alpha: Introduce target specific store_data_bypass_p function [PR105209]

This patch introduces alpha-specific version of store_data_bypass_p that
ignores TRAP_IF that would result in assertion failure (and internal
compiler error) in the generic store_data_bypass_p function.

While at it, also remove ev4_ist_c reservation, store_data_bypass_p
can handle the patterns with multiple sets since some time ago.

2022-06-17  Uroš Bizjak  <ubizjak@gmail.com>

gcc/ChangeLog:

	PR target/105209
	* config/alpha/alpha-protos.h (alpha_store_data_bypass_p): New.
	* config/alpha/alpha.cc (alpha_store_data_bypass_p): New function.
	(alpha_store_data_bypass_p_1): Ditto.
	* config/alpha/ev4.md: Use alpha_store_data_bypass_p instead
	of generic store_data_bypass_p.
	(ev4_ist_c): Remove insn reservation.

gcc/testsuite/ChangeLog:

	PR target/105209
	* gcc.target/alpha/pr105209.c: New test.

(cherry picked from commit cc378e6557)

gcc/config/alpha/alpha-protos.h

@@ -73,6 +73,8 @@ extern void alpha_end_function (FILE *, const char *, tree);
 
 extern bool alpha_find_lo_sum_using_gp (rtx);
 
+extern int alpha_store_data_bypass_p (rtx_insn *, rtx_insn *);
+
 #ifdef REAL_VALUE_TYPE
 extern int check_float_value (machine_mode, REAL_VALUE_TYPE *, int);
 #endif

gcc/config/alpha/alpha.cc

@@ -7564,6 +7564,75 @@ alpha_does_function_need_gp (void)
   return 0;
 }
 
+/* Helper function for alpha_store_data_bypass_p, handle just a single SET
+   IN_SET.  */
+
+static bool
+alpha_store_data_bypass_p_1 (rtx_insn *out_insn, rtx in_set)
+{
+  if (!MEM_P (SET_DEST (in_set)))
+    return false;
+
+  rtx out_set = single_set (out_insn);
+  if (out_set)
+    return !reg_mentioned_p (SET_DEST (out_set), SET_DEST (in_set));
+
+  rtx out_pat = PATTERN (out_insn);
+  if (GET_CODE (out_pat) != PARALLEL)
+    return false;
+
+  for (int i = 0; i < XVECLEN (out_pat, 0); i++)
+    {
+      rtx out_exp = XVECEXP (out_pat, 0, i);
+
+      if (GET_CODE (out_exp) == CLOBBER || GET_CODE (out_exp) == USE
+	  || GET_CODE (out_exp) == TRAP_IF)
+	continue;
+
+      gcc_assert (GET_CODE (out_exp) == SET);
+
+      if (reg_mentioned_p (SET_DEST (out_exp), SET_DEST (in_set)))
+	return false;
+    }
+
+  return true;
+}
+
+/* True if the dependency between OUT_INSN and IN_INSN is on the store
+   data not the address operand(s) of the store.  IN_INSN and OUT_INSN
+   must be either a single_set or a PARALLEL with SETs inside.
+
+   This alpha-specific version of store_data_bypass_p ignores TRAP_IF
+   that would result in assertion failure (and internal compiler error)
+   in the generic store_data_bypass_p function.  */
+
+int
+alpha_store_data_bypass_p (rtx_insn *out_insn, rtx_insn *in_insn)
+{
+  rtx in_set = single_set (in_insn);
+  if (in_set)
+    return alpha_store_data_bypass_p_1 (out_insn, in_set);
+
+  rtx in_pat = PATTERN (in_insn);
+  if (GET_CODE (in_pat) != PARALLEL)
+    return false;
+
+  for (int i = 0; i < XVECLEN (in_pat, 0); i++)
+    {
+      rtx in_exp = XVECEXP (in_pat, 0, i);
+
+      if (GET_CODE (in_exp) == CLOBBER || GET_CODE (in_exp) == USE
+	  || GET_CODE (in_exp) == TRAP_IF)
+	continue;
+
+      gcc_assert (GET_CODE (in_exp) == SET);
+
+      if (!alpha_store_data_bypass_p_1 (out_insn, in_exp))
+	return false;
+    }
+
+  return true;
+}
 
 /* Helper function to set RTX_FRAME_RELATED_P on instructions, including
    sequences.  */

gcc/config/alpha/ev4.md

@@ -44,14 +44,7 @@
 ; Stores can issue before the data (but not address) is ready.
 (define_insn_reservation "ev4_ist" 1
   (and (eq_attr "tune" "ev4")
-       (eq_attr "type" "ist"))
-  "ev4_ib1+ev4_abox")
-
-; ??? Separate from ev4_ist because store_data_bypass_p can't handle
-; the patterns with multiple sets, like store-conditional.
-(define_insn_reservation "ev4_ist_c" 1
-  (and (eq_attr "tune" "ev4")
-       (eq_attr "type" "st_c"))
+       (eq_attr "type" "ist,st_c"))
   "ev4_ib1+ev4_abox")
 
 (define_insn_reservation "ev4_fst" 1
@@ -110,7 +103,7 @@
 (define_bypass 0
   "ev4_iaddlog,ev4_shiftcm,ev4_icmp"
   "ev4_ist"
-  "store_data_bypass_p")
+  "alpha_store_data_bypass_p")
 
 ; Multiplies use a non-pipelined imul unit.  Also, "no [ebox] insn can
 ; be issued exactly three cycles before an integer multiply completes".
@@ -121,7 +114,7 @@
 	    (eq_attr "opsize" "si")))
   "ev4_ib0+ev4_imul,ev4_imul*18,ev4_ebox")
 
-(define_bypass 20 "ev4_imulsi" "ev4_ist" "store_data_bypass_p")
+(define_bypass 20 "ev4_imulsi" "ev4_ist" "alpha_store_data_bypass_p")
 
 (define_insn_reservation "ev4_imuldi" 23
   (and (eq_attr "tune" "ev4")
@@ -129,7 +122,7 @@
 	    (eq_attr "opsize" "!si")))
   "ev4_ib0+ev4_imul,ev4_imul*20,ev4_ebox")
 
-(define_bypass 22 "ev4_imuldi" "ev4_ist" "store_data_bypass_p")
+(define_bypass 22 "ev4_imuldi" "ev4_ist" "alpha_store_data_bypass_p")
 
 ; Most FP insns have a 6 cycle latency, but with a 4 cycle bypass back in.
 (define_insn_reservation "ev4_fpop" 6

gcc/testsuite/gcc.target/alpha/pr105209.c

@@ -0,0 +1,26 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftrapv -mcpu=ev4" } */
+
+typedef struct tnode_t {
+        struct tnode_t *tn_left, *tn_right;
+        int v_quad;
+} tnode_t;
+
+int constant_addr(const tnode_t *, long *);
+int constant_addr(const tnode_t *tn, long *offsp)
+{
+        long offs1 = 0, offs2 = 0;
+
+        if (tn->v_quad > 0) {
+                offs1 = tn->v_quad;
+                return 0;
+        } else if (tn->v_quad > -1) {
+                offs2 = tn->tn_right->v_quad;
+                if (!constant_addr(tn->tn_left, &offs1))
+                        return 0;
+        } else {
+                return 0;
+        }
+        *offsp = offs1 + offs2;
+        return 1;
+}