[PPCMergeStringPool] Avoid replacing constant with instruction

String pool merging currently, for a reason that's not entirely
clear to me, tries to create GEP instructions instead of GEP
constant expressions when replacing constant references. It only
uses constant expressions in cases where this is required. However,
it does not catch all cases where such a requirement exists. For
example, the landingpad catch clause has to be a constant.

Fix this by always using the constant expression variant, which
also makes the implementation simpler.

Additionally, there are some edge cases where even replacement
with a constant GEP is not legal. The one I am aware of is the
llvm.eh.typeid.for intrinsic, so add a special case to forbid
replacements for it.

Fixes #88844.

Author: nikic

llvm/lib/Target/PowerPC/PPCMergeStringPool.cpp

@@ -23,6 +23,7 @@
 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/ValueSymbolTable.h"
 #include "llvm/Pass.h"
@@ -117,9 +118,16 @@ class PPCMergeStringPool : public ModulePass {
 // sure that they can be replaced.
 static bool hasReplaceableUsers(GlobalVariable &GV) {
   for (User *CurrentUser : GV.users()) {
-    // Instruction users are always valid.
-    if (isa<Instruction>(CurrentUser))
+    if (isa<Instruction>(CurrentUser)) {
+      if (auto *II = dyn_cast<IntrinsicInst>(CurrentUser)) {
+        // Some intrinsics require a plain global.
+        if (II->getIntrinsicID() == Intrinsic::eh_typeid_for)
+          return false;
+      }
+
+      // Other instruction users are always valid.
       continue;
+    }
 
     // We cannot replace GlobalValue users because they are not just nodes
     // in IR. To replace a user like this we would need to create a new
@@ -330,38 +338,13 @@ void PPCMergeStringPool::replaceUsesWithGEP(GlobalVariable *GlobalToReplace,
     if (isa<GlobalValue>(CurrentUser))
       continue;
 
-    if (!UserInstruction) {
-      // User is a constant type.
-      Constant *ConstGEP = ConstantExpr::getInBoundsGetElementPtr(
-          PooledStructType, GPool, Indices);
-      UserConstant->handleOperandChange(GlobalToReplace, ConstGEP);
-      continue;
-    }
-
-    if (PHINode *UserPHI = dyn_cast<PHINode>(UserInstruction)) {
-      // GEP instructions cannot be added before PHI nodes.
-      // With getInBoundsGetElementPtr we create the GEP and then replace it
-      // inline into the PHI.
-      Constant *ConstGEP = ConstantExpr::getInBoundsGetElementPtr(
-          PooledStructType, GPool, Indices);
-      UserPHI->replaceUsesOfWith(GlobalToReplace, ConstGEP);
-      continue;
-    }
-    // The user is a valid instruction that is not a PHINode.
-    GetElementPtrInst *GEPInst =
-        GetElementPtrInst::Create(PooledStructType, GPool, Indices);
-    GEPInst->insertBefore(UserInstruction);
-
-    LLVM_DEBUG(dbgs() << "Inserting GEP before:\n");
-    LLVM_DEBUG(UserInstruction->dump());
-
+    Constant *ConstGEP = ConstantExpr::getInBoundsGetElementPtr(
+        PooledStructType, GPool, Indices);
     LLVM_DEBUG(dbgs() << "Replacing this global:\n");
     LLVM_DEBUG(GlobalToReplace->dump());
     LLVM_DEBUG(dbgs() << "with this:\n");
-    LLVM_DEBUG(GEPInst->dump());
-
-    // After the GEP is inserted the GV can be replaced.
-    CurrentUser->replaceUsesOfWith(GlobalToReplace, GEPInst);
+    LLVM_DEBUG(ConstGEP->dump());
+    GlobalToReplace->replaceAllUsesWith(ConstGEP);
   }
 }
 

llvm/test/CodeGen/PowerPC/merge-string-used-by-metadata.mir

@@ -14,16 +14,14 @@
 
   define noundef ptr @func1(ptr noundef nonnull align 8 dereferenceable(8) %this) #0 !dbg !6 {
   ; CHECK-LABEL: func1
-  ; CHECK:       %0 = getelementptr { [7 x i8], [7 x i8] }, ptr @__ModuleStringPool, i32 0, i32 1
-  ; CHECK-NEXT:  ret ptr %0, !dbg !14
+  ; CHECK:       ret ptr getelementptr inbounds ({ [7 x i8], [7 x i8] }, ptr @__ModuleStringPool, i32 0, i32 1), !dbg !14
   entry:
     ret ptr @const.2, !dbg !14
   }
 
   define noundef ptr @func2(ptr noundef nonnull align 8 dereferenceable(8) %this) #0 {
   ; CHECK-LABEL: func2
-  ; CHECK:       %0 = getelementptr { [7 x i8], [7 x i8] }, ptr @__ModuleStringPool, i32 0, i32 0
-  ; CHECK-NEXT:  ret ptr %0
+  ; CHECK:       ret ptr @__ModuleStringPool
   entry:
     ret ptr @const.1
   }

llvm/test/CodeGen/PowerPC/mergeable-string-pool-exceptions.ll

@@ -0,0 +1,47 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=ppc64le-unknown-linux-gnu < %s | FileCheck %s
+
+@id = private unnamed_addr constant [4 x i8] c"@id\00", align 1
+@id2 = private unnamed_addr constant [5 x i8] c"@id2\00", align 1
+
+; Higher-aligned dummy to make sure it is first in the string pool.
+@dummy = private unnamed_addr constant [1 x i32] [i32 42], align 4
+
+define ptr @test1() personality ptr @__gnu_objc_personality_v0 {
+; CHECK-LABEL: test1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    mflr 0
+; CHECK-NEXT:    stdu 1, -32(1)
+; CHECK-NEXT:    std 0, 48(1)
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset lr, 16
+; CHECK-NEXT:    addis 3, 2, .L__ModuleStringPool@toc@ha
+; CHECK-NEXT:    addi 3, 3, .L__ModuleStringPool@toc@l
+; CHECK-NEXT:    bl foo
+; CHECK-NEXT:    nop
+  invoke void @foo(ptr @dummy)
+          to label %cont unwind label %unwind
+
+cont:
+  unreachable
+
+unwind:
+  %lp = landingpad { ptr, i32 }
+          catch ptr @id
+  resume { ptr, i32 } %lp
+}
+
+define i32 @test2() personality ptr @__gnu_objc_personality_v0 {
+; CHECK-LABEL: test2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    li 3, 1
+; CHECK-NEXT:    blr
+  %id = tail call i32 @llvm.eh.typeid.for(ptr @id2)
+  ret i32 %id
+}
+
+declare i32 @__gnu_objc_personality_v0(...)
+
+declare i32 @llvm.eh.typeid.for(ptr)
+
+declare void @foo()

llvm/test/CodeGen/PowerPC/mergeable-string-pool-pass-only.mir

@@ -35,8 +35,7 @@
     ret i32 %call
 
   ; CHECK-LABEL: test1
-  ; CHECK:         %0 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 6
-  ; CHECK:         tail call signext i32 @calleeStr
+  ; CHECK:         %call = tail call signext i32 @calleeStr(ptr noundef nonnull getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 6))
   }
 
   define dso_local signext i32 @test2() local_unnamed_addr #0 {
@@ -49,7 +48,7 @@
     ret i32 %call
 
   ; CHECK-LABEL: test2
-  ; CHECK:         %0 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 2
+  ; CHECK:         call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 4 dereferenceable(24) %A, ptr noundef nonnull align 4 dereferenceable(24) getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 2), i64 24, i1 false)
   ; CHECK:         call signext i32 @calleeInt
   }
 
@@ -62,7 +61,7 @@
     call void @llvm.lifetime.end.p0(i64 28, ptr nonnull %A) #0
     ret i32 %call
   ; CHECK-LABEL: test3
-  ; CHECK:         %0 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 4
+  ; CHECK:         call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 4 dereferenceable(28) %A, ptr noundef nonnull align 4 dereferenceable(28) getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 4), i64 28, i1 false)
   ; CHECK:         call signext i32 @calleeFloat
   }
 
@@ -75,7 +74,7 @@
     call void @llvm.lifetime.end.p0(i64 56, ptr nonnull %A) #0
     ret i32 %call
   ; CHECK-LABEL: test4
-  ; CHECK:         %0 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 0
+  ; CHECK:         call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 8 dereferenceable(56) %A, ptr noundef nonnull align 8 dereferenceable(56) @__ModuleStringPool, i64 56, i1 false)
   ; CHECK:         call signext i32 @calleeDouble
   }
 
@@ -102,11 +101,10 @@
     call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %B) #0
     ret i32 %add7
   ; CHECK-LABEL: test5
-  ; CHECK:         %0 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 3
-  ; CHECK:         %1 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 5
-  ; CHECK:         %2 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 1
-  ; CHECK:         %3 = getelementptr { [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 7
-  ; CHECK:         call signext i32 @calleeStr
+  ; CHECK:         call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 4 dereferenceable(24) %B, ptr noundef nonnull align 4 dereferenceable(24) getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 3), i64 24, i1 false)
+  ; CHECK:         call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 4 dereferenceable(28) %C, ptr noundef nonnull align 4 dereferenceable(28) getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 5), i64 28, i1 false)
+  ; CHECK:         call void @llvm.memcpy.p0.p0.i64(ptr noundef nonnull align 8 dereferenceable(56) %D, ptr noundef nonnull align 8 dereferenceable(56) getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 1), i64 56, i1 false)
+  ; CHECK:         call signext i32 @calleeStr(ptr noundef nonnull getelementptr inbounds ({ [7 x double], [7 x double], [6 x i32], [6 x i32], [7 x float], [7 x float], [8 x i8], [16 x i8] }, ptr @__ModuleStringPool, i32 0, i32 7))
   ; CHECK:         call signext i32 @calleeInt
   ; CHECK:         call signext i32 @calleeFloat
   ; CHECK:         call signext i32 @calleeDouble