https://gcc.gnu.org/g:7d2496170af17d46323445ea4832e6c098977869
commit r17-5894-g7d2496170af17d46323445ea4832e6c098977869 Author: Karl Meakin <[email protected]> Date: Thu Aug 6 14:10:51 2026 +0000 aarch64: Port vabs intrinsics to pragma-based framework Port the `vabs` family of NEON intrinsics to the pragma-based framework, and add assembly tests for every variant. This also makes the `__builtin_aarch64_abs<mode>` builtin functions obsolete, so they have been deleted too. gcc/ChangeLog: * config/aarch64/aarch64-builtins.cc (aarch64_general_fold_builtin): Delete case to handle `abs`. * config/aarch64/aarch64-neon-builtins-base.cc (vabsd, vabsh, vabs, vabsq): New function bases. * config/aarch64/aarch64-neon-builtins-base.def (vabsd): (vabs): (vabsq): (vabsh): New function declarations. * config/aarch64/aarch64-simd-builtins.def: Delete `abs` builtin functions. * config/aarch64/aarch64-simd.md (aarch64_abs<mode><vczle><vczbe>): Replace UNSPEC_ABS with abs expression. (aarch64_abs_plus<mode>): Likewise. * config/aarch64/iterators.md (UNSPEC_ABS): Delete unspec. * config/aarch64/arm_neon.h (vabs_f32, vabs_f64, vabs_s8, vabs_s16, vabs_s32, vabs_s64, vabsq_f32, vabsq_f64, vabsq_s8, vabsq_s16, vabsq_s32, vabsq_s64, vabsd_s64, vabs_f16, vabsq_f16): Delete function definitions. * config/aarch64/arm_fp16.h (vabsh_f16): Use `__builtin_fabs` instead of `__builtin_aarch64_abshf` now that the latter has been deleted. gcc/testsuite/ChangeLog: * gcc.target/aarch64/neon/vabs.c: New test. * gcc.target/aarch64/neon/trapping_math.c: New tests. * gcc.target/aarch64/singleton_intrinsics_1.c: Mark `vabs_s64` test as `xfail` because of codegen regression. * gcc.target/aarch64/vabs_intrinsic_1.c: Fix test. The `asm volatile` trick for disabling constant folding stopped working. Replace with declaring variables `volatile`. * gcc.target/aarch64/fneg-abs_1.c: Add `-fno-trapping-math` to flags. * gcc.target/aarch64/sve/fneg-abs_1.c: Likewise. Diff: --- gcc/config/aarch64/aarch64-builtins.cc | 2 - gcc/config/aarch64/aarch64-neon-builtins-base.cc | 19 ++++ gcc/config/aarch64/aarch64-neon-builtins-base.def | 11 ++ gcc/config/aarch64/aarch64-simd-builtins.def | 6 -- gcc/config/aarch64/aarch64-simd.md | 18 +--- gcc/config/aarch64/arm_fp16.h | 2 +- gcc/config/aarch64/arm_neon.h | 112 ------------------- gcc/config/aarch64/iterators.md | 1 - gcc/testsuite/gcc.target/aarch64/fneg-abs_1.c | 2 +- .../gcc.target/aarch64/neon/trapping_math.c | 66 ++++++++++++ gcc/testsuite/gcc.target/aarch64/neon/vabs.c | 118 +++++++++++++++++++++ .../gcc.target/aarch64/singleton_intrinsics_1.c | 2 +- gcc/testsuite/gcc.target/aarch64/sve/fneg-abs_1.c | 2 +- .../gcc.target/aarch64/vabs_intrinsic_1.c | 9 +- 14 files changed, 223 insertions(+), 147 deletions(-) diff --git a/gcc/config/aarch64/aarch64-builtins.cc b/gcc/config/aarch64/aarch64-builtins.cc index 765743f55073..d94d75b9afc1 100644 --- a/gcc/config/aarch64/aarch64-builtins.cc +++ b/gcc/config/aarch64/aarch64-builtins.cc @@ -4717,8 +4717,6 @@ aarch64_general_fold_builtin (unsigned int fcode, tree type, { switch (fcode) { - BUILTIN_VDQF (UNOP, abs, 2, ALL) - return fold_build1 (ABS_EXPR, type, args[0]); VAR1 (UNOP, floatv2si, 2, ALL, v2sf) VAR1 (UNOP, floatv4si, 2, ALL, v4sf) VAR1 (UNOP, floatv2di, 2, ALL, v2df) diff --git a/gcc/config/aarch64/aarch64-neon-builtins-base.cc b/gcc/config/aarch64/aarch64-neon-builtins-base.cc index 729e3038f1c1..b8a8060b1622 100644 --- a/gcc/config/aarch64/aarch64-neon-builtins-base.cc +++ b/gcc/config/aarch64/aarch64-neon-builtins-base.cc @@ -884,6 +884,24 @@ struct gimple_reinterpret : public gimple_function_base } }; +struct gimple_abs : public gimple_function_base +{ + gimple *fold (gimple_folder &f) const override + { + auto arg = gimple_call_arg (f.call, 0); + auto arg_type = TREE_TYPE (arg); + + if (!FLOAT_TYPE_P (arg_type)) + return gimple_build_assign ( + f.lhs, build_cast (arg_type, + f.force_val (fold_build1 ( + ABSU_EXPR, unsigned_type_for (arg_type), arg)))); + + auto ret = fold_build1 (ABS_EXPR, arg_type, arg); + return this->maybe_optimization_barrier (f, ret); + } +}; + // Reinterpret NEON_FUNCTION (vreinterpret, vreinterpretq, gimple_reinterpret,) @@ -902,6 +920,7 @@ NEON_FUNCTION (vadd, vaddq, vaddd, gimple_arith, (PLUS_EXPR, PLUS_EXPR, BIT_XOR_ NEON_FUNCTION (vsub, vsubq, vsubd, gimple_arith, (MINUS_EXPR)) NEON_FUNCTION (vdiv, vdivq, gimple_arith, (RDIV_EXPR)) NEON_FUNCTION (vneg, vnegq, vnegd, gimple_arith, (NEGATE_EXPR)) +NEON_FUNCTION (vabs, vabsq, vabsd, gimple_abs,) NEON_FUNCTION (vmul, vmulq, gimple_arith, (MULT_EXPR)) NEON_FUNCTION (vmul_n, vmulq_n, gimple_arith_dup, (MULT_EXPR)) diff --git a/gcc/config/aarch64/aarch64-neon-builtins-base.def b/gcc/config/aarch64/aarch64-neon-builtins-base.def index f2c0d22ee3ab..00428ed42e22 100644 --- a/gcc/config/aarch64/aarch64-neon-builtins-base.def +++ b/gcc/config/aarch64/aarch64-neon-builtins-base.def @@ -101,6 +101,13 @@ DEF_NEON_FUNCTION (vneg, all_signed, ("D0,D0")) DEF_NEON_FUNCTION (vnegq, all_signed, ("Q0,Q0")) DEF_NEON_FUNCTION (vneg, sd_float, ("D0,D0")) DEF_NEON_FUNCTION (vnegq, sd_float, ("Q0,Q0")) + +// Absolute value +DEF_NEON_FUNCTION (vabsd, d_signed, ("s0,s0")) +DEF_NEON_FUNCTION (vabs, all_signed, ("D0,D0")) +DEF_NEON_FUNCTION (vabsq, all_signed, ("Q0,Q0")) +DEF_NEON_FUNCTION (vabs, sd_float, ("D0,D0")) +DEF_NEON_FUNCTION (vabsq, sd_float, ("Q0,Q0")) #undef REQUIRED_EXTENSIONS // Lanewise arithmetic (FP16) @@ -132,6 +139,10 @@ DEF_NEON_FUNCTION (vdivq, h_float, ("Q0,Q0,Q0")) // Negation DEF_NEON_FUNCTION (vneg, h_float, ("D0,D0")) DEF_NEON_FUNCTION (vnegq, h_float, ("Q0,Q0")) + +// Absolute value +DEF_NEON_FUNCTION (vabs, h_float, ("D0,D0")) +DEF_NEON_FUNCTION (vabsq, h_float, ("Q0,Q0")) #undef REQUIRED_EXTENSIONS // Saturating arithmetic diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def index 0df54066b7a9..3254f07af70a 100644 --- a/gcc/config/aarch64/aarch64-simd-builtins.def +++ b/gcc/config/aarch64/aarch64-simd-builtins.def @@ -668,12 +668,6 @@ BUILTIN_VHSDF (UNOP, frecpe, 0, FP) BUILTIN_VHSDF_HSDF (BINOP, frecps, 0, FP) - /* Implemented by a mixture of abs2 patterns. Note the DImode builtin is - only ever used for the int64x1_t intrinsic, there is no scalar version. */ - BUILTIN_VSDQ_I_DI (UNOP, abs, 0, QUIET) - BUILTIN_VHSDF (UNOP, abs, 2, QUIET) - VAR1 (UNOP, abs, 2, QUIET, hf) - BUILTIN_VQ_HSF (UNOP, vec_unpacks_hi_, 10, FP) VAR1 (BINOP, float_truncate_hi_, 0, FP, v4sf) VAR1 (BINOP, float_truncate_hi_, 0, FP, v8hf) diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index 53c2cbf0f687..1599f1fa349e 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -974,19 +974,6 @@ [(set_attr "type" "neon_abs<q>")] ) -;; The intrinsic version of integer ABS must not be allowed to -;; combine with any operation with an integrated ABS step, such -;; as SABD. -(define_insn "aarch64_abs<mode><vczle><vczbe>" - [(set (match_operand:VSDQ_I_DI 0 "register_operand" "=w") - (unspec:VSDQ_I_DI - [(match_operand:VSDQ_I_DI 1 "register_operand" "w")] - UNSPEC_ABS))] - "TARGET_SIMD" - "abs\t%<v>0<Vmtype>, %<v>1<Vmtype>" - [(set_attr "type" "neon_abs<q>")] -) - ;; It's tempting to represent SABD as ABS (MINUS op1 op2). ;; This isn't accurate as ABS treats always its input as a signed value. ;; So (ABS:QI (minus:QI 64 -128)) == (ABS:QI (192 or -64 signed)) == 64. @@ -1318,9 +1305,8 @@ (define_insn_and_split "*aarch64_abs_plus<mode>" [(set (match_operand:VDQ_BHSI 0 "register_operand" "=&w") (plus:VDQ_BHSI - (unspec:VDQ_BHSI - [(match_operand:VDQ_BHSI 1 "register_operand" "w")] - UNSPEC_ABS) + (abs:VDQ_BHSI + (match_operand:VDQ_BHSI 1 "register_operand" "w")) (match_operand:VDQ_BHSI 2 "register_operand" "w")))] "TARGET_SIMD && can_create_pseudo_p ()" "#" diff --git a/gcc/config/aarch64/arm_fp16.h b/gcc/config/aarch64/arm_fp16.h index c6356fdc78bf..f67a9e199880 100644 --- a/gcc/config/aarch64/arm_fp16.h +++ b/gcc/config/aarch64/arm_fp16.h @@ -40,7 +40,7 @@ __extension__ extern __inline float16_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vabsh_f16 (float16_t __a) { - return __builtin_aarch64_abshf (__a); + return __builtin_fabs (__a); } __extension__ extern __inline uint16_t diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h index af5294f81909..fba4a2d8c093 100644 --- a/gcc/config/aarch64/arm_neon.h +++ b/gcc/config/aarch64/arm_neon.h @@ -4509,104 +4509,6 @@ vabdq_f64 (float64x2_t __a, float64x2_t __b) return __builtin_aarch64_fabdv2df (__a, __b); } -/* vabs */ - -__extension__ extern __inline float32x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_f32 (float32x2_t __a) -{ - return __builtin_aarch64_absv2sf (__a); -} - -__extension__ extern __inline float64x1_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_f64 (float64x1_t __a) -{ - return (float64x1_t) {__builtin_fabs (__a[0])}; -} - -__extension__ extern __inline int8x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_s8 (int8x8_t __a) -{ - return __builtin_aarch64_absv8qi (__a); -} - -__extension__ extern __inline int16x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_s16 (int16x4_t __a) -{ - return __builtin_aarch64_absv4hi (__a); -} - -__extension__ extern __inline int32x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_s32 (int32x2_t __a) -{ - return __builtin_aarch64_absv2si (__a); -} - -__extension__ extern __inline int64x1_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_s64 (int64x1_t __a) -{ - return (int64x1_t) {__builtin_aarch64_absdi (__a[0])}; -} - -__extension__ extern __inline float32x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_f32 (float32x4_t __a) -{ - return __builtin_aarch64_absv4sf (__a); -} - -__extension__ extern __inline float64x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_f64 (float64x2_t __a) -{ - return __builtin_aarch64_absv2df (__a); -} - -__extension__ extern __inline int8x16_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_s8 (int8x16_t __a) -{ - return __builtin_aarch64_absv16qi (__a); -} - -__extension__ extern __inline int16x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_s16 (int16x8_t __a) -{ - return __builtin_aarch64_absv8hi (__a); -} - -__extension__ extern __inline int32x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_s32 (int32x4_t __a) -{ - return __builtin_aarch64_absv4si (__a); -} - -__extension__ extern __inline int64x2_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_s64 (int64x2_t __a) -{ - return __builtin_aarch64_absv2di (__a); -} - -/* Try to avoid moving between integer and vector registers. - For why the cast to unsigned is needed check the vnegd_s64 intrinsic. - There is a testcase related to this issue: - gcc.target/aarch64/vabsd_s64.c. */ - -__extension__ extern __inline int64_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsd_s64 (int64_t __a) -{ - return __a < 0 ? - (uint64_t) __a : __a; -} - /* ARMv8.1-A intrinsics. */ #pragma GCC push_options #pragma GCC target ("+nothing+rdma") @@ -17619,20 +17521,6 @@ vuqaddd_s64 (int64_t __a, uint64_t __b) /* ARMv8.2-A FP16 one operand vector intrinsics. */ -__extension__ extern __inline float16x4_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabs_f16 (float16x4_t __a) -{ - return __builtin_aarch64_absv4hf (__a); -} - -__extension__ extern __inline float16x8_t -__attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) -vabsq_f16 (float16x8_t __a) -{ - return __builtin_aarch64_absv8hf (__a); -} - __extension__ extern __inline uint16x4_t __attribute__ ((__always_inline__, __gnu_inline__, __artificial__)) vceqz_f16 (float16x4_t __a) diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md index ae78fbe5b5bd..dfee954a5d54 100644 --- a/gcc/config/aarch64/iterators.md +++ b/gcc/config/aarch64/iterators.md @@ -926,7 +926,6 @@ [ UNSPEC_ASHIFT_SIGNED ; Used in aarch-simd.md. UNSPEC_ASHIFT_UNSIGNED ; Used in aarch64-simd.md. - UNSPEC_ABS ; Used in aarch64-simd.md. UNSPEC_FCVTN_FP8 ; Used in aarch64-simd.md. UNSPEC_FCVTN2_FP8 ; Used in aarch64-builtins.cc. UNSPEC_F1CVTL_FP8 ; Used in aarch64-simd.md. diff --git a/gcc/testsuite/gcc.target/aarch64/fneg-abs_1.c b/gcc/testsuite/gcc.target/aarch64/fneg-abs_1.c index 43c37e21b50e..9af313f42203 100644 --- a/gcc/testsuite/gcc.target/aarch64/fneg-abs_1.c +++ b/gcc/testsuite/gcc.target/aarch64/fneg-abs_1.c @@ -1,5 +1,5 @@ /* { dg-do compile } */ -/* { dg-options "-O3" } */ +/* { dg-options "-O3 -fno-trapping-math" } */ /* { dg-final { check-function-bodies "**" "" "" { target lp64 } } } */ #pragma GCC target "+nosve" diff --git a/gcc/testsuite/gcc.target/aarch64/neon/trapping_math.c b/gcc/testsuite/gcc.target/aarch64/neon/trapping_math.c index 0f56da59c12e..9ad49b580422 100644 --- a/gcc/testsuite/gcc.target/aarch64/neon/trapping_math.c +++ b/gcc/testsuite/gcc.target/aarch64/neon/trapping_math.c @@ -72,3 +72,69 @@ test_vaddq_f64 (float64x2_t a, float64x2_t b) { vaddq_f64 (a, b); } + +/* +** test_vabs_f16: +** fabs v[0-9]+\.4h, v0\.4h +** ret +*/ +void +test_vabs_f16 (float16x4_t a) +{ + vabs_f16 (a); +} + +/* +** test_vabsq_f16: +** fabs v[0-9]+\.8h, v0\.8h +** ret +*/ +void +test_vabsq_f16 (float16x8_t a) +{ + vabsq_f16 (a); +} + +/* +** test_vabs_f32: +** fabs v[0-9]+\.2s, v0\.2s +** ret +*/ +void +test_vabs_f32 (float32x2_t a) +{ + vabs_f32 (a); +} + +/* +** test_vabsq_f32: +** fabs v[0-9]+\.4s, v0\.4s +** ret +*/ +void +test_vabsq_f32 (float32x4_t a) +{ + vabsq_f32 (a); +} + +/* +** test_vabs_f64: +** fabs d[0-9]+, d0 +** ret +*/ +void +test_vabs_f64 (float64x1_t a) +{ + vabs_f64 (a); +} + +/* +** test_vabsq_f64: +** fabs v[0-9]+\.2d, v0\.2d +** ret +*/ +void +test_vabsq_f64 (float64x2_t a) +{ + vabsq_f64 (a); +} diff --git a/gcc/testsuite/gcc.target/aarch64/neon/vabs.c b/gcc/testsuite/gcc.target/aarch64/neon/vabs.c new file mode 100644 index 000000000000..5b0bf5d48510 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/neon/vabs.c @@ -0,0 +1,118 @@ +/* { dg-do compile } */ +/* { dg-final { check-function-bodies "**" "" } } */ + +#include "arm_neon_test.h" + +/* +** test_vabs_f32: +** fabs v0\.2s, v0\.2s +** ret +*/ +TEST_UNIFORM_UNARY (vabs_f32, float32x2_t) + +/* +** test_vabs_f64: +** fabs d0, d0 +** ret +*/ +TEST_UNIFORM_UNARY (vabs_f64, float64x1_t) + +/* +** test_vabs_s8: +** abs v0\.8b, v0\.8b +** ret +*/ +TEST_UNIFORM_UNARY (vabs_s8, int8x8_t) + +/* +** test_vabs_s16: +** abs v0\.4h, v0\.4h +** ret +*/ +TEST_UNIFORM_UNARY (vabs_s16, int16x4_t) + +/* +** test_vabs_s32: +** abs v0\.2s, v0\.2s +** ret +*/ +TEST_UNIFORM_UNARY (vabs_s32, int32x2_t) + +/* FIXME (Karl Meakin): Performs abs in scalar register, even though it requires + extra moves: +fmov x31, d0 +cmp x31, #?0 +csneg x31, x31, x31, ge +fmov d0, x31 +*/ + +/* +** test_vabs_s64: { xfail *-*-* } +** abs d0, d0 +** ret +*/ +TEST_UNIFORM_UNARY (vabs_s64, int64x1_t) + +/* +** test_vabsd_s64: +** cmp x0, #?0 +** csneg x0, x0, x0, ge +** ret +*/ +TEST_UNIFORM_UNARY (vabsd_s64, int64_t) + +/* +** test_vabsq_f32: +** fabs v0\.4s, v0\.4s +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_f32, float32x4_t) + +/* +** test_vabsq_f64: +** fabs v0\.2d, v0\.2d +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_f64, float64x2_t) + +/* +** test_vabsq_s8: +** abs v0\.16b, v0\.16b +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_s8, int8x16_t) + +/* +** test_vabsq_s16: +** abs v0\.8h, v0\.8h +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_s16, int16x8_t) + +/* +** test_vabsq_s32: +** abs v0\.4s, v0\.4s +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_s32, int32x4_t) + +/* +** test_vabsq_s64: +** abs v0\.2d, v0\.2d +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_s64, int64x2_t) + +/* +** test_vabs_f16: +** fabs v0\.4h, v0\.4h +** ret +*/ +TEST_UNIFORM_UNARY (vabs_f16, float16x4_t) + +/* +** test_vabsq_f16: +** fabs v0\.8h, v0\.8h +** ret +*/ +TEST_UNIFORM_UNARY (vabsq_f16, float16x8_t) diff --git a/gcc/testsuite/gcc.target/aarch64/singleton_intrinsics_1.c b/gcc/testsuite/gcc.target/aarch64/singleton_intrinsics_1.c index 27360150b582..062d3aca375e 100644 --- a/gcc/testsuite/gcc.target/aarch64/singleton_intrinsics_1.c +++ b/gcc/testsuite/gcc.target/aarch64/singleton_intrinsics_1.c @@ -19,7 +19,7 @@ test_vadd_s64 (int64x1_t a, int64x1_t b) return vadd_s64 (a, b); } -/* { dg-final { scan-assembler-times "\\tabs\\td\[0-9\]+, d\[0-9\]+" 1 } } */ +/* { dg-final { scan-assembler-times "\\tabs\\td\[0-9\]+, d\[0-9\]+" 1 { xfail *-*-* } } } */ int64x1_t test_vabs_s64 (int64x1_t a) diff --git a/gcc/testsuite/gcc.target/aarch64/sve/fneg-abs_1.c b/gcc/testsuite/gcc.target/aarch64/sve/fneg-abs_1.c index 03560008fda1..67c49cabcf19 100644 --- a/gcc/testsuite/gcc.target/aarch64/sve/fneg-abs_1.c +++ b/gcc/testsuite/gcc.target/aarch64/sve/fneg-abs_1.c @@ -1,5 +1,5 @@ /* { dg-do compile } */ -/* { dg-options "-O3" } */ +/* { dg-options "-O3 -fno-trapping-math" } */ /* { dg-final { check-function-bodies "**" "" "" { target lp64 } } } */ #include <arm_neon.h> diff --git a/gcc/testsuite/gcc.target/aarch64/vabs_intrinsic_1.c b/gcc/testsuite/gcc.target/aarch64/vabs_intrinsic_1.c index b18db7ec6411..f21771bfe890 100644 --- a/gcc/testsuite/gcc.target/aarch64/vabs_intrinsic_1.c +++ b/gcc/testsuite/gcc.target/aarch64/vabs_intrinsic_1.c @@ -13,8 +13,9 @@ static void \ test_vabs##q##_##size (ETYPE (size) * res, \ const ETYPE (size) *in1) \ { \ - VTYPE (size, lanes) a = vld1##q##_s##size (res); \ - VTYPE (size, lanes) b = vld1##q##_s##size (in1); \ + /* Use volatile to prevent constant folding. */ \ + volatile VTYPE (size, lanes) a = vld1##q##_s##size (res); \ + volatile VTYPE (size, lanes) b = vld1##q##_s##size (in1); \ a = vabs##q##_s##size (b); \ vst1##q##_s##size (res, a); \ } @@ -54,15 +55,11 @@ test_##size (void) \ ETYPE (size) res2[lanes_128] = {0}; \ ETYPE (size) expected2[lanes_128] = EXPECTED##lanes_128; \ \ - /* Forcefully avoid optimization. */ \ - asm volatile ("" : : : "memory"); \ test_vabs_##size (res1, pool1); \ for (i = 0; i < lanes_64; i++) \ if (res1[i] != expected1[i]) \ abort (); \ \ - /* Forcefully avoid optimization. */ \ - asm volatile ("" : : : "memory"); \ test_vabsq_##size (res2, pool2); \ for (i = 0; i < lanes_128; i++) \ if (res2[i] != expected2[i]) \
