[AMDGPU][SDAG] Add tests for ISD::PTRADD DAG combines

[ritter-x2a]

Pre-committing tests to show improvements in a follow-up PR with the
combines.

[ritter-x2a]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• [AMDGPU][SDAG] Handle ISD::PTRADD in SelectionDAGAddressAnalysis #142778
• [AMDGPU][SDAG] Add test for ISD::PTRADD handling in SelectionDAGAddressAnalysis #142777
• [AMDGPU][SDAG] Add ISD::PTRADD DAG combines #142739
• [AMDGPU][SDAG] Add tests for ISD::PTRADD DAG combines #142738 👈 (View in Graphite)
• [AMDGPU][SDAG] Initial support for ISD::PTRADD #141725
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Fabian Ritter (ritter-x2a)

Changes

Pre-committing tests to show improvements in a follow-up PR with the
combines.

---

Full diff: https://github.com/llvm/llvm-project/pull/142738.diff

1 Files Affected:

• (added) llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll (+207)

diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
new file mode 100644
index 0000000000000..0241be9197e1a
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
@@ -0,0 +1,207 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -disable-separate-const-offset-from-gep=1 -amdgpu-use-sdag-ptradd=1 < %s | FileCheck --check-prefixes=GFX942,GFX942_PTRADD %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -disable-separate-const-offset-from-gep=1 -amdgpu-use-sdag-ptradd=0 < %s | FileCheck --check-prefixes=GFX942,GFX942_LEGACY %s
+
+; Tests for DAG combines and folds related to the ISD::PTRADD SelectionDAG
+; opcode. The RUN lines uses -disable-separate-const-offset-from-gep to disable
+; similar transformations in that pass.
+
+; Tests reassociation (ptradd N0:(ptradd p, c1), z) where N0 has only one use.
+define i64 @global_load_ZTwoUses(ptr addrspace(1) %base, i64 %voffset) {
+; GFX942_PTRADD-LABEL: global_load_ZTwoUses:
+; GFX942_PTRADD:       ; %bb.0:
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, 24
+; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
+; GFX942_PTRADD-NEXT:    global_load_dwordx2 v[0:1], v[0:1], off
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
+; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX942_LEGACY-LABEL: global_load_ZTwoUses:
+; GFX942_LEGACY:       ; %bb.0:
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
+; GFX942_LEGACY-NEXT:    global_load_dwordx2 v[0:1], v[0:1], off offset:24
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
+; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %base, i64 24
+  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 %voffset
+  %l = load i64, ptr addrspace(1) %gep1, align 8
+  %r = add i64 %l, %voffset
+  ret i64 %r
+}
+
+define i64 @global_load_gep_add_reassoc(ptr addrspace(1) %base, i64 %voffset) {
+; GFX942_PTRADD-LABEL: global_load_gep_add_reassoc:
+; GFX942_PTRADD:       ; %bb.0:
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[2:3], v[2:3], 0, 24
+; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 0, v[2:3]
+; GFX942_PTRADD-NEXT:    global_load_dwordx2 v[0:1], v[0:1], off
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX942_LEGACY-LABEL: global_load_gep_add_reassoc:
+; GFX942_LEGACY:       ; %bb.0:
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942_LEGACY-NEXT:    v_lshl_add_u64 v[0:1], v[2:3], 0, v[0:1]
+; GFX942_LEGACY-NEXT:    global_load_dwordx2 v[0:1], v[0:1], off offset:24
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+  %add0 = add nuw nsw i64 %voffset, 24
+  %gep0 = getelementptr nuw inbounds i8, ptr addrspace(1) %base, i64 %add0
+  %l = load i64, ptr addrspace(1) %gep0, align 8
+  ret i64 %l
+}
+
+; Tests reassociation (ptradd (ptradd p, c1), c2) with two constants. These
+; would be folded away in most cases, but the index computation introduced by
+; the legalization of wide vector stores can for example introduce them.
+define amdgpu_kernel void @store_v16i32(ptr addrspace(1) %out, <16 x i32> %a) {
+; GFX942_PTRADD-LABEL: store_v16i32:
+; GFX942_PTRADD:       ; %bb.0: ; %entry
+; GFX942_PTRADD-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX942_PTRADD-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x40
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v4, 0
+; GFX942_PTRADD-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942_PTRADD-NEXT:    s_add_u32 s2, s0, 32
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v0, s20
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v1, s21
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v2, s22
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v3, s23
+; GFX942_PTRADD-NEXT:    s_addc_u32 s3, s1, 0
+; GFX942_PTRADD-NEXT:    global_store_dwordx4 v4, v[0:3], s[2:3] offset:16
+; GFX942_PTRADD-NEXT:    s_nop 1
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v0, s16
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v1, s17
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v2, s18
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v3, s19
+; GFX942_PTRADD-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1] offset:32
+; GFX942_PTRADD-NEXT:    s_nop 1
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v0, s12
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v1, s13
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v2, s14
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v3, s15
+; GFX942_PTRADD-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1] offset:16
+; GFX942_PTRADD-NEXT:    s_nop 1
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v0, s8
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v1, s9
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v2, s10
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v3, s11
+; GFX942_PTRADD-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1]
+; GFX942_PTRADD-NEXT:    s_endpgm
+;
+; GFX942_LEGACY-LABEL: store_v16i32:
+; GFX942_LEGACY:       ; %bb.0: ; %entry
+; GFX942_LEGACY-NEXT:    s_load_dwordx16 s[8:23], s[4:5], 0x40
+; GFX942_LEGACY-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v4, 0
+; GFX942_LEGACY-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v0, s20
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v1, s21
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v2, s22
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v3, s23
+; GFX942_LEGACY-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1] offset:48
+; GFX942_LEGACY-NEXT:    s_nop 1
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v0, s16
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v1, s17
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v2, s18
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v3, s19
+; GFX942_LEGACY-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1] offset:32
+; GFX942_LEGACY-NEXT:    s_nop 1
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v0, s12
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v1, s13
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v2, s14
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v3, s15
+; GFX942_LEGACY-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1] offset:16
+; GFX942_LEGACY-NEXT:    s_nop 1
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v0, s8
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v1, s9
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v2, s10
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v3, s11
+; GFX942_LEGACY-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1]
+; GFX942_LEGACY-NEXT:    s_endpgm
+entry:
+  store <16 x i32> %a, ptr addrspace(1) %out
+  ret void
+}
+
+
+; Tests the (ptradd 0, x) -> x DAG combine.
+define void @baseptr_null(i64 %offset, i8 %v) {
+; GFX942_PTRADD-LABEL: baseptr_null:
+; GFX942_PTRADD:       ; %bb.0:
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942_PTRADD-NEXT:    v_lshl_add_u64 v[0:1], 0, 0, v[0:1]
+; GFX942_PTRADD-NEXT:    flat_store_byte v[0:1], v2
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942_PTRADD-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX942_LEGACY-LABEL: baseptr_null:
+; GFX942_LEGACY:       ; %bb.0:
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX942_LEGACY-NEXT:    flat_store_byte v[0:1], v2
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX942_LEGACY-NEXT:    s_setpc_b64 s[30:31]
+  %gep = getelementptr i8, ptr null, i64 %offset
+  store i8 %v, ptr %gep, align 1
+  ret void
+}
+
+; Taken from implicit-kernarg-backend-usage.ll, tests the PTRADD handling in the
+; assertalign DAG combine.
+define amdgpu_kernel void @llvm_amdgcn_queue_ptr(ptr addrspace(1) %ptr)  #0 {
+; GFX942_PTRADD-LABEL: llvm_amdgcn_queue_ptr:
+; GFX942_PTRADD:       ; %bb.0:
+; GFX942_PTRADD-NEXT:    s_add_u32 s8, s4, 8
+; GFX942_PTRADD-NEXT:    v_mov_b32_e32 v2, 0
+; GFX942_PTRADD-NEXT:    s_addc_u32 s9, s5, 0
+; GFX942_PTRADD-NEXT:    global_load_ubyte v0, v2, s[2:3] sc0 sc1
+; GFX942_PTRADD-NEXT:    global_load_ubyte v0, v2, s[8:9] sc0 sc1
+; GFX942_PTRADD-NEXT:    global_load_ubyte v0, v2, s[0:1] sc0 sc1
+; GFX942_PTRADD-NEXT:    ; kill: killed $sgpr0_sgpr1
+; GFX942_PTRADD-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_PTRADD-NEXT:    v_mov_b64_e32 v[0:1], s[6:7]
+; GFX942_PTRADD-NEXT:    ; kill: killed $sgpr8 killed $sgpr9
+; GFX942_PTRADD-NEXT:    ; kill: killed $sgpr2_sgpr3
+; GFX942_PTRADD-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942_PTRADD-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1] sc0 sc1
+; GFX942_PTRADD-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_PTRADD-NEXT:    s_endpgm
+;
+; GFX942_LEGACY-LABEL: llvm_amdgcn_queue_ptr:
+; GFX942_LEGACY:       ; %bb.0:
+; GFX942_LEGACY-NEXT:    v_mov_b32_e32 v2, 0
+; GFX942_LEGACY-NEXT:    global_load_ubyte v0, v2, s[2:3] sc0 sc1
+; GFX942_LEGACY-NEXT:    global_load_ubyte v0, v2, s[4:5] offset:8 sc0 sc1
+; GFX942_LEGACY-NEXT:    global_load_ubyte v0, v2, s[0:1] sc0 sc1
+; GFX942_LEGACY-NEXT:    ; kill: killed $sgpr0_sgpr1
+; GFX942_LEGACY-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_LEGACY-NEXT:    v_mov_b64_e32 v[0:1], s[6:7]
+; GFX942_LEGACY-NEXT:    ; kill: killed $sgpr2_sgpr3
+; GFX942_LEGACY-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX942_LEGACY-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1] sc0 sc1
+; GFX942_LEGACY-NEXT:    s_waitcnt vmcnt(0)
+; GFX942_LEGACY-NEXT:    s_endpgm
+  %queue.ptr = call ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+  %implicitarg.ptr = call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
+  %dispatch.ptr = call ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+  %dispatch.id = call i64 @llvm.amdgcn.dispatch.id()
+  %queue.load = load volatile i8, ptr addrspace(4) %queue.ptr
+  %implicitarg.load = load volatile i8, ptr addrspace(4) %implicitarg.ptr
+  %dispatch.load = load volatile i8, ptr addrspace(4) %dispatch.ptr
+  store volatile i64 %dispatch.id, ptr addrspace(1) %ptr
+  ret void
+}
+
+declare noalias ptr addrspace(4) @llvm.amdgcn.queue.ptr()
+declare noalias ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()
+declare i64 @llvm.amdgcn.dispatch.id()
+declare noalias ptr addrspace(4) @llvm.amdgcn.dispatch.ptr()
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GFX942: {{.*}}