genrecog relies on a predicate foo_operand (op, mode) checking that OP
really does have mode MODE, with VOIDmode acting as a wildcard. This was
true even with the old genrecog, but as Andreas found on s390x, new genrecog
is being a bit more aggressive about it.
The problem is that at the moment, genpreds drops the mode check for
anything that can accept a CONST_INT or CONST_DOUBLE, on the basis that
constant integers have a recorded mode of VOIDmode whatever their "real"
logical mode is. But that means that a predicate that accepts constant
integers and other rtxes like REG won't check the modes of those other
rtxes either. Even dropping the check for CONST_DOUBLE is wrong, since it
can be a floating-point constant with a real mode.
This patch therefore adjusts the automatic mode test to allow
GET_MODE (op) == VOIDmode in cases where that's necessary.
It also includes CONST_WIDE_INT in the list of constant integer codes.
Bootstrapped & regression-tested on x86_64-linux-gnu. OK to install?
Thanks,
Richard
gcc/
* genpreds.c (mark_mode_tests): Mark all MATCH_CODEs as
NO_MODE_TEST.
(add_mode_tests): Don't add mode tests if the predicate only
accepts scalar constant integers. Otherwise, allow the mode
of "op" to be VOIDmode if the predicate does accept such integers.
Index: gcc/genpreds.c
===================================================================
--- gcc/genpreds.c 2015-05-22 16:49:55.454319576 +0100
+++ gcc/genpreds.c 2015-05-23 10:46:58.187370110 +0100
@@ -218,11 +218,11 @@ needs_variable (rtx exp, const char *var
/* Given an RTL expression EXP, find all subexpressions which we may
assume to perform mode tests. Normal MATCH_OPERAND does;
- MATCH_CODE does if it applies to the whole expression and accepts
- CONST_INT or CONST_DOUBLE; and we have to assume that MATCH_TEST
- does not. These combine in almost-boolean fashion - the only
- exception is that (not X) must be assumed not to perform a mode
- test, whether or not X does.
+ MATCH_CODE doesn't as such (although certain codes always have
+ VOIDmode); and we have to assume that MATCH_TEST does not.
+ These combine in almost-boolean fashion - the only exception is
+ that (not X) must be assumed not to perform a mode test, whether
+ or not X does.
The mark is the RTL /v flag, which is true for subexpressions which
do *not* perform mode tests.
@@ -244,10 +244,7 @@ mark_mode_tests (rtx exp)
break;
case MATCH_CODE:
- if (XSTR (exp, 1)[0] != '\0'
- || (!strstr (XSTR (exp, 0), "const_int")
- && !strstr (XSTR (exp, 0), "const_double")))
- NO_MODE_TEST (exp) = 1;
+ NO_MODE_TEST (exp) = 1;
break;
case MATCH_TEST:
@@ -313,6 +310,33 @@ add_mode_tests (struct pred_data *p)
if (p->special)
return;
+ /* Check whether the predicate accepts const scalar ints (which always
+ have a stored mode of VOIDmode, but logically have a real mode)
+ and whether it matches anything besides const scalar ints. */
+ bool matches_const_scalar_int_p = false;
+ bool matches_other_p = false;
+ for (int i = 0; i < NUM_RTX_CODE; ++i)
+ if (p->codes[i])
+ switch (i)
+ {
+ CASE_CONST_SCALAR_INT:
+ matches_const_scalar_int_p = true;
+ break;
+
+ default:
+ matches_other_p = true;
+ break;
+ }
+
+ /* There's no need for a mode check if the predicate only accepts
+ constant integers. The code checks in the predicate are enough
+ to establish that the mode is VOIDmode.
+
+ Note that the predicate itself should check whether a scalar
+ integer is in range of the given mode. */
+ if (!matches_other_p && !p->codes[CONST_DOUBLE])
+ return;
+
mark_mode_tests (p->exp);
/* If the whole expression already tests the mode, we're done. */
@@ -320,7 +344,11 @@ add_mode_tests (struct pred_data *p)
return;
match_test_exp = rtx_alloc (MATCH_TEST);
- XSTR (match_test_exp, 0) = "mode == VOIDmode || GET_MODE (op) == mode";
+ if (matches_const_scalar_int_p)
+ XSTR (match_test_exp, 0) = ("mode == VOIDmode || GET_MODE (op) == mode"
+ " || GET_MODE (op) == VOIDmode");
+ else
+ XSTR (match_test_exp, 0) = "mode == VOIDmode || GET_MODE (op) == mode";
and_exp = rtx_alloc (AND);
XEXP (and_exp, 1) = match_test_exp;
