[llvm] [SLP]Reduce number of alternate instruction, where possible (PR #123360)

[Hans Wennborg via llvm-commits]

zmodem wrote:

The re-land (9d37e61fc77d3d6de891c30630f1c0227522031d) caused `Instruction does not dominate all uses!` errors. Here is a reproducer: 
[mask_mem_test-84ec17.cpp.gz](https://github.com/user-attachments/files/19184767/mask_mem_test-84ec17.cpp.gz)

```
$ build/bin/clang -cc1 -triple x86_64-unknown-linux-gnu -emit-obj --crel -disable-free -clear-ast-before-backend -main-file-name mask_mem_test.cc -mrelocation-model pic -pic-level 2 -fhalf-no-semantic-interposition -fmerge-all-constants -fno-delete-null-pointer-checks -mframe-pointer=all -relaxed-aliasing -ffp-contract=off -fno-rounding-math -mconstructor-aliases -funwind-tables=2 -target-cpu x86-64 -target-feature +sse3 -tune-cpu generic -debugger-tuning=gdb -gsimple-template-names=simple -debug-forward-template-params -fdebug-compilation-dir=. -mllvm -crash-diagnostics-dir=../../tools/clang/crashreports -ffunction-sections -fdata-sections -fno-unique-section-names -fcoverage-compilation-dir=. -nostdinc++ -D HWY_IS_TEST -D DCHECK_ALWAYS_ON=1 -D USE_UDEV -D USE_AURA=1 -D USE_GLIB=1 -D USE_OZONE=1 -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D _FORTIFY_SOURCE=2 -D _FILE_OFFSET_BITS=64 -D _LARGEFILE_SOURCE -D _LARGEFILE64_SOURCE -D _GNU_SOURCE -D _LIBCPP_HARDENING_MODE=_LIBCPP_HARDENING_MODE_EXTENSIVE -D _LIBCPP_DISABLE_VISIBILITY_ANNOTATIONS -D _LIBCXXABI_DISABLE_VISIBILITY_ANNOTATIONS -D _LIBCPP_INSTRUMENTED_WITH_ASAN=0 -D CR_LIBCXX_REVISION=7c3087b8a7563d8218b27b7183ba41cf52d41508 -D TMP_REBUILD_HACK -D CR_SYSROOT_KEY=20250129T203412Z-1 -D NDEBUG -D NVALGRIND -D DYNAMIC_ANNOTATIONS_ENABLED=0 -D GLIB_VERSION_MAX_ALLOWED=GLIB_VERSION_2_56 -D GLIB_VERSION_MIN_REQUIRED=GLIB_VERSION_2_56 -D BENCHMARK_STATIC_DEFINE -D UNIT_TEST -D GTEST_API_= -D GTEST_HAS_POSIX_RE=0 -D GTEST_LANG_CXX11=1 -D GTEST_HAS_TR1_TUPLE=0 -D GTEST_HAS_ABSL -D GTEST_NO_ABSL_FLAGS -D __DATE__= -D __TIME__= -D __TIMESTAMP__= -O2 -std=c++20 -fdeprecated-macro -ferror-limit 19 -fvisibility=hidden -fvisibility-inlines-hidden -fwrapv -fwrapv-pointer -pthread -stack-protector 1 -ftrivial-auto-var-init=pattern -fno-rtti -fgnuc-version=4.2.1 -fno-implicit-modules -fskip-odr-check-in-gmf -fno-sized-deallocation -Qn -fcolor-diagnostics -vectorize-loops -vectorize-slp -mllvm -instcombine-lower-dbg-declare=0 -mllvm -split-threshold-for-reg-with-hint=0 -fcomplete-member-pointers -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -x c++ mask_mem_test-84ec17.cpp -w
Instruction does not dominate all uses!
  %29 = insertelement <8 x i64> %28, i64 %xor6.i.5, i64 6
  %17 = shufflevector <8 x i64> %29, <8 x i64> poison, <6 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
in function _ZN3hwy6N_SSE412_GLOBAL__N_114TestMaskedLoadclIsNS0_4SimdIsLm8ELi0EEEEEvT_T0_
fatal error: error in backend: Broken function found, compilation aborted!
```

I'll prepare a revert.

https://github.com/llvm/llvm-project/pull/123360