[llvm] [GlobalISel][AArch64] Generate ptrtoint/inttoptr as opposed to bitcast in unmerge combine. (PR #115225)

[Madhur Amilkanthwar via llvm-commits]

================
@@ -0,0 +1,454 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=aarch64 -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64 -global-isel -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
+
+define ptr @s(ptr %p, i32 %q) {
+; CHECK-LABEL: s:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    add x0, x0, w1, sxtw #2
+; CHECK-NEXT:    ret
+  %d = getelementptr i32, ptr %p, i32 %q
+  ret ptr %d
+}
+
+define <2 x ptr> @v2(<2 x ptr> %p, i32 %q) {
+; CHECK-SD-LABEL: v2:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    dup v1.2s, w0
+; CHECK-SD-NEXT:    sshll v1.2d, v1.2s, #2
+; CHECK-SD-NEXT:    add v0.2d, v0.2d, v1.2d
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: v2:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    dup v1.2s, w0
+; CHECK-GI-NEXT:    adrp x8, .LCPI1_0
+; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI1_0]
+; CHECK-GI-NEXT:    fmov x9, d2
+; CHECK-GI-NEXT:    mov x11, v2.d[1]
+; CHECK-GI-NEXT:    sshll v1.2d, v1.2s, #0
+; CHECK-GI-NEXT:    fmov x8, d1
+; CHECK-GI-NEXT:    mov x10, v1.d[1]
+; CHECK-GI-NEXT:    mul x8, x8, x9
+; CHECK-GI-NEXT:    mul x9, x10, x11
+; CHECK-GI-NEXT:    mov v1.d[0], x8
+; CHECK-GI-NEXT:    mov v1.d[1], x9
+; CHECK-GI-NEXT:    add v0.2d, v0.2d, v1.2d
+; CHECK-GI-NEXT:    ret
+  %d = getelementptr i32, <2 x ptr> %p, i32 %q
----------------
madhur13490 wrote:

The generated code is pretty suboptimal compared to SDAG. Just curious, do you know why?

https://github.com/llvm/llvm-project/pull/115225