diff options
author | Hao Liu OS <hliu@os.amperecomputing.com> | 2023-07-06 10:04:46 -0600 |
---|---|---|
committer | Jeff Law <jlaw@ventanamicro.com> | 2023-07-06 10:04:46 -0600 |
commit | 224fd59b2dc8a5fa78a309a09863afe9b3cf2111 (patch) | |
tree | 4c78a652002dcf12e2e24c935e5e240fbcae5115 /gcc | |
parent | b90a70984a9beee39b41f842b56926f9db2069ca (diff) | |
download | gcc-224fd59b2dc8a5fa78a309a09863afe9b3cf2111.zip gcc-224fd59b2dc8a5fa78a309a09863afe9b3cf2111.tar.gz gcc-224fd59b2dc8a5fa78a309a09863afe9b3cf2111.tar.bz2 |
Vect: use a small step to calculate induction for the unrolled loop (PR tree-optimization/110449)
If a loop is unrolled by n times during vectoriation, two steps are used to
calculate the induction variable:
- The small step for the unrolled ith-copy: vec_1 = vec_iv + (VF/n * Step)
- The large step for the whole loop: vec_loop = vec_iv + (VF * Step)
This patch calculates an extra vec_n to replace vec_loop:
vec_n = vec_prev + (VF/n * S) = vec_iv + (VF/n * S) * n = vec_loop.
So that we can save the large step register and related operations.
gcc/ChangeLog:
PR tree-optimization/110449
* tree-vect-loop.cc (vectorizable_induction): use vec_n to replace
vec_loop for the unrolled loop.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/pr110449.c: New testcase.
Diffstat (limited to 'gcc')
-rw-r--r-- | gcc/testsuite/gcc.target/aarch64/pr110449.c | 40 | ||||
-rw-r--r-- | gcc/tree-vect-loop.cc | 21 |
2 files changed, 58 insertions, 3 deletions
diff --git a/gcc/testsuite/gcc.target/aarch64/pr110449.c b/gcc/testsuite/gcc.target/aarch64/pr110449.c new file mode 100644 index 0000000..bb3b6dc --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/pr110449.c @@ -0,0 +1,40 @@ +/* { dg-do compile } */ +/* { dg-options "-Ofast -mcpu=neoverse-n2 --param aarch64-vect-unroll-limit=2" } */ +/* { dg-final { scan-assembler-not "8.0e\\+0" } } */ + +/* Calcualte the vectorized induction with smaller step for an unrolled loop. + + before (suggested_unroll_factor=2): + fmov s30, 8.0e+0 + fmov s31, 4.0e+0 + dup v27.4s, v30.s[0] + dup v28.4s, v31.s[0] + .L6: + mov v30.16b, v31.16b + fadd v31.4s, v31.4s, v27.4s + fadd v29.4s, v30.4s, v28.4s + stp q30, q29, [x0] + add x0, x0, 32 + cmp x1, x0 + bne .L6 + + after: + fmov s31, 4.0e+0 + dup v29.4s, v31.s[0] + .L6: + fadd v30.4s, v31.4s, v29.4s + stp q31, q30, [x0] + add x0, x0, 32 + fadd v31.4s, v29.4s, v30.4s + cmp x0, x1 + bne .L6 */ + +void +foo2 (float *arr, float freq, float step) +{ + for (int i = 0; i < 1024; i++) + { + arr[i] = freq; + freq += step; + } +} diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc index 36d19a5..7d917bf 100644 --- a/gcc/tree-vect-loop.cc +++ b/gcc/tree-vect-loop.cc @@ -10098,7 +10098,7 @@ vectorizable_induction (loop_vec_info loop_vinfo, new_vec, step_vectype, NULL); vec_def = induc_def; - for (i = 1; i < ncopies; i++) + for (i = 1; i < ncopies + 1; i++) { /* vec_i = vec_prev + vec_step */ gimple_seq stmts = NULL; @@ -10108,8 +10108,23 @@ vectorizable_induction (loop_vec_info loop_vinfo, vec_def = gimple_convert (&stmts, vectype, vec_def); gsi_insert_seq_before (&si, stmts, GSI_SAME_STMT); - new_stmt = SSA_NAME_DEF_STMT (vec_def); - STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt); + if (i < ncopies) + { + new_stmt = SSA_NAME_DEF_STMT (vec_def); + STMT_VINFO_VEC_STMTS (stmt_info).safe_push (new_stmt); + } + else + { + /* vec_1 = vec_iv + (VF/n * S) + vec_2 = vec_1 + (VF/n * S) + ... + vec_n = vec_prev + (VF/n * S) = vec_iv + VF * S = vec_loop + + vec_n is used as vec_loop to save the large step register and + related operations. */ + add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop), + UNKNOWN_LOCATION); + } } } |