[llvm] [AMDGPU] Allow sinking of free vector ops (PR #162580)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,146 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 < %s | FileCheck -check-prefix=OPT %s
+
+; testing insert case
+define amdgpu_kernel void @runningSum(ptr addrspace(1) %out0, ptr addrspace(1) %out1, i32 %inputElement0, i32 %inputElement1, i32 %inputIter) {
+; OPT-LABEL: define amdgpu_kernel void @runningSum(
+; OPT-SAME: ptr addrspace(1) [[OUT0:%.*]], ptr addrspace(1) [[OUT1:%.*]], i32 [[INPUTELEMENT0:%.*]], i32 [[INPUTELEMENT1:%.*]], i32 [[INPUTITER:%.*]]) #[[ATTR0:[0-9]+]] {
+; OPT-NEXT:  [[PREHEADER:.*]]:
+; OPT-NEXT:    [[VECELEMENT0:%.*]] = insertelement <2 x i32> poison, i32 [[INPUTELEMENT0]], i64 0
+; OPT-NEXT:    [[BROADCAST0:%.*]] = shufflevector <2 x i32> [[VECELEMENT0]], <2 x i32> poison, <2 x i32> zeroinitializer
+; OPT-NEXT:    [[VECELEMENT1:%.*]] = insertelement <2 x i32> poison, i32 [[INPUTELEMENT1]], i64 0
+; OPT-NEXT:    [[BROADCAST1:%.*]] = shufflevector <2 x i32> [[VECELEMENT1]], <2 x i32> poison, <2 x i32> zeroinitializer
+; OPT-NEXT:    br label %[[LOOPBODY:.*]]
+; OPT:       [[LOOPBODY]]:
+; OPT-NEXT:    [[PREVIOUSSUM:%.*]] = phi <2 x i32> [ [[BROADCAST1]], %[[PREHEADER]] ], [ [[RUNNINGSUM:%.*]], %[[LOOPBODY]] ]
+; OPT-NEXT:    [[ITERCOUNT:%.*]] = phi i32 [ [[INPUTITER]], %[[PREHEADER]] ], [ [[ITERSLEFT:%.*]], %[[LOOPBODY]] ]
+; OPT-NEXT:    [[TMP0:%.*]] = shufflevector <2 x i32> [[VECELEMENT1]], <2 x i32> poison, <2 x i32> zeroinitializer
+; OPT-NEXT:    [[RUNNINGSUM]] = add <2 x i32> [[TMP0]], [[PREVIOUSSUM]]
+; OPT-NEXT:    [[ITERSLEFT]] = sub i32 [[ITERCOUNT]], 1
+; OPT-NEXT:    [[COND:%.*]] = icmp eq i32 [[ITERSLEFT]], 0
+; OPT-NEXT:    br i1 [[COND]], label %[[LOOPEXIT:.*]], label %[[LOOPBODY]]
+; OPT:       [[LOOPEXIT]]:
+; OPT-NEXT:    [[SUMELEMENT0:%.*]] = extractelement <2 x i32> [[RUNNINGSUM]], i64 0
+; OPT-NEXT:    [[SUMELEMENT1:%.*]] = extractelement <2 x i32> [[RUNNINGSUM]], i64 1
+; OPT-NEXT:    store i32 [[SUMELEMENT0]], ptr addrspace(1) [[OUT0]], align 4
+; OPT-NEXT:    store i32 [[SUMELEMENT1]], ptr addrspace(1) [[OUT1]], align 4
+; OPT-NEXT:    ret void
+;
+preheader:
+  %vecElement0 = insertelement <2 x i32> poison, i32 %inputElement0, i64 0
+  %broadcast0 = shufflevector <2 x i32> %vecElement0, <2 x i32> poison, <2 x i32> zeroinitializer
+  %vecElement1 = insertelement <2 x i32> poison, i32 %inputElement1, i64 0
+  %broadcast1 = shufflevector <2 x i32> %vecElement1, <2 x i32> poison, <2 x i32> zeroinitializer
+  br label %loopBody
+
+loopBody:
+  %previousSum = phi <2 x i32> [ %broadcast1, %preheader ], [ %runningSum, %loopBody ]
+  %iterCount = phi i32 [ %inputIter, %preheader ], [ %itersLeft, %loopBody ]
+  %runningSum = add <2 x i32> %broadcast1, %previousSum
+  %itersLeft = sub i32 %iterCount, 1
+  %cond = icmp eq i32 %itersLeft, 0
+  br i1 %cond, label %loopExit, label %loopBody
+
+loopExit:
+  %sumElement0 = extractelement <2 x i32> %runningSum, i64 0
+  %sumElement1 = extractelement <2 x i32> %runningSum, i64 1
+  store i32 %sumElement0, ptr addrspace(1) %out0
+  store i32 %sumElement1, ptr addrspace(1) %out1
+  ret void
+}
+
+; testing extract case with single use
+define amdgpu_kernel void @test_sink_extract_single_use_operands(ptr addrspace(1) %out0, <2 x i32> %inputVec) {
----------------
arsenm wrote:

Not sure how these tests are testing sinking since they have a single block 

https://github.com/llvm/llvm-project/pull/162580