From: Sudakshina Das Date: Mon, 19 Mar 2018 18:50:32 +0000 (+0000) Subject: [PR81647][AARCH64] Fix handling of Unordered Comparisons in aarch64-simd.md X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=f7d884d45b929c3f0ab65b0830cce4517b5f968d;p=gcc.git [PR81647][AARCH64] Fix handling of Unordered Comparisons in aarch64-simd.md This patch fixes the inconsistent behavior observed at -O3 for the unordered comparisons. According to the online docs (https://gcc.gnu.org/onlinedocs /gcc-7.2.0/gccint/Unary-and-Binary-Expressions.html), all of the following should not raise an FP exception: - UNGE_EXPR - UNGT_EXPR - UNLE_EXPR - UNLT_EXPR - UNEQ_EXPR Also ORDERED_EXPR and UNORDERED_EXPR should only return zero or one. The aarch64-simd.md handling of these were generating exception raising instructions such as fcmgt. This patch changes the instructions that are emitted in order to not give out the exceptions. We first check each operand for NaNs and force any elements containing NaN to zero before using them in the compare. Example: UN (a, b) -> UNORDERED (a, b) | (cm (isnan (a) ? 0.0 : a, isnan (b) ? 0.0 : b)) The ORDERED_EXPR is now handled as (cmeq (a, a) & cmeq (b, b)) and UNORDERED_EXPR as ~ORDERED_EXPR and UNEQ as (~ORDERED_EXPR | cmeq (a,b)). ChangeLog Entries: *** gcc/ChangeLog *** 2018-03-19 Sudakshina Das PR target/81647 * config/aarch64/aarch64-simd.md (vec_cmp): Modify instructions for UNLT, UNLE, UNGT, UNGE, UNEQ, UNORDERED and ORDERED. *** gcc/testsuite/ChangeLog *** 2018-03-19 Sudakshina Das PR target/81647 * gcc.target/aarch64/pr81647.c: New. From-SVN: r258653 --- diff --git a/gcc/ChangeLog b/gcc/ChangeLog index bed98d6b4ea..713e9642833 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,9 @@ +2018-03-19 Sudakshina Das + + PR target/81647 + * config/aarch64/aarch64-simd.md (vec_cmp): Modify + instructions for UNLT, UNLE, UNGT, UNGE, UNEQ, UNORDERED and ORDERED. + 2018-03-19 Jim Wilson PR bootstrap/84856 diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index 3d1f6a01cb7..1154fc3d58d 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -2730,10 +2730,10 @@ break; } /* Fall through. */ - case UNGE: + case UNLT: std::swap (operands[2], operands[3]); /* Fall through. */ - case UNLE: + case UNGT: case GT: comparison = gen_aarch64_cmgt; break; @@ -2744,10 +2744,10 @@ break; } /* Fall through. */ - case UNGT: + case UNLE: std::swap (operands[2], operands[3]); /* Fall through. */ - case UNLT: + case UNGE: case GE: comparison = gen_aarch64_cmge; break; @@ -2770,21 +2770,41 @@ case UNGT: case UNLE: case UNLT: - case NE: - /* FCM returns false for lanes which are unordered, so if we use - the inverse of the comparison we actually want to emit, then - invert the result, we will end up with the correct result. - Note that a NE NaN and NaN NE b are true for all a, b. - - Our transformations are: - a UNGE b -> !(b GT a) - a UNGT b -> !(b GE a) - a UNLE b -> !(a GT b) - a UNLT b -> !(a GE b) - a NE b -> !(a EQ b) */ - gcc_assert (comparison != NULL); - emit_insn (comparison (operands[0], operands[2], operands[3])); - emit_insn (gen_one_cmpl2 (operands[0], operands[0])); + { + /* All of the above must not raise any FP exceptions. Thus we first + check each operand for NaNs and force any elements containing NaN to + zero before using them in the compare. + Example: UN (a, b) -> UNORDERED (a, b) | + (cm (isnan (a) ? 0.0 : a, + isnan (b) ? 0.0 : b)) + We use the following transformations for doing the comparisions: + a UNGE b -> a GE b + a UNGT b -> a GT b + a UNLE b -> b GE a + a UNLT b -> b GT a. */ + + rtx tmp0 = gen_reg_rtx (mode); + rtx tmp1 = gen_reg_rtx (mode); + rtx tmp2 = gen_reg_rtx (mode); + emit_insn (gen_aarch64_cmeq (tmp0, operands[2], operands[2])); + emit_insn (gen_aarch64_cmeq (tmp1, operands[3], operands[3])); + emit_insn (gen_and3 (tmp2, tmp0, tmp1)); + emit_insn (gen_and3 (tmp0, tmp0, + lowpart_subreg (mode, + operands[2], + mode))); + emit_insn (gen_and3 (tmp1, tmp1, + lowpart_subreg (mode, + operands[3], + mode))); + gcc_assert (comparison != NULL); + emit_insn (comparison (operands[0], + lowpart_subreg (mode, + tmp0, mode), + lowpart_subreg (mode, + tmp1, mode))); + emit_insn (gen_orn3 (operands[0], tmp2, operands[0])); + } break; case LT: @@ -2792,25 +2812,19 @@ case GT: case GE: case EQ: + case NE: /* The easy case. Here we emit one of FCMGE, FCMGT or FCMEQ. As a LT b <=> b GE a && a LE b <=> b GT a. Our transformations are: a GE b -> a GE b a GT b -> a GT b a LE b -> b GE a a LT b -> b GT a - a EQ b -> a EQ b */ + a EQ b -> a EQ b + a NE b -> ~(a EQ b) */ gcc_assert (comparison != NULL); emit_insn (comparison (operands[0], operands[2], operands[3])); - break; - - case UNEQ: - /* We first check (a > b || b > a) which is !UNEQ, inverting - this result will then give us (a == b || a UNORDERED b). */ - emit_insn (gen_aarch64_cmgt (operands[0], - operands[2], operands[3])); - emit_insn (gen_aarch64_cmgt (tmp, operands[3], operands[2])); - emit_insn (gen_ior3 (operands[0], operands[0], tmp)); - emit_insn (gen_one_cmpl2 (operands[0], operands[0])); + if (code == NE) + emit_insn (gen_one_cmpl2 (operands[0], operands[0])); break; case LTGT: @@ -2822,21 +2836,22 @@ emit_insn (gen_ior3 (operands[0], operands[0], tmp)); break; - case UNORDERED: - /* Operands are ORDERED iff (a > b || b >= a), so we can compute - UNORDERED as !ORDERED. */ - emit_insn (gen_aarch64_cmgt (tmp, operands[2], operands[3])); - emit_insn (gen_aarch64_cmge (operands[0], - operands[3], operands[2])); - emit_insn (gen_ior3 (operands[0], operands[0], tmp)); - emit_insn (gen_one_cmpl2 (operands[0], operands[0])); - break; - case ORDERED: - emit_insn (gen_aarch64_cmgt (tmp, operands[2], operands[3])); - emit_insn (gen_aarch64_cmge (operands[0], - operands[3], operands[2])); - emit_insn (gen_ior3 (operands[0], operands[0], tmp)); + case UNORDERED: + case UNEQ: + /* cmeq (a, a) & cmeq (b, b). */ + emit_insn (gen_aarch64_cmeq (operands[0], + operands[2], operands[2])); + emit_insn (gen_aarch64_cmeq (tmp, operands[3], operands[3])); + emit_insn (gen_and3 (operands[0], operands[0], tmp)); + + if (code == UNORDERED) + emit_insn (gen_one_cmpl2 (operands[0], operands[0])); + else if (code == UNEQ) + { + emit_insn (gen_aarch64_cmeq (tmp, operands[2], operands[3])); + emit_insn (gen_orn3 (operands[0], operands[0], tmp)); + } break; default: diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog index 5fbedb4e460..2202df759b9 100644 --- a/gcc/testsuite/ChangeLog +++ b/gcc/testsuite/ChangeLog @@ -1,3 +1,8 @@ +2018-03-19 Sudakshina Das + + PR target/81647 + * gcc.target/aarch64/pr81647.c: New. + 2018-03-19 Richard Biener PR tree-optimization/84933 diff --git a/gcc/testsuite/gcc.target/aarch64/pr81647.c b/gcc/testsuite/gcc.target/aarch64/pr81647.c new file mode 100644 index 00000000000..f60dfba49d5 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/pr81647.c @@ -0,0 +1,44 @@ +/* { dg-do run } */ +/* { dg-options "-O3 -fdump-tree-ssa" } */ + +#include + +double x[28], y[28]; +int res[28]; + +int +main (void) +{ + int i; + for (i = 0; i < 28; ++i) + { + x[i] = __builtin_nan (""); + y[i] = i; + } + __asm__ volatile ("" ::: "memory"); + feclearexcept (FE_ALL_EXCEPT); + for (i = 0; i < 4; ++i) + res[i] = __builtin_isgreater (x[i], y[i]); + for (i = 4; i < 8; ++i) + res[i] = __builtin_isgreaterequal (x[i], y[i]); + for (i = 8; i < 12; ++i) + res[i] = __builtin_isless (x[i], y[i]); + for (i = 12; i < 16; ++i) + res[i] = __builtin_islessequal (x[i], y[i]); + for (i = 16; i < 20; ++i) + res[i] = __builtin_islessgreater (x[i], y[i]); + for (i = 20; i < 24; ++i) + res[i] = __builtin_isunordered (x[i], y[i]); + for (i = 24; i < 28; ++i) + res[i] = !(__builtin_isunordered (x[i], y[i])); + __asm__ volatile ("" ::: "memory"); + return fetestexcept (FE_ALL_EXCEPT) != 0; +} + +/* { dg-final { scan-tree-dump " u> " "ssa" } } */ +/* { dg-final { scan-tree-dump " u>= " "ssa" } } */ +/* { dg-final { scan-tree-dump " u< " "ssa" } } */ +/* { dg-final { scan-tree-dump " u<= " "ssa" } } */ +/* { dg-final { scan-tree-dump " u== " "ssa" } } */ +/* { dg-final { scan-tree-dump " unord " "ssa" } } */ +/* { dg-final { scan-tree-dump " ord " "ssa" } } */