[PATCH] D45748: [RISCV] Separate base from offset in lowerGlobalAddress

[Alex Bradbury via Phabricator via llvm-commits]

asb added a comment.

What are your thoughts on handling regressions such as those demonstrated in the test case below?

  @g = global [128 x i8] zeroinitializer, align 1
  
  define i8* @test() {
    ret i8* getelementptr inbounds ([128 x i8], [128 x i8]* @g, i32 0, i32 125)
  }

Before:

  	lui	a0, %hi(g+125)
  	addi	a0, a0, %lo(g+125)

After:

  	lui	a0, %hi(g)
  	addi	a0, a0, %lo(g)
  	addi	a0, a0, 125

It's tempting to have a peephole that converts the final addi to `addi a0, a0, %lo(g+125)` but that's not actually correct as the %hi relocation could be affected by the top bit of the %lo due to sign extension. A peephole might still be a sensible approach, but it would need to transform the lui as well.

You can also note that in the worst case with a large offset you end up with an additional two instructions

  @g = global [1048576 x i8] zeroinitializer, align 1  
  
  define i8* @test() {
    ret i8* getelementptr inbounds ([1048576 x i8], [1048576 x i8]* @g, i32 0, i32 524288)
  }

Before:

  	lui	a0, %hi(g)
  	addi	a0, a0, %lo(g)
  	lui	a1, 128
  	add	a0, a0, a1

After:

  	lui	a0, %hi(g+524288)
  	addi	a0, a0, %lo(g+524288)


https://reviews.llvm.org/D45748