Hi
This patch fixes the inconsistent behavior observed at -O3 for the
unordered comparisons. According to the online docs
(https://gcc.gnu.org/onlinedocs/gcc-7.2.0/gccint/Unary-and-Binary-Expressions.html),
all of the following should not raise an FP exception:
- UNGE_EXPR
- UNGT_EXPR
- UNLE_EXPR
- UNLT_EXPR
- UNEQ_EXPR
Also ORDERED_EXPR and UNORDERED_EXPR should only return zero or one.
The aarch64-simd.md handling of these were generating exception raising
instructions such as fcmgt. This patch changes the instructions that are
emitted to in order to not give out the exceptions. We first check each
operand for NaNs and force any elements containing NaN to zero before
using them in the compare.
Example: UN<cc> (a, b) -> UNORDERED (a, b) | (cm<cc> (isnan (a) ? 0.0 :
a, isnan (b) ? 0.0 : b))
The ORDERED_EXPR is now handled as (cmeq (a, a) & cmeq (b, b)) and
UNORDERED_EXPR as ~ORDERED_EXPR and UNEQ as (~ORDERED_EXPR | cmeq (a,b)).
Testing done: Checked for regressions on bootstrapped
aarch64-none-linux-gnu and added a new test case.
Is this ok for trunk? This will probably need a back-port to
gcc-7-branch as well.
Thanks
Sudi
ChangeLog Entries:
*** gcc/ChangeLog ***
2017-12-15 Sudakshina Das <sudi....@arm.com>
PR target/81647
* config/aarch64/aarch64-simd.md (vec_cmp<mode><v_int_equiv>): Modify
instructions for
UNLT, UNLE, UNGT, UNGE, UNEQ, UNORDERED and ORDERED.
*** gcc/testsuite/ChangeLog ***
2017-12-15 Sudakshina Das <sudi....@arm.com>
PR target/81647
* gcc.target/aarch64/pr81647.c: New.
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index f90f74fe7fd5990a97b9f4eb68f5735b7d4fb9aa..acff06c753b3e3aaa5775632929909afa4d3294b 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -2731,10 +2731,10 @@
break;
}
/* Fall through. */
- case UNGE:
+ case UNLT:
std::swap (operands[2], operands[3]);
/* Fall through. */
- case UNLE:
+ case UNGT:
case GT:
comparison = gen_aarch64_cmgt<mode>;
break;
@@ -2745,10 +2745,10 @@
break;
}
/* Fall through. */
- case UNGT:
+ case UNLE:
std::swap (operands[2], operands[3]);
/* Fall through. */
- case UNLT:
+ case UNGE:
case GE:
comparison = gen_aarch64_cmge<mode>;
break;
@@ -2771,21 +2771,35 @@
case UNGT:
case UNLE:
case UNLT:
- case NE:
- /* FCM returns false for lanes which are unordered, so if we use
- the inverse of the comparison we actually want to emit, then
- invert the result, we will end up with the correct result.
- Note that a NE NaN and NaN NE b are true for all a, b.
-
- Our transformations are:
- a UNGE b -> !(b GT a)
- a UNGT b -> !(b GE a)
- a UNLE b -> !(a GT b)
- a UNLT b -> !(a GE b)
- a NE b -> !(a EQ b) */
- gcc_assert (comparison != NULL);
- emit_insn (comparison (operands[0], operands[2], operands[3]));
- emit_insn (gen_one_cmpl<v_int_equiv>2 (operands[0], operands[0]));
+ {
+ /* All of the above must not raise any FP exceptions. Thus we first
+ check each operand for NaNs and force any elements containing NaN to
+ zero before using them in the compare.
+ Example: UN<cc> (a, b) -> UNORDERED (a, b) |
+ (cm<cc> (isnan (a) ? 0.0 : a,
+ isnan (b) ? 0.0 : b))
+ We use the following transformations for doing the comparisions:
+ a UNGE b -> a GE b
+ a UNGT b -> a GT b
+ a UNLE b -> b GE a
+ a UNLT b -> b GT a. */
+
+ rtx tmp0 = gen_reg_rtx (<V_INT_EQUIV>mode);
+ rtx tmp1 = gen_reg_rtx (<V_INT_EQUIV>mode);
+ rtx tmp2 = gen_reg_rtx (<V_INT_EQUIV>mode);
+ emit_insn (gen_aarch64_cmeq<mode> (tmp0, operands[2], operands[2]));
+ emit_insn (gen_aarch64_cmeq<mode> (tmp1, operands[3], operands[3]));
+ emit_insn (gen_and<v_int_equiv>3 (tmp2, tmp0, tmp1));
+ emit_insn (gen_and<v_int_equiv>3 (tmp0, tmp0,
+ lowpart_subreg (<V_INT_EQUIV>mode, operands[2], <MODE>mode)));
+ emit_insn (gen_and<v_int_equiv>3 (tmp1, tmp1,
+ lowpart_subreg (<V_INT_EQUIV>mode, operands[3], <MODE>mode)));
+ gcc_assert (comparison != NULL);
+ emit_insn (comparison (operands[0],
+ lowpart_subreg (<MODE>mode, tmp0, <V_INT_EQUIV>mode),
+ lowpart_subreg (<MODE>mode, tmp1, <V_INT_EQUIV>mode)));
+ emit_insn (gen_orn<v_int_equiv>3 (operands[0], tmp2, operands[0]));
+ }
break;
case LT:
@@ -2793,25 +2807,19 @@
case GT:
case GE:
case EQ:
+ case NE:
/* The easy case. Here we emit one of FCMGE, FCMGT or FCMEQ.
As a LT b <=> b GE a && a LE b <=> b GT a. Our transformations are:
a GE b -> a GE b
a GT b -> a GT b
a LE b -> b GE a
a LT b -> b GT a
- a EQ b -> a EQ b */
+ a EQ b -> a EQ b
+ a NE b -> ~(a EQ b) */
gcc_assert (comparison != NULL);
emit_insn (comparison (operands[0], operands[2], operands[3]));
- break;
-
- case UNEQ:
- /* We first check (a > b || b > a) which is !UNEQ, inverting
- this result will then give us (a == b || a UNORDERED b). */
- emit_insn (gen_aarch64_cmgt<mode> (operands[0],
- operands[2], operands[3]));
- emit_insn (gen_aarch64_cmgt<mode> (tmp, operands[3], operands[2]));
- emit_insn (gen_ior<v_int_equiv>3 (operands[0], operands[0], tmp));
- emit_insn (gen_one_cmpl<v_int_equiv>2 (operands[0], operands[0]));
+ if (code == NE)
+ emit_insn (gen_one_cmpl<v_int_equiv>2 (operands[0], operands[0]));
break;
case LTGT:
@@ -2823,21 +2831,22 @@
emit_insn (gen_ior<v_int_equiv>3 (operands[0], operands[0], tmp));
break;
- case UNORDERED:
- /* Operands are ORDERED iff (a > b || b >= a), so we can compute
- UNORDERED as !ORDERED. */
- emit_insn (gen_aarch64_cmgt<mode> (tmp, operands[2], operands[3]));
- emit_insn (gen_aarch64_cmge<mode> (operands[0],
- operands[3], operands[2]));
- emit_insn (gen_ior<v_int_equiv>3 (operands[0], operands[0], tmp));
- emit_insn (gen_one_cmpl<v_int_equiv>2 (operands[0], operands[0]));
- break;
-
case ORDERED:
- emit_insn (gen_aarch64_cmgt<mode> (tmp, operands[2], operands[3]));
- emit_insn (gen_aarch64_cmge<mode> (operands[0],
- operands[3], operands[2]));
- emit_insn (gen_ior<v_int_equiv>3 (operands[0], operands[0], tmp));
+ case UNORDERED:
+ case UNEQ:
+ /* cmeq (a, a) & cmeq (b, b). */
+ emit_insn (gen_aarch64_cmeq<mode> (operands[0],
+ operands[2], operands[2]));
+ emit_insn (gen_aarch64_cmeq<mode> (tmp, operands[3], operands[3]));
+ emit_insn (gen_and<v_int_equiv>3 (operands[0], operands[0], tmp));
+
+ if (code == UNORDERED)
+ emit_insn (gen_one_cmpl<v_int_equiv>2 (operands[0], operands[0]));
+ else if (code == UNEQ)
+ {
+ emit_insn (gen_aarch64_cmeq<mode> (tmp, operands[2], operands[3]));
+ emit_insn (gen_orn<v_int_equiv>3 (operands[0], operands[0], tmp));
+ }
break;
default:
diff --git a/gcc/testsuite/gcc.target/aarch64/pr81647.c b/gcc/testsuite/gcc.target/aarch64/pr81647.c
new file mode 100644
index 0000000000000000000000000000000000000000..f60dfba49d538e3b2164b11392ab8dbfdba6546e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/pr81647.c
@@ -0,0 +1,44 @@
+/* { dg-do run } */
+/* { dg-options "-O3 -fdump-tree-ssa" } */
+
+#include <fenv.h>
+
+double x[28], y[28];
+int res[28];
+
+int
+main (void)
+{
+ int i;
+ for (i = 0; i < 28; ++i)
+ {
+ x[i] = __builtin_nan ("");
+ y[i] = i;
+ }
+ __asm__ volatile ("" ::: "memory");
+ feclearexcept (FE_ALL_EXCEPT);
+ for (i = 0; i < 4; ++i)
+ res[i] = __builtin_isgreater (x[i], y[i]);
+ for (i = 4; i < 8; ++i)
+ res[i] = __builtin_isgreaterequal (x[i], y[i]);
+ for (i = 8; i < 12; ++i)
+ res[i] = __builtin_isless (x[i], y[i]);
+ for (i = 12; i < 16; ++i)
+ res[i] = __builtin_islessequal (x[i], y[i]);
+ for (i = 16; i < 20; ++i)
+ res[i] = __builtin_islessgreater (x[i], y[i]);
+ for (i = 20; i < 24; ++i)
+ res[i] = __builtin_isunordered (x[i], y[i]);
+ for (i = 24; i < 28; ++i)
+ res[i] = !(__builtin_isunordered (x[i], y[i]));
+ __asm__ volatile ("" ::: "memory");
+ return fetestexcept (FE_ALL_EXCEPT) != 0;
+}
+
+/* { dg-final { scan-tree-dump " u> " "ssa" } } */
+/* { dg-final { scan-tree-dump " u>= " "ssa" } } */
+/* { dg-final { scan-tree-dump " u< " "ssa" } } */
+/* { dg-final { scan-tree-dump " u<= " "ssa" } } */
+/* { dg-final { scan-tree-dump " u== " "ssa" } } */
+/* { dg-final { scan-tree-dump " unord " "ssa" } } */
+/* { dg-final { scan-tree-dump " ord " "ssa" } } */
