[llvm] [FunctionAttrs] Improve handling of alias-preserving intrinsic calls (PR #68453)

Krzysztof Drewniak via llvm-commits llvm-commits at lists.llvm.org
Fri Oct 20 07:58:22 PDT 2023


https://github.com/krzysz00 updated https://github.com/llvm/llvm-project/pull/68453

>From 649da847a4fa39756e9d803084e400889917bf8d Mon Sep 17 00:00:00 2001
From: Krzysztof Drewniak <Krzysztof.Drewniak at amd.com>
Date: Fri, 6 Oct 2023 21:14:41 +0000
Subject: [PATCH] [FunctionAttrs] Improve handling of alias-preserving
 intrinsic calls

Fixes #68270

The function attribute analysis handles many instructions, like
addrspacecast, which do not themselves read or write memory but which
transform pointers into other values in the same alias set.

There are intrinsic functions, such as ptrmask or the AMDGPU-specific
make.buffer.rsrc, which also preserve membership in alias sets without
capturing. This commit adds the addrspacecast-like behavior to these
calls.
---
 llvm/lib/Transforms/IPO/FunctionAttrs.cpp     | 10 +++-
 .../FunctionAttrs/AMDGPU/lit.local.cfg        |  2 +
 .../FunctionAttrs/AMDGPU/make-buffer-rsrc.ll  | 47 +++++++++++++++++++
 .../Transforms/FunctionAttrs/writeonly.ll     | 41 ++++++++++++----
 4 files changed, 89 insertions(+), 11 deletions(-)
 create mode 100644 llvm/test/Transforms/FunctionAttrs/AMDGPU/lit.local.cfg
 create mode 100644 llvm/test/Transforms/FunctionAttrs/AMDGPU/make-buffer-rsrc.ll

diff --git a/llvm/lib/Transforms/IPO/FunctionAttrs.cpp b/llvm/lib/Transforms/IPO/FunctionAttrs.cpp
index 95b3204d02beb81..2240ef86194ffc2 100644
--- a/llvm/lib/Transforms/IPO/FunctionAttrs.cpp
+++ b/llvm/lib/Transforms/IPO/FunctionAttrs.cpp
@@ -654,7 +654,15 @@ determinePointerAccessAttrs(Argument *A,
       // must be a data operand (e.g. argument or operand bundle)
       const unsigned UseIndex = CB.getDataOperandNo(U);
 
-      if (!CB.doesNotCapture(UseIndex)) {
+      // Some intrinsics (for instance ptrmask) do not capture their results,
+      // but return results thas alias their pointer argument, and thus should
+      // be handled like GEP or addrspacecast above.
+      if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
+              &CB, /*MustPreserveNullness=*/false)) {
+        for (Use &UU : CB.uses())
+          if (Visited.insert(&UU).second)
+            Worklist.push_back(&UU);
+      } else if (!CB.doesNotCapture(UseIndex)) {
         if (!CB.onlyReadsMemory())
           // If the callee can save a copy into other memory, then simply
           // scanning uses of the call is insufficient.  We have no way
diff --git a/llvm/test/Transforms/FunctionAttrs/AMDGPU/lit.local.cfg b/llvm/test/Transforms/FunctionAttrs/AMDGPU/lit.local.cfg
new file mode 100644
index 000000000000000..7c492428aec7614
--- /dev/null
+++ b/llvm/test/Transforms/FunctionAttrs/AMDGPU/lit.local.cfg
@@ -0,0 +1,2 @@
+if not "AMDGPU" in config.root.targets:
+    config.unsupported = True
diff --git a/llvm/test/Transforms/FunctionAttrs/AMDGPU/make-buffer-rsrc.ll b/llvm/test/Transforms/FunctionAttrs/AMDGPU/make-buffer-rsrc.ll
new file mode 100644
index 000000000000000..5f9f2571e796e48
--- /dev/null
+++ b/llvm/test/Transforms/FunctionAttrs/AMDGPU/make-buffer-rsrc.ll
@@ -0,0 +1,47 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-attributes
+; RUN: opt -passes=function-attrs -S < %s | FileCheck --check-prefixes=COMMON,FNATTRS %s
+; RUN: opt -passes=attributor-light -S < %s | FileCheck --check-prefixes=COMMON,ATTRIBUTOR %s
+
+target triple = "amdgcn-amd-amdhsa"
+target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-ni:7:8"
+
+define amdgpu_kernel void @test_make_buffer_rsrc(ptr %p, ptr %q) {
+; FNATTRS: Function Attrs: mustprogress nofree nosync nounwind willreturn memory(argmem: readwrite)
+; FNATTRS-LABEL: define {{[^@]+}}@test_make_buffer_rsrc
+; FNATTRS-SAME: (ptr nocapture readonly [[P:%.*]], ptr nocapture writeonly [[Q:%.*]]) #[[ATTR0:[0-9]+]] {
+; FNATTRS-NEXT:    [[P_RSRC:%.*]] = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr [[P]], i16 0, i32 4, i32 822243328)
+; FNATTRS-NEXT:    [[Q_RSRC:%.*]] = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr [[Q]], i16 0, i32 4, i32 822243328)
+; FNATTRS-NEXT:    [[V:%.*]] = call i8 @llvm.amdgcn.raw.ptr.buffer.load.i8(ptr addrspace(8) [[P_RSRC]], i32 0, i32 0, i32 0)
+; FNATTRS-NEXT:    call void @llvm.amdgcn.raw.ptr.buffer.store.i8(i8 [[V]], ptr addrspace(8) [[Q_RSRC]], i32 0, i32 0, i32 0)
+; FNATTRS-NEXT:    ret void
+;
+; ATTRIBUTOR: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: readwrite)
+; ATTRIBUTOR-LABEL: define {{[^@]+}}@test_make_buffer_rsrc
+; ATTRIBUTOR-SAME: (ptr nocapture nofree readonly [[P:%.*]], ptr nocapture nofree writeonly [[Q:%.*]]) #[[ATTR0:[0-9]+]] {
+; ATTRIBUTOR-NEXT:    [[P_RSRC:%.*]] = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr [[P]], i16 0, i32 4, i32 822243328) #[[ATTR4:[0-9]+]]
+; ATTRIBUTOR-NEXT:    [[Q_RSRC:%.*]] = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr [[Q]], i16 0, i32 4, i32 822243328) #[[ATTR4]]
+; ATTRIBUTOR-NEXT:    [[V:%.*]] = call i8 @llvm.amdgcn.raw.ptr.buffer.load.i8(ptr addrspace(8) nocapture readonly [[P_RSRC]], i32 0, i32 0, i32 0) #[[ATTR5:[0-9]+]]
+; ATTRIBUTOR-NEXT:    call void @llvm.amdgcn.raw.ptr.buffer.store.i8(i8 [[V]], ptr addrspace(8) nocapture writeonly [[Q_RSRC]], i32 0, i32 0, i32 0) #[[ATTR6:[0-9]+]]
+; ATTRIBUTOR-NEXT:    ret void
+;
+  %p.rsrc = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr %p, i16 0, i32 4, i32 822243328)
+  %q.rsrc = call ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr %q, i16 0, i32 4, i32 822243328)
+  %v = call i8 @llvm.amdgcn.raw.ptr.buffer.load.i8(ptr addrspace(8) %p.rsrc, i32 0, i32 0, i32 0)
+  call void @llvm.amdgcn.raw.ptr.buffer.store.i8(i8 %v, ptr addrspace(8) %q.rsrc, i32 0, i32 0, i32 0)
+  ret void
+}
+
+; Function Attrs: mustprogress nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare ptr addrspace(8) @llvm.amdgcn.make.buffer.rsrc.p0(ptr nocapture readnone, i16, i32, i32) #0
+
+; Function Attrs: mustprogress nocallback nofree nosync nounwind willreturn memory(argmem: read)
+declare i8 @llvm.amdgcn.raw.ptr.buffer.load.i8(ptr addrspace(8) nocapture readonly, i32, i32, i32 immarg) #1
+
+; Function Attrs: mustprogress nocallback nofree nosync nounwind willreturn memory(argmem: write)
+declare void @llvm.amdgcn.raw.ptr.buffer.store.i8(i8, ptr addrspace(8) nocapture writeonly, i32, i32, i32 immarg) #2
+
+attributes #0 = { mustprogress nocallback nofree nosync nounwind speculatable willreturn memory(none) }
+attributes #1 = { mustprogress nocallback nofree nosync nounwind willreturn memory(argmem: read) }
+attributes #2 = { mustprogress nocallback nofree nosync nounwind willreturn memory(argmem: write) }
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; COMMON: {{.*}}
diff --git a/llvm/test/Transforms/FunctionAttrs/writeonly.ll b/llvm/test/Transforms/FunctionAttrs/writeonly.ll
index 633068129eac697..9d919b4892031aa 100644
--- a/llvm/test/Transforms/FunctionAttrs/writeonly.ll
+++ b/llvm/test/Transforms/FunctionAttrs/writeonly.ll
@@ -178,6 +178,27 @@ define void @test_atomicrmw(ptr %p) {
   ret void
 }
 
+define void @test_ptrmask(ptr %p) {
+; FNATTRS: Function Attrs: mustprogress nofree nosync nounwind willreturn memory(argmem: write)
+; FNATTRS-LABEL: define {{[^@]+}}@test_ptrmask
+; FNATTRS-SAME: (ptr writeonly [[P:%.*]]) #[[ATTR8:[0-9]+]] {
+; FNATTRS-NEXT:    [[MASK:%.*]] = call ptr @llvm.ptrmask.p0.i64(ptr [[P]], i64 -5)
+; FNATTRS-NEXT:    store i8 0, ptr [[MASK]], align 1
+; FNATTRS-NEXT:    ret void
+;
+; ATTRIBUTOR: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; ATTRIBUTOR-LABEL: define {{[^@]+}}@test_ptrmask
+; ATTRIBUTOR-SAME: (ptr nofree writeonly [[P:%.*]]) #[[ATTR3]] {
+; ATTRIBUTOR-NEXT:    [[MASK:%.*]] = call ptr @llvm.ptrmask.p0.i64(ptr [[P]], i64 -5) #[[ATTR9:[0-9]+]]
+; ATTRIBUTOR-NEXT:    store i8 0, ptr [[MASK]], align 1
+; ATTRIBUTOR-NEXT:    ret void
+;
+  %mask = call ptr @llvm.ptrmask.p0.i64(ptr %p, i64 -5)
+  store i8 0, ptr %mask
+  ret void
+}
+
+declare ptr @llvm.ptrmask.p0.i64(ptr, i64)
 
 declare void @direct1_callee(ptr %p)
 
@@ -197,14 +218,14 @@ declare void @direct2_callee(ptr %p) writeonly
 define void @direct2(ptr %p) {
 ; FNATTRS: Function Attrs: memory(write)
 ; FNATTRS-LABEL: define {{[^@]+}}@direct2
-; FNATTRS-SAME: (ptr [[P:%.*]]) #[[ATTR8:[0-9]+]] {
+; FNATTRS-SAME: (ptr [[P:%.*]]) #[[ATTR10:[0-9]+]] {
 ; FNATTRS-NEXT:    call void @direct2_callee(ptr [[P]])
 ; FNATTRS-NEXT:    ret void
 ;
 ; ATTRIBUTOR: Function Attrs: memory(write)
 ; ATTRIBUTOR-LABEL: define {{[^@]+}}@direct2
-; ATTRIBUTOR-SAME: (ptr writeonly [[P:%.*]]) #[[ATTR7:[0-9]+]] {
-; ATTRIBUTOR-NEXT:    call void @direct2_callee(ptr [[P]]) #[[ATTR7]]
+; ATTRIBUTOR-SAME: (ptr writeonly [[P:%.*]]) #[[ATTR8:[0-9]+]] {
+; ATTRIBUTOR-NEXT:    call void @direct2_callee(ptr [[P]]) #[[ATTR8]]
 ; ATTRIBUTOR-NEXT:    ret void
 ;
   call void @direct2_callee(ptr %p)
@@ -215,14 +236,14 @@ define void @direct2(ptr %p) {
 define void @direct2b(ptr %p) {
 ; FNATTRS: Function Attrs: memory(write)
 ; FNATTRS-LABEL: define {{[^@]+}}@direct2b
-; FNATTRS-SAME: (ptr nocapture [[P:%.*]]) #[[ATTR8]] {
+; FNATTRS-SAME: (ptr nocapture [[P:%.*]]) #[[ATTR10]] {
 ; FNATTRS-NEXT:    call void @direct2_callee(ptr nocapture [[P]])
 ; FNATTRS-NEXT:    ret void
 ;
 ; ATTRIBUTOR: Function Attrs: memory(write)
 ; ATTRIBUTOR-LABEL: define {{[^@]+}}@direct2b
-; ATTRIBUTOR-SAME: (ptr nocapture writeonly [[P:%.*]]) #[[ATTR7]] {
-; ATTRIBUTOR-NEXT:    call void @direct2_callee(ptr nocapture writeonly [[P]]) #[[ATTR7]]
+; ATTRIBUTOR-SAME: (ptr nocapture writeonly [[P:%.*]]) #[[ATTR8]] {
+; ATTRIBUTOR-NEXT:    call void @direct2_callee(ptr nocapture writeonly [[P]]) #[[ATTR8]]
 ; ATTRIBUTOR-NEXT:    ret void
 ;
   call void @direct2_callee(ptr nocapture %p)
@@ -304,14 +325,14 @@ define void @fptr_test2(ptr %p, ptr %f) {
 define void @fptr_test3(ptr %p, ptr %f) {
 ; FNATTRS: Function Attrs: memory(write)
 ; FNATTRS-LABEL: define {{[^@]+}}@fptr_test3
-; FNATTRS-SAME: (ptr nocapture [[P:%.*]], ptr nocapture readonly [[F:%.*]]) #[[ATTR8]] {
-; FNATTRS-NEXT:    call void [[F]](ptr nocapture [[P]]) #[[ATTR8]]
+; FNATTRS-SAME: (ptr nocapture [[P:%.*]], ptr nocapture readonly [[F:%.*]]) #[[ATTR10]] {
+; FNATTRS-NEXT:    call void [[F]](ptr nocapture [[P]]) #[[ATTR10]]
 ; FNATTRS-NEXT:    ret void
 ;
 ; ATTRIBUTOR: Function Attrs: memory(write)
 ; ATTRIBUTOR-LABEL: define {{[^@]+}}@fptr_test3
-; ATTRIBUTOR-SAME: (ptr nocapture writeonly [[P:%.*]], ptr nocapture nofree nonnull writeonly [[F:%.*]]) #[[ATTR7]] {
-; ATTRIBUTOR-NEXT:    call void [[F]](ptr nocapture [[P]]) #[[ATTR7]]
+; ATTRIBUTOR-SAME: (ptr nocapture writeonly [[P:%.*]], ptr nocapture nofree nonnull writeonly [[F:%.*]]) #[[ATTR8]] {
+; ATTRIBUTOR-NEXT:    call void [[F]](ptr nocapture [[P]]) #[[ATTR8]]
 ; ATTRIBUTOR-NEXT:    ret void
 ;
   call void %f(ptr nocapture %p) writeonly



More information about the llvm-commits mailing list