[llvm] Update the base and index value for masked gather (PR #130920)

[Rohit Aggarwal via llvm-commits]

rohitaggarwal007 wrote:

> Here is my experiment: [main...RKSimon:llvm-project:x86-gather-scatter-addr-scale](https://github.com/llvm/llvm-project/compare/main...RKSimon:llvm-project:x86-gather-scatter-addr-scale)
> 
> It still needs work: it stupidly moves 1 bit of shift left per combine instead of doing something smarter, plus we should probably only bother trying to do this if we think it will eventually allow truncation.
> 
> Anyway, take a look and see if there's anything you can use that will help.

I commented my changes and applied the patch. It worked for the reduced test case. In the test case attached in the commit, it is working.
But in the case where there are two GEPs in the test case, it is failing
```
define {<16 x float>, <16 x float>} @test_gather_16f32_3(ptr %x, ptr %arr, <16 x i1> %mask, <16 x float> %src0)  {
; CHECK-LABEL: test_gather_16f32_2:
  %wide.load = load <16 x i32>, ptr %arr, align 4
  %and = and <16 x i32> %wide.load, <i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911, i32 536870911>
  %zext = zext <16 x i32> %and to <16 x i64>
  %ptrs1 = getelementptr inbounds %struct.pt, ptr %x, <16 x i64> %zext
  %res1 = call <16 x float> @llvm.masked.gather.v16f32.v16p0(<16 x ptr> %ptrs1, i32 4, <16 x i1> %mask, <16 x float> %src0)
  %ptrs = getelementptr inbounds %struct.pt, ptr %x, <16 x i64> %zext, i32 1
  %res = call <16 x float> @llvm.masked.gather.v16f32.v16p0(<16 x ptr> %ptrs, i32 4, <16 x i1> %mask, <16 x float> %src0)
  %pair1 = insertvalue {<16 x float>, <16 x float>} undef, <16 x float> %res1, 0
  %pair2 = insertvalue {<16 x float>, <16 x float>} %pair1, <16 x float> %res, 1
  ret {<16 x float>, <16 x float>} %pair2
  }
```
In the DAGcombiner when refineUniformBase is called to combine the gather again. For %res, when DAG is trying to simplify base and separate the Operand from index and, update the base. It is bailing out.
```
  if (!isNullConstant(BasePtr) && !Index.hasOneUse())
    return false;
```
In our case index has multiple uses and bailing out. 

> Is it good to relax this condition? Combiner create new node on combing!

Additionally, When i debug the logs, I found out that instruction which are already combined and process,  are still referring to these uses. Looks a wired behavior.



https://github.com/llvm/llvm-project/pull/130920