[X86] Assume alignment of movdir64b dst argument

[craig.topper]

What effect does this have?

[GBuella]

In D46683#1094662, @craig.topper wrote:

What effect does this have?

Nothing important really, I just guessed it doesn't cost us much.
One contrived example I could come up with in 2 minutes:

#include <x86intrin.h>

void x(char *restrict a __attribute__((align_value(64))), char *restrict b, const char *c)
{
        _movdir64b(b, c);

        for (int i = 0; i < 32; ++i)
                a[i] = b[i];
}

The silly memcpy loop above becomes 2 movaps instructions, instead of two movups instructions.

[craig.topper]

LGTM

cfe/trunk/lib/Headers/movdirintrin.h

 __attribute__((__always_inline__, __nodebug__,  __target__("movdiri")))
 _directstoreu_u64 (void *__dst, unsigned long __value)
 {
   __builtin_ia32_directstore_u64((unsigned long *)__dst, __value);
 }
 
 #endif /* __x86_64__ */
 
-// Move 64 bytes as direct store
+/*
+ * movdir64b - Move 64 bytes as direct store.
+ * The destination must be 64 byte aligned, and the store is atomic.
+ * The source address has no alignment requirement, and the load from
+ * the source address is not atomic.
+ */
 static __inline__ void
 __attribute__((__always_inline__, __nodebug__,  __target__("movdir64b")))
-_movdir64b (void *__dst, const void *__src)
+_movdir64b (void *__dst __attribute__((align_value(64))), const void *__src)
 {
   __builtin_ia32_movdir64b(__dst, __src);
 }
 
 #endif /* _MOVDIRINTRIN_H */

cfe/trunk/test/CodeGen/builtin-movdir.c

 // RUN: %clang_cc1 -ffreestanding -Wall -pedantic -triple x86_64-unknown-unknown -target-feature +movdiri -target-feature +movdir64b %s -emit-llvm -o - | FileCheck %s --check-prefix=X86_64 --check-prefix=CHECK
-// RUN: %clang_cc1 -ffreestanding -Wall -pedantic -triple i386-unknown-unknown -target-feature +movdiri -target-feature +movdir64b %s -emit-llvm -o - | FileCheck %s --check-prefix=CHECK
+// RUN: %clang_cc1 -ffreestanding -Wall -pedantic -triple i386-unknown-unknown -target-feature +movdiri -target-feature +movdir64b %s -emit-llvm -o - | FileCheck %s --check-prefix=X86 --check-prefix=CHECK
 
 #include <x86intrin.h>
 #include <stdint.h>
 
 void test_directstore32(void *dst, uint32_t value) {
   // CHECK-LABEL: test_directstore32
   // CHECK: call void @llvm.x86.directstore32
   _directstoreu_u32(dst, value);
 }
 
 #ifdef __x86_64__
 
 void test_directstore64(void *dst, uint64_t value) {
   // X86_64-LABEL: test_directstore64
   // X86_64: call void @llvm.x86.directstore64
   _directstoreu_u64(dst, value);
 }
 
 #endif
 
 void test_dir64b(void *dst, const void *src) {
   // CHECK-LABEL: test_dir64b
+  // CHECK: [[PTRINT1:%.+]] = ptrtoint
+  // X86: [[MASKEDPTR1:%.+]] = and i32 [[PTRINT1]], 63
+  // X86: [[MASKCOND1:%.+]] = icmp eq i32 [[MASKEDPTR1]], 0
+  // X86_64: [[MASKEDPTR1:%.+]] = and i64 [[PTRINT1]], 63
+  // X86_64: [[MASKCOND1:%.+]] = icmp eq i64 [[MASKEDPTR1]], 0
   // CHECK: call void @llvm.x86.movdir64b
   _movdir64b(dst, src);
 }
 
 // CHECK: declare void @llvm.x86.directstore32(i8*, i32)
 // X86_64: declare void @llvm.x86.directstore64(i8*, i64)
 // CHECK: declare void @llvm.x86.movdir64b(i8*, i8*)