Missed optimization: fold a chain of getelementptr operations with the same index into a single getelementptr with a multiplied index

[dreamking60]

define ptr @src(i64 %arg0, ptr %arg1) {
  %v0 = getelementptr inbounds double, ptr %arg1, i64 %arg0
  %v1 = getelementptr inbounds double, ptr %v0, i64 %arg0
  %v2 = getelementptr inbounds double, ptr %v1, i64 %arg0
  %v3 = getelementptr inbounds double, ptr %v2, i64 %arg0
  ret ptr %v3
}

define ptr @tgt(i64 %arg0, ptr %arg1) {
  %idx = mul i64 4, %arg0
  %v3 = getelementptr inbounds double, ptr %arg1, i64 %idx
  ret ptr %v3
}

godbolt: https://godbolt.org/z/jEqE81K71
alive2: https://alive2.llvm.org/ce/z/CSfwcA

Pattern found in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/openblas/optimized/dgemv_n.ll

[EugeneZelenko]

@dreamking60: Please specify language for code blocks.

[dreamking60]

@dreamking60: Please specify language for code blocks.

Thanks for the reminder — I’ll specify the language for code blocks going forward.