On 2020/3/15 13:16, Richard Henderson wrote:
On 3/12/20 7:58 AM, LIU Zhiwei wrote:
+#define GEN_VEXT_VSLIDEUP_VX(NAME, ETYPE, H, CLEAR_FN) \
+void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2, \
+ CPURISCVState *env, uint32_t desc) \
+{ \
+ uint32_t mlen = vext_mlen(desc); \
+ uint32_t vlmax = env_archcpu(env)->cfg.vlen / mlen; \
+ uint32_t vm = vext_vm(desc); \
+ uint32_t vl = env->vl; \
+ uint32_t offset = s1, i; \
+ \
+ if (offset > vl) { \
+ offset = vl; \
+ } \
This isn't right.
That's to process a corner case. As you can see the behavior of
vslideup.vx from Section 17.4.1
0 < i < max(vstart, OFFSET) unchanged
max(vstart, OFFSET) <= i < vl vd[i] = vs2[i-OFFSET] if mask enabled,
unchanged if not
vl <= i < VLMAX
tail elements, vd[i] = 0
The spec v0.7.1 or v0.8 does not specified when OFFSET > vl.
Should The elements (vl <= i < OFFSET) be seen as tail elements, or
unchanged?
And it is possible because OFFSET is from a scalar register.
Here (vl <= i < OFFSET) elements are seen as tail elements.
+ for (i = 0; i < vl; i++) { \
+ if (((i < offset)) || (!vm && !vext_elem_mask(v0, mlen, i))) { \
+ continue; \
+ } \
+ *((ETYPE *)vd + H(i)) = *((ETYPE *)vs2 + H(i - offset)); \
+ } \
+ if (i == 0) { \
+ return; \
+ } \
You need to eliminate vl == 0 first, not last.
Then
for (i = offset; i < vl; i++)
The types of i and vl need to be extended to target_ulong, so that you don't
incorrectly crop the input offset.
Yes, I should.
It may be worth special-casing vm=1, or hoisting it out of the loop. The
operation becomes a memcpy (at least for little-endian) at that point. See
swap_memmove in arm/sve_helper.c.
+#define GEN_VEXT_VSLIDEDOWN_VX(NAME, ETYPE, H, CLEAR_FN) \
+void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2, \
+ CPURISCVState *env, uint32_t desc) \
+{ \
+ uint32_t mlen = vext_mlen(desc); \
+ uint32_t vlmax = env_archcpu(env)->cfg.vlen / mlen; \
+ uint32_t vm = vext_vm(desc); \
+ uint32_t vl = env->vl; \
+ uint32_t offset = s1, i; \
+ \
+ for (i = 0; i < vl; i++) { \
+ if (!vm && !vext_elem_mask(v0, mlen, i)) { \
+ continue; \
+ } \
+ if (i + offset < vlmax) { \
+ *((ETYPE *)vd + H(i)) = *((ETYPE *)vs2 + H(i + offset)); \
Again, eliminate vl == 0 first. In fact, why don't we make that a global
request for all of the patches for the next revision.
I don't get it.
Check vl == 0 first for all patches. Is it right?
Checking for i == 0 last
is silly, and checks for the zero twice: once in the loop bounds and again at
the end.
It is probably worth changing the loop bounds to
if (offset >= vlmax) {
max = 0;
} else {
max = MIN(vl, vlmax - offset);
}
for (i = 0; i < max; ++i)
Yes.
+ } else { \
+ *((ETYPE *)vd + H(i)) = 0; \
+ }
Which lets these zeros merge into...
It's a mistake here.
+ for (; i < vlmax; i++) { \
+ CLEAR_FN(vd, vl, vl * sizeof(ETYPE), vlmax * sizeof(ETYPE)); \
+ } \
These zeros.
+#define GEN_VEXT_VSLIDE1UP_VX(NAME, ETYPE, H, CLEAR_FN) \
+void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2, \
+ CPURISCVState *env, uint32_t desc) \
+{ \
+ uint32_t mlen = vext_mlen(desc); \
+ uint32_t vlmax = env_archcpu(env)->cfg.vlen / mlen; \
+ uint32_t vm = vext_vm(desc); \
+ uint32_t vl = env->vl; \
+ uint32_t i; \
+ \
+ for (i = 0; i < vl; i++) { \
+ if (!vm && !vext_elem_mask(v0, mlen, i)) { \
+ continue; \
+ } \
+ if (i == 0) { \
+ *((ETYPE *)vd + H(i)) = s1; \
+ } else { \
+ *((ETYPE *)vd + H(i)) = *((ETYPE *)vs2 + H(i - 1)); \
+ } \
+ } \
+ if (i == 0) { \
+ return; \
+ } \
+ for (; i < vlmax; i++) { \
+ CLEAR_FN(vd, vl, vl * sizeof(ETYPE), vlmax * sizeof(ETYPE)); \
+ } \
+}
As a preference, I think you can do away with this helper.
Simply use the slideup helper with argument 1, and then
afterwards store the integer register into element 0. You should be able to
re-use code from vmv.s.x for that.
I will try just in line.
Zhiwei
+#define GEN_VEXT_VSLIDE1DOWN_VX(NAME, ETYPE, H, CLEAR_FN) \
+void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2, \
+ CPURISCVState *env, uint32_t desc) \
+{ \
Likewise.
r~
