[llvm] AMDGPU: Add baseline tests for gep flag handling (PR #110814)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Wed Oct 2 03:01:33 PDT 2024
https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/110814
We need to know the address computation won't overflow on
older subtargets to match the addressing mode of stack instructions.
>From 95b7c370afacac58559149286845aa7513013f83 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Wed, 2 Oct 2024 13:49:18 +0400
Subject: [PATCH] AMDGPU: Add baseline tests for gep flag handling
We need to know the address computation won't overflow on
older subtargets to match the addressing mode of stack instructions.
---
.../CodeGen/AMDGPU/gep-flags-stack-offsets.ll | 197 ++++++++++++++++++
1 file changed, 197 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/gep-flags-stack-offsets.ll
diff --git a/llvm/test/CodeGen/AMDGPU/gep-flags-stack-offsets.ll b/llvm/test/CodeGen/AMDGPU/gep-flags-stack-offsets.ll
new file mode 100644
index 00000000000000..782894976c711c
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/gep-flags-stack-offsets.ll
@@ -0,0 +1,197 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; gfx8 required knowing no overflow happened to fold the addressing mode
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx803 < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
+
+define void @gep_noflags_alloca(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_noflags_alloca:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_noflags_alloca:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, v2
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [32 x i32], addrspace(5)
+ %gep0 = getelementptr [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx
+ %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1
+ ret void
+}
+
+define void @gep_inbounds_alloca(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_inbounds_alloca:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_inbounds_alloca:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, v2
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [32 x i32], addrspace(5)
+ %gep0 = getelementptr inbounds [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx
+ %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1
+ ret void
+}
+
+define void @gep_nuw_alloca(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_nuw_alloca:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_nuw_alloca:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, v2
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [32 x i32], addrspace(5)
+ %gep0 = getelementptr nuw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx
+ %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1
+ ret void
+}
+
+define void @gep_nusw_alloca(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_nusw_alloca:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_nusw_alloca:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, v2
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [32 x i32], addrspace(5)
+ %gep0 = getelementptr nusw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx
+ %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1
+ ret void
+}
+
+define void @gep_inbounds_nuw_alloca(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_inbounds_nuw_alloca:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_inbounds_nuw_alloca:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, v2
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [32 x i32], addrspace(5)
+ %gep0 = getelementptr inbounds nuw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx
+ %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1
+ ret void
+}
+
+define void @gep_nusw_nuw_alloca(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_nusw_nuw_alloca:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_nusw_nuw_alloca:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_lshl_add_u32 v0, v0, 2, v2
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [32 x i32], addrspace(5)
+ %gep0 = getelementptr nusw nuw [32 x i32], ptr addrspace(5) %alloca, i32 0, i32 %idx
+ %gep1 = getelementptr i32, ptr addrspace(5) %gep0, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1
+ ret void
+}
+
+define void @gep_inbounds_nuw_alloca_nonpow2_scale(i32 %idx, i32 %val) #0 {
+; GFX8-LABEL: gep_inbounds_nuw_alloca_nonpow2_scale:
+; GFX8: ; %bb.0:
+; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT: s_movk_i32 s4, 0x84
+; GFX8-NEXT: v_mul_lo_u32 v0, v0, s4
+; GFX8-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, v2, v0
+; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
+; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
+; GFX8-NEXT: s_waitcnt vmcnt(0)
+; GFX8-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: gep_inbounds_nuw_alloca_nonpow2_scale:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT: s_movk_i32 s4, 0x84
+; GFX9-NEXT: v_mul_lo_u32 v0, v0, s4
+; GFX9-NEXT: v_lshrrev_b32_e64 v2, 6, s32
+; GFX9-NEXT: v_add_u32_e32 v0, v2, v0
+; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: s_setpc_b64 s[30:31]
+ %alloca = alloca [5 x [33 x i32]], align 4, addrspace(5)
+ %gep1 = getelementptr inbounds nuw [5 x [33 x i32]], ptr addrspace(5) %alloca, i32 0, i32 %idx, i32 4
+ store volatile i32 %val, ptr addrspace(5) %gep1, align 4
+ ret void
+}
+
+attributes #0 = { nounwind }
More information about the llvm-commits
mailing list