[X86] Refactor GetSSETypeAtOffset to fix pr51813

[pengfei]

Summary

D105263 adds support for _Float16 type. It introduced a bug (pr51813) that generates a <4 x half> type instead the default double when passing blank structure by SSE registers.

Although I doubt it may expose a bug somewhere other than D105263, it's good to avoid return half type when no half type in arguments.

[pengfei]

Fix lit fails.

[pengfei]

Fix typo.

[pengfei]

Change to use getFPTypeAtOffset.

[pengfei]

Fix format.

[pengfei]

Ues getTypeAllocSize.

[LuoYuanke]

clang/lib/CodeGen/TargetInfo.cpp

3416

Would you add comments on each case like previous code?

3454

Is this the major change?

[pengfei]

Add more comments.

[pengfei]

clang/lib/CodeGen/TargetInfo.cpp

3454

No, this logic is not changed. When T1 is nullptr, it means IRType is simply a half or float. So we return T0 that equals to the code here.

[LuoYuanke]

clang/lib/CodeGen/TargetInfo.cpp

3421

Not quite understanding why "+4". Would you comments on it?

[LuoYuanke]

clang/test/CodeGen/X86/avx512fp16-abi.c

153

Add a test case for "{ struct {}; half; struct {}; half;}?

[pengfei]

Address Yuanke's comments.

[LuoYuanke]

LGTM, but pls wait 1 or 2 days for the comments from others.

[Meinersbur]

This patch seem to have broken GCC-C-execute-pr44575 from the llvm-test-suite. See http://meinersbur.de:8011/#/builders/76/builds/761 (this builder compiles with Polly, but it also crashes without Polly)

[pengfei]

In D109607#3007860, @Meinersbur wrote:

This patch seem to have broken GCC-C-execute-pr44575 from the llvm-test-suite. See http://meinersbur.de:8011/#/builders/76/builds/761 (this builder compiles with Polly, but it also crashes without Polly)

Thanks @Meinersbur for reporting this. Do you have a small reproducer or the crash log? I didn't find any detail about the crash on the bot.

[pengfei]

In D109607#3007860, @Meinersbur wrote:

This patch seem to have broken GCC-C-execute-pr44575 from the llvm-test-suite. See http://meinersbur.de:8011/#/builders/76/builds/761 (this builder compiles with Polly, but it also crashes without Polly)

@Meinersbur, sorry for the late response. I just managed to reproduce the failure. I create D110037 to try to fix this problem. The test passed locally.
By the way, does this bot send notification to authors when it fails? I didn't receive this fail, so I'm not aware of it at the first time.

[Meinersbur]

In D109607#3008290, @pengfei wrote:

By the way, does this bot send notification to authors when it fails? I didn't receive this fail, so I'm not aware of it at the first time.

This is my personal staging buildbot server to park my worker until my zorg patches are greenlit. Hence, I disabled sending any email notification. However, my other one that is connected to lab.llvm.org has failed as well and should have sent an email: https://lab.llvm.org/buildbot/#/builders/102/builds/2722. Unfortunately it is slow and packing to many commits together, which I am trying to improve: D110048

Independent of that, it seems no other builder running the test-suite on x86_64 (there are armv7, s390x and ppc64le ones)

[pengfei]

In D109607#3008486, @Meinersbur wrote:

However, my other one that is connected to lab.llvm.org has failed as well and should have sent an email: https://lab.llvm.org/buildbot/#/builders/102/builds/2722. Unfortunately it is slow and packing to many commits together, which I am trying to improve: D110048

I didn't receive it either. I once suspected my mailbox but haven't had any fortune for now. :(

Independent of that, it seems no other builder running the test-suite on x86_64 (there are armv7, s390x and ppc64le ones)

Thanks for your information.

[Meinersbur]

In D109607#3009377, @pengfei wrote:

In D109607#3008486, @Meinersbur wrote:

However, my other one that is connected to lab.llvm.org has failed as well and should have sent an email: https://lab.llvm.org/buildbot/#/builders/102/builds/2722. Unfortunately it is slow and packing to many commits together, which I am trying to improve: D110048

I didn't receive it either. I once suspected my mailbox but haven't had any fortune for now. :(

Might have because that somehow the master thinks the job before the one that is breaking is still building, ie. cannot identify whether the failure is new.

[pengfei]

In D109607#3009412, @Meinersbur wrote:

In D109607#3009377, @pengfei wrote:

In D109607#3008486, @Meinersbur wrote:

However, my other one that is connected to lab.llvm.org has failed as well and should have sent an email: https://lab.llvm.org/buildbot/#/builders/102/builds/2722. Unfortunately it is slow and packing to many commits together, which I am trying to improve: D110048

I didn't receive it either. I once suspected my mailbox but haven't had any fortune for now. :(

Might have because that somehow the master thinks the job before the one that is breaking is still building, ie. cannot identify whether the failure is new.

Got it, thanks Michael.

[dyung]

Hi, our internal testing started to hit an assertion failure in one of our tests after this commit. I have put the details in PR52011, can you please take a look?

[pengfei]

In D109607#3030478, @dyung wrote:

Hi, our internal testing started to hit an assertion failure in one of our tests after this commit. I have put the details in PR52011, can you please take a look?

Sure. Thanks for reporting it.

clang/lib/CodeGen/TargetInfo.cpp

 //===---- TargetInfo.cpp - Encapsulate target details -----------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
if (const RecordType *RT = Ty->getAs<RecordType>()) {
     // If nothing in this record overlapped the area of interest, then we're
     // clean.
     return true;
   }
 
   return false;
 }
 
-/// ContainsFloatAtOffset - Return true if the specified LLVM IR type has a
-/// float member at the specified offset.  For example, {int,{float}} has a
-/// float at offset 4.  It is conservatively correct for this routine to return
-/// false.
-static bool ContainsFloatAtOffset(llvm::Type *IRType, unsigned IROffset,
+/// getFPTypeAtOffset - Return a floating point type at the specified offset.
+static llvm::Type *getFPTypeAtOffset(llvm::Type *IRType, unsigned IROffset,
                                      const llvm::DataLayout &TD) {
-  // Base case if we find a float.
-  if (IROffset == 0 && IRType->isFloatTy())
-    return true;
-
-  // If this is a struct, recurse into the field at the specified offset.
-  if (llvm::StructType *STy = dyn_cast<llvm::StructType>(IRType)) {
-    const llvm::StructLayout *SL = TD.getStructLayout(STy);
-    unsigned Elt = SL->getElementContainingOffset(IROffset);
-    IROffset -= SL->getElementOffset(Elt);
-    return ContainsFloatAtOffset(STy->getElementType(Elt), IROffset, TD);
-  }
-
-  // If this is an array, recurse into the field at the specified offset.
-  if (llvm::ArrayType *ATy = dyn_cast<llvm::ArrayType>(IRType)) {
-    llvm::Type *EltTy = ATy->getElementType();
-    unsigned EltSize = TD.getTypeAllocSize(EltTy);
-    IROffset -= IROffset/EltSize*EltSize;
-    return ContainsFloatAtOffset(EltTy, IROffset, TD);
-  }
-
-  return false;
-}
-
-/// ContainsHalfAtOffset - Return true if the specified LLVM IR type has a
-/// half member at the specified offset.  For example, {int,{half}} has a
-/// half at offset 4.  It is conservatively correct for this routine to return
-/// false.
-/// FIXME: Merge with ContainsFloatAtOffset
-static bool ContainsHalfAtOffset(llvm::Type *IRType, unsigned IROffset,
-                                 const llvm::DataLayout &TD) {
-  // Base case if we find a float.
-  if (IROffset == 0 && IRType->isHalfTy())
-    return true;
+  if (IROffset == 0 && IRType->isFloatingPointTy())
+    return IRType;
 
   // If this is a struct, recurse into the field at the specified offset.
   if (llvm::StructType *STy = dyn_cast<llvm::StructType>(IRType)) {
     const llvm::StructLayout *SL = TD.getStructLayout(STy);
     unsigned Elt = SL->getElementContainingOffset(IROffset);
     IROffset -= SL->getElementOffset(Elt);
-    return ContainsHalfAtOffset(STy->getElementType(Elt), IROffset, TD);
+    return getFPTypeAtOffset(STy->getElementType(Elt), IROffset, TD);
   }
 
   // If this is an array, recurse into the field at the specified offset.
   if (llvm::ArrayType *ATy = dyn_cast<llvm::ArrayType>(IRType)) {
     llvm::Type *EltTy = ATy->getElementType();
     unsigned EltSize = TD.getTypeAllocSize(EltTy);
     IROffset -= IROffset / EltSize * EltSize;
-    return ContainsHalfAtOffset(EltTy, IROffset, TD);
+    return getFPTypeAtOffset(EltTy, IROffset, TD);
   }
 
-  return false;
+  return nullptr;
 }
 
 /// GetSSETypeAtOffset - Return a type that will be passed by the backend in the
 /// low 8 bytes of an XMM register, corresponding to the SSE class.
 llvm::Type *X86_64ABIInfo::
 GetSSETypeAtOffset(llvm::Type *IRType, unsigned IROffset,
                    QualType SourceTy, unsigned SourceOffset) const {
-  // If the high 32 bits are not used, we have three choices. Single half,
-  // single float or two halfs.
-  if (BitsContainNoUserData(SourceTy, SourceOffset * 8 + 32,
-                            SourceOffset * 8 + 64, getContext())) {
-    if (ContainsFloatAtOffset(IRType, IROffset, getDataLayout()))
-      return llvm::Type::getFloatTy(getVMContext());
-    if (ContainsHalfAtOffset(IRType, IROffset + 2, getDataLayout()))
-      return llvm::FixedVectorType::get(llvm::Type::getHalfTy(getVMContext()),
-                                        2);
-
-    return llvm::Type::getHalfTy(getVMContext());
-  }
-
-  // We want to pass as <2 x float> if the LLVM IR type contains a float at
-  // offset+0 and offset+4. Walk the LLVM IR type to find out if this is the
-  // case.
-  if (ContainsFloatAtOffset(IRType, IROffset, getDataLayout()) &&
-      ContainsFloatAtOffset(IRType, IROffset + 4, getDataLayout()))
-    return llvm::FixedVectorType::get(llvm::Type::getFloatTy(getVMContext()),
-                                      2);
-
-  // We want to pass as <4 x half> if the LLVM IR type contains a half at
-  // offset+0, +2, +4. Walk the LLVM IR type to find out if this is the case.
-  if (ContainsHalfAtOffset(IRType, IROffset, getDataLayout()) &&
-      ContainsHalfAtOffset(IRType, IROffset + 2, getDataLayout()) &&
-      ContainsHalfAtOffset(IRType, IROffset + 4, getDataLayout()))
-    return llvm::FixedVectorType::get(llvm::Type::getHalfTy(getVMContext()), 4);
+  const llvm::DataLayout &TD = getDataLayout();
+  llvm::Type *T0 = getFPTypeAtOffset(IRType, IROffset, TD);
+  if (!T0 || T0->isDoubleTy())
+    return llvm::Type::getDoubleTy(getVMContext());
+
+  // Get the adjacent FP type.
+  llvm::Type *T1 =
+      getFPTypeAtOffset(IRType, IROffset + TD.getTypeAllocSize(T0), TD);
+  if (T1 == nullptr) {
+    // Check if IRType is a half + float. float type will be in IROffset+4 due
+    // to its alignment.
+    if (T0->isHalfTy())
+      T1 = getFPTypeAtOffset(IRType, IROffset + 4, TD);
+    // If we can't get a second FP type, return a simple half or float.
+    // avx512fp16-abi.c:pr51813_2 shows it works to return float for
+    // {float, i8} too.
+    if (T1 == nullptr)
+      return T0;
+  }
+
+  if (T0->isFloatTy() && T1->isFloatTy())
+    return llvm::FixedVectorType::get(T0, 2);
+
+  if (T0->isHalfTy() && T1->isHalfTy()) {
+    llvm::Type *T2 = getFPTypeAtOffset(IRType, IROffset + 4, TD);
+    if (T2 == nullptr)
+      return llvm::FixedVectorType::get(T0, 2);
+    return llvm::FixedVectorType::get(T0, 4);
+  }
 
-  // We want to pass as <4 x half> if the LLVM IR type contains a mix of float
-  // and half.
-  // FIXME: Do we have a better representation for the mixed type?
-  if (ContainsFloatAtOffset(IRType, IROffset, getDataLayout()) ||
-      ContainsFloatAtOffset(IRType, IROffset + 4, getDataLayout()))
+  if (T0->isHalfTy() || T1->isHalfTy())
     return llvm::FixedVectorType::get(llvm::Type::getHalfTy(getVMContext()), 4);
 
   return llvm::Type::getDoubleTy(getVMContext());
 }
 
 
 /// GetINTEGERTypeAtOffset - The ABI specifies that a value should be passed in
 /// an 8-byte GPR.  This means that we either have a scalar or we are talking

clang/test/CodeGen/X86/avx512fp16-abi.c

-// RUN: %clang_cc1 -triple x86_64-linux -emit-llvm  -target-feature +avx512fp16 < %s | FileCheck %s --check-prefixes=CHECK
+// RUN: %clang_cc1 -triple x86_64-linux -emit-llvm  -target-feature +avx512fp16 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-C
+// RUN: %clang_cc1 -triple x86_64-linux -emit-llvm  -target-feature +avx512fp16 -x c++ -std=c++11 < %s | FileCheck %s --check-prefixes=CHECK,CHECK-CPP
 
 struct half1 {
   _Float16 a;
 };
 
 struct half1 h1(_Float16 a) {
-  // CHECK: define{{.*}}half @h1
+  // CHECK: define{{.*}}half @
   struct half1 x;
   x.a = a;
   return x;
 }
 
 struct half2 {
   _Float16 a;
   _Float16 b;
 };
 
 struct half2 h2(_Float16 a, _Float16 b) {
-  // CHECK: define{{.*}}<2 x half> @h2
+  // CHECK: define{{.*}}<2 x half> @
   struct half2 x;
   x.a = a;
   x.b = b;
   return x;
 }
 
 struct half3 {
   _Float16 a;
   _Float16 b;
   _Float16 c;
 };
 
 struct half3 h3(_Float16 a, _Float16 b, _Float16 c) {
-  // CHECK: define{{.*}}<4 x half> @h3
+  // CHECK: define{{.*}}<4 x half> @
   struct half3 x;
   x.a = a;
   x.b = b;
   x.c = c;
   return x;
 }
 
 struct half4 {
   _Float16 a;
   _Float16 b;
   _Float16 c;
   _Float16 d;
 };
 
 struct half4 h4(_Float16 a, _Float16 b, _Float16 c, _Float16 d) {
-  // CHECK: define{{.*}}<4 x half> @h4
+  // CHECK: define{{.*}}<4 x half> @
   struct half4 x;
   x.a = a;
   x.b = b;
   x.c = c;
   x.d = d;
   return x;
 }
 
 struct floathalf {
   float a;
   _Float16 b;
 };
 
 struct floathalf fh(float a, _Float16 b) {
-  // CHECK: define{{.*}}<4 x half> @fh
+  // CHECK: define{{.*}}<4 x half> @
   struct floathalf x;
   x.a = a;
   x.b = b;
   return x;
 }
 
 struct floathalf2 {
   float a;
   _Float16 b;
   _Float16 c;
 };
 
 struct floathalf2 fh2(float a, _Float16 b, _Float16 c) {
-  // CHECK: define{{.*}}<4 x half> @fh2
+  // CHECK: define{{.*}}<4 x half> @
   struct floathalf2 x;
   x.a = a;
   x.b = b;
   x.c = c;
   return x;
 }
 
 struct halffloat {
   _Float16 a;
   float b;
 };
 
 struct halffloat hf(_Float16 a, float b) {
-  // CHECK: define{{.*}}<4 x half> @hf
+  // CHECK: define{{.*}}<4 x half> @
   struct halffloat x;
   x.a = a;
   x.b = b;
   return x;
 }
 
 struct half2float {
   _Float16 a;
   _Float16 b;
   float c;
 };
 
 struct half2float h2f(_Float16 a, _Float16 b, float c) {
-  // CHECK: define{{.*}}<4 x half> @h2f
+  // CHECK: define{{.*}}<4 x half> @
   struct half2float x;
   x.a = a;
   x.b = b;
   x.c = c;
   return x;
 }
 
 struct floathalf3 {
   float a;
   _Float16 b;
   _Float16 c;
   _Float16 d;
 };
 
 struct floathalf3 fh3(float a, _Float16 b, _Float16 c, _Float16 d) {
-  // CHECK: define{{.*}}{ <4 x half>, half } @fh3
+  // CHECK: define{{.*}}{ <4 x half>, half } @
   struct floathalf3 x;
   x.a = a;
   x.b = b;
   x.c = c;
   x.d = d;
   return x;
 }
 
 struct half5 {
   _Float16 a;
   _Float16 b;
   _Float16 c;
   _Float16 d;
   _Float16 e;
 };
 
 struct half5 h5(_Float16 a, _Float16 b, _Float16 c, _Float16 d, _Float16 e) {
-  // CHECK: define{{.*}}{ <4 x half>, half } @h5
+  // CHECK: define{{.*}}{ <4 x half>, half } @
   struct half5 x;
   x.a = a;
   x.b = b;
   x.c = c;
   x.d = d;
   x.e = e;
   return x;
 }
+
+struct float2 {
+  struct {} s;
+  float a;
+  float b;
+};
+
+float pr51813(struct float2 s) {
+  // CHECK-C: define{{.*}} @pr51813(<2 x float>
+  // CHECK-CPP: define{{.*}} @_Z7pr518136float2(double {{.*}}, float
+  return s.a;
+}
+
+struct float3 {
+  float a;
+  struct {} s;
+  float b;
+};
+
+float pr51813_2(struct float3 s) {
+  // CHECK-C: define{{.*}} @pr51813_2(<2 x float>
+  // CHECK-CPP: define{{.*}} @_Z9pr51813_26float3(double {{.*}}, float
+  return s.a;
+}
+
+struct shalf2 {
+  struct {} s;
+  _Float16 a;
+  _Float16 b;
+};
+
+_Float16 sf2(struct shalf2 s) {
+  // CHECK-C: define{{.*}} @sf2(<2 x half>
+  // CHECK-CPP: define{{.*}} @_Z3sf26shalf2(double {{.*}}
+  return s.a;
+};
+
+struct halfs2 {
+  _Float16 a;
+  struct {} s1;
+  _Float16 b;
+  struct {} s2;
+};
+
+_Float16 fs2(struct shalf2 s) {
+  // CHECK-C: define{{.*}} @fs2(<2 x half>
+  // CHECK-CPP: define{{.*}} @_Z3fs26shalf2(double {{.*}}
+  return s.a;
+};