[ms] Reintroduce feature guards in intrinsic headers in Microsoft mode

[thakis]

Summary

Visual Studio's C++ standard library headers include intrin.h, so the intrinsic headers get included a lot more often in Microsoft mode than elsewhere. The AVX512 intrinsics are a lot of code (0.7 MB, causing 30% compile time overhead for small programs including e.g. <string> and 6% compile time overhead for larger projects like e.g. v8). Since multiversioning can't be relied on in Microsoft mode (cl.exe doesn't support it), having faster compiles seems like the much better tradeoff until we have a better intrinsic story going forward (which we'll need for e.g. PR19898).

Actually using intrinsics on Windows already requires the right /arch: settings, so this patch should have no big behavior change.

See also thread "The intrinsics headers (especially avx512) are too big. What to do about it?" on cfe-dev.

[hans]

lgtm

[thakis]

r269675, thanks!

[Neumann-A]

Should this be reverted due to:

https://github.com/llvm/llvm-project/issues/53520
https://github.com/llvm/llvm-project/issues/63492

I mean: yes faster compile is nice to have but having diverging behavior from MSVC seems like the wrong way to solve it.

lib/Headers/immintrin.h

  * THE SOFTWARE.
  *
  *===-----------------------------------------------------------------------===
  */
 
 #ifndef __IMMINTRIN_H
 #define __IMMINTRIN_H
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__MMX__)
 #include <mmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__SSE__)
 #include <xmmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__SSE2__)
 #include <emmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__SSE3__)
 #include <pmmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__SSSE3__)
 #include <tmmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__SSE4_2__) || defined(__SSE4_1__))
 #include <smmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__AES__) || defined(__PCLMUL__))
 #include <wmmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX__)
 #include <avxintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX2__)
 #include <avx2intrin.h>
 
 /* The 256-bit versions of functions in f16cintrin.h.
    Intel documents these as being in immintrin.h, and
    they depend on typedefs from avxintrin.h. */
 
 #define _mm256_cvtps_ph(a, imm) __extension__ ({ \
  (__m128i)__builtin_ia32_vcvtps2ph256((__v8sf)(__m256)(a), (imm)); })
 
 static __inline __m256 __attribute__((__always_inline__, __nodebug__, __target__("f16c")))
 _mm256_cvtph_ps(__m128i __a)
 {
   return (__m256)__builtin_ia32_vcvtph2ps256((__v8hi)__a);
 }
+#endif /* __AVX2__ */
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__BMI__)
 #include <bmiintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__BMI2__)
 #include <bmi2intrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__LZCNT__)
 #include <lzcntintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__FMA__)
 #include <fmaintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512F__)
 #include <avx512fintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512VL__)
 #include <avx512vlintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512BW__)
 #include <avx512bwintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512CD__)
 #include <avx512cdintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512DQ__)
 #include <avx512dqintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__AVX512VL__) && defined(__AVX512BW__))
 #include <avx512vlbwintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__AVX512VL__) && defined(__AVX512CD__))
 #include <avx512vlcdintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__AVX512VL__) && defined(__AVX512DQ__))
 #include <avx512vldqintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512ER__)
 #include <avx512erintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512IFMA__)
 #include <avx512ifmaintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__AVX512IFMA__) && defined(__AVX512VL__))
 #include <avx512ifmavlintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512VBMI__)
 #include <avx512vbmiintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || \
+    (defined(__AVX512VBMI__) && defined(__AVX512VL__))
 #include <avx512vbmivlintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__AVX512PF__)
 #include <avx512pfintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__PKU__)
 #include <pkuintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__RDRND__)
 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
 _rdrand16_step(unsigned short *__p)
 {
   return __builtin_ia32_rdrand16_step(__p);
 }
 
 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
 _rdrand32_step(unsigned int *__p)
 {
   return __builtin_ia32_rdrand32_step(__p);
 }
 
 #ifdef __x86_64__
 static __inline__ int __attribute__((__always_inline__, __nodebug__, __target__("rdrnd")))
 _rdrand64_step(unsigned long long *__p)
 {
   return __builtin_ia32_rdrand64_step(__p);
 }
 #endif
+#endif /* __RDRND__ */
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__FSGSBASE__)
 #ifdef __x86_64__
 static __inline__ unsigned int __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
 _readfsbase_u32(void)
 {
   return __builtin_ia32_rdfsbase32();
 }
 
 static __inline__ unsigned long long __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))

 }
 
 static __inline__ void __attribute__((__always_inline__, __nodebug__, __target__("fsgsbase")))
 _writegsbase_u64(unsigned long long __V)
 {
   return __builtin_ia32_wrgsbase64(__V);
 }
 #endif
+#endif /* __FSGSBASE__ */
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__RTM__)
 #include <rtmintrin.h>
-
 #include <xtestintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__SHA__)
 #include <shaintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__FXSR__)
 #include <fxsrintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__XSAVE__)
 #include <xsaveintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__XSAVEOPT__)
 #include <xsaveoptintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__XSAVEC__)
 #include <xsavecintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__XSAVES__)
 #include <xsavesintrin.h>
+#endif
 
 /* Some intrinsics inside adxintrin.h are available only on processors with ADX,
  * whereas others are also available at all times. */
 #include <adxintrin.h>
 
 #endif /* __IMMINTRIN_H */

lib/Headers/x86intrin.h

 
 #ifndef __X86INTRIN_H
 #define __X86INTRIN_H
 
 #include <ia32intrin.h>
 
 #include <immintrin.h>
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__3dNOW__)
 #include <mm3dnow.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__BMI__)
 #include <bmiintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__BMI2__)
 #include <bmi2intrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__LZCNT__)
 #include <lzcntintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__POPCNT__)
 #include <popcntintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__RDSEED__)
 #include <rdseedintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__PRFCHW__)
 #include <prfchwintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__SSE4A__)
 #include <ammintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__FMA4__)
 #include <fma4intrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__XOP__)
 #include <xopintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__TBM__)
 #include <tbmintrin.h>
+#endif
 
+#if !defined(_MSC_VER) || __has_feature(modules) || defined(__F16C__)
 #include <f16cintrin.h>
+#endif
 
 /* FIXME: LWP */
 
 #endif /* __X86INTRIN_H */

test/CodeGen/ms-mm-align.c

-// RUN: %clang_cc1 -ffreestanding -fms-extensions -fms-compatibility -fms-compatibility-version=17.00 \
+// RUN: %clang_cc1 -ffreestanding -fms-extensions -fms-compatibility -fms-compatibility-version=17.00 -target-feature +sse \
 // RUN:         -triple i686--windows -emit-llvm %s -o - \
 // RUN:         | FileCheck %s -check-prefix CHECK
 
 // Intrin.h needs size_t, but -ffreestanding prevents us from getting it from
 // stddef.h.  Work around it with this typedef.
 typedef __SIZE_TYPE__ size_t;
 #include <Intrin.h>
 
 void capture_ptr(int* i);
 void test_mm_align16(int p) {
   _MM_ALIGN16 int i;
   capture_ptr(&i);
 }
 
 // CHECK: alloca i32, align 16