[llvm] [MachineLICM] Allow hoisting loads from invariant address (PR #70796)

[David Sherwood via llvm-commits]

================
@@ -0,0 +1,428 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=aarch64-linux-gnu < %s | FileCheck %s
+
+define i64 @one_dimensional(ptr %a, ptr %b, i64 %N, i64 %M, i64 %K) {
+; CHECK-LABEL: one_dimensional:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:    cbz x2, .LBB0_2
+; CHECK-NEXT:  .LBB0_1: // %for.body
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    ldr x9, [x0], #8
+; CHECK-NEXT:    ldr w10, [x1]
+; CHECK-NEXT:    ldr w9, [x9]
+; CHECK-NEXT:    cmp w9, w10
+; CHECK-NEXT:    cinc x8, x8, ne
+; CHECK-NEXT:    subs x2, x2, #1
+; CHECK-NEXT:    b.ne .LBB0_1
+; CHECK-NEXT:  .LBB0_2: // %for.cond.cleanup
+; CHECK-NEXT:    mov x0, x8
+; CHECK-NEXT:    ret
+entry:
+  %cmp4 = icmp eq i64 %N, 0
+  br i1 %cmp4, label %for.cond.cleanup, label %for.body
+
+for.cond.cleanup:                                 ; preds = %for.body, %entry
+  %sum.0.lcssa = phi i64 [ 0, %entry ], [ %spec.select, %for.body ]
+  ret i64 %sum.0.lcssa
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.06 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
+  %sum.05 = phi i64 [ %spec.select, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.06
+  %0 = load ptr, ptr %arrayidx, align 8
+  %bcmp = tail call i32 @bcmp(ptr %0, ptr %b, i64 4)
+  %tobool = icmp ne i32 %bcmp, 0
+  %add = zext i1 %tobool to i64
+  %spec.select = add i64 %sum.05, %add
+  %inc = add nuw i64 %i.06, 1
+  %exitcond = icmp eq i64 %inc, %N
+  br i1 %exitcond, label %for.cond.cleanup, label %for.body
+}
+
+define i64 @two_dimensional(ptr %a, ptr %b, i64 %N, i64 %M, i64 %K) {
+; CHECK-LABEL: two_dimensional:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:    cbz x2, .LBB1_6
+; CHECK-NEXT:  // %bb.1: // %entry
+; CHECK-NEXT:    cbz x3, .LBB1_6
+; CHECK-NEXT:  // %bb.2: // %for.cond1.preheader.preheader
+; CHECK-NEXT:    mov x9, xzr
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:  .LBB1_3: // %for.cond1.preheader
+; CHECK-NEXT:    // =>This Loop Header: Depth=1
+; CHECK-NEXT:    // Child Loop BB1_4 Depth 2
+; CHECK-NEXT:    ldr x10, [x0, x9, lsl #3]
+; CHECK-NEXT:    mov x11, x3
+; CHECK-NEXT:  .LBB1_4: // %for.body4
+; CHECK-NEXT:    // Parent Loop BB1_3 Depth=1
+; CHECK-NEXT:    // => This Inner Loop Header: Depth=2
+; CHECK-NEXT:    ldr x12, [x10], #8
+; CHECK-NEXT:    ldr w13, [x1]
+; CHECK-NEXT:    ldr w12, [x12]
+; CHECK-NEXT:    cmp w12, w13
+; CHECK-NEXT:    cinc x8, x8, ne
+; CHECK-NEXT:    subs x11, x11, #1
+; CHECK-NEXT:    b.ne .LBB1_4
+; CHECK-NEXT:  // %bb.5: // %for.cond1.for.cond.cleanup3_crit_edge
+; CHECK-NEXT:    // in Loop: Header=BB1_3 Depth=1
+; CHECK-NEXT:    add x9, x9, #1
+; CHECK-NEXT:    cmp x9, x2
+; CHECK-NEXT:    b.ne .LBB1_3
+; CHECK-NEXT:  .LBB1_6: // %for.cond.cleanup
+; CHECK-NEXT:    mov x0, x8
+; CHECK-NEXT:    ret
+entry:
+  %cmp17 = icmp eq i64 %N, 0
+  %cmp214 = icmp eq i64 %M, 0
+  %or.cond = or i1 %cmp17, %cmp214
+  br i1 %or.cond, label %for.cond.cleanup, label %for.cond1.preheader
+
+for.cond1.preheader:                           ; preds = %entry, %for.cond1.for.cond.cleanup3_crit_edge
+  %i.019 = phi i64 [ %inc7, %for.cond1.for.cond.cleanup3_crit_edge ], [ 0, %entry ]
+  %sum.018 = phi i64 [ %spec.select, %for.cond1.for.cond.cleanup3_crit_edge ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds ptr, ptr %a, i64 %i.019
+  %0 = load ptr, ptr %arrayidx, align 8
+  br label %for.body4
+
+for.body4:                                     ; preds = %for.cond1.preheader, %for.body4
+  %j.016 = phi i64 [ 0, %for.cond1.preheader ], [ %inc, %for.body4 ]
+  %sum.115 = phi i64 [ %sum.018, %for.cond1.preheader ], [ %spec.select, %for.body4 ]
+  %arrayidx5 = getelementptr inbounds ptr, ptr %0, i64 %j.016
+  %1 = load ptr, ptr %arrayidx5, align 8
+  %bcmp = tail call i32 @bcmp(ptr %1, ptr %b, i64 4)
+  %tobool = icmp ne i32 %bcmp, 0
+  %add = zext i1 %tobool to i64
+  %spec.select = add i64 %sum.115, %add
+  %inc = add nuw i64 %j.016, 1
+  %exitcond = icmp eq i64 %inc, %M
+  br i1 %exitcond, label %for.cond1.for.cond.cleanup3_crit_edge, label %for.body4
+
+for.cond1.for.cond.cleanup3_crit_edge:         ; preds = %for.body4
+  %inc7 = add nuw i64 %i.019, 1
+  %exitcond22 = icmp eq i64 %inc7, %N
+  br i1 %exitcond22, label %for.cond.cleanup, label %for.cond1.preheader
+
+for.cond.cleanup:                                 ; preds = %for.cond1.for.cond.cleanup3_crit_edge, %entry
+  %sum.0.lcssa = phi i64 [ 0, %entry ], [ %spec.select, %for.cond1.for.cond.cleanup3_crit_edge ]
+  ret i64 %sum.0.lcssa
+}
+
+define i64 @three_dimensional_middle(ptr %a, ptr %b, i64 %N, i64 %M, i64 %K) {
----------------
david-arm wrote:

Might be useful to have a comment explaining how this is different to `@three_dimensional`, since it's not immediately obvious. Or you could even have simple C equivalents in comments above the functions?

https://github.com/llvm/llvm-project/pull/70796