llvm-project/clang/test/CodeGen/arm-abi-vector.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

330 lines
13 KiB
C
Raw Normal View History

// RUN: %clang_cc1 -triple armv7-apple-darwin -target-abi aapcs -emit-llvm -o - %s | FileCheck %s
// RUN: %clang_cc1 -triple armv7-apple-darwin -target-abi apcs-gnu -emit-llvm -o - %s | FileCheck -check-prefix=APCS-GNU %s
// RUN: %clang_cc1 -triple arm-linux-androideabi -emit-llvm -o - %s | FileCheck -check-prefix=ANDROID %s
#include <stdarg.h>
typedef __attribute__(( ext_vector_type(2) )) int __int2;
typedef __attribute__(( ext_vector_type(3) )) char __char3;
typedef __attribute__(( ext_vector_type(5) )) char __char5;
typedef __attribute__(( ext_vector_type(9) )) char __char9;
typedef __attribute__(( ext_vector_type(19) )) char __char19;
typedef __attribute__(( ext_vector_type(3) )) short __short3;
typedef __attribute__(( ext_vector_type(5) )) short __short5;
// Passing legal vector types as varargs.
double varargs_vec_2i(int fixed, ...) {
// CHECK: varargs_vec_2i
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR:%.*]] = alloca <2 x i32>, align 8
// CHECK: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// CHECK: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 8
// CHECK: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <2 x i32>*
// CHECK: [[VEC:%.*]] = load <2 x i32>, <2 x i32>* [[AP_CAST]], align 8
// CHECK: store <2 x i32> [[VEC]], <2 x i32>* [[VAR]], align 8
// APCS-GNU: varargs_vec_2i
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[VAR:%.*]] = alloca <2 x i32>, align 8
// APCS-GNU: [[AP:%.*]] = load i8*,
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP]], i32 8
// APCS-GNU: [[AP_CAST:%.*]] = bitcast i8* [[AP]] to <2 x i32>*
// APCS-GNU: [[VEC:%.*]] = load <2 x i32>, <2 x i32>* [[AP_CAST]], align 4
// APCS-GNU: store <2 x i32> [[VEC]], <2 x i32>* [[VAR]], align 8
// ANDROID: varargs_vec_2i
// ANDROID: [[VAR:%.*]] = alloca <2 x i32>, align 8
// ANDROID: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// ANDROID: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 8
// ANDROID: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <2 x i32>*
// ANDROID: [[VEC:%.*]] = load <2 x i32>, <2 x i32>* [[AP_CAST]], align 8
// ANDROID: store <2 x i32> [[VEC]], <2 x i32>* [[VAR]], align 8
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__int2 c3 = va_arg(ap, __int2);
sum = sum + c3.x + c3.y;
va_end(ap);
return sum;
}
double test_2i(__int2 *in) {
// CHECK: test_2i
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_2i(i32 3, <2 x i32> {{%.*}})
// APCS-GNU: test_2i
// APCS-GNU: call double (i32, ...) @varargs_vec_2i(i32 3, <2 x i32> {{%.*}})
// ANDROID: test_2i
// ANDROID: call double (i32, ...) @varargs_vec_2i(i32 3, <2 x i32> {{%.*}})
return varargs_vec_2i(3, *in);
}
double varargs_vec_3c(int fixed, ...) {
// CHECK: varargs_vec_3c
// CHECK: alloca <3 x i8>, align 4
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP:%.*]], i32 4
// CHECK: bitcast i8* [[AP]] to <3 x i8>*
// APCS-GNU: varargs_vec_3c
// APCS-GNU: alloca <3 x i8>, align 4
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP:%.*]], i32 4
// APCS-GNU: bitcast i8* [[AP]] to <3 x i8>*
// ANDROID: varargs_vec_3c
// ANDROID: alloca <3 x i8>, align 4
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP:%.*]], i32 4
// ANDROID: bitcast i8* [[AP]] to <3 x i8>*
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__char3 c3 = va_arg(ap, __char3);
sum = sum + c3.x + c3.y;
va_end(ap);
return sum;
}
double test_3c(__char3 *in) {
// CHECK: test_3c
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_3c(i32 3, i32 {{%.*}})
// APCS-GNU: test_3c
// APCS-GNU: call double (i32, ...) @varargs_vec_3c(i32 3, i32 {{%.*}})
// ANDROID: test_3c
// ANDROID: call double (i32, ...) @varargs_vec_3c(i32 3, <3 x i8> {{%.*}})
return varargs_vec_3c(3, *in);
}
double varargs_vec_5c(int fixed, ...) {
// CHECK: varargs_vec_5c
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[VAR:%.*]] = alloca <5 x i8>, align 8
// CHECK: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// CHECK: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 8
// CHECK: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <5 x i8>*
// CHECK: [[VEC:%.*]] = load <5 x i8>, <5 x i8>* [[AP_CAST]], align 8
// CHECK: store <5 x i8> [[VEC]], <5 x i8>* [[VAR]], align 8
// APCS-GNU: varargs_vec_5c
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[VAR:%.*]] = alloca <5 x i8>, align 8
// APCS-GNU: [[AP:%.*]] = load i8*,
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP]], i32 8
// APCS-GNU: [[AP_CAST:%.*]] = bitcast i8* [[AP]] to <5 x i8>*
// APCS-GNU: [[VEC:%.*]] = load <5 x i8>, <5 x i8>* [[AP_CAST]], align 4
// APCS-GNU: store <5 x i8> [[VEC]], <5 x i8>* [[VAR]], align 8
// ANDROID: varargs_vec_5c
// ANDROID: [[VAR:%.*]] = alloca <5 x i8>, align 8
// ANDROID: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// ANDROID: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 8
// ANDROID: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <5 x i8>*
// ANDROID: [[VEC:%.*]] = load <5 x i8>, <5 x i8>* [[AP_CAST]], align 8
// ANDROID: store <5 x i8> [[VEC]], <5 x i8>* [[VAR]], align 8
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__char5 c5 = va_arg(ap, __char5);
sum = sum + c5.x + c5.y;
va_end(ap);
return sum;
}
double test_5c(__char5 *in) {
// CHECK: test_5c
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_5c(i32 5, <2 x i32> {{%.*}})
// APCS-GNU: test_5c
// APCS-GNU: call double (i32, ...) @varargs_vec_5c(i32 5, <2 x i32> {{%.*}})
// ANDROID: test_5c
// ANDROID: call double (i32, ...) @varargs_vec_5c(i32 5, <2 x i32> {{%.*}})
return varargs_vec_5c(5, *in);
}
double varargs_vec_9c(int fixed, ...) {
// CHECK: varargs_vec_9c
// CHECK: [[VAR:%.*]] = alloca <9 x i8>, align 16
// CHECK: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// CHECK: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 16
// CHECK: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <9 x i8>*
// CHECK: [[T0:%.*]] = load <9 x i8>, <9 x i8>* [[AP_CAST]], align 8
// CHECK: store <9 x i8> [[T0]], <9 x i8>* [[VAR]], align 16
// APCS-GNU: varargs_vec_9c
// APCS-GNU: [[VAR:%.*]] = alloca <9 x i8>, align 16
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[AP:%.*]] = load i8*,
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP]], i32 16
// APCS-GNU: [[AP_CAST:%.*]] = bitcast i8* [[AP]] to <9 x i8>*
// APCS-GNU: [[VEC:%.*]] = load <9 x i8>, <9 x i8>* [[AP_CAST]], align 4
// APCS-GNU: store <9 x i8> [[VEC]], <9 x i8>* [[VAR]], align 16
// ANDROID: varargs_vec_9c
// ANDROID: [[VAR:%.*]] = alloca <9 x i8>, align 16
// ANDROID: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// ANDROID: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 16
// ANDROID: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <9 x i8>*
// ANDROID: [[T0:%.*]] = load <9 x i8>, <9 x i8>* [[AP_CAST]], align 8
// ANDROID: store <9 x i8> [[T0]], <9 x i8>* [[VAR]], align 16
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__char9 c9 = va_arg(ap, __char9);
sum = sum + c9.x + c9.y;
va_end(ap);
return sum;
}
double test_9c(__char9 *in) {
// CHECK: test_9c
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_9c(i32 9, <4 x i32> {{%.*}})
// APCS-GNU: test_9c
// APCS-GNU: call double (i32, ...) @varargs_vec_9c(i32 9, <4 x i32> {{%.*}})
// ANDROID: test_9c
// ANDROID: call double (i32, ...) @varargs_vec_9c(i32 9, <4 x i32> {{%.*}})
return varargs_vec_9c(9, *in);
}
double varargs_vec_19c(int fixed, ...) {
// CHECK: varargs_vec_19c
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP:%.*]], i32 4
// CHECK: [[VAR:%.*]] = bitcast i8* [[AP]] to <19 x i8>**
// CHECK: [[VAR2:%.*]] = load <19 x i8>*, <19 x i8>** [[VAR]]
// APCS-GNU: varargs_vec_19c
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP:%.*]], i32 4
// APCS-GNU: [[VAR:%.*]] = bitcast i8* [[AP]] to <19 x i8>**
// APCS-GNU: [[VAR2:%.*]] = load <19 x i8>*, <19 x i8>** [[VAR]]
// ANDROID: varargs_vec_19c
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP:%.*]], i32 4
// ANDROID: [[VAR:%.*]] = bitcast i8* [[AP]] to <19 x i8>**
// ANDROID: [[VAR2:%.*]] = load <19 x i8>*, <19 x i8>** [[VAR]]
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__char19 c19 = va_arg(ap, __char19);
sum = sum + c19.x + c19.y;
va_end(ap);
return sum;
}
double test_19c(__char19 *in) {
// CHECK: test_19c
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_19c(i32 19, <19 x i8>* {{%.*}})
// APCS-GNU: test_19c
// APCS-GNU: call double (i32, ...) @varargs_vec_19c(i32 19, <19 x i8>* {{%.*}})
// ANDROID: test_19c
// ANDROID: call double (i32, ...) @varargs_vec_19c(i32 19, <19 x i8>* {{%.*}})
return varargs_vec_19c(19, *in);
}
double varargs_vec_3s(int fixed, ...) {
// CHECK: varargs_vec_3s
// CHECK: alloca <3 x i16>, align 8
// CHECK: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// CHECK: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 8
// CHECK: bitcast i8* [[AP_ALIGN]] to <3 x i16>*
// APCS-GNU: varargs_vec_3s
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[VAR:%.*]] = alloca <3 x i16>, align 8
// APCS-GNU: [[AP:%.*]] = load i8*,
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP]], i32 8
// APCS-GNU: [[AP_CAST:%.*]] = bitcast i8* [[AP]] to <3 x i16>*
// APCS-GNU: [[VEC:%.*]] = load <3 x i16>, <3 x i16>* [[AP_CAST]], align 4
// ANDROID: varargs_vec_3s
// ANDROID: alloca <3 x i16>, align 8
// ANDROID: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// ANDROID: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 8
// ANDROID: bitcast i8* [[AP_ALIGN]] to <3 x i16>*
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__short3 c3 = va_arg(ap, __short3);
sum = sum + c3.x + c3.y;
va_end(ap);
return sum;
}
double test_3s(__short3 *in) {
// CHECK: test_3s
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_3s(i32 3, <2 x i32> {{%.*}})
// APCS-GNU: test_3s
// APCS-GNU: call double (i32, ...) @varargs_vec_3s(i32 3, <2 x i32> {{%.*}})
// ANDROID: test_3s
// ANDROID: call double (i32, ...) @varargs_vec_3s(i32 3, <3 x i16> {{%.*}})
return varargs_vec_3s(3, *in);
}
double varargs_vec_5s(int fixed, ...) {
// CHECK: varargs_vec_5s
// CHECK: [[VAR_ALIGN:%.*]] = alloca <5 x i16>, align 16
// CHECK: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// CHECK: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 16
// CHECK: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <5 x i16>*
// CHECK: [[VEC:%.*]] = load <5 x i16>, <5 x i16>* [[AP_CAST]], align 8
// CHECK: store <5 x i16> [[VEC]], <5 x i16>* [[VAR_ALIGN]], align 16
// APCS-GNU: varargs_vec_5s
// APCS-GNU: [[VAR:%.*]] = alloca <5 x i16>, align 16
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[AP:%.*]] = load i8*,
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP]], i32 16
// APCS-GNU: [[AP_CAST:%.*]] = bitcast i8* [[AP]] to <5 x i16>*
// APCS-GNU: [[VEC:%.*]] = load <5 x i16>, <5 x i16>* [[AP_CAST]], align 4
// ANDROID: varargs_vec_5s
// ANDROID: [[VAR_ALIGN:%.*]] = alloca <5 x i16>, align 16
// ANDROID: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// ANDROID: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 16
// ANDROID: [[AP_CAST:%.*]] = bitcast i8* [[AP_ALIGN]] to <5 x i16>*
// ANDROID: [[VEC:%.*]] = load <5 x i16>, <5 x i16>* [[AP_CAST]], align 8
// ANDROID: store <5 x i16> [[VEC]], <5 x i16>* [[VAR_ALIGN]], align 16
va_list ap;
double sum = fixed;
va_start(ap, fixed);
__short5 c5 = va_arg(ap, __short5);
sum = sum + c5.x + c5.y;
va_end(ap);
return sum;
}
double test_5s(__short5 *in) {
// CHECK: test_5s
// CHECK: call arm_aapcscc double (i32, ...) @varargs_vec_5s(i32 5, <4 x i32> {{%.*}})
// APCS-GNU: test_5s
// APCS-GNU: call double (i32, ...) @varargs_vec_5s(i32 5, <4 x i32> {{%.*}})
// ANDROID: test_5s
// ANDROID: call double (i32, ...) @varargs_vec_5s(i32 5, <4 x i32> {{%.*}})
return varargs_vec_5s(5, *in);
}
// Pass struct as varargs.
typedef struct
{
__int2 i2;
float f;
} StructWithVec;
double varargs_struct(int fixed, ...) {
// CHECK: varargs_struct
// CHECK: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// CHECK: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// CHECK: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 16
// CHECK: bitcast i8* [[AP_ALIGN]] to %struct.StructWithVec*
// APCS-GNU: varargs_struct
// APCS-GNU: [[VAR_ALIGN:%.*]] = alloca %struct.StructWithVec
Compute and preserve alignment more faithfully in IR-generation. Introduce an Address type to bundle a pointer value with an alignment. Introduce APIs on CGBuilderTy to work with Address values. Change core APIs on CGF/CGM to traffic in Address where appropriate. Require alignments to be non-zero. Update a ton of code to compute and propagate alignment information. As part of this, I've promoted CGBuiltin's EmitPointerWithAlignment helper function to CGF and made use of it in a number of places in the expression emitter. The end result is that we should now be significantly more correct when performing operations on objects that are locally known to be under-aligned. Since alignment is not reliably tracked in the type system, there are inherent limits to this, but at least we are no longer confused by standard operations like derived-to-base conversions and array-to-pointer decay. I've also fixed a large number of bugs where we were applying the complete-object alignment to a pointer instead of the non-virtual alignment, although most of these were hidden by the very conservative approach we took with member alignment. Also, because IRGen now reliably asserts on zero alignments, we should no longer be subject to an absurd but frustrating recurring bug where an incomplete type would report a zero alignment and then we'd naively do a alignmentAtOffset on it and emit code using an alignment equal to the largest power-of-two factor of the offset. We should also now be emitting much more aggressive alignment attributes in the presence of over-alignment. In particular, field access now uses alignmentAtOffset instead of min. Several times in this patch, I had to change the existing code-generation pattern in order to more effectively use the Address APIs. For the most part, this seems to be a strict improvement, like doing pointer arithmetic with GEPs instead of ptrtoint. That said, I've tried very hard to not change semantics, but it is likely that I've failed in a few places, for which I apologize. ABIArgInfo now always carries the assumed alignment of indirect and indirect byval arguments. In order to cut down on what was already a dauntingly large patch, I changed the code to never set align attributes in the IR on non-byval indirect arguments. That is, we still generate code which assumes that indirect arguments have the given alignment, but we don't express this information to the backend except where it's semantically required (i.e. on byvals). This is likely a minor regression for those targets that did provide this information, but it'll be trivial to add it back in a later patch. I partially punted on applying this work to CGBuiltin. Please do not add more uses of the CreateDefaultAligned{Load,Store} APIs; they will be going away eventually. llvm-svn: 246985
2015-09-08 16:05:57 +08:00
// APCS-GNU: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* {{%.*}}, i32 16
// APCS-GNU: bitcast %struct.StructWithVec* [[VAR_ALIGN]] to i8*
// APCS-GNU: call void @llvm.memcpy
// ANDROID: varargs_struct
// ANDROID: [[ALIGN:%.*]] = and i32 {{%.*}}, -8
// ANDROID: [[AP_ALIGN:%.*]] = inttoptr i32 [[ALIGN]] to i8*
// ANDROID: [[AP_NEXT:%.*]] = getelementptr inbounds i8, i8* [[AP_ALIGN]], i32 16
// ANDROID: bitcast i8* [[AP_ALIGN]] to %struct.StructWithVec*
va_list ap;
double sum = fixed;
va_start(ap, fixed);
StructWithVec c3 = va_arg(ap, StructWithVec);
sum = sum + c3.i2.x + c3.i2.y + c3.f;
va_end(ap);
return sum;
}
double test_struct(StructWithVec* d) {
// CHECK: test_struct
// CHECK: call arm_aapcscc double (i32, ...) @varargs_struct(i32 3, [2 x i64] {{%.*}})
// APCS-GNU: test_struct
// APCS-GNU: call double (i32, ...) @varargs_struct(i32 3, [2 x i64] {{%.*}})
// ANDROID: test_struct
// ANDROID: call double (i32, ...) @varargs_struct(i32 3, [2 x i64] {{%.*}})
return varargs_struct(3, *d);
}