[llvm] [AArch64] merge index address with large offset into base address (PR #72187)

[David Green via llvm-commits]

davemgreen wrote:

It looks like it was running with `-O3 -flto -mcpu=neoverse-v1`, although I can reproduce without the -mcpu.
`bin/clang -target aarch64 -O3 llvm-test-suite/MultiSource/Benchmarks/TSVC/StatementReordering-flt/tsc.c`

I think it is in these loops that look like:
```
.LBB6_3:                               |  10079 .LBB6_3:                          
        ldr     s1, [x9, x23]          |  10082         add     x10, x9, #93, lsl #12   
        ldr     s2, [x9, x24]          |  10083         add     x11, x9, #62, lsl #12   
        subs    x8, x8, #1             |  10084         subs    x8, x8, #1              
---------------------------------------|  10085         ldr     s1, [x11, #2100]        
---------------------------------------|  10086         ldr     s2, [x10, #3156]        
        fmadd   s0, s1, s2, s0         |  10087         fmadd   s0, s1, s2, s0          
        str     s0, [x9, #4]!          |  10088         str     s0, [x9, #4]!           
        add     x10, x9, x22           |  10089         add     x10, x9, #125, lsl #12  
        ldr     s0, [x10, #4]          |  10090         add     x11, x9, x22            
        ldr     s1, [x9, x25]          |  10091         ldr     s1, [x10, #128]         
---------------------------------------|  10092         add     x10, x9, #31, lsl #12   
---------------------------------------|  10093         ldr     s0, [x11, #4]           
        fmsub   s0, s1, s2, s0         |  10094         fmsub   s0, s1, s2, s0          
        str     s0, [x10]              |  10095         str     s0, [x10, #1040]        
        b.ne    .LBB6_3                |  10096         b.ne    .LBB6_3                 
```

https://github.com/llvm/llvm-project/pull/72187