]
+{
+ operands[2] = gen_reg_rtx (<MODE>mode);
+})
+
;; Convert HImode/SImode test instructions with immediate to QImode ones.
;; i386 does not allow to encode test with 8bit sign extended immediate, so
;; this is relatively important trick.
@@ -1,8 +1,8 @@
/* PR target/91400 */
/* { dg-do compile } */
/* { dg-options "-O2" } */
-/* { dg-final { scan-assembler-times "andl" 1 } } */
-/* { dg-final { scan-assembler-times "cmpl" 1 } } */
+/* { dg-final { scan-assembler-times "notl" 1 } } */
+/* { dg-final { scan-assembler-times "testb" 1 } } */
/* { dg-final { scan-assembler-times "sete" 1 } } */
/* { dg-final { scan-assembler-not "cmove" } } */
@@ -1,8 +1,8 @@
/* PR target/91400 */
/* { dg-do compile } */
/* { dg-options "-O2" } */
-/* { dg-final { scan-assembler-times "andl" 1 } } */
-/* { dg-final { scan-assembler-times "cmpl" 1 } } */
+/* { dg-final { scan-assembler-times "notl" 1 } } */
+/* { dg-final { scan-assembler-times "testb" 1 } } */
/* { dg-final { scan-assembler-times "sete" 1 } } */
/* { dg-final { scan-assembler-not "cmove" } } */
new file mode 100644
@@ -0,0 +1,24 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+int foo(int x)
+{
+ return (x & 1234) == 1234;
+}
+
+int foos(short x)
+{
+ return (x & 1234) == 1234;
+}
+
+int fooc(char x)
+{
+ return (x & 123) == 123;
+}
+
+int fool(long long x)
+{
+ return (x & 1234) == 1234;
+}
+
+/* { dg-final { scan-assembler-not "cmp" } } */
new file mode 100644
@@ -0,0 +1,24 @@
+/* { dg-do compile { target { ! ia32 } } } */
+/* { dg-options "-O2" } */
+
+int foo(int x, int y)
+{
+ return (x & y) == y;
+}
+
+int foos(short x, short y)
+{
+ return (x & y) == y;
+}
+
+int fooc(char x, char y)
+{
+ return (x & y) == y;
+}
+
+int fool(long long x, long long y)
+{
+ return (x & y) == y;
+}
+
+/* { dg-final { scan-assembler-not "cmp" } } */