s390x: `__builtin_reduce_and` does not optimize well

[folkertdev]

given this C code

https://godbolt.org/z/WvfG8TTxf

#include <vecintrin.h>
#include <stdbool.h>

bool vectors_equal_builtin(vector int a, vector int b) {
    return vec_all_eq(a, b);
}

typedef int vec4i __attribute__((vector_size(16)));

bool vectors_equal_manual(vec4i a, vec4i b) {
    return __builtin_reduce_and(a == b);
}

The manual implementation fails to optimize to the builtin one.

vectors_equal_builtin:
        vceqfs  %v0, %v24, %v26
        lghi    %r2, 0
        locghie %r2, 1
        br      %r14

vectors_equal_manual:
        aghi    %r15, -168
        vceqf   %v0, %v24, %v26
        vno     %v0, %v0, %v0
        vlgvf   %r1, %v0, 0
        vlgvf   %r0, %v0, 1
        sll     %r1, 3
        rosbg   %r1, %r0, 61, 61, 2
        vlgvf   %r0, %v0, 2
        rosbg   %r1, %r0, 62, 62, 1
        vlgvf   %r0, %v0, 3
        rosbg   %r1, %r0, 63, 63, 0
        tmll    %r1, 15
        lghi    %r2, 0
        locghie %r2, 1
        aghi    %r15, 168
        br      %r14

define dso_local noundef zeroext i1 @vectors_equal_manual(<4 x i32> noundef %a, <4 x i32> noundef %b) local_unnamed_addr {
entry:
  %0 = icmp ne <4 x i32> %a, %b
  %1 = bitcast <4 x i1> %0 to i4
  %2 = icmp eq i4 %1, 0
  ret i1 %2
}

There are many varitions on vec_all_eq (see https://www.ibm.com/docs/en/zos/2.4.0?topic=functions-any-predicates), and it would be neat if those all optimized. It might be possible to simplify clang's vecintrin.h too.

This came up while implementing vec_all_eq in the rust standard library, where fewer custom intrinsics are better in every way.

cc @uweigand (posted here so it can be linked to)

[folkertdev]

I suspect that there is also some relation go vec_test_mask (the vtm instruction), in that some form of reduce should optimize to it, but I'm not sure because I can't find any official documentation on that instruction.

[uweigand]

I suspect that there is also some relation go vec_test_mask (the vtm instruction), in that some form of reduce should optimize to it, but I'm not sure because I can't find any official documentation on that instruction.

Possibly, in particular when used in conditional context (VTM does not have any output, it only sets the condition code). Here's the official documentation (z/Architecture Principles of Operation): https://publibfp.dhe.ibm.com/epubs/pdf/a227832d.pdf - VTM is on page 22-31.