[llvm] e6d9542 - [X86][Inline] Check correct function for target feature check (#152515)

[via llvm-commits]

Author: Nikita Popov
Date: 2025-08-19T09:44:00+02:00
New Revision: e6d9542b7768b8b5c1ddffcc5b79879bfee3cdc5

URL: https://github.com/llvm/llvm-project/commit/e6d9542b7768b8b5c1ddffcc5b79879bfee3cdc5
DIFF: https://github.com/llvm/llvm-project/commit/e6d9542b7768b8b5c1ddffcc5b79879bfee3cdc5.diff

LOG: [X86][Inline] Check correct function for target feature check (#152515)

The check for ABI differences for inlined calls involves the caller, the
callee and the nested callee. Before inlining, the ABI is determined by
the target features of the callee. After inlining it is determined by
the caller. The features of the nested callee should never actually
matter.

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86TargetTransformInfo.cpp
    llvm/test/Transforms/Inline/X86/call-abi-compatibility.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
index cae6bb99d9637..0c63bf1893126 100644
--- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -6526,8 +6526,8 @@ bool X86TTIImpl::areInlineCompatible(const Function *Caller,
 
   for (const Instruction &I : instructions(Callee)) {
     if (const auto *CB = dyn_cast<CallBase>(&I)) {
-      // Having more target features is fine for inline ASM.
-      if (CB->isInlineAsm())
+      // Having more target features is fine for inline ASM and intrinsics.
+      if (CB->isInlineAsm() || CB->getIntrinsicID() != Intrinsic::not_intrinsic)
         continue;
 
       SmallVector<Type *, 8> Types;
@@ -6543,19 +6543,9 @@ bool X86TTIImpl::areInlineCompatible(const Function *Caller,
       if (all_of(Types, IsSimpleTy))
         continue;
 
-      if (Function *NestedCallee = CB->getCalledFunction()) {
-        // Assume that intrinsics are always ABI compatible.
-        if (NestedCallee->isIntrinsic())
-          continue;
-
-        // Do a precise compatibility check.
-        if (!areTypesABICompatible(Caller, NestedCallee, Types))
-          return false;
-      } else {
-        // We don't know the target features of the callee,
-        // assume it is incompatible.
+      // Do a precise compatibility check.
+      if (!areTypesABICompatible(Caller, Callee, Types))
         return false;
-      }
     }
   }
   return true;

diff  --git a/llvm/test/Transforms/Inline/X86/call-abi-compatibility.ll b/llvm/test/Transforms/Inline/X86/call-abi-compatibility.ll
index 6f582cab2f145..9d0981cd63756 100644
--- a/llvm/test/Transforms/Inline/X86/call-abi-compatibility.ll
+++ b/llvm/test/Transforms/Inline/X86/call-abi-compatibility.ll
@@ -34,8 +34,7 @@ define i64 @callee_not_avx(<4 x i64> %arg) noinline {
   ret i64 %v
 }
 
-; This call also shouldn't be inlined, as we don't know whether callee_unknown
-; is ABI compatible or not.
+; This call also shouldn't be inlined, as caller_not_avx2 is not ABI compatible.
 define void @caller_avx2() "target-features"="+avx" {
 ; CHECK-LABEL: define {{[^@]+}}@caller_avx2
 ; CHECK-SAME: () #[[ATTR0]] {
@@ -55,6 +54,29 @@ define internal void @caller_not_avx2() {
   ret void
 }
 
+; Should be inlined, as caller_avx7 is ABI compatible. The fact that we don't
+; know anything about callee_unknown doesn't matter, as it is the caller that
+; determines the ABI as far as target features are concerned.
+define void @caller_avx6() "target-features"="+avx" {
+; CHECK-LABEL: define {{[^@]+}}@caller_avx6
+; CHECK-SAME: () #[[ATTR0]] {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @callee_unknown(<4 x i64> <i64 0, i64 1, i64 2, i64 3>)
+; CHECK-NEXT:    ret void
+;
+  call void @caller_avx7()
+  ret void
+}
+
+define void @caller_avx7() "target-features"="+avx" {
+; CHECK-LABEL: define {{[^@]+}}@caller_avx7
+; CHECK-SAME: () #[[ATTR0]] {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @callee_unknown(<4 x i64> <i64 0, i64 1, i64 2, i64 3>)
+; CHECK-NEXT:    ret void
+;
+  call i64 @callee_unknown(<4 x i64> <i64 0, i64 1, i64 2, i64 3>)
+  ret void
+}
+
 declare i64 @callee_unknown(<4 x i64>)
 
 ; This call should get inlined, because we assume that intrinsics are always
@@ -62,20 +84,24 @@ declare i64 @callee_unknown(<4 x i64>)
 define void @caller_avx3() "target-features"="+avx" {
 ; CHECK-LABEL: define {{[^@]+}}@caller_avx3
 ; CHECK-SAME: () #[[ATTR0]] {
-; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.some_intrinsic(<4 x i64> <i64 0, i64 1, i64 2, i64 3>)
+; CHECK-NEXT:    [[V_I:%.*]] = load <4 x i64>, ptr @g, align 32
+; CHECK-NEXT:    [[V2_I:%.*]] = call <4 x i64> @llvm.abs.v4i64(<4 x i64> [[V_I]], i1 false)
+; CHECK-NEXT:    store <4 x i64> [[V2_I]], ptr @g, align 32
 ; CHECK-NEXT:    ret void
 ;
   call void @caller_not_avx3()
   ret void
 }
 
+ at g = external global <4 x i64>
+
 define internal void @caller_not_avx3() {
-  call i64 @llvm.some_intrinsic(<4 x i64> <i64 0, i64 1, i64 2, i64 3>)
+  %v = load <4 x i64>, ptr @g
+  %v2 = call <4 x i64> @llvm.abs(<4 x i64> %v, i1 false)
+  store <4 x i64> %v2, ptr @g
   ret void
 }
 
-declare i64 @llvm.some_intrinsic(<4 x i64>)
-
 ; This call should get inlined, because only simple types are involved.
 define void @caller_avx4() "target-features"="+avx" {
 ; CHECK-LABEL: define {{[^@]+}}@caller_avx4