[llvm] c831d83 - [InferAddrSpaces] Correctly replace identical operands of insts (#82610)

via llvm-commits llvm-commits at lists.llvm.org
Thu Feb 22 04:59:08 PST 2024


Author: Pierre van Houtryve
Date: 2024-02-22T13:59:04+01:00
New Revision: c831d83bb17caa3a8f137052559cb6c54b21b7c1

URL: https://github.com/llvm/llvm-project/commit/c831d83bb17caa3a8f137052559cb6c54b21b7c1
DIFF: https://github.com/llvm/llvm-project/commit/c831d83bb17caa3a8f137052559cb6c54b21b7c1.diff

LOG: [InferAddrSpaces] Correctly replace identical operands of insts (#82610)

It's important for PHI nodes because if a PHI node has multiple edges
coming from the same block, we can have the same incoming value multiple
times in the list of incoming values. All of those need to be consistent
(exact same Value*) otherwise verifier complains.

Fixes SWDEV-445797

Added: 
    llvm/test/Transforms/InferAddressSpaces/AMDGPU/multiple-uses-of-val.ll

Modified: 
    llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp b/llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp
index 1bf50d79e5331e..851eab04c8dbb2 100644
--- a/llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp
+++ b/llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp
@@ -1221,6 +1221,7 @@ bool InferAddressSpacesImpl::rewriteWithNewAddressSpaces(
     Value::use_iterator I, E, Next;
     for (I = V->use_begin(), E = V->use_end(); I != E;) {
       Use &U = *I;
+      User *CurUser = U.getUser();
 
       // Some users may see the same pointer operand in multiple operands. Skip
       // to the next instruction.
@@ -1231,11 +1232,10 @@ bool InferAddressSpacesImpl::rewriteWithNewAddressSpaces(
         // If V is used as the pointer operand of a compatible memory operation,
         // sets the pointer operand to NewV. This replacement does not change
         // the element type, so the resultant load/store is still valid.
-        U.set(NewV);
+        CurUser->replaceUsesOfWith(V, NewV);
         continue;
       }
 
-      User *CurUser = U.getUser();
       // Skip if the current user is the new value itself.
       if (CurUser == NewV)
         continue;
@@ -1311,10 +1311,13 @@ bool InferAddressSpacesImpl::rewriteWithNewAddressSpaces(
 
           while (isa<PHINode>(InsertPos))
             ++InsertPos;
-          U.set(new AddrSpaceCastInst(NewV, V->getType(), "", &*InsertPos));
+          // This instruction may contain multiple uses of V, update them all.
+          CurUser->replaceUsesOfWith(
+              V, new AddrSpaceCastInst(NewV, V->getType(), "", &*InsertPos));
         } else {
-          U.set(ConstantExpr::getAddrSpaceCast(cast<Constant>(NewV),
-                                               V->getType()));
+          CurUser->replaceUsesOfWith(
+              V, ConstantExpr::getAddrSpaceCast(cast<Constant>(NewV),
+                                                V->getType()));
         }
       }
     }

diff  --git a/llvm/test/Transforms/InferAddressSpaces/AMDGPU/multiple-uses-of-val.ll b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/multiple-uses-of-val.ll
new file mode 100644
index 00000000000000..717bd09897732b
--- /dev/null
+++ b/llvm/test/Transforms/InferAddressSpaces/AMDGPU/multiple-uses-of-val.ll
@@ -0,0 +1,69 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -S -passes=infer-address-spaces --verify-each %s | FileCheck %s
+
+; Inst can use a value multiple time. When we're inserting an addrspacecast to flat,
+; it's important all the identical uses use an indentical replacement, especially
+; for PHIs.
+
+define amdgpu_kernel void @test_phi() {
+; CHECK-LABEL: @test_phi(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LOADED_PTR:%.*]] = load ptr, ptr addrspace(4) null, align 8
+; CHECK-NEXT:    [[TMP0:%.*]] = addrspacecast ptr [[LOADED_PTR]] to ptr addrspace(1)
+; CHECK-NEXT:    br label [[BB0:%.*]]
+; CHECK:       bb0:
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i64, ptr addrspace(1) [[TMP0]], i64 3
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr addrspace(1) [[GEP]] to ptr
+; CHECK-NEXT:    switch i32 0, label [[END:%.*]] [
+; CHECK-NEXT:      i32 1, label [[END]]
+; CHECK-NEXT:      i32 4, label [[END]]
+; CHECK-NEXT:      i32 5, label [[BB1:%.*]]
+; CHECK-NEXT:    ]
+; CHECK:       bb1:
+; CHECK-NEXT:    [[TMP2:%.*]] = load double, ptr addrspace(1) [[GEP]], align 16
+; CHECK-NEXT:    br label [[END]]
+; CHECK:       end:
+; CHECK-NEXT:    [[RETVAL_SROA_0_0_I569_PH:%.*]] = phi ptr [ null, [[BB1]] ], [ [[TMP1]], [[BB0]] ], [ [[TMP1]], [[BB0]] ], [ [[TMP1]], [[BB0]] ]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %loaded.ptr = load ptr, ptr addrspace(4) null, align 8
+  br label %bb0
+
+bb0:
+  %gep = getelementptr i64, ptr %loaded.ptr, i64 3
+  switch i32 0, label %end [
+  i32 1, label %end
+  i32 4, label %end
+  i32 5, label %bb1
+  ]
+
+bb1:
+  %0 = load double, ptr %gep, align 16
+  br label %end
+
+end:
+  %retval.sroa.0.0.i569.ph = phi ptr [ null, %bb1 ], [ %gep, %bb0 ], [ %gep, %bb0 ], [ %gep, %bb0 ]
+  ret void
+}
+
+declare void @uses_ptrs(ptr, ptr, ptr)
+
+; We shouldn't treat PHIs 
diff erently, even other users should have the same treatment.
+; All occurences of %gep are replaced with an identical value.
+define amdgpu_kernel void @test_other() {
+; CHECK-LABEL: @test_other(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[LOADED_PTR:%.*]] = load ptr, ptr addrspace(4) null, align 8
+; CHECK-NEXT:    [[TMP0:%.*]] = addrspacecast ptr [[LOADED_PTR]] to ptr addrspace(1)
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr addrspace(1) [[TMP0]] to ptr
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i64, ptr [[TMP1]], i64 3
+; CHECK-NEXT:    call void @uses_ptrs(ptr [[GEP]], ptr [[GEP]], ptr [[GEP]])
+; CHECK-NEXT:    ret void
+;
+entry:
+  %loaded.ptr = load ptr, ptr addrspace(4) null, align 8
+  %gep = getelementptr i64, ptr %loaded.ptr, i64 3
+  call void @uses_ptrs(ptr %gep, ptr %gep, ptr %gep)
+  ret void
+}


        


More information about the llvm-commits mailing list