llvmbot
commented
9 months ago @llvm/issue-subscribers-backend-aarch64
Author: m-saito-fj (m-saito-fj)
Clang cannot SVE vectorize TSVC s114, 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 s115()
{
for (int nl = 0; nl < 1000*(ntimes/LEN2); nl++) {
for (int j = 0; j < LEN2; j++) {
for (int i = j+1; i < LEN2; i++) {
a[i] -= aa[j][i] * a[j];
}
}
dummy(a, b, c, d, e, aa, bb, cc, 0.);
}
return 0;
}
```
See also (Clang vs GCC):
https://godbolt.org/z/h8WE5raja
for.body IR:
```llvm
for.body8.lr.ph: ; preds = %for.body4
%arrayidx12 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv37
br label %for.body8, !dbg !26
for.body8: ; preds = %for.body8.lr.ph, %for.body8
%indvars.iv34 = phi i64 [ %indvars.iv, %for.body8.lr.ph ], [ %indvars.iv.next35, %for.body8 ]
%arrayidx10 = getelementptr inbounds [256 x [256 x float]], ptr @aa, i64 0, i64 %indvars.iv37, i64 %indvars.iv34, !dbg !27
%0 = load float, ptr %arrayidx10, align 4, !dbg !27, !tbaa !28
%1 = load float, ptr %arrayidx12, align 4, !dbg !32, !tbaa !28
%mul13 = fmul fast float %1, %0, !dbg !33
%arrayidx15 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv34, !dbg !34
%2 = load float, ptr %arrayidx15, align 4, !dbg !35, !tbaa !28
%sub = fsub fast float %2, %mul13, !dbg !35
store float %sub, ptr %arrayidx15, align 4, !dbg !35, !tbaa !28
%indvars.iv.next35 = add nuw nsw i64 %indvars.iv34, 1, !dbg !36
%exitcond.not = icmp eq i64 %indvars.iv.next35, 256, !dbg !25
br i1 %exitcond.not, label %for.cond1.loopexit.loopexit, label %for.body8, !dbg !26, !llvm.loop !37
}
```
`-mllvm -debug-only=loop-accesses` messages:
```
LAA: Found a runtime check ptr: %arrayidx15 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv34, !dbg !34
LAA: Found a runtime check ptr: %arrayidx12 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv37
LAA: We need to do 0 pointer comparisons.
LAA: May be able to perform a memory runtime check if needed.
LAA: Checking memory dependencies
LAA: Bad stride - Not striding over innermost loop %arrayidx12 = getelementptr inbounds [32000 x float], ptr @a, i64 0, i64 %indvars.iv37 SCEV: {@a,+,4}<nw><%for.body4>
LAA: Src Scev: {@a,+,4}<nw><%for.body4>Sink Scev: {{(4 + @a)<nuw>,+,4}<nw><%for.body4>,+,4}<nuw><%for.body8>(Induction step: 0)
LAA: Distance for %1 = load float, ptr %arrayidx12, align 4, !dbg !32, !tbaa !28 to store float %sub, ptr %arrayidx15, align 4, !dbg !35, !tbaa !28: {4,+,4}<nw><%for.body8>
Pointer access with non-constant stride
LAA: Src Scev: {{(4 + @a)<nuw>,+,4}<nw><%for.body4>,+,4}<nuw><%for.body8>Sink Scev: {{(4 + @a)<nuw>,+,4}<nw><%for.body4>,+,4}<nuw><%for.body8>(Induction step: 1)
LAA: Distance for %2 = load float, ptr %arrayidx15, align 4, !dbg !35, !tbaa !28 to store float %sub, ptr %arrayidx15, align 4, !dbg !35, !tbaa !28: 0
Total Dependences: 2
LAA: unsafe dependent memory operations in loop
```
There seems to be a problem with the two dependencies.
1) Load and store for a[i]
2) Store for a[i] and load for a[j]
GCC13 is able to vectorize by vector masked load/store.
```asm
.L4:
ld1w z1.s, p0/z, [x8, x0, lsl 2]
ld1w z0.s, p0/z, [x10, x0, lsl 2]
fmsb z0.s, p1/m, z2.s, z1.s
st1w z0.s, p0, [x8, x0, lsl 2]
add x0, x0, x20
whilelo p0.s, w0, w9
b.any .L4
```
fhahn
Clang cannot SVE vectorize TSVC s115, but GCC13.2.0 can.
Option:
-Ofast -march=armv8.2-a+sveSee also (Clang vs GCC): https://godbolt.org/z/h8WE5raja
for.body IR:
-mllvm -debug-only=loop-accessesmessages:There seems to be a problem with the two dependencies. 1) Load and store for a[i] 2) Store for a[i] and load for a[j]
GCC13 is able to vectorize by vector masked load/store.