AMDGPU: Do not infer implicit inputs for !nocallback intrinsics

[arsenm]

This isn't really the right check, we want to know that the intrinsic
does not perform a true function call to any code (in the module or not). nocallback
appears to be the closest thing to this property we have now though. Fixes theoretically
miscompiles with intrinsics like statepoint, which hide a call to a real function.

Also do the same for inferring no-agpr usage.

[arsenm]

• AMDGPU: Do not infer implicit inputs for !nocallback intrinsics  #131759 👈 (View in Graphite)
• AMDGPU: Fix attributor not handling all trap intrinsics #131758
• AMDGPU: Fix broken check prefix and degraded cov4 test coverage #131757
• main

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

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

This isn't really the right check, we want to know that the intrinsic
does not perform a true function call to any code (in the module or not). nocallback
appears to be the closest thing to this property we have now though.

Also do the same for inferring no-agpr usage.

---

Full diff: https://github.com/llvm/llvm-project/pull/131759.diff

3 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp (+24-5)
• (added) llvm/test/CodeGen/AMDGPU/amdgpu-attributor-intrinsic-missing-nocallback.ll (+31)
• (added) llvm/test/CodeGen/AMDGPU/amdgpu-attributor-nocallback-intrinsics.ll (+75)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
index 0cee3c3cb5e92..9d410ae5a55e6 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp
@@ -48,9 +48,10 @@ enum ImplicitArgumentPositions {
 #define AMDGPU_ATTRIBUTE(Name, Str) Name = 1 << Name##_POS,
 
 enum ImplicitArgumentMask {
-  NOT_IMPLICIT_INPUT = 0,
+  UNKNOWN_INTRINSIC = 0,
 #include "AMDGPUAttributes.def"
-  ALL_ARGUMENT_MASK = (1 << LAST_ARG_POS) - 1
+  ALL_ARGUMENT_MASK = (1 << LAST_ARG_POS) - 1,
+  NOT_IMPLICIT_INPUT
 };
 
 #define AMDGPU_ATTRIBUTE(Name, Str) {Name, Str},
@@ -118,7 +119,7 @@ intrinsicToAttrMask(Intrinsic::ID ID, bool &NonKernelOnly, bool &NeedsImplicit,
     NeedsImplicit = (CodeObjectVersion >= AMDGPU::AMDHSA_COV5);
     return QUEUE_PTR;
   default:
-    return NOT_IMPLICIT_INPUT;
+    return UNKNOWN_INTRINSIC;
   }
 }
 
@@ -522,6 +523,21 @@ struct AAAMDAttributesFunction : public AAAMDAttributes {
       ImplicitArgumentMask AttrMask =
           intrinsicToAttrMask(IID, NonKernelOnly, NeedsImplicit,
                               HasApertureRegs, SupportsGetDoorbellID, COV);
+
+      if (AttrMask == UNKNOWN_INTRINSIC) {
+        // Assume not-nocallback intrinsics may invoke a function which accesses
+        // implicit arguments.
+        //
+        // FIXME: This isn't really the correct check. We want to ensure it
+        // isn't calling any function that may use implicit arguments regardless
+        // of whether it's internal to the module or not.
+        //
+        // TODO: Ignoring callsite attributes.
+        if (!Callee->hasFnAttribute(Attribute::NoCallback))
+          return indicatePessimisticFixpoint();
+        continue;
+      }
+
       if (AttrMask != NOT_IMPLICIT_INPUT) {
         if ((IsNonEntryFunc || !NonKernelOnly))
           removeAssumedBits(AttrMask);
@@ -1282,8 +1298,11 @@ struct AAAMDGPUNoAGPR
 
       // Some intrinsics may use AGPRs, but if we have a choice, we are not
       // required to use AGPRs.
-      if (Callee->isIntrinsic())
-        return true;
+      if (Callee->isIntrinsic()) {
+        // Assume !nocallback intrinsics may call a function which requires
+        // AGPRs.
+        return CB.hasFnAttr(Attribute::NoCallback);
+      }
 
       // TODO: Handle callsite attributes
       const auto *CalleeInfo = A.getAAFor<AAAMDGPUNoAGPR>(
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-intrinsic-missing-nocallback.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-intrinsic-missing-nocallback.ll
new file mode 100644
index 0000000000000..892bfa12140d4
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-intrinsic-missing-nocallback.ll
@@ -0,0 +1,31 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-attributes --check-globals all --version 5
+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx90a -passes=amdgpu-attributor %s | FileCheck %s
+
+; Make sure we do not infer anything about implicit inputs through an
+; intrinsic call which is not nocallback.
+
+declare zeroext i32 @return_i32()
+
+define i32 @test_i32_return() gc "statepoint-example" {
+; CHECK-LABEL: define i32 @test_i32_return(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] gc "statepoint-example" {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[SAFEPOINT_TOKEN:%.*]] = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(i32 ()) @return_i32, i32 0, i32 0, i32 0, i32 0)
+; CHECK-NEXT:    [[CALL1:%.*]] = call zeroext i32 @llvm.experimental.gc.result.i32(token [[SAFEPOINT_TOKEN]])
+; CHECK-NEXT:    ret i32 [[CALL1]]
+;
+entry:
+  %safepoint_token = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(i32 ()) @return_i32, i32 0, i32 0, i32 0, i32 0)
+  %call1 = call zeroext i32 @llvm.experimental.gc.result.i32(token %safepoint_token)
+  ret i32 %call1
+}
+
+declare token @llvm.experimental.gc.statepoint.p0(i64 immarg, i32 immarg, ptr, i32 immarg, i32 immarg, ...)
+declare i32 @llvm.experimental.gc.result.i32(token) #0
+
+attributes #0 = { nocallback nofree nosync nounwind willreturn memory(none) }
+;.
+; CHECK: attributes #[[ATTR0]] = { "target-cpu"="gfx90a" "uniform-work-group-size"="false" }
+; CHECK: attributes #[[ATTR1:[0-9]+]] = { "target-cpu"="gfx90a" }
+; CHECK: attributes #[[ATTR2:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(none) "target-cpu"="gfx90a" }
+;.
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-nocallback-intrinsics.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-nocallback-intrinsics.ll
new file mode 100644
index 0000000000000..b607c6cd8e720
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-attributor-nocallback-intrinsics.ll
@@ -0,0 +1,75 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-attributes --check-globals all --version 5
+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -passes=amdgpu-attributor -mcpu=gfx90a %s | FileCheck %s
+
+; Make sure we infer no inputs are used through some intrinsics
+
+define void @use_fake_use(i32 %arg) {
+; CHECK-LABEL: define void @use_fake_use(
+; CHECK-SAME: i32 [[ARG:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:    call void (...) @llvm.fake.use(i32 [[ARG]])
+; CHECK-NEXT:    ret void
+;
+  call void (...) @llvm.fake.use(i32 %arg)
+  ret void
+}
+
+define void @use_donothing() {
+; CHECK-LABEL: define void @use_donothing(
+; CHECK-SAME: ) #[[ATTR1:[0-9]+]] {
+; CHECK-NEXT:    call void @llvm.donothing()
+; CHECK-NEXT:    ret void
+;
+  call void @llvm.donothing()
+  ret void
+}
+
+define void @use_assume(i1 %arg) {
+; CHECK-LABEL: define void @use_assume(
+; CHECK-SAME: i1 [[ARG:%.*]]) #[[ATTR1]] {
+; CHECK-NEXT:    call void @llvm.assume(i1 [[ARG]])
+; CHECK-NEXT:    ret void
+;
+  call void @llvm.assume(i1 %arg)
+  ret void
+}
+
+define void @use_trap() {
+; CHECK-LABEL: define void @use_trap(
+; CHECK-SAME: ) #[[ATTR2:[0-9]+]] {
+; CHECK-NEXT:    call void @llvm.trap()
+; CHECK-NEXT:    ret void
+;
+  call void @llvm.trap()
+  ret void
+}
+
+define void @use_debugtrap() {
+; CHECK-LABEL: define void @use_debugtrap(
+; CHECK-SAME: ) #[[ATTR2]] {
+; CHECK-NEXT:    call void @llvm.debugtrap()
+; CHECK-NEXT:    ret void
+;
+  call void @llvm.debugtrap()
+  ret void
+}
+
+define void @use_ubsantrap() {
+; CHECK-LABEL: define void @use_ubsantrap(
+; CHECK-SAME: ) #[[ATTR2]] {
+; CHECK-NEXT:    call void @llvm.ubsantrap(i8 0)
+; CHECK-NEXT:    ret void
+;
+  call void @llvm.ubsantrap(i8 0)
+  ret void
+}
+
+;.
+; CHECK: attributes #[[ATTR0]] = { "target-cpu"="gfx90a" "uniform-work-group-size"="false" }
+; CHECK: attributes #[[ATTR1]] = { "amdgpu-agpr-alloc"="0" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "target-cpu"="gfx90a" "uniform-work-group-size"="false" }
+; CHECK: attributes #[[ATTR2]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "target-cpu"="gfx90a" "uniform-work-group-size"="false" }
+; CHECK: attributes #[[ATTR3:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: write) "target-cpu"="gfx90a" }
+; CHECK: attributes #[[ATTR4:[0-9]+]] = { nounwind "target-cpu"="gfx90a" }
+; CHECK: attributes #[[ATTR5:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(none) "target-cpu"="gfx90a" }
+; CHECK: attributes #[[ATTR6:[0-9]+]] = { cold noreturn nounwind memory(inaccessiblemem: write) "target-cpu"="gfx90a" }
+; CHECK: attributes #[[ATTR7:[0-9]+]] = { cold noreturn nounwind "target-cpu"="gfx90a" }
+;.

[shiltian]

This needs an iterative approach, similar to other AAs, and then propagate the "attribute".