Skip to content

[AMDGPU] Check for nonnull loads feeding addrspacecast #138184

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 2, 2025

Conversation

jayfoad
Copy link
Contributor

@jayfoad jayfoad commented May 1, 2025

Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.

Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.
@llvmbot
Copy link
Member

llvmbot commented May 1, 2025

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

Changes

Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.


Full diff: https://github.com/llvm/llvm-project/pull/138184.diff

2 Files Affected:

  • (modified) llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp (+5)
  • (modified) llvm/test/CodeGen/AMDGPU/codegen-prepare-addrspacecast-non-null.ll (+120-4)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
index 6617373f89c8b..3915028b557d7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
@@ -2130,6 +2130,11 @@ static bool isPtrKnownNeverNull(const Value *V, const DataLayout &DL,
   if (const auto *Arg = dyn_cast<Argument>(V); Arg && Arg->hasNonNullAttr())
     return true;
 
+  // Check nonnull loads.
+  if (const auto *Load = dyn_cast<LoadInst>(V);
+      Load && Load->hasMetadata(LLVMContext::MD_nonnull))
+    return true;
+
   // getUnderlyingObject may have looked through another addrspacecast, although
   // the optimizable situations most likely folded out by now.
   if (AS != cast<PointerType>(V->getType())->getAddressSpace())
diff --git a/llvm/test/CodeGen/AMDGPU/codegen-prepare-addrspacecast-non-null.ll b/llvm/test/CodeGen/AMDGPU/codegen-prepare-addrspacecast-non-null.ll
index 3216e71e6221a..31c23b94a8de8 100644
--- a/llvm/test/CodeGen/AMDGPU/codegen-prepare-addrspacecast-non-null.ll
+++ b/llvm/test/CodeGen/AMDGPU/codegen-prepare-addrspacecast-non-null.ll
@@ -88,6 +88,122 @@ define void @flat_to_private_nonnull_arg(ptr nonnull %ptr) {
   ret void
 }
 
+define void @local_to_flat_nonnull_load(ptr %p) {
+; OPT-LABEL: define void @local_to_flat_nonnull_load(
+; OPT-SAME: ptr [[P:%.*]]) {
+; OPT-NEXT:    [[PTR:%.*]] = load ptr addrspace(3), ptr [[P]], align 4, !nonnull [[META0:![0-9]+]]
+; OPT-NEXT:    [[X:%.*]] = call ptr @llvm.amdgcn.addrspacecast.nonnull.p0.p3(ptr addrspace(3) [[PTR]])
+; OPT-NEXT:    store volatile i32 7, ptr [[X]], align 4
+; OPT-NEXT:    ret void
+;
+; ASM-LABEL: local_to_flat_nonnull_load:
+; ASM:       ; %bb.0:
+; ASM-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; ASM-NEXT:    flat_load_dword v0, v[0:1]
+; ASM-NEXT:    s_mov_b64 s[4:5], src_shared_base
+; ASM-NEXT:    v_mov_b32_e32 v1, s5
+; ASM-NEXT:    v_mov_b32_e32 v2, 7
+; ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; ASM-NEXT:    flat_store_dword v[0:1], v2
+; ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; ASM-NEXT:    s_setpc_b64 s[30:31]
+  %ptr = load ptr addrspace(3), ptr %p, !nonnull !{}
+  %x = addrspacecast ptr addrspace(3) %ptr to ptr
+  store volatile i32 7, ptr %x
+  ret void
+}
+
+define void @private_to_flat_nonnull_load(ptr %p) {
+; OPT-LABEL: define void @private_to_flat_nonnull_load(
+; OPT-SAME: ptr [[P:%.*]]) {
+; OPT-NEXT:    [[PTR:%.*]] = load ptr addrspace(5), ptr [[P]], align 4, !nonnull [[META0]]
+; OPT-NEXT:    [[X:%.*]] = call ptr @llvm.amdgcn.addrspacecast.nonnull.p0.p5(ptr addrspace(5) [[PTR]])
+; OPT-NEXT:    store volatile i32 7, ptr [[X]], align 4
+; OPT-NEXT:    ret void
+;
+; ASM-LABEL: private_to_flat_nonnull_load:
+; ASM:       ; %bb.0:
+; ASM-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; ASM-NEXT:    flat_load_dword v0, v[0:1]
+; ASM-NEXT:    s_mov_b64 s[4:5], src_private_base
+; ASM-NEXT:    v_mov_b32_e32 v1, s5
+; ASM-NEXT:    v_mov_b32_e32 v2, 7
+; ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; ASM-NEXT:    flat_store_dword v[0:1], v2
+; ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; ASM-NEXT:    s_setpc_b64 s[30:31]
+  %ptr = load ptr addrspace(5), ptr %p, !nonnull !{}
+  %x = addrspacecast ptr addrspace(5) %ptr to ptr
+  store volatile i32 7, ptr %x
+  ret void
+}
+
+define void @flat_to_local_nonnull_load(ptr %p) {
+; OPT-LABEL: define void @flat_to_local_nonnull_load(
+; OPT-SAME: ptr [[P:%.*]]) {
+; OPT-NEXT:    [[PTR:%.*]] = load ptr, ptr [[P]], align 8, !nonnull [[META0]]
+; OPT-NEXT:    [[X:%.*]] = call ptr addrspace(3) @llvm.amdgcn.addrspacecast.nonnull.p3.p0(ptr [[PTR]])
+; OPT-NEXT:    store volatile i32 7, ptr addrspace(3) [[X]], align 4
+; OPT-NEXT:    ret void
+;
+; DAGISEL-ASM-LABEL: flat_to_local_nonnull_load:
+; DAGISEL-ASM:       ; %bb.0:
+; DAGISEL-ASM-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; DAGISEL-ASM-NEXT:    flat_load_dword v0, v[0:1]
+; DAGISEL-ASM-NEXT:    v_mov_b32_e32 v1, 7
+; DAGISEL-ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; DAGISEL-ASM-NEXT:    ds_write_b32 v0, v1
+; DAGISEL-ASM-NEXT:    s_waitcnt lgkmcnt(0)
+; DAGISEL-ASM-NEXT:    s_setpc_b64 s[30:31]
+;
+; GISEL-ASM-LABEL: flat_to_local_nonnull_load:
+; GISEL-ASM:       ; %bb.0:
+; GISEL-ASM-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-ASM-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GISEL-ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GISEL-ASM-NEXT:    v_mov_b32_e32 v1, 7
+; GISEL-ASM-NEXT:    ds_write_b32 v0, v1
+; GISEL-ASM-NEXT:    s_waitcnt lgkmcnt(0)
+; GISEL-ASM-NEXT:    s_setpc_b64 s[30:31]
+  %ptr = load ptr, ptr %p, !nonnull !{}
+  %x = addrspacecast ptr %ptr to ptr addrspace(3)
+  store volatile i32 7, ptr addrspace(3) %x
+  ret void
+}
+
+define void @flat_to_private_nonnull_load(ptr %p) {
+; OPT-LABEL: define void @flat_to_private_nonnull_load(
+; OPT-SAME: ptr [[P:%.*]]) {
+; OPT-NEXT:    [[PTR:%.*]] = load ptr, ptr [[P]], align 8, !nonnull [[META0]]
+; OPT-NEXT:    [[X:%.*]] = call ptr addrspace(5) @llvm.amdgcn.addrspacecast.nonnull.p5.p0(ptr [[PTR]])
+; OPT-NEXT:    store volatile i32 7, ptr addrspace(5) [[X]], align 4
+; OPT-NEXT:    ret void
+;
+; DAGISEL-ASM-LABEL: flat_to_private_nonnull_load:
+; DAGISEL-ASM:       ; %bb.0:
+; DAGISEL-ASM-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; DAGISEL-ASM-NEXT:    flat_load_dword v0, v[0:1]
+; DAGISEL-ASM-NEXT:    v_mov_b32_e32 v1, 7
+; DAGISEL-ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; DAGISEL-ASM-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
+; DAGISEL-ASM-NEXT:    s_waitcnt vmcnt(0)
+; DAGISEL-ASM-NEXT:    s_setpc_b64 s[30:31]
+;
+; GISEL-ASM-LABEL: flat_to_private_nonnull_load:
+; GISEL-ASM:       ; %bb.0:
+; GISEL-ASM-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-ASM-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GISEL-ASM-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GISEL-ASM-NEXT:    v_mov_b32_e32 v1, 7
+; GISEL-ASM-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
+; GISEL-ASM-NEXT:    s_waitcnt vmcnt(0)
+; GISEL-ASM-NEXT:    s_setpc_b64 s[30:31]
+  %ptr = load ptr, ptr %p, !nonnull !{}
+  %x = addrspacecast ptr %ptr to ptr addrspace(5)
+  store volatile i32 7, ptr addrspace(5) %x
+  ret void
+}
+
 define void @private_alloca_to_flat(ptr %ptr) {
 ; OPT-LABEL: define void @private_alloca_to_flat(
 ; OPT-SAME: ptr [[PTR:%.*]]) {
@@ -218,7 +334,7 @@ define void @recursive_phis(i1 %cond, ptr addrspace(5) %ptr) {
 ; DAGISEL-ASM-NEXT:    s_mov_b64 s[4:5], 0
 ; DAGISEL-ASM-NEXT:    s_mov_b64 s[8:9], src_private_base
 ; DAGISEL-ASM-NEXT:    v_mov_b32_e32 v2, 7
-; DAGISEL-ASM-NEXT:  .LBB7_3: ; %finally
+; DAGISEL-ASM-NEXT:  .LBB11_3: ; %finally
 ; DAGISEL-ASM-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; DAGISEL-ASM-NEXT:    s_and_b64 s[10:11], exec, s[6:7]
 ; DAGISEL-ASM-NEXT:    s_or_b64 s[4:5], s[10:11], s[4:5]
@@ -226,7 +342,7 @@ define void @recursive_phis(i1 %cond, ptr addrspace(5) %ptr) {
 ; DAGISEL-ASM-NEXT:    flat_store_dword v[0:1], v2
 ; DAGISEL-ASM-NEXT:    s_waitcnt vmcnt(0)
 ; DAGISEL-ASM-NEXT:    s_andn2_b64 exec, exec, s[4:5]
-; DAGISEL-ASM-NEXT:    s_cbranch_execnz .LBB7_3
+; DAGISEL-ASM-NEXT:    s_cbranch_execnz .LBB11_3
 ; DAGISEL-ASM-NEXT:  ; %bb.4: ; %end
 ; DAGISEL-ASM-NEXT:    s_or_b64 exec, exec, s[4:5]
 ; DAGISEL-ASM-NEXT:    s_waitcnt lgkmcnt(0)
@@ -249,14 +365,14 @@ define void @recursive_phis(i1 %cond, ptr addrspace(5) %ptr) {
 ; GISEL-ASM-NEXT:    s_mov_b64 s[6:7], 0
 ; GISEL-ASM-NEXT:    v_mov_b32_e32 v1, s9
 ; GISEL-ASM-NEXT:    v_mov_b32_e32 v2, 7
-; GISEL-ASM-NEXT:  .LBB7_3: ; %finally
+; GISEL-ASM-NEXT:  .LBB11_3: ; %finally
 ; GISEL-ASM-NEXT:    ; =>This Inner Loop Header: Depth=1
 ; GISEL-ASM-NEXT:    s_and_b64 s[8:9], exec, s[4:5]
 ; GISEL-ASM-NEXT:    s_or_b64 s[6:7], s[8:9], s[6:7]
 ; GISEL-ASM-NEXT:    flat_store_dword v[0:1], v2
 ; GISEL-ASM-NEXT:    s_waitcnt vmcnt(0)
 ; GISEL-ASM-NEXT:    s_andn2_b64 exec, exec, s[6:7]
-; GISEL-ASM-NEXT:    s_cbranch_execnz .LBB7_3
+; GISEL-ASM-NEXT:    s_cbranch_execnz .LBB11_3
 ; GISEL-ASM-NEXT:  ; %bb.4: ; %end
 ; GISEL-ASM-NEXT:    s_or_b64 exec, exec, s[6:7]
 ; GISEL-ASM-NEXT:    s_waitcnt lgkmcnt(0)

Copy link
Contributor

@arsenm arsenm left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@jayfoad jayfoad merged commit 9060ca0 into llvm:main May 2, 2025
13 checks passed
@jayfoad jayfoad deleted the nonnull-loads branch May 2, 2025 11:54
IanWood1 pushed a commit to IanWood1/llvm-project that referenced this pull request May 6, 2025
Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.
IanWood1 pushed a commit to IanWood1/llvm-project that referenced this pull request May 6, 2025
Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.
IanWood1 pushed a commit to IanWood1/llvm-project that referenced this pull request May 6, 2025
Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.
GeorgeARM pushed a commit to GeorgeARM/llvm-project that referenced this pull request May 7, 2025
Handle nonnull loads just like nonnull arguments when checking for
addrspacecasts that are known never null.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants