* [PATCH] RISC-V: split to allow formation of sh[123]add before divw
@ 2022-11-08 19:56 Philipp Tomsich
2022-11-18 19:37 ` Jeff Law
0 siblings, 1 reply; 3+ messages in thread
From: Philipp Tomsich @ 2022-11-08 19:56 UTC (permalink / raw)
To: gcc-patches
Cc: Kito Cheng, Vineet Gupta, Palmer Dabbelt, Christoph Muellner,
Jeff Law, Philipp Tomsich
When using strength-reduction, we will reduce a multiplication to a
sequence of shifts and adds. If this is performed with 32-bit types
and followed by a division, the lack of w-form sh[123]add will make
combination impossible and lead to a slli + addw being generated.
Split the sequence with the knowledge that a w-form div will perform
implicit sign-extensions.
gcc/ChangeLog:
* config/riscv/bitmanip.md: Add a define_split to optimize
slliw + addiw + divw into sh[123]add + divw.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/zba-shNadd-05.c: New test.
Signed-off-by: Philipp Tomsich <philipp.tomsich@vrull.eu>
---
gcc/config/riscv/bitmanip.md | 17 +++++++++++++++++
gcc/testsuite/gcc.target/riscv/zba-shNadd-05.c | 11 +++++++++++
2 files changed, 28 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/riscv/zba-shNadd-05.c
diff --git a/gcc/config/riscv/bitmanip.md b/gcc/config/riscv/bitmanip.md
index 30dabdf8ddc..726a07b0d90 100644
--- a/gcc/config/riscv/bitmanip.md
+++ b/gcc/config/riscv/bitmanip.md
@@ -39,6 +39,23 @@
[(set_attr "type" "bitmanip")
(set_attr "mode" "<X:MODE>")])
+; When using strength-reduction, we will reduce a multiplication to a
+; sequence of shifts and adds. If this is performed with 32-bit types
+; and followed by a division, the lack of w-form sh[123]add will make
+; combination impossible and lead to a slli + addw being generated.
+; Split the sequence with the knowledge that a w-form div will perform
+; implicit sign-extensions.
+(define_split
+ [(set (match_operand:DI 0 "register_operand")
+ (sign_extend:DI (div:SI (plus:SI (subreg:SI (ashift:DI (match_operand:DI 1 "register_operand")
+ (match_operand:QI 2 "imm123_operand")) 0)
+ (subreg:SI (match_operand:DI 3 "register_operand") 0))
+ (subreg:SI (match_operand:DI 4 "register_operand") 0))))
+ (clobber (match_operand:DI 5 "register_operand"))]
+ "TARGET_64BIT && TARGET_ZBA"
+ [(set (match_dup 5) (plus:DI (ashift:DI (match_dup 1) (match_dup 2)) (match_dup 3)))
+ (set (match_dup 0) (sign_extend:DI (div:SI (subreg:SI (match_dup 5) 0) (subreg:SI (match_dup 4) 0))))])
+
(define_insn "*shNadduw"
[(set (match_operand:DI 0 "register_operand" "=r")
(plus:DI
diff --git a/gcc/testsuite/gcc.target/riscv/zba-shNadd-05.c b/gcc/testsuite/gcc.target/riscv/zba-shNadd-05.c
new file mode 100644
index 00000000000..271c3a8c0ac
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/zba-shNadd-05.c
@@ -0,0 +1,11 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gc_zba -mabi=lp64" } */
+/* { dg-skip-if "" { *-*-* } { "-O0" "-O1" "-Os" "-Oz" "-Og" } } */
+
+long long f(int a, int b)
+{
+ return (a * 3) / b;
+}
+
+/* { dg-final { scan-assembler-times "sh1add\t" 1 } } */
+/* { dg-final { scan-assembler-times "divw\t" 1 } } */
--
2.34.1
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] RISC-V: split to allow formation of sh[123]add before divw
2022-11-08 19:56 [PATCH] RISC-V: split to allow formation of sh[123]add before divw Philipp Tomsich
@ 2022-11-18 19:37 ` Jeff Law
2022-11-18 19:56 ` Philipp Tomsich
0 siblings, 1 reply; 3+ messages in thread
From: Jeff Law @ 2022-11-18 19:37 UTC (permalink / raw)
To: Philipp Tomsich, gcc-patches
Cc: Kito Cheng, Vineet Gupta, Palmer Dabbelt, Christoph Muellner, Jeff Law
On 11/8/22 12:56, Philipp Tomsich wrote:
> When using strength-reduction, we will reduce a multiplication to a
> sequence of shifts and adds. If this is performed with 32-bit types
> and followed by a division, the lack of w-form sh[123]add will make
> combination impossible and lead to a slli + addw being generated.
>
> Split the sequence with the knowledge that a w-form div will perform
> implicit sign-extensions.
>
> gcc/ChangeLog:
>
> * config/riscv/bitmanip.md: Add a define_split to optimize
> slliw + addiw + divw into sh[123]add + divw.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/zba-shNadd-05.c: New test.
OK. I won't complain about the subregs on this one :-)
jeff
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] RISC-V: split to allow formation of sh[123]add before divw
2022-11-18 19:37 ` Jeff Law
@ 2022-11-18 19:56 ` Philipp Tomsich
0 siblings, 0 replies; 3+ messages in thread
From: Philipp Tomsich @ 2022-11-18 19:56 UTC (permalink / raw)
To: Jeff Law
Cc: gcc-patches, Kito Cheng, Vineet Gupta, Palmer Dabbelt,
Christoph Muellner, Jeff Law
[-- Attachment #1: Type: text/plain, Size: 881 bytes --]
Applied to master. Thanks!
--Philipp.
On Fri, 18 Nov 2022 at 20:37, Jeff Law <jeffreyalaw@gmail.com> wrote:
>
> On 11/8/22 12:56, Philipp Tomsich wrote:
> > When using strength-reduction, we will reduce a multiplication to a
> > sequence of shifts and adds. If this is performed with 32-bit types
> > and followed by a division, the lack of w-form sh[123]add will make
> > combination impossible and lead to a slli + addw being generated.
> >
> > Split the sequence with the knowledge that a w-form div will perform
> > implicit sign-extensions.
> >
> > gcc/ChangeLog:
> >
> > * config/riscv/bitmanip.md: Add a define_split to optimize
> > slliw + addiw + divw into sh[123]add + divw.
> >
> > gcc/testsuite/ChangeLog:
> >
> > * gcc.target/riscv/zba-shNadd-05.c: New test.
>
> OK. I won't complain about the subregs on this one :-)
>
>
> jeff
>
>
>
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2022-11-18 19:56 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-11-08 19:56 [PATCH] RISC-V: split to allow formation of sh[123]add before divw Philipp Tomsich
2022-11-18 19:37 ` Jeff Law
2022-11-18 19:56 ` Philipp Tomsich
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).