[Headers][X86] Add rounding and exception notes to conversions

[pogo59]

Consistently describe rounding/truncating on convert intrinsics. Add notes where an out-of-range result can raise an exception.

[pogo59]

FYI @cflores

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-clang

Author: Paul T Robinson (pogo59)

Changes

Consistently describe rounding/truncating on convert intrinsics. Add notes where an out-of-range result can raise an exception.

---

Patch is 23.00 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/83447.diff

3 Files Affected:

• (modified) clang/lib/Headers/avxintrin.h (+27-9)
• (modified) clang/lib/Headers/emmintrin.h (+66-25)
• (modified) clang/lib/Headers/xmmintrin.h (+89-36)

diff --git a/clang/lib/Headers/avxintrin.h b/clang/lib/Headers/avxintrin.h
index f116d8bc3a94c7..51c7d76e75ca1a 100644
--- a/clang/lib/Headers/avxintrin.h
+++ b/clang/lib/Headers/avxintrin.h
@@ -2180,7 +2180,8 @@ _mm256_cvtepi32_pd(__m128i __a)
   return (__m256d)__builtin_convertvector((__v4si)__a, __v4df);
 }
 
-/// Converts a vector of [8 x i32] into a vector of [8 x float].
+/// Converts a vector of [8 x i32] into a vector of [8 x float]. Rounds inexact
+///    results according to the rounding control bits in the MXCSR register.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -2196,7 +2197,8 @@ _mm256_cvtepi32_ps(__m256i __a)
 }
 
 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of
-///    [4 x float].
+///    [4 x float]. Rounds inexact results according to the rounding control
+///    bits in the MXCSR register.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -2211,7 +2213,12 @@ _mm256_cvtpd_ps(__m256d __a)
   return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) __a);
 }
 
-/// Converts a vector of [8 x float] into a vector of [8 x i32].
+/// Converts a vector of [8 x float] into a vector of [8 x i32]. Rounds inexact
+///    results according to the rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -2243,8 +2250,11 @@ _mm256_cvtps_pd(__m128 __a)
 }
 
 /// Converts a 256-bit vector of [4 x double] into a 128-bit vector of [4
-///    x i32], truncating the result by rounding towards zero when it is
-///    inexact.
+///    x i32], truncating inexact results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -2259,9 +2269,13 @@ _mm256_cvttpd_epi32(__m256d __a)
   return (__m128i)__builtin_ia32_cvttpd2dq256((__v4df) __a);
 }
 
-/// Converts a 256-bit vector of [4 x double] into a 128-bit vector of [4
-///    x i32]. When a conversion is inexact, the value returned is rounded
-///    according to the rounding control bits in the MXCSR register.
+/// Converts a 256-bit vector of [4 x double] into a 128-bit vector of
+///    [4 x i32]. Rounds inexact results according to the rounding control bits
+///    in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -2277,7 +2291,11 @@ _mm256_cvtpd_epi32(__m256d __a)
 }
 
 /// Converts a vector of [8 x float] into a vector of [8 x i32],
-///    truncating the result by rounding towards zero when it is inexact.
+///    truncating inexact results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
diff --git a/clang/lib/Headers/emmintrin.h b/clang/lib/Headers/emmintrin.h
index 96e3ebdecbdf83..9ae7fe2c809ac6 100644
--- a/clang/lib/Headers/emmintrin.h
+++ b/clang/lib/Headers/emmintrin.h
@@ -1245,7 +1245,8 @@ static __inline__ int __DEFAULT_FN_ATTRS _mm_ucomineq_sd(__m128d __a,
 /// Converts the two double-precision floating-point elements of a
 ///    128-bit vector of [2 x double] into two single-precision floating-point
 ///    values, returned in the lower 64 bits of a 128-bit vector of [4 x float].
-///    The upper 64 bits of the result vector are set to zero.
+///    Rounds inexact results according to the rounding control bits in the
+///    MXCSR register. The upper 64 bits of the result vector are set to zero.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1300,9 +1301,13 @@ static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_cvtepi32_pd(__m128i __a) {
 }
 
 /// Converts the two double-precision floating-point elements of a
-///    128-bit vector of [2 x double] into two signed 32-bit integer values,
-///    returned in the lower 64 bits of a 128-bit vector of [4 x i32]. The upper
-///    64 bits of the result vector are set to zero.
+///    128-bit vector of [2 x double] into two signed 32-bit integer values.
+///    Rounds inexact results according to the rounding control bits in the
+///    MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1317,7 +1322,12 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cvtpd_epi32(__m128d __a) {
 }
 
 /// Converts the low-order element of a 128-bit vector of [2 x double]
-///    into a 32-bit signed integer value.
+///    into a 32-bit signed integer value. Rounds inexact results according to
+///    the rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1333,9 +1343,10 @@ static __inline__ int __DEFAULT_FN_ATTRS _mm_cvtsd_si32(__m128d __a) {
 
 /// Converts the lower double-precision floating-point element of a
 ///    128-bit vector of [2 x double], in the second parameter, into a
-///    single-precision floating-point value, returned in the lower 32 bits of a
-///    128-bit vector of [4 x float]. The upper 96 bits of the result vector are
-///    copied from the upper 96 bits of the first parameter.
+///    single-precision floating-point value, returned in the lower 32 bits of
+///    a 128-bit vector of [4 x float]. Rounds inexact results according to the
+///    rounding control bits in the MXCSR register. The upper 96 bits of the
+///    result vector are copied from the upper 96 bits of the first parameter.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1405,11 +1416,11 @@ static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_cvtss_sd(__m128d __a,
 
 /// Converts the two double-precision floating-point elements of a
 ///    128-bit vector of [2 x double] into two signed 32-bit integer values,
-///    returned in the lower 64 bits of a 128-bit vector of [4 x i32].
+///    truncating inexact results.
 ///
-///    If the result of either conversion is inexact, the result is truncated
-///    (rounded towards zero) regardless of the current MXCSR setting. The upper
-///    64 bits of the result vector are set to zero.
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1425,7 +1436,11 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cvttpd_epi32(__m128d __a) {
 }
 
 /// Converts the low-order element of a [2 x double] vector into a 32-bit
-///    signed integer value, truncating the result when it is inexact.
+///    signed integer value, truncating inexact results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1441,8 +1456,13 @@ static __inline__ int __DEFAULT_FN_ATTRS _mm_cvttsd_si32(__m128d __a) {
 }
 
 /// Converts the two double-precision floating-point elements of a
-///    128-bit vector of [2 x double] into two signed 32-bit integer values,
-///    returned in a 64-bit vector of [2 x i32].
+///    128-bit vector of [2 x double] into two signed 32-bit integer values.
+///    Rounds inexact results according to the rounding control bits in the
+///    MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1457,10 +1477,11 @@ static __inline__ __m64 __DEFAULT_FN_ATTRS_MMX _mm_cvtpd_pi32(__m128d __a) {
 
 /// Converts the two double-precision floating-point elements of a
 ///    128-bit vector of [2 x double] into two signed 32-bit integer values,
-///    returned in a 64-bit vector of [2 x i32].
+///    truncating inexact results.
 ///
-///    If the result of either conversion is inexact, the result is truncated
-///    (rounded towards zero) regardless of the current MXCSR setting.
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -3185,8 +3206,9 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cmplt_epi32(__m128i __a,
 #ifdef __x86_64__
 /// Converts a 64-bit signed integer value from the second operand into a
 ///    double-precision value and returns it in the lower element of a [2 x
-///    double] vector; the upper element of the returned vector is copied from
-///    the upper element of the first operand.
+///    double] vector. Rounds inexact results according to the rounding control
+///    bits in the MXCSR register. The upper element of the returned vector is
+///    copied from the upper element of the first operand.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -3207,7 +3229,12 @@ static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_cvtsi64_sd(__m128d __a,
 }
 
 /// Converts the first (lower) element of a vector of [2 x double] into a
-///    64-bit signed integer value, according to the current rounding mode.
+///    64-bit signed integer value. Rounds inexact results according to the
+///    rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -3222,7 +3249,11 @@ static __inline__ long long __DEFAULT_FN_ATTRS _mm_cvtsd_si64(__m128d __a) {
 }
 
 /// Converts the first (lower) element of a vector of [2 x double] into a
-///    64-bit signed integer value, truncating the result when it is inexact.
+///    64-bit signed integer value, truncating inexact results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -3238,7 +3269,8 @@ static __inline__ long long __DEFAULT_FN_ATTRS _mm_cvttsd_si64(__m128d __a) {
 }
 #endif
 
-/// Converts a vector of [4 x i32] into a vector of [4 x float].
+/// Converts a vector of [4 x i32] into a vector of [4 x float]. Rounds inexact
+///    results according to the rounding control bits in the MXCSR register.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -3251,7 +3283,12 @@ static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_cvtepi32_ps(__m128i __a) {
   return (__m128) __builtin_convertvector((__v4si)__a, __v4sf);
 }
 
-/// Converts a vector of [4 x float] into a vector of [4 x i32].
+/// Converts a vector of [4 x float] into a vector of [4 x i32]. Rounds inexact
+///    results according to the rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -3266,7 +3303,11 @@ static __inline__ __m128i __DEFAULT_FN_ATTRS _mm_cvtps_epi32(__m128 __a) {
 }
 
 /// Converts a vector of [4 x float] into a vector of [4 x i32],
-///    truncating the result when it is inexact.
+///    truncating inexact results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
diff --git a/clang/lib/Headers/xmmintrin.h b/clang/lib/Headers/xmmintrin.h
index 47368f3c23d2d6..0feee669e1e818 100644
--- a/clang/lib/Headers/xmmintrin.h
+++ b/clang/lib/Headers/xmmintrin.h
@@ -1295,7 +1295,12 @@ _mm_ucomineq_ss(__m128 __a, __m128 __b)
 }
 
 /// Converts a float value contained in the lower 32 bits of a vector of
-///    [4 x float] into a 32-bit integer.
+///    [4 x float] into a 32-bit integer. Rounds an inexact result according to
+///    the rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1313,7 +1318,12 @@ _mm_cvtss_si32(__m128 __a)
 }
 
 /// Converts a float value contained in the lower 32 bits of a vector of
-///    [4 x float] into a 32-bit integer.
+///    [4 x float] into a 32-bit integer. Rounds an inexact result according to
+///    the rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1333,7 +1343,12 @@ _mm_cvt_ss2si(__m128 __a)
 #ifdef __x86_64__
 
 /// Converts a float value contained in the lower 32 bits of a vector of
-///    [4 x float] into a 64-bit integer.
+///    [4 x float] into a 64-bit integer. Rounds an inexact result according to
+///    the rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1352,8 +1367,13 @@ _mm_cvtss_si64(__m128 __a)
 
 #endif
 
-/// Converts two low-order float values in a 128-bit vector of
-///    [4 x float] into a 64-bit vector of [2 x i32].
+/// Converts two low-order float values in a 128-bit vector of [4 x float] into
+///    a 64-bit vector of [2 x i32]. Rounds inexact results according to the
+///    rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1368,8 +1388,13 @@ _mm_cvtps_pi32(__m128 __a)
   return (__m64)__builtin_ia32_cvtps2pi((__v4sf)__a);
 }
 
-/// Converts two low-order float values in a 128-bit vector of
-///    [4 x float] into a 64-bit vector of [2 x i32].
+/// Converts two low-order float values in a 128-bit vector of [4 x float] into
+///    a 64-bit vector of [2 x i32]. Rounds inexact results according to the
+///    rounding control bits in the MXCSR register.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1385,8 +1410,11 @@ _mm_cvt_ps2pi(__m128 __a)
 }
 
 /// Converts a float value contained in the lower 32 bits of a vector of
-///    [4 x float] into a 32-bit integer, truncating the result when it is
-///    inexact.
+///    [4 x float] into a 32-bit integer, truncating an inexact result.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1404,8 +1432,11 @@ _mm_cvttss_si32(__m128 __a)
 }
 
 /// Converts a float value contained in the lower 32 bits of a vector of
-///    [4 x float] into a 32-bit integer, truncating the result when it is
-///    inexact.
+///    [4 x float] into a 32-bit integer, truncating an inexact result.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1424,8 +1455,11 @@ _mm_cvtt_ss2si(__m128 __a)
 
 #ifdef __x86_64__
 /// Converts a float value contained in the lower 32 bits of a vector of
-///    [4 x float] into a 64-bit integer, truncating the result when it is
-///    inexact.
+///    [4 x float] into a 64-bit integer, truncating an inexact result.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1444,8 +1478,12 @@ _mm_cvttss_si64(__m128 __a)
 #endif
 
 /// Converts two low-order float values in a 128-bit vector of
-///    [4 x float] into a 64-bit vector of [2 x i32], truncating the result
-///    when it is inexact.
+///    [4 x float] into a 64-bit vector of [2 x i32], truncating inexact
+///    results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1462,8 +1500,11 @@ _mm_cvttps_pi32(__m128 __a)
 }
 
 /// Converts two low-order float values in a 128-bit vector of [4 x
-///    float] into a 64-bit vector of [2 x i32], truncating the result when it
-///    is inexact.
+///    float] into a 64-bit vector of [2 x i32], truncating inexact results.
+///
+///    If a converted value is larger than the maximum possible result,
+///    raises a floating-point invalid exception. If the exception is
+///    masked, returns the most negative integer.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1478,10 +1519,11 @@ _mm_cvtt_ps2pi(__m128 __a)
   return _mm_cvttps_pi32(__a);
 }
 
-/// Converts a 32-bit signed integer value into a floating point value
-///    and writes it to the lower 32 bits of the destination. The remaining
-///    higher order elements of the destination vector are copied from the
-///    corresponding elements in the first operand.
+/// Converts a 32-bit signed integer value into a floating point value and
+///    writes it to the lower 32 bits of the destination. Rounds an inexact
+///    result according to the rounding control bits in the MXCSR register.
+///    The remaining higher order elements of the destination vector are copied
+///    from the corresponding elements in the first operand.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1501,10 +1543,11 @@ _mm_cvtsi32_ss(__m128 __a, int __b)
   return __a;
 }
 
-/// Converts a 32-bit signed integer value into a floating point value
-///    and writes it to the lower 32 bits of the destination. The remaining
-///    higher order elements of the destination are copied from the
-///    corresponding elements in the first operand.
+/// Converts a 32-bit signed integer value into a floating point value and
+///    writes it to the lower 32 bits of the destination. Rounds an inexact
+///    result according to the rounding control bits in the MXCSR register.
+///    The remaining higher order elements of the destination are copied from
+///    the corresponding elements in the first operand.
 ///
 /// \headerfile <x86intrin.h>
 ///
@@ -1525,10 +1568,11 @@ _mm_cvt_si2ss(__m128 __a, int __b)
 
 #ifdef __x86_64__
 
-/// Converts a 64-bit signed integer value into a floating point value
-///    and writes it to the lower 32 bits of the destination. The remaining
-///    higher order elements of the destination are copied from the
-///    corresponding elements in the first operand.
+/// Converts a 64-bit signed integer value into a floating point value and
+///    writes it to the lower 32 bits of the destination. Rounds an inexact
+///    result according to the rounding control bits in the MXCSR register.
+///    The remaining higher order elements of the destina...
[truncated]

[phoebewang]

I think we don't need to mention it. This is the default behavior if not mentioned otherwise.

[pogo59]

Discussed this internally and agreed that it is not necessary.

[phoebewang]

Not only larger than, considering the negative number.

[phoebewang]

The last sentence is not necessary and has ambiguity in truncating the fractional part or truncating higher bits to 32-bit.

[pogo59]

Truncation is what distinguishes this from _mm_cvtpd_epi32 so I think it is necessary? That information used to be in the second paragraph, I shortened it and moved it up to the first paragraph, to make that difference more obvious.
Behavior on overflow was added to the second paragraph.

Do you have a different suggestion for how to organize this information?

[phoebewang]

Oh, I took that simply as wrong infromation.

Do you have a different suggestion for how to organize this information?

Not ideal, something like

/// Converts the two double-precision floating-point elements of a
///    128-bit vector of [2 x double] into two signed truncated
///    (rounded towards zero) 32-bit integer values,returned in the
///    most negative integer.

[craigflores]

Test from "craig.flores@sony.com".

[pogo59]

Reworded the parts about truncating results.

[phoebewang]

Looks great, thanks!