This patch refactors and fixes an issue where arm_mve_dlstp_check_dec_counter was making an assumption about the form of what a candidate for a dec_insn
    should be, which caused an ICE.
This dec_insn is the instruction that decreases the loop counter inside a
    decrementing loop and we expect it to have the following form:
    (set (reg CONDCOUNT)
         (plus (reg CONDCOUNT)
               (const_int)))

Where CONDCOUNT is the loop counter, and const int is the negative constant
    used to decrement it.

This patch also improves our search for a valid dec_insn. Before this patch we'd only look for a dec_insn inside the loop header if the loop latch was empty. We now also search the loop header if the loop latch is not empty but the last instruction is not a valid dec_insn. This could potentially be improved
    to search all instructions inside the loop latch.

    gcc/ChangeLog:

* config/arm/arm.cc (check_dec_insn): New helper function containing
            code hoisted from...
(arm_mve_dlstp_check_dec_counter): ... here. Use check_dec_insn to
            check the validity of the candidate dec_insn.

    gcc/testsuite/ChangeLog:

            * gcc.targer/arm/mve/dlstp-loop-form.c: New test.

On 31/07/2024 15:15, Christophe Lyon wrote:
Because I tested with a toolchain configured for cortex-m85, which has mve.fp enabled by default, which means I didn't realize the testcase required arm_v8_1m_mve_fp_ok instead of arm_v8_1m_mve_ok.

Addressed that now.

Thanks, I thought you meant you ran the testsuite with -mcpu=cortex-m85 in RUNTESTFLAGS.

To be fair, that's not a terrible assumption. But what I did was I configured the toolchain (and single multilib) for I ran them in a build configured for armv8.1-m.main+mve.fp+fp.dp and fpu=auto (and float-abi=hard).


Regarding the patch, did you consider making the new check_dec_insn helper return an rtx (NULL or dec_set) instead of bool? I think it would save a call to single_set when computing decrementnum, but that's nitpicking.

Yeah I had also contemplated that, I'm OK either way, doesn't look too bad with the rtx return. See attached.


Thanks,

Christophe
diff --git a/gcc/config/arm/arm.cc b/gcc/config/arm/arm.cc
index 
92cd168e65937ef7350477464e8b0becf85bceed..363a972170b37275372bb8bf30d510876021c8c0
 100644
--- a/gcc/config/arm/arm.cc
+++ b/gcc/config/arm/arm.cc
@@ -35214,6 +35214,32 @@ arm_mve_dlstp_check_inc_counter (loop *loop, rtx_insn* 
vctp_insn,
   return vctp_insn;
 }
 
+/* Helper function to 'arm_mve_dlstp_check_dec_counter' to make sure DEC_INSN
+   is of the expected form:
+   (set (reg a) (plus (reg a) (const_int)))
+   where (reg a) is the same as CONDCOUNT.
+   Return a rtx with the set if it is in the right format or NULL_RTX
+   otherwise.  */
+
+static rtx
+check_dec_insn (rtx_insn *dec_insn, rtx condcount)
+{
+  if (!NONDEBUG_INSN_P (dec_insn))
+    return NULL_RTX;
+  rtx dec_set = single_set (dec_insn);
+  if (!dec_set
+      || !REG_P (SET_DEST (dec_set))
+      || GET_CODE (SET_SRC (dec_set)) != PLUS
+      || !REG_P (XEXP (SET_SRC (dec_set), 0))
+      || !CONST_INT_P (XEXP (SET_SRC (dec_set), 1))
+      || REGNO (SET_DEST (dec_set))
+         != REGNO (XEXP (SET_SRC (dec_set), 0))
+      || REGNO (SET_DEST (dec_set)) != REGNO (condcount))
+    return NULL_RTX;
+
+  return dec_set;
+}
+
 /* Helper function to `arm_mve_loop_valid_for_dlstp`.  In the case of a
    counter that is decrementing, ensure that it is decrementing by the
    right amount in each iteration and that the target condition is what
@@ -35230,30 +35256,19 @@ arm_mve_dlstp_check_dec_counter (loop *loop, 
rtx_insn* vctp_insn,
      loop latch.  Here we simply need to verify that this counter is the same
      reg that is also used in the vctp_insn and that it is not otherwise
      modified.  */
-  rtx_insn *dec_insn = BB_END (loop->latch);
+  rtx dec_set = check_dec_insn (BB_END (loop->latch), condcount);
   /* If not in the loop latch, try to find the decrement in the loop header.  
*/
-  if (!NONDEBUG_INSN_P (dec_insn))
+  if (dec_set == NULL_RTX)
   {
     df_ref temp = df_bb_regno_only_def_find (loop->header, REGNO (condcount));
     /* If we haven't been able to find the decrement, bail out.  */
     if (!temp)
       return NULL;
-    dec_insn = DF_REF_INSN (temp);
-  }
-
-  rtx dec_set = single_set (dec_insn);
+    dec_set = check_dec_insn (DF_REF_INSN (temp), condcount);
 
-  /* Next, ensure that it is a PLUS of the form:
-     (set (reg a) (plus (reg a) (const_int)))
-     where (reg a) is the same as condcount.  */
-  if (!dec_set
-      || !REG_P (SET_DEST (dec_set))
-      || !REG_P (XEXP (SET_SRC (dec_set), 0))
-      || !CONST_INT_P (XEXP (SET_SRC (dec_set), 1))
-      || REGNO (SET_DEST (dec_set))
-         != REGNO (XEXP (SET_SRC (dec_set), 0))
-      || REGNO (SET_DEST (dec_set)) != REGNO (condcount))
-    return NULL;
+    if (dec_set == NULL_RTX)
+      return NULL;
+  }
 
   decrementnum = INTVAL (XEXP (SET_SRC (dec_set), 1));
 
diff --git a/gcc/testsuite/gcc.target/arm/mve/dlstp-loop-form.c 
b/gcc/testsuite/gcc.target/arm/mve/dlstp-loop-form.c
new file mode 100644
index 
0000000000000000000000000000000000000000..a1b26873d7908035c726e3724c91b186c697bc60
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/mve/dlstp-loop-form.c
@@ -0,0 +1,27 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_v8_1m_mve_fp_ok } */
+/* { dg-options "-Ofast" } */
+/* { dg-add-options arm_v8_1m_mve_fp } */
+#pragma GCC arm "arm_mve_types.h"
+#pragma GCC arm "arm_mve.h" false
+typedef __attribute__((aligned(2))) float16x8_t e;
+mve_pred16_t c(long d) { return __builtin_mve_vctp16qv8bi(d); }
+int f();
+void n() {
+  int g, h, *i, j;
+  mve_pred16_t k;
+  e acc;
+  e l;
+  e m;
+  for (;;) {
+    j = g;
+    acc[g];
+    for (; h < g; h += 8) {
+      k = c(j);
+      acc = vfmsq_m(acc, l, m, k);
+      j -= 8;
+    }
+    i[g] = f(acc);
+  }
+}
+

Reply via email to