This patch improves code generation for the PRU backend.  Committed to
trunk.

Manually expanding into 32-bit comparisons is much more efficient than
the default expansion into word-size comparisons.  Note that word for PRU
is 8-bit.

        PR target/106562

gcc/ChangeLog:

        * config/pru/pru-protos.h (pru_noteq_condition): New
        function declaration.
        * config/pru/pru.cc (pru_noteq_condition): New function.
        * config/pru/pru.md (cbranchdi4): Define new pattern.

gcc/testsuite/ChangeLog:

        * gcc.target/pru/pr106562-1.c: New test.
        * gcc.target/pru/pr106562-2.c: New test.
        * gcc.target/pru/pr106562-3.c: New test.
        * gcc.target/pru/pr106562-4.c: New test.

Signed-off-by: Dimitar Dimitrov <dimi...@dinux.eu>
---
 gcc/config/pru/pru-protos.h               |   1 +
 gcc/config/pru/pru.cc                     |  21 +++
 gcc/config/pru/pru.md                     | 180 ++++++++++++++++++++++
 gcc/testsuite/gcc.target/pru/pr106562-1.c |   9 ++
 gcc/testsuite/gcc.target/pru/pr106562-2.c |   9 ++
 gcc/testsuite/gcc.target/pru/pr106562-3.c |   9 ++
 gcc/testsuite/gcc.target/pru/pr106562-4.c | 159 +++++++++++++++++++
 7 files changed, 388 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/pru/pr106562-1.c
 create mode 100644 gcc/testsuite/gcc.target/pru/pr106562-2.c
 create mode 100644 gcc/testsuite/gcc.target/pru/pr106562-3.c
 create mode 100644 gcc/testsuite/gcc.target/pru/pr106562-4.c

diff --git a/gcc/config/pru/pru-protos.h b/gcc/config/pru/pru-protos.h
index 4b190c98206..517fa02e272 100644
--- a/gcc/config/pru/pru-protos.h
+++ b/gcc/config/pru/pru-protos.h
@@ -52,6 +52,7 @@ extern const char *pru_output_signed_cbranch (rtx *, bool);
 extern const char *pru_output_signed_cbranch_ubyteop2 (rtx *, bool);
 extern const char *pru_output_signed_cbranch_zeroop2 (rtx *, bool);
 
+extern enum rtx_code pru_noteq_condition (enum rtx_code code);
 extern rtx pru_expand_fp_compare (rtx comparison, machine_mode mode);
 
 extern void pru_emit_doloop (rtx *, int);
diff --git a/gcc/config/pru/pru.cc b/gcc/config/pru/pru.cc
index 04eca90b255..0029dcbc6aa 100644
--- a/gcc/config/pru/pru.cc
+++ b/gcc/config/pru/pru.cc
@@ -895,6 +895,27 @@ pru_init_libfuncs (void)
   set_optab_libfunc (udivmod_optab, DImode, "__pruabi_divremull");
 }
 
+/* Given a comparison CODE, return a similar comparison but without
+   the "equals" condition.  In other words, it strips GE/GEU/LE/LEU
+   and instead returns GT/GTU/LT/LTU.  */
+
+enum rtx_code
+pru_noteq_condition (enum rtx_code code)
+{
+  switch (code)
+    {
+    case GT: return GT;
+    case GTU: return GTU;
+    case GE: return GT;
+    case GEU: return GTU;
+    case LT: return LT;
+    case LTU: return LTU;
+    case LE: return LT;
+    case LEU: return LTU;
+    default:
+      gcc_unreachable ();
+    }
+}
 
 /* Emit comparison instruction if necessary, returning the expression
    that holds the compare result in the proper mode.  Return the comparison
diff --git a/gcc/config/pru/pru.md b/gcc/config/pru/pru.md
index 53ffff07708..bdc5ad79ba0 100644
--- a/gcc/config/pru/pru.md
+++ b/gcc/config/pru/pru.md
@@ -1309,6 +1309,186 @@ (define_expand "cbranch<mode>4"
   operands[2] = XEXP (t, 1);
 })
 
+;; Expand the cbranchdi pattern in order to avoid the default
+;; expansion into word_mode operations, which is not efficient for PRU.
+;; In pseudocode this expansion outputs:
+;;
+;; /* EQ */
+;; if (OP1_hi {reverse_condition (cmp)} OP2_hi)
+;;     goto fallthrough
+;; if (OP1_lo {cmp} OP2_lo)
+;;     goto label3
+;; fallthrough:
+;;
+;; /* NE */
+;; if (OP1_hi {cmp} OP2_hi)
+;;     goto label3
+;; if (OP1_lo {cmp} OP2_lo)
+;;     goto label3
+;;
+;; The LT comparisons with zero take one machine instruction to simply
+;; check the sign bit.  The GT comparisons with zero take two - one
+;; to check the sign bit, and one to check for zero.  Hence arrange
+;; the expand such that only LT comparison is used for OP1_HI, because
+;; OP2_HI is const0_rtx.
+;;
+;; The LTU comparisons with zero will be removed by subsequent passes.
+;;
+;;  /* LT/LTU/LE/LEU */
+;;  if (OP1_hi {noteq_condition (cmp)} OP2_hi)
+;;     goto label3             /* DI comparison obviously true.  */
+;;  if (OP1_hi != OP2_hi)
+;;     goto fallthrough                /* DI comparison obviously not true.  */
+;;  if (OP1_lo {unsigned_condition (cmp)} OP2_lo)
+;;     goto label3             /* Comparison was deferred to lo parts.  */
+;;  fallthrough:
+
+;;  /* GT/GTU/GE/GEU */
+;;  if (OP1_hi {reverse_condition (noteq_condition (cmp))} OP2_hi)
+;;     goto fallthrough        /* DI comparison obviously not true.  */
+;;  if (OP1_hi != OP2_hi)
+;;     goto label3             /* DI comparison obviously true.  */
+;;  if (OP1_lo {unsigned_condition (cmp)} OP2_lo)
+;;     goto label3             /* Comparison was deferred to lo parts.  */
+;;  fallthrough:
+
+(define_expand "cbranchdi4"
+  [(set (pc)
+     (if_then_else
+       (match_operator 0 "ordered_comparison_operator"
+        [(match_operand:DI 1 "register_operand")
+         (match_operand:DI 2 "reg_or_ubyte_operand")])
+       (label_ref (match_operand 3 ""))
+       (pc)))]
+  ""
+{
+  const enum rtx_code code = GET_CODE (operands[0]);
+  rtx label3 = operands[3];
+  rtx op1_lo = simplify_gen_subreg (SImode, operands[1], DImode, 0);
+  rtx op1_hi = simplify_gen_subreg (SImode, operands[1], DImode, 4);
+  rtx op2_lo = simplify_gen_subreg (SImode, operands[2], DImode, 0);
+  rtx op2_hi = simplify_gen_subreg (SImode, operands[2], DImode, 4);
+  rtx j;
+
+  if (code == EQ)
+    {
+      rtx label_fallthrough = gen_label_rtx ();
+      rtx label_fallthrough_ref = gen_rtx_LABEL_REF (Pmode, label_fallthrough);
+
+      rtx cond_hi = gen_rtx_fmt_ee (NE, VOIDmode, op1_hi, op2_hi);
+      rtx check_hi = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_hi,
+                                          label_fallthrough_ref, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_hi));
+      JUMP_LABEL (j) = label_fallthrough;
+      LABEL_NUSES (label_fallthrough)++;
+
+      rtx label3_ref = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_lo = gen_rtx_fmt_ee (EQ, VOIDmode, op1_lo, op2_lo);
+      rtx check_lo = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_lo,
+                                          label3_ref, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_lo));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      emit_label (label_fallthrough);
+      DONE;
+    }
+  if (code == NE)
+    {
+      rtx label3_ref1 = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_hi = gen_rtx_fmt_ee (NE, VOIDmode, op1_hi, op2_hi);
+      rtx check_hi = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_hi,
+                                          label3_ref1, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_hi));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      rtx label3_ref2 = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_lo = gen_rtx_fmt_ee (NE, VOIDmode, op1_lo, op2_lo);
+      rtx check_lo = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_lo,
+                                          label3_ref2, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_lo));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      DONE;
+    }
+
+  if (code == LT || code == LTU || code == LE || code == LEU)
+    {
+      /* Check for "DI comparison obviously true".  */
+      rtx label3_ref1 = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_hi = gen_rtx_fmt_ee (pru_noteq_condition (code),
+                                   VOIDmode, op1_hi, op2_hi);
+      rtx check_hi = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_hi,
+                                          label3_ref1, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_hi));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      /* Check for "DI comparison obviously not true".  */
+      rtx label_fallthrough = gen_label_rtx ();
+      rtx label_fallthrough_ref = gen_rtx_LABEL_REF (Pmode, label_fallthrough);
+      rtx cond_hine = gen_rtx_fmt_ee (NE, VOIDmode, op1_hi, op2_hi);
+      rtx check_hine = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_hine,
+                                            label_fallthrough_ref, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_hine));
+      JUMP_LABEL (j) = label_fallthrough;
+      LABEL_NUSES (label_fallthrough)++;
+
+      /* Comparison deferred to the lo parts.  */
+      rtx label3_ref2 = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_lo = gen_rtx_fmt_ee (unsigned_condition (code),
+                                   VOIDmode, op1_lo, op2_lo);
+      rtx check_lo = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_lo,
+                                          label3_ref2, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_lo));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      emit_label (label_fallthrough);
+      DONE;
+    }
+
+  if (code == GT || code == GTU || code == GE || code == GEU)
+    {
+      /* Check for "DI comparison obviously not true".  */
+      const enum rtx_code reversed_code = reverse_condition (code);
+      rtx label_fallthrough = gen_label_rtx ();
+      rtx label_fallthrough_ref = gen_rtx_LABEL_REF (Pmode, label_fallthrough);
+      rtx cond_hi = gen_rtx_fmt_ee (pru_noteq_condition (reversed_code),
+                                   VOIDmode, op1_hi, op2_hi);
+      rtx check_hi = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_hi,
+                                          label_fallthrough_ref, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_hi));
+      JUMP_LABEL (j) = label_fallthrough;
+      LABEL_NUSES (label_fallthrough)++;
+
+      /* Check for "DI comparison obviously true".  */
+      rtx label3_ref1 = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_hine = gen_rtx_fmt_ee (NE, VOIDmode, op1_hi, op2_hi);
+      rtx check_hine = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_hine,
+                                            label3_ref1, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_hine));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      /* Comparison deferred to the lo parts.  */
+      rtx label3_ref2 = gen_rtx_LABEL_REF (Pmode, label3);
+      rtx cond_lo = gen_rtx_fmt_ee (unsigned_condition (code),
+                                   VOIDmode, op1_lo, op2_lo);
+      rtx check_lo = gen_rtx_IF_THEN_ELSE (VOIDmode, cond_lo,
+                                          label3_ref2, pc_rtx);
+      j = emit_jump_insn (gen_rtx_SET (pc_rtx, check_lo));
+      JUMP_LABEL (j) = label3;
+      LABEL_NUSES (label3)++;
+
+      emit_label (label_fallthrough);
+      DONE;
+    }
+    gcc_unreachable ();
+})
+
 ;
 ; Bit test branch
 
diff --git a/gcc/testsuite/gcc.target/pru/pr106562-1.c 
b/gcc/testsuite/gcc.target/pru/pr106562-1.c
new file mode 100644
index 00000000000..5bfbe52bb01
--- /dev/null
+++ b/gcc/testsuite/gcc.target/pru/pr106562-1.c
@@ -0,0 +1,9 @@
+/* { dg-do assemble } */
+/* { dg-options "-Os" } */
+/* { dg-final { object-size text <= 40 } } */
+
+
+char test(unsigned long long a, unsigned long long b)
+{
+        return a && b;
+}
diff --git a/gcc/testsuite/gcc.target/pru/pr106562-2.c 
b/gcc/testsuite/gcc.target/pru/pr106562-2.c
new file mode 100644
index 00000000000..ec5f82ad89f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/pru/pr106562-2.c
@@ -0,0 +1,9 @@
+/* { dg-do assemble } */
+/* { dg-options "-Os" } */
+/* { dg-final { object-size text <= 32 } } */
+
+
+char test(long long a)
+{
+        return a > 10;
+}
diff --git a/gcc/testsuite/gcc.target/pru/pr106562-3.c 
b/gcc/testsuite/gcc.target/pru/pr106562-3.c
new file mode 100644
index 00000000000..d0980581f1d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/pru/pr106562-3.c
@@ -0,0 +1,9 @@
+/* { dg-do assemble } */
+/* { dg-options "-Os" } */
+/* { dg-final { object-size text <= 32 } } */
+
+
+char test(long long a)
+{
+        return a < 10;
+}
diff --git a/gcc/testsuite/gcc.target/pru/pr106562-4.c 
b/gcc/testsuite/gcc.target/pru/pr106562-4.c
new file mode 100644
index 00000000000..b29e426b012
--- /dev/null
+++ b/gcc/testsuite/gcc.target/pru/pr106562-4.c
@@ -0,0 +1,159 @@
+/* Functional test for DI comparisons.  */
+
+/* { dg-do run } */
+/* { dg-options "-pedantic-errors" } */
+
+/* The default test suite options use "-ansi", which
+   generates spurious errors by enabling "-Wlong-long".
+   Thus override the options and drop "-ansi", in order
+   to freely use 64-bit (long long) types for PRU.  */
+
+#include <stddef.h>
+#include <stdint.h>
+
+extern void abort (void);
+
+char __attribute__((noinline)) test_lt (int64_t a, int64_t b)
+{
+  return a < b;
+}
+
+char __attribute__((noinline)) test_ltu (uint64_t a, uint64_t b)
+{
+  return a < b;
+}
+
+char __attribute__((noinline)) test_le (int64_t a, int64_t b)
+{
+  return a <= b;
+}
+
+char __attribute__((noinline)) test_leu (uint64_t a, uint64_t b)
+{
+  return a <= b;
+}
+
+char __attribute__((noinline)) test_gt (int64_t a, int64_t b)
+{
+  return a > b;
+}
+
+char __attribute__((noinline)) test_gtu (uint64_t a, uint64_t b)
+{
+  return a > b;
+}
+
+char __attribute__((noinline)) test_ge (int64_t a, int64_t b)
+{
+  return a >= b;
+}
+
+char __attribute__((noinline)) test_geu (uint64_t a, uint64_t b)
+{
+  return a >= b;
+}
+
+char __attribute__((noinline)) test_eq (uint64_t a, uint64_t b)
+{
+  return a == b;
+}
+
+char __attribute__((noinline)) test_ne (uint64_t a, uint64_t b)
+{
+  return a != b;
+}
+
+struct test_case {
+    uint64_t a;
+    uint64_t b;
+    char lt;
+    char ltu;
+    char le;
+    char leu;
+    char gt;
+    char gtu;
+    char ge;
+    char geu;
+    char eq;
+    char ne;
+};
+
+const struct test_case cases[] = {
+  /*                       LT,LTU,LE,LEU,GT,GTU,GE,GEU,EQ,NE */
+  { 0x1234567800112233ULL,
+    0x1234567800112233ULL, 0,  0, 1,  1, 0,  0, 1,  1, 1, 0 },
+  { 0x0000000000000000ULL,
+    0x0000000000000000ULL, 0,  0, 1,  1, 0,  0, 1,  1, 1, 0 },
+  { 0xffffffffffffffffULL,
+    0xffffffffffffffffULL, 0,  0, 1,  1, 0,  0, 1,  1, 1, 0 },
+
+  { 0xffffffffffffffefULL,
+    0xffffffffffffffffULL, 1,  1, 1,  1, 0,  0, 0,  0, 0, 1 },
+  { 0x8000000000000000ULL,
+    0xffffffffffffffffULL, 1,  1, 1,  1, 0,  0, 0,  0, 0, 1 },
+  { 0x80000000ffffffffULL,
+    0xffffffffffffffffULL, 1,  1, 1,  1, 0,  0, 0,  0, 0, 1 },
+  { 0x80000000ffffffffULL,
+    0xffffffff00000000ULL, 1,  1, 1,  1, 0,  0, 0,  0, 0, 1 },
+  { 0xffefffffffffffffULL,
+    0xffffffffffffffffULL, 1,  1, 1,  1, 0,  0, 0,  0, 0, 1 },
+
+  { 0x0000000000000000ULL,
+    0xffffffffffffffffULL, 0,  1, 0,  1, 1,  0, 1,  0, 0, 1 },
+  { 0x0000000000000001ULL,
+    0xffffffffffffffffULL, 0,  1, 0,  1, 1,  0, 1,  0, 0, 1 },
+  { 0x0000000000000001ULL,
+    0x8000000000000000ULL, 0,  1, 0,  1, 1,  0, 1,  0, 0, 1 },
+  { 0x7fffffffffffffffULL,
+    0x8000000000000000ULL, 0,  1, 0,  1, 1,  0, 1,  0, 0, 1 },
+
+  /* Ensure lo uses unsigned comparison if hi parts are same.  */
+  { 0x12345678ffffffffULL,
+    0x1234567800000001ULL, 0,  0, 0,  0, 1,  1, 1,  1, 0, 1 },
+  { 0xf23456780fffffffULL,
+    0xf234567800000001ULL, 0,  0, 0,  0, 1,  1, 1,  1, 0, 1 },
+  { 0xf2345678ffffffffULL,
+    0xf234567800000001ULL, 0,  0, 0,  0, 1,  1, 1,  1, 0, 1 },
+  { 0x1234567800000002ULL,
+    0x1234567800000001ULL, 0,  0, 0,  0, 1,  1, 1,  1, 0, 1 },
+  { 0x1234567800000002ULL,
+    0x1234567800000003ULL, 1,  1, 1,  1, 0,  0, 0,  0, 0, 1 },
+};
+
+int
+main (void)
+{
+  size_t i;
+
+  for (i = 0; i < (sizeof (cases)/sizeof (cases[0])); i++)
+    {
+      const int64_t sa = (int64_t)cases[i].a;
+      const int64_t sb = (int64_t)cases[i].b;
+      const uint64_t ua = cases[i].a;
+      const uint64_t ub = cases[i].b;
+
+      if (cases[i].lt != test_lt (sa, sb))
+       abort ();
+      if (cases[i].ltu != test_ltu (ua, ub))
+       abort ();
+      if (cases[i].le != test_le (sa, sb))
+       abort ();
+      if (cases[i].leu != test_leu (ua, ub))
+       abort ();
+      if (cases[i].gt != test_gt (sa, sb))
+       abort ();
+      if (cases[i].gtu != test_gtu (ua, ub))
+       abort ();
+      if (cases[i].ge != test_ge (sa, sb))
+       abort ();
+      if (cases[i].geu != test_geu (ua, ub))
+       abort ();
+      if (cases[i].eq != test_eq (ua, ub))
+       abort ();
+      if (cases[i].ne != test_ne (ua, ub))
+       abort ();
+    }
+
+  return 0;
+}
+
-- 
2.37.3

Reply via email to