llvmbot
commented
7 months ago @llvm/issue-subscribers-backend-aarch64
Author: m-saito-fj (m-saito-fj)
Clang cannot SVE vectorize TSVC s1161, but GCC13.2.0 can.
Option:
`-Ofast -march=armv8.2-a+sve`
```c
#define LEN 32000
#define LEN2 256
static int ntimes = 200000;
float a[LEN], b[LEN], c[LEN], d[LEN], e[LEN];
float aa[LEN2][LEN2], bb[LEN2][LEN2], cc[LEN2][LEN2], dd[LEN2][LEN2];
int dummy(float[LEN], float[LEN], float[LEN], float[LEN], float[LEN],
float[LEN2][LEN2], float[LEN2][LEN2], float[LEN2][LEN2], float);
int s1161()
{
for (int nl = 0; nl < ntimes; nl++) {
for (int i = 0; i < LEN-1; ++i) {
if (c[i] < (float)0.) {
goto L20;
}
a[i] = c[i] + d[i] * e[i];
goto L10;
L20:
b[i] = a[i] + d[i] * d[i];
L10:
;
}
dummy(a, b, c, d, e, aa, bb, cc, 0.);
}
return 0;
}
```
See also (Clang vs GCC):
https://godbolt.org/z/EEq8Mj5zs
GCC result:
```asm
.L2:
ld1w z2.s, p0/z, [x27, x5, lsl 2]
lsl x0, x5, 2
fcmge p1.s, p2/z, z2.s, #0.0
ld1w z1.s, p0/z, [x26, x5, lsl 2]
add x6, x25, x0
add x7, x20, x0
ld1w z3.s, p0/z, [x6]
ld1w z5.s, p0/z, [x7]
add x0, x24, x0
sel z4.s, p1, z2.s, z3.s
ld1w z0.s, p0/z, [x0]
fmla z4.s, p1/m, z1.s, z5.s
fcmlt p1.s, p2/z, z2.s, #0.0
movprfx z0.s, p1/m, z3.s
fmla z0.s, p1/m, z1.s, z1.s
st1w z4.s, p0, [x6]
st1w z0.s, p0, [x0]
add x5, x5, x28
whilelo p0.s, w5, w19
b.any .L2
```
Loop Body IR:
```llvm
for.body4: ; preds = %for.cond1.preheader, %for.inc
%indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %for.inc ]
%arrayidx = getelementptr inbounds [32000 x float], ptr @c, i64 0, i64 %indvars.iv, !dbg !21
%0 = load float, ptr %arrayidx, align 4, !dbg !21, !tbaa !22
%cmp5 = fcmp fast olt float %0, 0.000000e+00, !dbg !26
br i1 %cmp5, label %L20, label %if.end, !dbg !21
if.end: ; preds = %for.body4
%arrayidx9 = getelementptr inbounds [32000 x float], ptr @d, i64 0, i64 %indvars.iv, !dbg !27
%1 = load float, ptr %arrayidx9, align 4, !dbg !27, !tbaa !22
%arrayidx11 = getelementptr inbounds [32000 x float], ptr @e, i64 0, i64 %indvars.iv, !dbg !28
%2 = load float, ptr %arrayidx11, align 4, !dbg !28, !tbaa !22
%mul = fmul fast float %2, %1, !dbg !29
%add = fadd fast float %mul, %0, !dbg !30
br label %for.inc, !dbg !31
L20: ; preds = %for.body4
%arrayidx15 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv, !dbg !32
%3 = load float, ptr %arrayidx15, align 4, !dbg !32, !tbaa !22
%arrayidx17 = getelementptr inbounds [32000 x float], ptr @d, i64 0, i64 %indvars.iv, !dbg !33
%4 = load float, ptr %arrayidx17, align 4, !dbg !33, !tbaa !22
%mul20 = fmul fast float %4, %4, !dbg !34
%add21 = fadd fast float %mul20, %3, !dbg !35
br label %for.inc, !dbg !36
for.inc: ; preds = %if.end, %L20
%a.sink = phi ptr [ @a, %if.end ], [ @b, %L20 ]
%add.sink = phi float [ %add, %if.end ], [ %add21, %L20 ]
%arrayidx13 = getelementptr inbounds [32000 x float], ptr %a.sink, i64 0, i64 %indvars.iv, !dbg !37
store float %add.sink, ptr %arrayidx13, align 4, !dbg !37, !tbaa !22
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !38
%exitcond.not = icmp eq i64 %indvars.iv.next, 31999, !dbg !39
br i1 %exitcond.not, label %for.cond.cleanup3, label %for.body4, !dbg !13, !llvm.loop !40
```
`-mllvm -debug-only=loop-accesses` messages:
```
LAA: Can't find bounds for ptr: %arrayidx13 = getelementptr inbounds [32000 x float], ptr %a.sink, i64 0, i64 %indvars.iv, !dbg !37
LAA: Found a runtime check ptr: %arrayidx15 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv, !dbg !32
LAA: We need to do 0 pointer comparisons.
LAA: We can't vectorize because we can't find the array bounds.
```
The direct factor is the missing ptr boundary.
Phi for selecting base address of store in "%a.sink" of "for.inc" block seems to be affected.
Clang cannot SVE vectorize TSVC s1161, but GCC13.2.0 can.
Option:
-Ofast -march=armv8.2-a+sveSee also (Clang vs GCC): https://godbolt.org/z/EEq8Mj5zs
GCC result:
Loop Body IR:
-mllvm -debug-only=loop-accessesmessages:The direct factor is the missing ptr boundary. Phi for selecting base address of store in "%a.sink" of "for.inc" block seems to be affected.