public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: juzhe.zhong@rivai.ai
To: gcc-patches@gcc.gnu.org
Cc: kito.cheng@gmail.com, Ju-Zhe Zhong <juzhe.zhong@rivai.ai>,
	kito-cheng <kito.cheng@sifive.com>
Subject: [PATCH] RISC-V: Fix bugs of ternary integer and floating-point ternary intrinsics.
Date: Tue, 14 Mar 2023 10:23:31 +0800	[thread overview]
Message-ID: <20230314022331.105558-1-juzhe.zhong@rivai.ai> (raw)

From: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>

Co-authored-by: kito-cheng <kito.cheng@sifive.com>
Co-authored-by: kito-cheng <kito.cheng@gmail.com>

This patch is fixing the bugs reported by @kito.

  // vnmsub.vx vd, rs1, vs2, vm    # vd[i] = -(x[rs1] * vd[i]) + vs2[i]
  // vd = -(vb * a) + vc
  //    = -(3 * 1) + 10
  //    = 7
  // GCC wrongly optmize this pattern to `3 - 10` due to we write wrong RTL
  // pattern.
  // vd = (3 * 1) - 10
  //    = 3 - 10
  //    = -7
  // NOTE: GCC optimized (vb * a) - vc to vb - vc

Signed-off-by: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
Co-authored-by: kito-cheng <kito.cheng@sifive.com>
Co-authored-by: kito-cheng <kito.cheng@gmail.com>

gcc/ChangeLog:

        * config/riscv/vector.md: Correct ternary patterns.

gcc/testsuite/ChangeLog:

        * gcc.target/riscv/rvv/base/bug-3.c: New test.
        * gcc.target/riscv/rvv/base/bug-4.c: New test.
        * gcc.target/riscv/rvv/base/bug-5.c: New test.

---
 gcc/config/riscv/vector.md                    | 80 +++++++++----------
 .../gcc.target/riscv/rvv/base/bug-3.c         | 22 +++++
 .../gcc.target/riscv/rvv/base/bug-4.c         | 22 +++++
 .../gcc.target/riscv/rvv/base/bug-5.c         | 22 +++++
 4 files changed, 106 insertions(+), 40 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/base/bug-3.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/base/bug-4.c
 create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/base/bug-5.c

diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md
index 5f765cdbacb..27c5cccb451 100644
--- a/gcc/config/riscv/vector.md
+++ b/gcc/config/riscv/vector.md
@@ -4160,10 +4160,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "register_operand")
 	    (mult:VI
 	      (match_operand:VI 2 "register_operand")
-	      (match_operand:VI 3 "register_operand"))
-	    (match_operand:VI 4 "register_operand"))
+	      (match_operand:VI 3 "register_operand")))
 	  (match_operand:VI 5 "register_operand")))]
   "TARGET_VECTOR"
 {
@@ -4185,10 +4185,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "register_operand"       " vr, vr,   vr")
 	    (mult:VI
 	      (match_operand:VI 2 "register_operand"     "  0,  0,   vr")
-	      (match_operand:VI 3 "register_operand"     " vr, vr,   vr"))
-	    (match_operand:VI 4 "register_operand"       " vr, vr,   vr"))
+	      (match_operand:VI 3 "register_operand"     " vr, vr,   vr")))
 	  (match_dup 2)))]
   "TARGET_VECTOR"
   "@
@@ -4215,10 +4215,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "register_operand"       "  0,  0,   vr")
 	    (mult:VI
 	      (match_operand:VI 2 "register_operand"     " vr, vr,   vr")
-	      (match_operand:VI 3 "register_operand"     " vr, vr,   vr"))
-	    (match_operand:VI 4 "register_operand"       "  0,  0,   vr"))
+	      (match_operand:VI 3 "register_operand"     " vr, vr,   vr")))
 	  (match_dup 4)))]
   "TARGET_VECTOR"
   "@
@@ -4245,10 +4245,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "vector_arith_operand"   "   vr,   vi,   vr,   vr,   vr")
 	    (mult:VI
 	      (match_operand:VI 2 "register_operand"     "   vr,   vr,   vi,   vr,   vr")
-	      (match_operand:VI 3 "register_operand"     "   vr,   vr,   vr,   vi,   vr"))
-	    (match_operand:VI 4 "vector_arith_operand"   "   vr,   vi,   vr,   vr,   vr"))
+	      (match_operand:VI 3 "register_operand"     "   vr,   vr,   vr,   vi,   vr")))
 	  (match_operand:VI 5 "register_operand"         "    0,   vr,   vr,   vr,   vr")))]
   "TARGET_VECTOR
    && !rtx_equal_p (operands[2], operands[5])
@@ -4296,11 +4296,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI_QHS
+	    (match_operand:VI_QHS 4 "register_operand")
 	    (mult:VI_QHS
 	      (vec_duplicate:VI_QHS
 	        (match_operand:<VEL> 2 "reg_or_int_operand"))
-	      (match_operand:VI_QHS 3 "register_operand"))
-	    (match_operand:VI_QHS 4 "register_operand"))
+	      (match_operand:VI_QHS 3 "register_operand")))
 	  (match_operand:VI_QHS 5 "register_operand")))]
   "TARGET_VECTOR"
 {
@@ -4319,11 +4319,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "register_operand"        " vr, vr,   vr")
 	    (mult:VI
 	      (vec_duplicate:VI
 	        (match_operand:<VEL> 2 "register_operand" "  r,  r,   vr"))
-	      (match_operand:VI 3 "register_operand"      "  0,  0,   vr"))
-	    (match_operand:VI 4 "register_operand"        " vr, vr,   vr"))
+	      (match_operand:VI 3 "register_operand"      "  0,  0,   vr")))
 	  (match_dup 3)))]
   "TARGET_VECTOR"
   "@
@@ -4350,11 +4350,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "register_operand"        "  0,  0,   vr")
 	    (mult:VI
 	      (vec_duplicate:VI
 	        (match_operand:<VEL> 2 "register_operand" "  r,  r,   vr"))
-	      (match_operand:VI 3 "register_operand"      " vr, vr,   vr"))
-	    (match_operand:VI 4 "register_operand"        "  0,  0,   vr"))
+	      (match_operand:VI 3 "register_operand"      " vr, vr,   vr")))
 	  (match_dup 4)))]
   "TARGET_VECTOR"
   "@
@@ -4381,11 +4381,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI
+	    (match_operand:VI 4 "vector_arith_operand"    "   vr,   vi,   vr,   vr")
 	    (mult:VI
 	      (vec_duplicate:VI
 	        (match_operand:<VEL> 2 "register_operand" "    r,    r,    r,    r"))
-	      (match_operand:VI 3 "register_operand"      "   vr,   vr,   vi,   vr"))
-	    (match_operand:VI 4 "vector_arith_operand"    "   vr,   vi,   vr,   vr"))
+	      (match_operand:VI 3 "register_operand"      "   vr,   vr,   vi,   vr")))
 	  (match_operand:VI 5 "register_operand"          "    0,   vr,   vr,   vr")))]
   "TARGET_VECTOR
    && !rtx_equal_p (operands[3], operands[5])
@@ -4428,11 +4428,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI_D
+	    (match_operand:VI_D 4 "register_operand")
 	    (mult:VI_D
 	      (vec_duplicate:VI_D
 	        (match_operand:<VEL> 2 "reg_or_int_operand"))
-	      (match_operand:VI_D 3 "register_operand"))
-	    (match_operand:VI_D 4 "register_operand"))
+	      (match_operand:VI_D 3 "register_operand")))
 	  (match_operand:VI_D 5 "register_operand")))]
   "TARGET_VECTOR"
 {
@@ -4463,12 +4463,12 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI_D
+	    (match_operand:VI_D 4 "register_operand"           " vr, vr,   vr")
 	    (mult:VI_D
 	      (vec_duplicate:VI_D
 	        (sign_extend:<VEL>
 	          (match_operand:<VSUBEL> 2 "register_operand" "  r,  r,   vr")))
-	      (match_operand:VI_D 3 "register_operand"         "  0,  0,   vr"))
-	    (match_operand:VI_D 4 "register_operand"           " vr, vr,   vr"))
+	      (match_operand:VI_D 3 "register_operand"         "  0,  0,   vr")))
 	  (match_dup 3)))]
   "TARGET_VECTOR"
   "@
@@ -4495,12 +4495,12 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI_D
+	    (match_operand:VI_D 4 "register_operand"           "  0,  0,   vr")
 	    (mult:VI_D
 	      (vec_duplicate:VI_D
 	        (sign_extend:<VEL>
 	          (match_operand:<VSUBEL> 2 "register_operand" "  r,  r,   vr")))
-	      (match_operand:VI_D 3 "register_operand"         " vr, vr,   vr"))
-	    (match_operand:VI_D 4 "register_operand"           "  0,  0,   vr"))
+	      (match_operand:VI_D 3 "register_operand"         " vr, vr,   vr")))
 	  (match_dup 4)))]
   "TARGET_VECTOR"
   "@
@@ -4527,12 +4527,12 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VI_D
+	    (match_operand:VI_D 4 "vector_arith_operand"       "   vr,   vr,   vr,   vr")
 	    (mult:VI_D
 	      (vec_duplicate:VI_D
 	        (sign_extend:<VEL>
 	          (match_operand:<VSUBEL> 2 "register_operand" "    r,    r,    r,    r")))
-	      (match_operand:VI_D 3 "register_operand"         "   vr,   vr,   vr,   vr"))
-	    (match_operand:VI_D 4 "vector_arith_operand"       "   vr,   vr,   vr,   vr"))
+	      (match_operand:VI_D 3 "register_operand"         "   vr,   vr,   vr,   vr")))
 	  (match_operand:VI_D 5 "register_operand"             "    0,   vr,   vr,   vr")))]
   "TARGET_VECTOR
    && !rtx_equal_p (operands[3], operands[5])
@@ -5033,10 +5033,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "register_operand")
 	    (mult:VF
 	      (match_operand:VF 2 "register_operand")
-	      (match_operand:VF 3 "register_operand"))
-	    (match_operand:VF 4 "register_operand"))
+	      (match_operand:VF 3 "register_operand")))
 	  (match_operand:VF 5 "register_operand")))]
   "TARGET_VECTOR"
 {
@@ -5058,10 +5058,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "register_operand"       " vr, vr,   vr")
 	    (mult:VF
 	      (match_operand:VF 2 "register_operand"     "  0,  0,   vr")
-	      (match_operand:VF 3 "register_operand"     " vr, vr,   vr"))
-	    (match_operand:VF 4 "register_operand"       " vr, vr,   vr"))
+	      (match_operand:VF 3 "register_operand"     " vr, vr,   vr")))
 	  (match_dup 2)))]
   "TARGET_VECTOR"
   "@
@@ -5088,10 +5088,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "register_operand"       "  0,  0,   vr")
 	    (mult:VF
 	      (match_operand:VF 2 "register_operand"     " vr, vr,   vr")
-	      (match_operand:VF 3 "register_operand"     " vr, vr,   vr"))
-	    (match_operand:VF 4 "register_operand"       "  0,  0,   vr"))
+	      (match_operand:VF 3 "register_operand"     " vr, vr,   vr")))
 	  (match_dup 4)))]
   "TARGET_VECTOR"
   "@
@@ -5118,10 +5118,10 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "vector_arith_operand"   "   vr,   vr")
 	    (mult:VF
 	      (match_operand:VF 2 "register_operand"     "   vr,   vr")
-	      (match_operand:VF 3 "register_operand"     "   vr,   vr"))
-	    (match_operand:VF 4 "vector_arith_operand"   "   vr,   vr"))
+	      (match_operand:VF 3 "register_operand"     "   vr,   vr")))
 	  (match_operand:VF 5 "register_operand"         "    0,   vr")))]
   "TARGET_VECTOR
    && !rtx_equal_p (operands[2], operands[5])
@@ -5153,11 +5153,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "register_operand")
 	    (mult:VF
 	      (vec_duplicate:VF
 	        (match_operand:<VEL> 2 "register_operand"))
-	      (match_operand:VF 3 "register_operand"))
-	    (match_operand:VF 4 "register_operand"))
+	      (match_operand:VF 3 "register_operand")))
 	  (match_operand:VF 5 "register_operand")))]
   "TARGET_VECTOR"
 {})
@@ -5174,11 +5174,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "register_operand"        " vr, vr,   vr")
 	    (mult:VF
 	      (vec_duplicate:VF
 	        (match_operand:<VEL> 2 "register_operand" "  f,  f,   vr"))
-	      (match_operand:VF 3 "register_operand"      "  0,  0,   vr"))
-	    (match_operand:VF 4 "register_operand"        " vr, vr,   vr"))
+	      (match_operand:VF 3 "register_operand"      "  0,  0,   vr")))
 	  (match_dup 3)))]
   "TARGET_VECTOR"
   "@
@@ -5205,11 +5205,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "register_operand"        "  0,  0,   vr")
 	    (mult:VF
 	      (vec_duplicate:VF
 	        (match_operand:<VEL> 2 "register_operand" "  f,  f,   vr"))
-	      (match_operand:VF 3 "register_operand"      " vr, vr,   vr"))
-	    (match_operand:VF 4 "register_operand"        "  0,  0,   vr"))
+	      (match_operand:VF 3 "register_operand"      " vr, vr,   vr")))
 	  (match_dup 4)))]
   "TARGET_VECTOR"
   "@
@@ -5236,11 +5236,11 @@
 	     (reg:SI VL_REGNUM)
 	     (reg:SI VTYPE_REGNUM)] UNSPEC_VPREDICATE)
 	  (plus_minus:VF
+	    (match_operand:VF 4 "vector_arith_operand"    "   vr,  vr")
 	    (mult:VF
 	      (vec_duplicate:VF
 	        (match_operand:<VEL> 2 "register_operand" "    f,   f"))
-	      (match_operand:VF 3 "register_operand"      "   vr,  vr"))
-	    (match_operand:VF 4 "vector_arith_operand"    "   vr,  vr"))
+	      (match_operand:VF 3 "register_operand"      "   vr,  vr")))
 	  (match_operand:VF 5 "register_operand"          "    0,  vr")))]
   "TARGET_VECTOR
    && !rtx_equal_p (operands[3], operands[5])
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/base/bug-3.c b/gcc/testsuite/gcc.target/riscv/rvv/base/bug-3.c
new file mode 100644
index 00000000000..35b76892598
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/base/bug-3.c
@@ -0,0 +1,22 @@
+/* { dg-do run } */
+/* { dg-options "-O2" } */
+
+#include "riscv_vector.h"
+#include <stdio.h>
+
+int main()
+{
+  int32_t a = 1;
+  int32_t b[1] = {3};
+  int32_t c[1] = {10};
+  int32_t d[1] = {0};
+  vint32m1_t vb = __riscv_vle32_v_i32m1 (b, 1);
+  vint32m1_t vc = __riscv_vle32_v_i32m1 (c, 1);
+  vint32m1_t vd = __riscv_vnmsub_vx_i32m1 (vb, a, vc, 1);
+  __riscv_vse32_v_i32m1 (d, vd, 1);
+  if (d[0] != 7){
+      printf("d[0] should be 7, but got %d\n", d[0]);
+      __builtin_abort ();
+  }
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/base/bug-4.c b/gcc/testsuite/gcc.target/riscv/rvv/base/bug-4.c
new file mode 100644
index 00000000000..62dd3f50e44
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/base/bug-4.c
@@ -0,0 +1,22 @@
+/* { dg-do run } */
+/* { dg-options "-O2" } */
+
+#include "riscv_vector.h"
+#include <stdio.h>
+
+int main()
+{
+  float a = 1.0;
+  float b[1] = {3.0};
+  float c[1] = {10.0};
+  float d[1] = {0.0};
+  vfloat32m1_t vb = __riscv_vle32_v_f32m1 (b, 1);
+  vfloat32m1_t vc = __riscv_vle32_v_f32m1 (c, 1);
+  vfloat32m1_t vd = __riscv_vfnmsub_vf_f32m1 (vb, a, vc, 1);
+  __riscv_vse32_v_f32m1 (d, vd, 1);
+  if (d[0] != 7.0){
+      printf("d[0] should be 7.0, but got %f\n", d[0]);
+      __builtin_abort ();
+  }
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/base/bug-5.c b/gcc/testsuite/gcc.target/riscv/rvv/base/bug-5.c
new file mode 100644
index 00000000000..e43f85a0730
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/base/bug-5.c
@@ -0,0 +1,22 @@
+/* { dg-do run } */
+/* { dg-options "-O2" } */
+
+#include "riscv_vector.h"
+#include <stdio.h>
+
+int main()
+{
+  float a = 1.0;
+  float b[1] = {3.0};
+  float c[1] = {10.0};
+  float d[1] = {0.0};
+  vfloat32m1_t vb = __riscv_vle32_v_f32m1 (b, 1);
+  vfloat32m1_t vc = __riscv_vle32_v_f32m1 (c, 1);
+  vfloat32m1_t vd = __riscv_vfmsub_vf_f32m1 (vb, a, vc, 1);
+  __riscv_vse32_v_f32m1 (d, vd, 1);
+  if (d[0] != -7.0){
+      printf("d[0] should be -7.0, but got %f\n", d[0]);
+      __builtin_abort ();
+  }
+  return 0;
+}
-- 
2.36.3


             reply	other threads:[~2023-03-14  2:23 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-03-14  2:23 juzhe.zhong [this message]
2023-03-14  2:37 ` Kito Cheng
2023-03-14 14:50   ` Jeff Law
2023-03-15  5:23 juzhe.zhong
2023-03-15  6:37 juzhe.zhong
2023-03-19 17:03 ` Jeff Law

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20230314022331.105558-1-juzhe.zhong@rivai.ai \
    --to=juzhe.zhong@rivai.ai \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=kito.cheng@gmail.com \
    --cc=kito.cheng@sifive.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
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).