Hi, This is a rebased version of patch that adds a pattern to neon.md for implementing division with multiplication by reciprocal using vrecpe/vrecps with -funsafe-math-optimizations excluding -Os. The newly added test-cases are not vectorized on armeb target with -O2. I posted the analysis for that here: https://gcc.gnu.org/ml/gcc-patches/2016-05/msg01765.html
Briefly, the difference between little and big-endian vectorizer is in arm_builtin_support_vector_misalignment() which calls default_builtin_support_vector_misalignment() for big-endian case, and that returns false because movmisalign_optab does not exist for V2SF mode. This isn't observed with -O3 because loop peeling for alignment gets enabled. It seems that the test cases in patch appear unsupported on armeb, after r221677 thus this patch requires no changes to target-supports.exp to adjust for armeb (unlike last time which stalled the patch). Bootstrap+tested on arm-linux-gnueabihf. Cross-tested on arm*-*-* variants. OK for trunk ? Thanks, Prathamesh
2018-10-26 Prathamesh Kulkarni <prathamesh.kulka...@linaro.org> * config/arm/neon.md (div<mode>3): New pattern. testsuite/ * gcc.target/arm/neon-vect-div-1.c: New test. * gcc.target/arm/neon-vect-div-2.c: Likewise. diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md index 5aeee4b08c1..25ed45d381a 100644 --- a/gcc/config/arm/neon.md +++ b/gcc/config/arm/neon.md @@ -620,6 +620,38 @@ (const_string "neon_mul_<V_elem_ch><q>")))] ) +/* Perform division using multiply-by-reciprocal. + Reciprocal is calculated using Newton-Raphson method. + Enabled with -funsafe-math-optimizations -freciprocal-math + and disabled for -Os since it increases code size . */ + +(define_expand "div<mode>3" + [(set (match_operand:VCVTF 0 "s_register_operand" "=w") + (div:VCVTF (match_operand:VCVTF 1 "s_register_operand" "w") + (match_operand:VCVTF 2 "s_register_operand" "w")))] + "TARGET_NEON && !optimize_size + && flag_unsafe_math_optimizations && flag_reciprocal_math" + { + rtx rec = gen_reg_rtx (<MODE>mode); + rtx vrecps_temp = gen_reg_rtx (<MODE>mode); + + /* Reciprocal estimate. */ + emit_insn (gen_neon_vrecpe<mode> (rec, operands[2])); + + /* Perform 2 iterations of newton-raphson method. */ + for (int i = 0; i < 2; i++) + { + emit_insn (gen_neon_vrecps<mode> (vrecps_temp, rec, operands[2])); + emit_insn (gen_mul<mode>3 (rec, rec, vrecps_temp)); + } + + /* We now have reciprocal in rec, perform operands[0] = operands[1] * rec. */ + emit_insn (gen_mul<mode>3 (operands[0], operands[1], rec)); + DONE; + } +) + + (define_insn "mul<mode>3add<mode>_neon" [(set (match_operand:VDQW 0 "s_register_operand" "=w") (plus:VDQW (mult:VDQW (match_operand:VDQW 2 "s_register_operand" "w") diff --git a/gcc/testsuite/gcc.target/arm/neon-vect-div-1.c b/gcc/testsuite/gcc.target/arm/neon-vect-div-1.c new file mode 100644 index 00000000000..50d04b4175b --- /dev/null +++ b/gcc/testsuite/gcc.target/arm/neon-vect-div-1.c @@ -0,0 +1,16 @@ +/* Test pattern div<mode>3. */ +/* { dg-do compile } */ +/* { dg-require-effective-target arm_neon_ok } */ +/* { dg-require-effective-target vect_hw_misalign } */ +/* { dg-options "-O2 -ftree-vectorize -funsafe-math-optimizations -fdump-tree-vect-details" } */ +/* { dg-add-options arm_neon } */ + +void +foo (int len, float * __restrict p, float *__restrict x) +{ + len = len & ~31; + for (int i = 0; i < len; i++) + p[i] = p[i] / x[i]; +} + +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */ diff --git a/gcc/testsuite/gcc.target/arm/neon-vect-div-2.c b/gcc/testsuite/gcc.target/arm/neon-vect-div-2.c new file mode 100644 index 00000000000..606f54b4e0e --- /dev/null +++ b/gcc/testsuite/gcc.target/arm/neon-vect-div-2.c @@ -0,0 +1,16 @@ +/* Test pattern div<mode>3. */ +/* { dg-do compile } */ +/* { dg-require-effective-target arm_neon_ok } */ +/* { dg-require-effective-target vect_hw_misalign } */ +/* { dg-options "-O3 -ftree-vectorize -funsafe-math-optimizations -fdump-tree-vect-details -fno-reciprocal-math" } */ +/* { dg-add-options arm_neon } */ + +void +foo (int len, float * __restrict p, float *__restrict x) +{ + len = len & ~31; + for (int i = 0; i < len; i++) + p[i] = p[i] / x[i]; +} + +/* { dg-final { scan-tree-dump-not "vectorized 1 loops" "vect" } } */