[PATCH, v2] Recognize a missed usage of a sbfiz instruction
Luis Machado
luis.machado@linaro.org
Tue Feb 6 15:04:00 GMT 2018
Thanks for the feedback Kyrill. I've adjusted the v2 patch based on your
suggestions and re-tested the changes. Everything is still sane.
Since this is ARM-specific and fairly specific, i wonder if it would be
reasonable to consider it for inclusion at the current stage.
Regards,
Luis
Changes in v2:
- Added more restrictive predicates to operands 2, 3 and 4.
- Removed pattern conditional.
- Switched to long long for 64-bit signed integer for the testcase.
---
A customer reported the following missed opportunities to combine a couple
instructions into a sbfiz.
int sbfiz32 (int x)
{
return x << 29 >> 10;
}
long long sbfiz64 (long long x)
{
return x << 58 >> 20;
}
This gets converted to the following pattern:
(set (reg:SI 98)
(ashift:SI (sign_extend:SI (reg:HI 0 x0 [ xD.3334 ]))
(const_int 6 [0x6])))
Currently, gcc generates the following:
sbfiz32:
lsl x0, x0, 29
asr x0, x0, 10
ret
sbfiz64:
lsl x0, x0, 58
asr x0, x0, 20
ret
It could generate this instead:
sbfiz32:
sbfiz w0, w0, 19, 3
ret
sbfiz64::
sbfiz x0, x0, 38, 6
ret
The unsigned versions already generate ubfiz for the same code, so the lack of
a sbfiz pattern may have been an oversight.
This particular sbfiz pattern shows up in both CPU2006 (~ 80 hits) and
CPU2017 (~ 280 hits). It's not a lot, but seems beneficial in any case. No
significant performance differences, probably due to the small number of
occurrences.
2018-02-06 Luis Machado <luis.machado@linaro.org>
gcc/
* config/aarch64/aarch64.md (*ashift<mode>_extv_bfiz): New pattern.
2018-02-06 Luis Machado <luis.machado@linaro.org>
gcc/testsuite/
* gcc.target/aarch64/lsl_asr_sbfiz.c: New test.
---
gcc/config/aarch64/aarch64.md | 13 +++++++++++++
gcc/testsuite/gcc.target/aarch64/lsl_asr_sbfiz.c | 24 ++++++++++++++++++++++++
2 files changed, 37 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/aarch64/lsl_asr_sbfiz.c
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
index 5a2a930..e8284ae 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -4828,6 +4828,19 @@
[(set_attr "type" "bfx")]
)
+;; Match sbfiz pattern in a shift left + shift right operation.
+
+(define_insn "*ashift<mode>_extv_bfiz"
+ [(set (match_operand:GPI 0 "register_operand" "=r")
+ (ashift:GPI (sign_extract:GPI (match_operand:GPI 1 "register_operand" "r")
+ (match_operand 2 "aarch64_simd_shift_imm_offset_<mode>" "n")
+ (match_operand 3 "aarch64_simd_shift_imm_<mode>" "n"))
+ (match_operand 4 "aarch64_simd_shift_imm_<mode>" "n")))]
+ ""
+ "sbfiz\\t%<w>0, %<w>1, %4, %2"
+ [(set_attr "type" "bfx")]
+)
+
;; When the bit position and width of the equivalent extraction add up to 32
;; we can use a W-reg LSL instruction taking advantage of the implicit
;; zero-extension of the X-reg.
diff --git a/gcc/testsuite/gcc.target/aarch64/lsl_asr_sbfiz.c b/gcc/testsuite/gcc.target/aarch64/lsl_asr_sbfiz.c
new file mode 100644
index 0000000..106433d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/lsl_asr_sbfiz.c
@@ -0,0 +1,24 @@
+/* { dg-do compile } */
+/* { dg-options "-O3" } */
+
+/* Check that a LSL followed by an ASR can be combined into a single SBFIZ
+ instruction. */
+
+/* Using W-reg */
+
+int
+sbfiz32 (int x)
+{
+ return x << 29 >> 10;
+}
+
+/* Using X-reg */
+
+long long
+sbfiz64 (long long x)
+{
+ return x << 58 >> 20;
+}
+
+/* { dg-final { scan-assembler "sbfiz\tw" } } */
+/* { dg-final { scan-assembler "sbfiz\tx" } } */
--
2.7.4
More information about the Gcc-patches
mailing list