[AMDGPU] LRO: allow same-BB non-lookthrough users for PHI (#160909)

Loop headers frequently consume the loop-carried value in the header
block via non-lookthrough ops (e.g. byte-wise vector binops).
LiveRegOptimizer’s same-BB filter currently prunes these users, so the
loop-carried PHI is not coerced to i32 and the intended packed form is
lost.

Relax the filter: when the def is a PHI, allow same-BB non-lookthrough
users. Also fix the check to look at the user (CII) rather than the def
(II) so the walk does not terminate prematurely.

Author: michaelselehov

llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp

@@ -150,7 +150,10 @@ class LiveRegOptimizer {
       if (!CVisited.insert(CII).second)
         continue;
 
-      if (CII->getParent() == II->getParent() && !IsLookThru(II))
+      // Same-BB filter must look at the *user*; and allow non-lookthrough
+      // users when the def is a PHI (loop-header pattern).
+      if (CII->getParent() == II->getParent() && !IsLookThru(CII) &&
+          !isa<PHINode>(II))
         continue;
 
       if (isOpLegal(CII))

llvm/test/CodeGen/AMDGPU/lro-phi-samebb-nonlookthrough-store.ll

@@ -0,0 +1,46 @@
+; RUN: opt -S -passes=amdgpu-late-codegenprepare \
+; RUN:   -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a %s | FileCheck %s
+
+; Goal: With a loop-header PHI in illegal vector type and a same-BB
+; non-lookthrough user (vector add) in the header, LRO should still coerce
+; the PHI to i32 because a profitable sink (store) exists across BB.
+
+define amdgpu_kernel void @phi_samebb_nonlookthrough_store(
+    ptr addrspace(1) %out, <4 x i8> %v, i1 %exit) {
+; CHECK-LABEL: @phi_samebb_nonlookthrough_store(
+entry:
+  br label %loop
+
+loop:                                             ; preds = %entry, %loop
+  ; Loop-carried PHI in illegal vector type.
+  %acc = phi <4 x i8> [ zeroinitializer, %entry ], [ %acc.next, %loop ]
+
+  ; Same-BB non-lookthrough use in header.
+  %acc.next = add <4 x i8> %acc, %v
+
+  ; Make it a real loop: either iterate or exit to the sink block.
+  br i1 %exit, label %store, label %loop
+
+store:                                            ; preds = %loop
+  ; The across-BB sink: storing the PHI coerced to i32.
+  %acc.bc = bitcast <4 x i8> %acc to i32
+  store i32 %acc.bc, ptr addrspace(1) %out, align 4
+  ret void
+}
+
+; After AMDGPULateCodeGenPrepare we expect:
+;  - PHI is coerced to i32
+;  - A header bitcast materializes for the add
+; This proves the same-BB non-lookthrough user (add) did not get pruned
+; when the def is a PHI.
+
+; CHECK: loop:
+; CHECK:   %[[ACC_TC:[^ ]+]] = phi i32
+; CHECK:   %[[ACC_TC_BC:[^ ]+]] = bitcast i32 %[[ACC_TC]] to <4 x i8>
+; CHECK:   %[[ACC_NEXT:[^ ]+]] = add <4 x i8> %[[ACC_TC_BC]], %v
+; CHECK:   br i1 %exit, label %store, label %loop
+; CHECK: store:
+; CHECK:   %[[ACC_TC_BC2:[^ ]+]] = bitcast i32 %[[ACC_TC]] to <4 x i8>
+; CHECK:   %[[ST_I32:[^ ]+]] = bitcast <4 x i8> %[[ACC_TC_BC2]] to i32
+; CHECK:   store i32 %[[ST_I32]],
+