| Summary: | [12/13/14/15 Regression] suboptimal code for bool bitfield tests | ||
|---|---|---|---|
| Product: | gcc | Reporter: | Martin Sebor <msebor> |
| Component: | tree-optimization | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | NEW --- | ||
| Severity: | normal | CC: | fkorta, pinskia |
| Priority: | P2 | Keywords: | missed-optimization |
| Version: | 11.0 | ||
| Target Milestone: | 12.5 | ||
| See Also: |
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97588 https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99919 https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110637 |
||
| Host: | Target: | ||
| Build: | Known to work: | ||
| Known to fail: | 10.2.0, 11.0, 6.3.0, 7.0.1, 8.3.0, 9.3.0 | Last reconfirmed: | 2021-04-06 00:00:00 |
| Bug Depends on: | 99919 | ||
| Bug Blocks: | 19466, 85316 | ||
This comes down to lowering bitfields too soon. my bet it will happen even integer bitfields will have a problem. Bisection points to r225825 as the revision where GCC started to fail to fold the code in g(). This only seems to affect C _Bool bit-fields and not C++ bool. (In reply to Andrew Pinski from comment #1) > This comes down to lowering bitfields too soon. > my bet it will happen even integer bitfields will have a problem. Yes, unsigned bit-fields suffer the same problem but unlike for _Bool, GCC never emitted optimal code for those for this test case. The main issue is optimize_bit_field_compare in fold-const.c which produces
during GENERIC folding in .005t.original:
if ((BIT_FIELD_REF <b, 8, 0> & 1) != 0)
{
b.j = 0;
}
else
{
b.j = b.i;
}
return b.j;
that's premature in this place. For f() it also takes until DOM3 to do
the folding unless you disable SRA which then makes EVRP recognize the
second store as a.j = 0. With SRA we fail to derive ranges for a_10 in
a_10 = MEM <unsigned char> [(struct A *)&a];
a$1_11 = MEM <unsigned char> [(struct A *)&a + 1B];
_1 = VIEW_CONVERT_EXPR<_Bool>(a_10);
if (_1 != 0)
goto <bb 4>; [INV]
else
goto <bb 3>; [INV]
<bb 3> :
<bb 4> :
# a$1_9 = PHI <0(2), a_10(3)>
_7 = VIEW_CONVERT_EXPR<_Bool>(a$1_9);
thus we're missing looking through VIEW_CONVERT_EXPR in register_assert_for.
Amending that would eventually also allow optimizing the prematurely folded vairant.
GCC 9.4 is being released, retargeting bugs to GCC 9.5. GCC 9 branch is being closed GCC 10.4 is being released, retargeting bugs to GCC 10.5. Another simple example:
#include <cstdint>
struct SomeClass {
bool cfg1 : 1;
bool cfg2 : 1;
bool cfg3 : 1;
bool check() const noexcept { return cfg1 || cfg2 || cfg3; }
};
bool check(const SomeClass& rt) {
return rt.check();
}
Emits:
check(SomeClass const&):
movzx edx, BYTE PTR [rdi]
mov eax, edx
and eax, 1
jne .L1
mov eax, edx
shr al
and eax, 1
je .L4
.L1:
ret
.L4:
mov eax, edx
shr al, 2
and eax, 1
ret
While it should:
check(SomeClass const&):
test byte ptr [rdi], 7
setne al
ret
GCC 10 branch is being closed. (In reply to Martin Sebor from comment #2) > Bisection points to r225825 as the revision where GCC started to fail to > fold the code in g(). the fold-const didn't check `types_match (type, TREE_TYPE (@0))` but rather just did the equivalent to: (simplify (ne @0 integer_zerop@1) (if (TREE_CODE (TREE_TYPE (@0)) == BOOLEAN_TYPE) (non_lvalue (convert @0)))) While match now does not do the convert and checks the types_match check instead. GCC 11 branch is being closed. |
GCC does a better job folding operations involving plain Booleans than it does with bool bit-fields. The example below shows that in f() the return statement is folded to zero while in g() it's not. This is behind a class of -Wmaybe-uninitialized warnings. $ cat z.c && gcc -O2 -S -Wall -fdump-tree-optimized=/dev/stdout z.c struct A { _Bool i, j; }; _Bool f (struct A a) { if (a.i) a.j = 0; else a.j = a.i; return a.j; // folded to 0 } struct B { _Bool i: 1, j: 1; }; _Bool g (struct B b) { if (b.i) b.j = 0; else b.j = b.i; return b.j; // not folded } ;; Function f (f, funcdef_no=0, decl_uid=1946, cgraph_uid=1, symbol_order=0) _Bool f (struct A a) { <bb 2> [local count: 1073741824]: return 0; } ;; Function g (g, funcdef_no=1, decl_uid=1953, cgraph_uid=2, symbol_order=1) Removing basic block 5 _Bool g (struct B b) { _Bool b$j; unsigned char _1; unsigned char _2; _Bool _3; <bb 2> [local count: 1073741824]: _1 = VIEW_CONVERT_EXPR<unsigned char>(b); _2 = _1 & 1; if (_2 != 0) goto <bb 4>; [50.00%] else goto <bb 3>; [50.00%] <bb 3> [local count: 536870913]: _3 = b.i; <bb 4> [local count: 1073741824]: # b$j_5 = PHI <0(2), _3(3)> return b$j_5; }