/* PR rtl-optimization/28982. Function foo() does the equivalent of:
|
/* PR rtl-optimization/28982. Function foo() does the equivalent of:
|
|
|
float tmp_results[NVARS];
|
float tmp_results[NVARS];
|
for (int i = 0; i < NVARS; i++)
|
for (int i = 0; i < NVARS; i++)
|
{
|
{
|
int inc = incs[i];
|
int inc = incs[i];
|
float *ptr = ptrs[i], result = 0;
|
float *ptr = ptrs[i], result = 0;
|
for (int j = 0; j < n; j++)
|
for (int j = 0; j < n; j++)
|
result += *ptr, ptr += inc;
|
result += *ptr, ptr += inc;
|
tmp_results[i] = result;
|
tmp_results[i] = result;
|
}
|
}
|
memcpy (results, tmp_results, sizeof (results));
|
memcpy (results, tmp_results, sizeof (results));
|
|
|
but without the outermost loop. The idea is to create high register
|
but without the outermost loop. The idea is to create high register
|
pressure and ensure that some INC and PTR variables are spilled.
|
pressure and ensure that some INC and PTR variables are spilled.
|
|
|
On ARM targets, sequences like "result += *ptr, ptr += inc" can
|
On ARM targets, sequences like "result += *ptr, ptr += inc" can
|
usually be implemented using (mem (post_modify ...)), and we do
|
usually be implemented using (mem (post_modify ...)), and we do
|
indeed create such MEMs before reload for this testcase. However,
|
indeed create such MEMs before reload for this testcase. However,
|
(post_modify ...) is not a valid address for coprocessor loads, so
|
(post_modify ...) is not a valid address for coprocessor loads, so
|
for -mfloat-abi=softfp, reload reloads the POST_MODIFY into a base
|
for -mfloat-abi=softfp, reload reloads the POST_MODIFY into a base
|
register. GCC did not deal correctly with cases where the base and
|
register. GCC did not deal correctly with cases where the base and
|
index of the POST_MODIFY are themselves reloaded. */
|
index of the POST_MODIFY are themselves reloaded. */
|
#define NITER 4
|
#define NITER 4
|
#define NVARS 20
|
#define NVARS 20
|
#define MULTI(X) \
|
#define MULTI(X) \
|
X( 0), X( 1), X( 2), X( 3), X( 4), X( 5), X( 6), X( 7), X( 8), X( 9), \
|
X( 0), X( 1), X( 2), X( 3), X( 4), X( 5), X( 6), X( 7), X( 8), X( 9), \
|
X(10), X(11), X(12), X(13), X(14), X(15), X(16), X(17), X(18), X(19)
|
X(10), X(11), X(12), X(13), X(14), X(15), X(16), X(17), X(18), X(19)
|
|
|
#define DECLAREI(INDEX) inc##INDEX = incs[INDEX]
|
#define DECLAREI(INDEX) inc##INDEX = incs[INDEX]
|
#define DECLAREF(INDEX) *ptr##INDEX = ptrs[INDEX], result##INDEX = 0
|
#define DECLAREF(INDEX) *ptr##INDEX = ptrs[INDEX], result##INDEX = 0
|
#define LOOP(INDEX) result##INDEX += *ptr##INDEX, ptr##INDEX += inc##INDEX
|
#define LOOP(INDEX) result##INDEX += *ptr##INDEX, ptr##INDEX += inc##INDEX
|
#define COPYOUT(INDEX) results[INDEX] = result##INDEX
|
#define COPYOUT(INDEX) results[INDEX] = result##INDEX
|
|
|
float *ptrs[NVARS];
|
float *ptrs[NVARS];
|
float results[NVARS];
|
float results[NVARS];
|
int incs[NVARS];
|
int incs[NVARS];
|
|
|
void __attribute__((noinline))
|
void __attribute__((noinline))
|
foo (int n)
|
foo (int n)
|
{
|
{
|
int MULTI (DECLAREI);
|
int MULTI (DECLAREI);
|
float MULTI (DECLAREF);
|
float MULTI (DECLAREF);
|
while (n--)
|
while (n--)
|
MULTI (LOOP);
|
MULTI (LOOP);
|
MULTI (COPYOUT);
|
MULTI (COPYOUT);
|
}
|
}
|
|
|
float input[NITER * NVARS];
|
float input[NITER * NVARS];
|
|
|
int
|
int
|
main (void)
|
main (void)
|
{
|
{
|
int i;
|
int i;
|
|
|
for (i = 0; i < NVARS; i++)
|
for (i = 0; i < NVARS; i++)
|
ptrs[i] = input + i, incs[i] = i;
|
ptrs[i] = input + i, incs[i] = i;
|
for (i = 0; i < NITER * NVARS; i++)
|
for (i = 0; i < NITER * NVARS; i++)
|
input[i] = i;
|
input[i] = i;
|
foo (NITER);
|
foo (NITER);
|
for (i = 0; i < NVARS; i++)
|
for (i = 0; i < NVARS; i++)
|
if (results[i] != i * NITER * (NITER + 1) / 2)
|
if (results[i] != i * NITER * (NITER + 1) / 2)
|
return 1;
|
return 1;
|
return 0;
|
return 0;
|
}
|
}
|
|
|
