[X86] ABI mismatch between Clang and GCC/ICC when arguments are structures with more than one element

[phoebewang]

This is found during working on D122104. https://godbolt.org/z/5z4eG6a5P
The comments and the code below seems not correct:

llvm-project/clang/lib/CodeGen/TargetInfo.cpp

Lines 3050 to 3056 in d9cea8d

	// The only case a 256-bit wide vector could be used is when the array
	// contains a single 256-bit element. Since Lo and Hi logic isn't extended
	// to work for sizes wider than 128, early check and fallback to memory.
	//
	if (Size > 128 &&
	(Size != EltSize || Size > getNativeVectorSizeForAVXABI(AVXLevel)))
	return;

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

This looks suspicious:

typedef struct {
  __m256d r1[2];
} __sVector;

double bar(__sVector a) {
    return a.r1[0][4];
}

define dso_local double @bar(%struct.__sVector* nocapture noundef readnone byval(%struct.__sVector) align 32 %0) local_unnamed_addr #2 {
  ret double poison
}

[phoebewang]

I didn't intend to use the index 4. It should use 0~3 instead.
I'm not sure whether it's a problem of not. The 4 is out of the range of the vector. But if the array is continuous, it should be equal to a.r1[1][0]. What do you think?

[phoebewang]

Just found it's not a problem if I change it to 3. So it's a mistake of the test case.

[phoebewang]

https://godbolt.org/z/vnEh6evxf