RISC-V: Optimize a special case of VLA SLP

Message ID 20231123125910.1889955-1-juzhe.zhong@rivai.ai
State Unresolved
Headers
Series RISC-V: Optimize a special case of VLA SLP |

Checks

Context Check Description
snail/gcc-patch-check warning Git am fail log

Commit Message

juzhe.zhong@rivai.ai Nov. 23, 2023, 12:59 p.m. UTC
  When working on fixing bugs of zvl1024b. I notice a special VLA SLP case
can be better optimized.

v = vec_perm (op1, op2, { nunits - 1, nunits, nunits + 1, ... })

Before this patch, we are using genriec approach (vrgather):

vid
vadd.vx
vrgather
vmsgeu
vrgather

With this patch, we use vec_extract + slide1up:

scalar = vec_extract (last element of op1)
v = slide1up (op2, scalar)

I am gonna to run testing zvl128b/zvl256b/zvl512b/zvl1024b of both RV32 and RV64.

Ok for trunk if no regression on those testing above ?

	PR target/112599

gcc/ChangeLog:

	* config/riscv/riscv-v.cc (shuffle_extract_and_slide1up_patterns):
	(expand_vec_perm_const_1):

gcc/testsuite/ChangeLog:

	* gcc.target/riscv/rvv/autovec/pr112599-2.c: New test.

---
 gcc/config/riscv/riscv-v.cc                   | 33 ++++++++++++
 .../gcc.target/riscv/rvv/autovec/pr112599-2.c | 51 +++++++++++++++++++
 2 files changed, 84 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
  

Comments

juzhe.zhong@rivai.ai Nov. 23, 2023, 11:20 p.m. UTC | #1
Thanks Robin.

Send V2: 
https://gcc.gnu.org/pipermail/gcc-patches/2023-November/638033.html 
with adding changeLog since I realize changlog issue in V1:
gcc/ChangeLog:
	* config/riscv/riscv-v.cc (shuffle_extract_and_slide1up_patterns):
	(expand_vec_perm_const_1):


Tested on zvl128b/zvl256b/zvl512b/zvl1024b on both RV32 and RV64 no regression.

Hope we can land it on GCC-14.



juzhe.zhong@rivai.ai
 
From: Robin Dapp
Date: 2023-11-23 22:58
To: Juzhe-Zhong; gcc-patches
CC: rdapp.gcc; kito.cheng; kito.cheng; jeffreyalaw
Subject: Re: [PATCH] RISC-V: Optimize a special case of VLA SLP
LGTM (and harmless enough) but I'd rather wait for a second look or a
maintainer's OK as we're past stage 1 and it's not a real bugfix.
(On top, it's Thanksgiving so not many people will even notice).
 
On a related note, this should probably be a middle-end optimization
but before a variable-index vec extract most likely nobody bothered. 
 
Regards
Robin
  
juzhe.zhong@rivai.ai Nov. 24, 2023, 3:04 a.m. UTC | #2
I don't think loop vectorizer can do more optimization here.

GCC pass to vec_perm_const targethook vec_perm <,,(nunits - 1, nunits , nuits + 1, ....)>
to handle that. It's very target dependent. We can't do more about that.

For RVV, it's better transform this case into vec_extract + vec_shl_insert.
However, for ARM SVE, it's not. ARM SVE has a dedicated instruction to handle that (trn),
it's better to pass vec_perm_const with this permute indice for ARM SVE.



juzhe.zhong@rivai.ai
 
From: Robin Dapp
Date: 2023-11-23 22:58
To: Juzhe-Zhong; gcc-patches
CC: rdapp.gcc; kito.cheng; kito.cheng; jeffreyalaw
Subject: Re: [PATCH] RISC-V: Optimize a special case of VLA SLP
LGTM (and harmless enough) but I'd rather wait for a second look or a
maintainer's OK as we're past stage 1 and it's not a real bugfix.
(On top, it's Thanksgiving so not many people will even notice).
 
On a related note, this should probably be a middle-end optimization
but before a variable-index vec extract most likely nobody bothered. 
 
Regards
Robin
  

Patch

diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index 72b96d8339d..2ee1bf5191b 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -3232,6 +3232,37 @@  shuffle_bswap_pattern (struct expand_vec_perm_d *d)
   return true;
 }
 
+/* Recognize the pattern that can be shuffled by vec_extract and slide1up
+   approach.  */
+
+static bool
+shuffle_extract_and_slide1up_patterns (struct expand_vec_perm_d *d)
+{
+  poly_uint64 nunits = GET_MODE_NUNITS (d->vmode);
+
+  /* Recognize { nunits - 1, nunits, nunits + 1, ... }.  */
+  if (!d->perm.series_p (0, 2, nunits - 1, 2)
+      || !d->perm.series_p (1, 2, nunits, 2))
+    return false;
+
+  /* Success! */
+  if (d->testing_p)
+    return true;
+
+  /* Extract the last element of the first vector.  */
+  scalar_mode smode = GET_MODE_INNER (d->vmode);
+  rtx tmp = gen_reg_rtx (smode);
+  emit_vec_extract (tmp, d->op0, nunits - 1);
+
+  /* Insert the scalar into element 0.  */
+  unsigned int unspec
+    = FLOAT_MODE_P (d->vmode) ? UNSPEC_VFSLIDE1UP : UNSPEC_VSLIDE1UP;
+  insn_code icode = code_for_pred_slide (unspec, d->vmode);
+  rtx ops[] = {d->target, d->op1, tmp};
+  emit_vlmax_insn (icode, BINARY_OP, ops);
+  return true;
+}
+
 /* Recognize the pattern that can be shuffled by generic approach.  */
 
 static bool
@@ -3310,6 +3341,8 @@  expand_vec_perm_const_1 (struct expand_vec_perm_d *d)
 	    return true;
 	  if (shuffle_bswap_pattern (d))
 	    return true;
+	  if (shuffle_extract_and_slide1up_patterns (d))
+	    return true;
 	  if (shuffle_generic_patterns (d))
 	    return true;
 	  return false;
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
new file mode 100644
index 00000000000..fd87565b054
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112599-2.c
@@ -0,0 +1,51 @@ 
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv_zvl1024b -mabi=lp64d -O3" } */
+
+struct s { struct s *n; } *p;
+struct s ss;
+#define MAX     10
+struct s sss[MAX];
+int count = 0;
+
+int look( struct s *p, struct s **pp )
+{
+    for ( ; p; p = p->n )
+        ;
+    *pp = p;
+    count++;
+    return( 1 );
+}
+
+void sub( struct s *p, struct s **pp )
+{
+   for ( ; look( p, pp ); ) {
+        if ( p )
+            p = p->n;
+        else
+            break;
+   }
+}
+
+int
+foo(void)
+{
+    struct s *pp;
+    struct s *next;
+    int i;
+
+    p = &ss;
+    next = p;
+    for ( i = 0; i < MAX; i++ ) {
+        next->n = &sss[i];
+        next = next->n;
+    }
+    next->n = 0;
+
+    sub( p, &pp );
+    if (count != MAX+2)
+      __builtin_abort ();
+  return 0;
+}
+
+/* { dg-final { scan-assembler-not {vrgather} } } */
+/* { dg-final { scan-assembler-times {vslide1up\.vx} 1 } } */