[PowerPC] Provide XL-compatible builtins in altivec.h

[nemanjai]

Summary

There are some interfaces in altivec.h that are not compatible between Clang and XL (although Clang is compatible with GCC). Currently, we have found 3 but there may be others.

Clang/GCC signatures:

vector double vec_ctf(vector signed long long)
vector double vec_ctf(vector unsigned long long)
vector signed long long vec_cts(vector double)
vector unsigned long long vec_ctu(vector double)

XL signatures:

vector float vec_ctf(vector signed long long)
vector float vec_ctf(vector unsigned long long)
vector signed int vec_cts(vector double)
vector unsigned int vec_ctu(vector double)

This patch provides the XL behaviour under the __XL_COMPAT_ALTIVEC__ macro for users that rely on XL behaviour.

[cebowleratibm]

clang/test/CodeGen/builtins-ppc-xlcompat.c

3

Suggest adding an explicit -mcpu.

[bmahjour]

LGTM

clang/test/CodeGen/builtins-ppc-xlcompat.c

7

[nit] remove extra space

[jsji]

clang/lib/Headers/altivec.h

3055

This only affects VSX version right? If so, can we move #ifdef __XL_COMPAT_ALTIVEC__ into #ifdef __VSX__？
Or even better only into the part of affected types, eg: vector unsinged long long?

3065

Can we also add comments about what is the major difference with and without __XL_COMPAT_ALTIVEC__ , for long term maintenance?

clang/test/CodeGen/builtins-ppc-xlcompat.c

9

Can we add a RUN line to test that novsx are not affected?

[jsji]

clang/lib/Headers/altivec.h

3065

Can we also add comments about what is the major difference with and without __XL_COMPAT_ALTIVEC__ , for long term maintenance?

I know they are in commit message, but I think comments in source code would be more helpful than in commit message. Thanks.

[nemanjai]

clang/lib/Headers/altivec.h

3055

Sure, I'll make the non-vsx version unconditional on __XL_COMPAT_ALTIVEC__.

3065

Will do.

clang/test/CodeGen/builtins-ppc-xlcompat.c

3

I'm not sure that works with cc1. It would probably neeed to be something like -target-cpu.

7

Oops, I'll fix it.

9

I don't think this is useful as that would be a duplicate of the tests in builtins-ppc-vsx.c.

clang/lib/Headers/altivec.h

 }
 
 static __inline__ vector double __ATTRS_o_ai vec_cpsgn(vector double __a,
                                                        vector double __b) {
   return __builtin_vsx_xvcpsgndp(__b, __a);
 }
 #endif
 
 /* vec_ctf */
 
 #ifdef __VSX__
+// There are some functions that have different signatures with the XL compiler
+// from those in Clang/GCC and documented in the PVIPR. This macro ensures that
+// the XL-compatible signatures are used for those functions.
+#ifdef __XL_COMPAT_ALTIVEC__
+#define vec_ctf(__a, __b)                                                      \
+  _Generic((__a), vector int                                                   \
+           : (vector float)__builtin_altivec_vcfsx((vector int)(__a), (__b)),  \
+             vector unsigned int                                               \
+           : (vector float)__builtin_altivec_vcfux((vector unsigned int)(__a), \
+                                                   (__b)),                     \
+             vector unsigned long long                                         \
+           : (__builtin_vsx_xvcvuxdsp((vector unsigned long long)(__a)) *      \
+              (vector float)(vector unsigned)((0x7f - (__b)) << 23)),          \
+             vector signed long long                                           \
+           : (__builtin_vsx_xvcvsxdsp((vector signed long long)(__a)) *        \
+              (vector float)(vector unsigned)((0x7f - (__b)) << 23)))
+#else // __XL_COMPAT_ALTIVEC__
 #define vec_ctf(__a, __b)                                                      \
   _Generic((__a), vector int                                                   \
            : (vector float)__builtin_altivec_vcfsx((vector int)(__a), (__b)),  \
              vector unsigned int                                               \
            : (vector float)__builtin_altivec_vcfux((vector unsigned int)(__a), \
                                                    (__b)),                     \
              vector unsigned long long                                         \
            : (__builtin_convertvector((vector unsigned long long)(__a),        \
                                       vector double) *                         \
               (vector double)(vector unsigned long long)((0x3ffULL - (__b))    \
                                                          << 52)),              \
              vector signed long long                                           \
            : (__builtin_convertvector((vector signed long long)(__a),          \
                                       vector double) *                         \
               (vector double)(vector unsigned long long)((0x3ffULL - (__b))    \
                                                          << 52)))
+#endif // __XL_COMPAT_ALTIVEC__
 #else
 #define vec_ctf(__a, __b)                                                      \
   _Generic((__a), vector int                                                   \
            : (vector float)__builtin_altivec_vcfsx((vector int)(__a), (__b)),  \
              vector unsigned int                                               \
            : (vector float)__builtin_altivec_vcfux((vector unsigned int)(__a), \
                                                    (__b)))
 #endif

 #define vec_vcfux __builtin_altivec_vcfux
 /* vec_vcfux */
 
 #define vec_vcfsx(__a, __b) __builtin_altivec_vcfsx((vector int)(__a), (__b))
 
 /* vec_cts */
 
 #ifdef __VSX__
+#ifdef __XL_COMPAT_ALTIVEC__
+#define vec_cts(__a, __b)                                                      \
+  _Generic((__a), vector float                                                 \
+           : __builtin_altivec_vctsxs((vector float)(__a), (__b)),             \
+             vector double                                                     \
+           : __extension__({                                                   \
+             vector double __ret =                                             \
+                 (vector double)(__a) *                                        \
+                 (vector double)(vector unsigned long long)((0x3ffULL + (__b)) \
+                                                            << 52);            \
+             __builtin_vsx_xvcvdpsxws(__ret);                                  \
+           }))
+#else // __XL_COMPAT_ALTIVEC__
 #define vec_cts(__a, __b)                                                      \
   _Generic((__a), vector float                                                 \
            : __builtin_altivec_vctsxs((vector float)(__a), (__b)),             \
              vector double                                                     \
            : __extension__({                                                   \
              vector double __ret =                                             \
                  (vector double)(__a) *                                        \
                  (vector double)(vector unsigned long long)((0x3ffULL + (__b)) \
                                                             << 52);            \
              __builtin_convertvector(__ret, vector signed long long);          \
            }))
+#endif // __XL_COMPAT_ALTIVEC__
 #else
 #define vec_cts __builtin_altivec_vctsxs
 #endif
 
 /* vec_vctsxs */
 
 #define vec_vctsxs __builtin_altivec_vctsxs
 
 /* vec_ctu */
 
 #ifdef __VSX__
+#ifdef __XL_COMPAT_ALTIVEC__
+#define vec_ctu(__a, __b)                                                      \
+  _Generic((__a), vector float                                                 \
+           : __builtin_altivec_vctuxs((vector float)(__a), (__b)),             \
+             vector double                                                     \
+           : __extension__({                                                   \
+             vector double __ret =                                             \
+                 (vector double)(__a) *                                        \
+                 (vector double)(vector unsigned long long)((0x3ffULL + __b)   \
+                                                            << 52);            \
+             __builtin_vsx_xvcvdpuxws(__ret);                                  \
+           }))
+#else // __XL_COMPAT_ALTIVEC__
 #define vec_ctu(__a, __b)                                                      \
   _Generic((__a), vector float                                                 \
            : __builtin_altivec_vctuxs((vector float)(__a), (__b)),             \
              vector double                                                     \
            : __extension__({                                                   \
              vector double __ret =                                             \
                  (vector double)(__a) *                                        \
                  (vector double)(vector unsigned long long)((0x3ffULL + __b)   \
                                                             << 52);            \
              __builtin_convertvector(__ret, vector unsigned long long);        \
            }))
+#endif // __XL_COMPAT_ALTIVEC__
 #else
 #define vec_ctu __builtin_altivec_vctuxs
 #endif
 
 #ifdef __LITTLE_ENDIAN__
 /* vec_ctsl */
 
 #ifdef __VSX__

clang/test/CodeGen/builtins-ppc-xlcompat.c

+// REQUIRES: powerpc-registered-target
+// RUN: %clang_cc1 -target-feature +altivec -target-feature +vsx \
+// RUN:   -triple powerpc64-unknown-unknown -emit-llvm %s -o - \
+// RUN:   -D__XL_COMPAT_ALTIVEC__ -target-cpu pwr7 | FileCheck %s
+// RUN: %clang_cc1 -target-feature +altivec -target-feature +vsx \
+// RUN:   -triple powerpc64le-unknown-unknown -emit-llvm %s -o - \
+// RUN:   -D__XL_COMPAT_ALTIVEC__ -target-cpu pwr8 | FileCheck %s
+#include <altivec.h>
+vector double vd = { 3.4e22, 1.8e-3 };
+vector signed long long vsll = { -12345678999ll, 12345678999 };
+vector unsigned long long vull = { 11547229456923630743llu, 18014402265226391llu };
+vector float res_vf;
+vector signed int res_vsi;
+vector unsigned int res_vui;
+
+void test() {
+// CHECK-LABEL: @test(
+// CHECK-NEXT:  entry:
+// CHECK-LE-LABEL: @test(
+// CHECK-LE-NEXT:  entry:
+
+  res_vf = vec_ctf(vsll, 4);
+// CHECK:         [[TMP0:%.*]] = load <2 x i64>, <2 x i64>* @vsll, align 16
+// CHECK-NEXT:    [[TMP1:%.*]] = call <4 x float> @llvm.ppc.vsx.xvcvsxdsp(<2 x i64> [[TMP0]])
+// CHECK-NEXT:    fmul <4 x float> [[TMP1]], <float 6.250000e-02, float 6.250000e-02, float 6.250000e-02, float 6.250000e-02>
+
+  res_vf = vec_ctf(vull, 4);
+// CHECK:         [[TMP2:%.*]] = load <2 x i64>, <2 x i64>* @vull, align 16
+// CHECK-NEXT:    [[TMP3:%.*]] = call <4 x float> @llvm.ppc.vsx.xvcvuxdsp(<2 x i64> [[TMP2]])
+// CHECK-NEXT:    fmul <4 x float> [[TMP3]], <float 6.250000e-02, float 6.250000e-02, float 6.250000e-02, float 6.250000e-02>
+
+  res_vsi = vec_cts(vd, 4);
+// CHECK:         [[TMP4:%.*]] = load <2 x double>, <2 x double>* @vd, align 16
+// CHECK-NEXT:    fmul <2 x double> [[TMP4]], <double 1.600000e+01, double 1.600000e+01>
+// CHECK:         call <4 x i32> @llvm.ppc.vsx.xvcvdpsxws(<2 x double>
+
+  res_vui = vec_ctu(vd, 4);
+// CHECK:         [[TMP8:%.*]] = load <2 x double>, <2 x double>* @vd, align 16
+// CHECK-NEXT:    fmul <2 x double> [[TMP8]], <double 1.600000e+01, double 1.600000e+01>
+// CHECK:         call <4 x i32> @llvm.ppc.vsx.xvcvdpuxws(<2 x double>
+}