http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58838
Bug ID: 58838
Summary: mullw sets condition code incorrectly.
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: dougkwan at google dot com
It was observed in 4.7 and trunk that the Power backend generated bad code
involving condition setting mullw. Below is a test case for the problem:
-----
#include <cassert>
#include <cstdlib>
#include <vector>
struct b88 {
char data[88];
};
namespace {
inline int int_size(const std::vector<b88>& v) {
return v.size();
}
}
int __attribute__ ((noinline))
foo(const std::vector<b88>& v) {
if (int_size(v) > 0) {
return atoi("1");
}
return 0;
}
int main() {
std::vector<b88> v;
v.push_back(b88());
assert(foo(v) != 0);
return 0;
}
----
The above test failed with gcc 4.7 and trunk for target powerpc64-linux-gnu
We seem to be fine getting out of the middle-end of gcc-4.7:
;; Function int foo(const std::vector<b88>&) (_Z3fooRKSt6vectorI3b88SaIS0_EE,
funcdef_no=473, decl_uid=9136, cgraph_uid=102)
int foo(const std::vector<b88>&) (const struct vector & v)
{
int _1;
struct b88 * _4;
struct b88 * _5;
long int _6;
long int _7;
long int _8;
long int _9;
int _10;
long int _11;
int _12;
<bb 2>:
_4 = MEM[(const struct vector *)v_3(D)];
_5 = MEM[(const struct vector *)v_3(D) + 8B];
_6 = (long int) _5;
_7 = (long int) _4;
_8 = _6 - _7;
_9 = _8 /[ex] 88;
_10 = (int) _9;
if (_10 > 0)
goto <bb 3>;
else
goto <bb 4>;
<bb 3>:
_11 = strtol ("1", 0B, 10);
_12 = (int) _11;
<bb 4>:
# _1 = PHI <_12(3), 0(2)>
return _1;
}
RTL expansion looks correct, gcc expands "_8 /[ex] 88;" into shift and
multiplication.
(_8 >> 3) * 0x2e8ba2e8ba2e8ba3.
Since we know that the _8 is a multiple of 88, we can do this.
0x2e8ba2e8ba2e8ba3 is the multiplicative-inverse of 11 in Z{2^64}.
But the selected PPC instruction is bad:
0x10000b80 <._Z3fooRKSt6vectorI3b88SaIS0_EE>: lis r9,11915
0x10000b84 <._Z3fooRKSt6vectorI3b88SaIS0_EE+4>: ld r8,8(r3)
0x10000b88 <._Z3fooRKSt6vectorI3b88SaIS0_EE+8>: ld r10,0(r3)
0x10000b8c <._Z3fooRKSt6vectorI3b88SaIS0_EE+12>: ori r9,r9,41704
0x10000b90 <._Z3fooRKSt6vectorI3b88SaIS0_EE+16>: rldicr r9,r9,32,31
0x10000b94 <._Z3fooRKSt6vectorI3b88SaIS0_EE+20>: subf r10,r10,r8
0x10000b98 <._Z3fooRKSt6vectorI3b88SaIS0_EE+24>: oris r9,r9,47662
0x10000b9c <._Z3fooRKSt6vectorI3b88SaIS0_EE+28>: sradi r10,r10,3
0x10000ba0 <._Z3fooRKSt6vectorI3b88SaIS0_EE+32>: ori r9,r9,35747
0x10000ba4 <._Z3fooRKSt6vectorI3b88SaIS0_EE+36>: mullw. r8,r10,r9
<================ note the "dot"
0x10000ba8 <._Z3fooRKSt6vectorI3b88SaIS0_EE+40>: ble 0x10000bf0
<._Z3fooRKSt6vectorI3b88SaIS0_EE+112>
mullw correctly computes the lower 32-bits of "_9 = _8 /[ex] 88;" but it sets
the condition code incorrectly, since signed-comparison is done in 64 bits.
Instead there should be sign extension, ie.
mullw r8,r10,r9
extsw. r8,r8
