[llvm] [SeparateConstOffsetFromGEP] Decompose constant xor operand if possible (PR #150438)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,435 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; Test the xor with constant operand is decomposed in to gep.
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -passes=separate-const-offset-from-gep \
+; RUN: -S < %s | FileCheck %s
+; Test the gvn pass eliminates the redundant xor instructions from decomposition.
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -passes=separate-const-offset-from-gep,gvn \
+; RUN: -S < %s | FileCheck --check-prefix=GVN %s
+
+; Check that disjoint constants are properly extracted and folded into GEP
+; addressing modes and GVN to eliminate redundant computations
+define amdgpu_kernel void @test1(i1 %0, ptr addrspace(3) %1) {
+; CHECK-LABEL: define amdgpu_kernel void @test1(
+; CHECK-SAME: i1 [[TMP0:%.*]], ptr addrspace(3) [[TMP1:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP2:%.*]] = select i1 [[TMP0]], i32 0, i32 288
+; CHECK-NEXT:    [[TMP3:%.*]] = xor i32 [[TMP2]], 32
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr half, ptr addrspace(3) [[TMP1]], i32 [[TMP3]]
+; CHECK-NEXT:    [[TMP5:%.*]] = xor i32 [[TMP2]], 32
+; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr half, ptr addrspace(3) [[TMP1]], i32 [[TMP5]]
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr addrspace(3) [[TMP6]], i32 8192
+; CHECK-NEXT:    [[TMP8:%.*]] = xor i32 [[TMP2]], 32
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr half, ptr addrspace(3) [[TMP1]], i32 [[TMP8]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr addrspace(3) [[TMP9]], i32 16384
+; CHECK-NEXT:    [[TMP11:%.*]] = xor i32 [[TMP2]], 32
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr half, ptr addrspace(3) [[TMP1]], i32 [[TMP11]]
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr addrspace(3) [[TMP12]], i32 24576
+; CHECK-NEXT:    [[TMP14:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP4]], align 16
+; CHECK-NEXT:    [[TMP15:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP7]], align 16
+; CHECK-NEXT:    [[TMP16:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP10]], align 16
+; CHECK-NEXT:    [[TMP17:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP13]], align 16
+; CHECK-NEXT:    [[TMP18:%.*]] = fadd <8 x half> [[TMP14]], [[TMP15]]
+; CHECK-NEXT:    [[TMP19:%.*]] = fadd <8 x half> [[TMP16]], [[TMP17]]
+; CHECK-NEXT:    [[TMP20:%.*]] = fadd <8 x half> [[TMP18]], [[TMP19]]
+; CHECK-NEXT:    store <8 x half> [[TMP20]], ptr addrspace(3) [[TMP1]], align 16
+; CHECK-NEXT:    ret void
+;
+; GVN-LABEL: define amdgpu_kernel void @test1(
+; GVN-SAME: i1 [[TMP0:%.*]], ptr addrspace(3) [[TMP1:%.*]]) {
+; GVN-NEXT:  [[ENTRY:.*:]]
+; GVN-NEXT:    [[TMP2:%.*]] = select i1 [[TMP0]], i32 0, i32 288
+; GVN-NEXT:    [[TMP3:%.*]] = xor i32 [[TMP2]], 32
+; GVN-NEXT:    [[TMP4:%.*]] = getelementptr half, ptr addrspace(3) [[TMP1]], i32 [[TMP3]]
+; GVN-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr addrspace(3) [[TMP4]], i32 8192
+; GVN-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr addrspace(3) [[TMP4]], i32 16384
+; GVN-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr addrspace(3) [[TMP4]], i32 24576
+; GVN-NEXT:    [[TMP8:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP4]], align 16
+; GVN-NEXT:    [[TMP9:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP5]], align 16
+; GVN-NEXT:    [[TMP10:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP6]], align 16
+; GVN-NEXT:    [[TMP11:%.*]] = load <8 x half>, ptr addrspace(3) [[TMP7]], align 16
+; GVN-NEXT:    [[TMP12:%.*]] = fadd <8 x half> [[TMP8]], [[TMP9]]
+; GVN-NEXT:    [[TMP13:%.*]] = fadd <8 x half> [[TMP10]], [[TMP11]]
+; GVN-NEXT:    [[TMP14:%.*]] = fadd <8 x half> [[TMP12]], [[TMP13]]
+; GVN-NEXT:    store <8 x half> [[TMP14]], ptr addrspace(3) [[TMP1]], align 16
+; GVN-NEXT:    ret void
+;
+entry:
+  %2 = select i1 %0, i32 0, i32 288
----------------
arsenm wrote:

Use named values in tests 

https://github.com/llvm/llvm-project/pull/150438