Hexagon (target/hexagon) fix bug in fLSBNEW*

Change fLSBNEW/fLSBNEW0/fLSBNEW1 from copy to "x & 1"
Remove gen_logical_not function
Clean up fLSBNEWNOT to use andi-1 followed by xori-1

Test cases added to tests/tcg/hexagon/misc.c

Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <1622589584-22571-2-git-send-email-tsimpson@quicinc.com>

target/hexagon/macros.h

@@ -239,33 +239,26 @@ static inline void gen_pred_cancel(TCGv pred, int slot_num)
 #endif
 
 #ifdef QEMU_GENERATE
-#define fLSBNEW(PVAL)   tcg_gen_mov_tl(LSB, (PVAL))
+#define fLSBNEW(PVAL)   tcg_gen_andi_tl(LSB, (PVAL), 1)
-#define fLSBNEW0        tcg_gen_mov_tl(LSB, hex_new_pred_value[0])
+#define fLSBNEW0        tcg_gen_andi_tl(LSB, hex_new_pred_value[0], 1)
-#define fLSBNEW1        tcg_gen_mov_tl(LSB, hex_new_pred_value[1])
+#define fLSBNEW1        tcg_gen_andi_tl(LSB, hex_new_pred_value[1], 1)
 #else
-#define fLSBNEW(PVAL)   (PVAL)
+#define fLSBNEW(PVAL)   ((PVAL) & 1)
-#define fLSBNEW0        new_pred_value(env, 0)
+#define fLSBNEW0        (env->new_pred_value[0] & 1)
-#define fLSBNEW1        new_pred_value(env, 1)
+#define fLSBNEW1        (env->new_pred_value[1] & 1)
 #endif
 
 #ifdef QEMU_GENERATE
-static inline void gen_logical_not(TCGv dest, TCGv src)
-{
-    TCGv one = tcg_const_tl(1);
-    TCGv zero = tcg_const_tl(0);
-
-    tcg_gen_movcond_tl(TCG_COND_NE, dest, src, zero, zero, one);
-
-    tcg_temp_free(one);
-    tcg_temp_free(zero);
-}
 #define fLSBOLDNOT(VAL) \
     do { \
         tcg_gen_andi_tl(LSB, (VAL), 1); \
         tcg_gen_xori_tl(LSB, LSB, 1); \
     } while (0)
 #define fLSBNEWNOT(PNUM) \
-    gen_logical_not(LSB, (PNUM))
+    do { \
+        tcg_gen_andi_tl(LSB, (PNUM), 1); \
+        tcg_gen_xori_tl(LSB, LSB, 1); \
+    } while (0)
 #else
 #define fLSBNEWNOT(PNUM) (!fLSBNEW(PNUM))
 #define fLSBOLDNOT(VAL) (!fLSBOLD(VAL))

target/hexagon/op_helper.c

@@ -128,11 +128,6 @@ void HELPER(debug_start_packet)(CPUHexagonState *env)
     }
 }
 
-static int32_t new_pred_value(CPUHexagonState *env, int pnum)
-{
-    return env->new_pred_value[pnum];
-}
-
 /* Checks for bookkeeping errors between disassembly context and runtime */
 void HELPER(debug_check_store_width)(CPUHexagonState *env, int slot, int check)
 {

tests/tcg/hexagon/misc.c

@@ -181,6 +181,19 @@ static inline void S4_storeirifnew_io(void *p, int pred)
                : "p0", "memory");
 }
 
+static int L2_ploadrifnew_pi(void *p, int pred)
+{
+  int result;
+  asm volatile("%0 = #31\n\t"
+               "{\n\t"
+               "    p0 = cmp.eq(%1, #1)\n\t"
+               "    if (!p0.new) %0 = memw(%2++#4)\n\t"
+               "}\n\t"
+               : "=r"(result) : "r"(pred), "r"(p)
+               : "p0");
+  return result;
+}
+
 /*
  * Test that compound-compare-jump is executed in 2 parts
  * First we have to do all the compares in the packet and
@@ -298,8 +311,24 @@ static int auto_and(void)
     return retval;
 }
 
+void test_lsbnew(void)
+{
+    int result;
+
+    asm("r0 = #2\n\t"
+        "r1 = #5\n\t"
+        "{\n\t"
+        "    p0 = r0\n\t"
+        "    if (p0.new) r1 = #3\n\t"
+        "}\n\t"
+        "%0 = r1\n\t"
+        : "=r"(result) :: "r0", "r1", "p0");
+    check(result, 5);
+}
+
 int main()
 {
+    int res;
     long long res64;
     int pred;
 
@@ -394,6 +423,12 @@ int main()
     S4_storeirifnew_io(&array[8], 1);
     check(array[9], 9);
 
+    memcpy(array, init, sizeof(array));
+    res = L2_ploadrifnew_pi(&array[6], 0);
+    check(res, 6);
+    res = L2_ploadrifnew_pi(&array[7], 1);
+    check(res, 31);
+
     int x = cmpnd_cmp_jump();
     check(x, 12);
 
@@ -406,7 +441,7 @@ int main()
     check((int)pair, 5);
     check((int)(pair >> 32), 7);
 
-    int res = test_clrtnew(1, 7);
+    res = test_clrtnew(1, 7);
     check(res, 0);
     res = test_clrtnew(2, 7);
     check(res, 7);
@@ -422,6 +457,8 @@ int main()
     res = auto_and();
     check(res, 0);
 
+    test_lsbnew();
+
     puts(err ? "FAIL" : "PASS");
     return err;
 }