Hi!
As the testcase shows, expand_divmod doesn't handle properly
division/remainder in modes larger than HOST_BITS_PER_WIDE_INT like
TImode - if op1 is "negative" CONST_INT, it means it has 65 most significant
bits set, but EXACT_POWER_OF_2_OR_ZERO_P will happily return true on that
and optimize the division into a right shift.
Fixed by taking into account that EXACT_POWER_OF_2_OR_ZERO_P works only
if either the mode bitsize/precision are smaller or equal than HWI, or
if the value doesn't have MSB bit set.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-11-18 Jakub Jelinek <ja...@redhat.com>
PR middle-end/78416
* expmed.c (expand_divmod): For modes wider than HWI, take into
account implicit 1 bits above msb for EXACT_POWER_OF_2_OR_ZERO_P.
Formatting fixes. Use size <= HOST_BITS_PER_WIDE_INT instead of
HOST_BITS_PER_WIDE_INT >= size.
* gcc.dg/torture/pr78416.c: New test.
--- gcc/expmed.c.jj 2016-10-31 13:28:12.000000000 +0100
+++ gcc/expmed.c 2016-11-18 16:57:57.608703605 +0100
@@ -3998,7 +3998,15 @@ expand_divmod (int rem_flag, enum tree_c
if (unsignedp)
ext_op1 &= GET_MODE_MASK (mode);
op1_is_pow2 = ((EXACT_POWER_OF_2_OR_ZERO_P (ext_op1)
- || (! unsignedp && EXACT_POWER_OF_2_OR_ZERO_P
(-ext_op1))));
+ /* If mode is wider than HWI and op1 has msb set,
+ then it has there extra implicit 1 bits above it. */
+ && (GET_MODE_PRECISION (mode) <= HOST_BITS_PER_WIDE_INT
+ || INTVAL (op1) >= 0))
+ || (! unsignedp
+ && EXACT_POWER_OF_2_OR_ZERO_P (-ext_op1)
+ && (GET_MODE_PRECISION (mode)
+ <= HOST_BITS_PER_WIDE_INT
+ || INTVAL (op1) < 0)));
}
/*
@@ -4141,8 +4149,17 @@ expand_divmod (int rem_flag, enum tree_c
op1_is_constant = CONST_INT_P (op1);
op1_is_pow2 = (op1_is_constant
&& ((EXACT_POWER_OF_2_OR_ZERO_P (INTVAL (op1))
- || (! unsignedp
- && EXACT_POWER_OF_2_OR_ZERO_P (-UINTVAL
(op1))))));
+ /* If mode is wider than HWI and op1 has msb set,
+ then it has there extra implicit 1 bits above
+ it. */
+ && (GET_MODE_PRECISION (compute_mode)
+ <= HOST_BITS_PER_WIDE_INT
+ || INTVAL (op1) >= 0))
+ || (! unsignedp
+ && EXACT_POWER_OF_2_OR_ZERO_P (-UINTVAL (op1))
+ && (GET_MODE_PRECISION (compute_mode)
+ <= HOST_BITS_PER_WIDE_INT
+ || INTVAL (op1) < 0))));
}
/* If one of the operands is a volatile MEM, copy it into a register. */
@@ -4185,7 +4202,8 @@ expand_divmod (int rem_flag, enum tree_c
unsigned HOST_WIDE_INT d = (INTVAL (op1)
& GET_MODE_MASK (compute_mode));
- if (EXACT_POWER_OF_2_OR_ZERO_P (d))
+ if (EXACT_POWER_OF_2_OR_ZERO_P (d)
+ && (INTVAL (op1) >= 0 || size <= HOST_BITS_PER_WIDE_INT))
{
pre_shift = floor_log2 (d);
if (rem_flag)
@@ -4325,7 +4343,7 @@ expand_divmod (int rem_flag, enum tree_c
else if (d == -1)
quotient = expand_unop (compute_mode, neg_optab, op0,
tquotient, 0);
- else if (HOST_BITS_PER_WIDE_INT >= size
+ else if (size <= HOST_BITS_PER_WIDE_INT
&& abs_d == HOST_WIDE_INT_1U << (size - 1))
{
/* This case is not handled correctly below. */
@@ -4335,6 +4353,7 @@ expand_divmod (int rem_flag, enum tree_c
goto fail1;
}
else if (EXACT_POWER_OF_2_OR_ZERO_P (d)
+ && (size <= HOST_BITS_PER_WIDE_INT || d >= 0)
&& (rem_flag
? smod_pow2_cheap (speed, compute_mode)
: sdiv_pow2_cheap (speed, compute_mode))
@@ -4348,7 +4367,9 @@ expand_divmod (int rem_flag, enum tree_c
compute_mode)
!= CODE_FOR_nothing)))
;
- else if (EXACT_POWER_OF_2_OR_ZERO_P (abs_d))
+ else if (EXACT_POWER_OF_2_OR_ZERO_P (abs_d)
+ && (size <= HOST_BITS_PER_WIDE_INT
+ || abs_d != (unsigned HOST_WIDE_INT) d))
{
if (rem_flag)
{
@@ -4483,7 +4504,7 @@ expand_divmod (int rem_flag, enum tree_c
case FLOOR_DIV_EXPR:
case FLOOR_MOD_EXPR:
/* We will come here only for signed operations. */
- if (op1_is_constant && HOST_BITS_PER_WIDE_INT >= size)
+ if (op1_is_constant && size <= HOST_BITS_PER_WIDE_INT)
{
unsigned HOST_WIDE_INT mh, ml;
int pre_shift, lgup, post_shift;
@@ -4552,9 +4573,8 @@ expand_divmod (int rem_flag, enum tree_c
op0, constm1_rtx), NULL_RTX);
t2 = expand_binop (compute_mode, ior_optab, op0, t1, NULL_RTX,
0, OPTAB_WIDEN);
- nsign = expand_shift
- (RSHIFT_EXPR, compute_mode, t2,
- size - 1, NULL_RTX, 0);
+ nsign = expand_shift (RSHIFT_EXPR, compute_mode, t2,
+ size - 1, NULL_RTX, 0);
t3 = force_operand (gen_rtx_MINUS (compute_mode, t1, nsign),
NULL_RTX);
t4 = expand_divmod (0, TRUNC_DIV_EXPR, compute_mode, t3, op1,
@@ -4663,7 +4683,10 @@ expand_divmod (int rem_flag, enum tree_c
case CEIL_MOD_EXPR:
if (unsignedp)
{
- if (op1_is_constant && EXACT_POWER_OF_2_OR_ZERO_P (INTVAL (op1)))
+ if (op1_is_constant
+ && EXACT_POWER_OF_2_OR_ZERO_P (INTVAL (op1))
+ && (size <= HOST_BITS_PER_WIDE_INT
+ || INTVAL (op1) >= 0))
{
rtx t1, t2, t3;
unsigned HOST_WIDE_INT d = INTVAL (op1);
@@ -4876,7 +4899,7 @@ expand_divmod (int rem_flag, enum tree_c
break;
case EXACT_DIV_EXPR:
- if (op1_is_constant && HOST_BITS_PER_WIDE_INT >= size)
+ if (op1_is_constant && size <= HOST_BITS_PER_WIDE_INT)
{
HOST_WIDE_INT d = INTVAL (op1);
unsigned HOST_WIDE_INT ml;
--- gcc/testsuite/gcc.dg/torture/pr78416.c.jj 2016-11-18 16:43:35.010448325
+0100
+++ gcc/testsuite/gcc.dg/torture/pr78416.c 2016-11-18 16:44:40.388634204
+0100
@@ -0,0 +1,17 @@
+/* PR middle-end/78416 */
+/* { dg-do run { target int128 } } */
+
+int
+main ()
+{
+ unsigned __int128 x;
+ x = 0xFFFFFFFFFFFFFFFFULL;
+ x /= ~0x7FFFFFFFFFFFFFFFLL;
+ if (x != 0)
+ __builtin_abort ();
+ x = ~0x7FFFFFFFFFFFFFFELL;
+ x /= ~0x7FFFFFFFFFFFFFFFLL;
+ if (x != 1)
+ __builtin_abort ();
+ return 0;
+}
Jakub
