banach-space wrote:
> Hi @banach-space followed the incubator
>
> This is the IR generated with -fclangir
>
> ```
> define dso_local <2 x i64> @test_vshlq_n_s64(<2 x i64> noundef %0) #0 {
> br label %2
>
> 2: ; preds = %1
> %3 = bitcast <2 x i64> %0 to <16 x i8>
> %4 = bitcast <16 x i8> %3 to <2 x i64>
> %5 = shl <2 x i64> %4, splat (i64 3)
> br label %6
>
> 6: ; preds = %2
> ret <2 x i64> %5
> }
> ```
>
> this is the IR generated without -fclangir
>
> ```
> define dso_local <2 x i64> @test_vshlq_n_s64(<2 x i64> noundef %a) #0 {
> entry:
> %0 = bitcast <2 x i64> %a to <16 x i8>
> %1 = bitcast <16 x i8> %0 to <2 x i64>
> %vshl_n = shl <2 x i64> %1, splat (i64 3)
> ret <2 x i64> %vshl_n
> }
> ```
>
> hence the slight difference in the test cases between neon-intrinsics.c and
> intrinsics.c
Are you worried about the extra branch instructions? From what I can tell, you
can just ignore them and still match using the original `CHECK` lines, no?
The extra instructions will disappear once we enable `-simplifycfg` (or
similar) in the CIR pipeline. Since that is orthogonal to lowering the
builtins, it is OK to ignore that.
https://github.com/llvm/llvm-project/pull/187516
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
