Skip to content

Commit

Permalink
erlang: add aarch64 div/rem JIT fix
Browse files Browse the repository at this point in the history 
See erlang/otp#7567 for details.
  • Loading branch information
@fhunleth
fhunleth committed Aug 25, 2023
1 parent 86b0663 commit b1c2382
Showing 1 changed file with 312 additions and 0 deletions.
312 changes: 312 additions & 0 deletions patches/buildroot/0016-erlang-add-aarch64-div-rem-JIT-fix.patch
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,312 @@
From 1aaad4a32e6fcf5df01d9f8a5447afd0e29ea8dd Mon Sep 17 00:00:00 2001
From: Frank Hunleth <fhunleth@troodon-software.com>
Date: Fri, 25 Aug 2023 15:18:39 -0400
Subject: [PATCH] erlang: add aarch64 div/rem JIT fix

---
...-Fix-incorrect-fusion-of-div-and-rem.patch | 293 ++++++++++++++++++
1 file changed, 293 insertions(+)
create mode 100644 package/erlang/26.0.2/0004-Fix-incorrect-fusion-of-div-and-rem.patch

diff --git a/package/erlang/26.0.2/0004-Fix-incorrect-fusion-of-div-and-rem.patch b/package/erlang/26.0.2/0004-Fix-incorrect-fusion-of-div-and-rem.patch
new file mode 100644
index 0000000000..6f2f5957a3
--- /dev/null
+++ b/package/erlang/26.0.2/0004-Fix-incorrect-fusion-of-div-and-rem.patch
@@ -0,0 +1,293 @@
+From 9f46c61840c699228694716edbd31ee002e733ba Mon Sep 17 00:00:00 2001
+From: =?UTF-8?q?Bj=C3=B6rn=20Gustavsson?= <bjorn@erlang.org>
+Date: Wed, 16 Aug 2023 18:17:31 +0200
+Subject: [PATCH] Fix incorrect fusion of `div` and `rem`
+
+The JIT would generate code that calculated the remainder incorrectly
+for the following example:
+
+ bug(Bin) ->
+ N = byte_size(Bin),
+ {N rem 128, N div 128}.
+
+Essentially, the JIT would rewrite the code like this:
+
+ bug(Bin) ->
+ N = byte_size(Bin),
+ Q = N bsr 7,
+ {Q band 16#7f, Q}.
+
+That is, the remainder would be calculated as `(N div 128) rem 128`.
+
+Fixes #7566
+---
+ erts/emulator/beam/jit/arm/instr_arith.cpp | 7 +-
+ erts/emulator/test/small_SUITE.erl | 198 +++++++++++++++++++--
+ 2 files changed, 188 insertions(+), 17 deletions(-)
+
+diff --git a/erts/emulator/beam/jit/arm/instr_arith.cpp b/erts/emulator/beam/jit/arm/instr_arith.cpp
+index 1fbec6dbb2..485f93956d 100644
+--- a/erts/emulator/beam/jit/arm/instr_arith.cpp
++++ b/erts/emulator/beam/jit/arm/instr_arith.cpp
+@@ -696,9 +696,14 @@ void BeamModuleAssembler::emit_div_rem(const ArgLabel &Fail,
+ if (Support::isPowerOf2(divisor) &&
+ std::get<0>(getClampedRange(LHS)) >= 0) {
+ int trailing_bits = Support::ctz<Eterm>(divisor);
++ arm::Gp LHS_reg = lhs.reg;
+ if (need_div) {
+ comment("optimized div by replacing with right shift");
+ ERTS_CT_ASSERT(_TAG_IMMED1_SMALL == _TAG_IMMED1_MASK);
++ if (need_rem && quotient.reg == lhs.reg) {
++ LHS_reg = TMP1;
++ a.mov(LHS_reg, lhs.reg);
++ }
+ a.lsr(quotient.reg, lhs.reg, imm(trailing_bits));
+ a.orr(quotient.reg, quotient.reg, imm(_TAG_IMMED1_SMALL));
+ }
+@@ -706,7 +711,7 @@ void BeamModuleAssembler::emit_div_rem(const ArgLabel &Fail,
+ comment("optimized rem by replacing with masking");
+ auto mask = Support::lsbMask<Uint>(trailing_bits +
+ _TAG_IMMED1_SIZE);
+- a.and_(remainder.reg, lhs.reg, imm(mask));
++ a.and_(remainder.reg, LHS_reg, imm(mask));
+ }
+ } else {
+ a.asr(TMP1, lhs.reg, imm(_TAG_IMMED1_SIZE));
+diff --git a/erts/emulator/test/small_SUITE.erl b/erts/emulator/test/small_SUITE.erl
+index c3ca540cd5..7d49522f05 100644
+--- a/erts/emulator/test/small_SUITE.erl
++++ b/erts/emulator/test/small_SUITE.erl
+@@ -521,17 +521,21 @@ div_gen_pairs() ->
+ {_, MaxSmall} = determine_small_limits(0),
+ NumBitsMaxSmall = num_bits(MaxSmall),
+
++ Divisors = [-8,-2,-1,1,2,3,4,5,8,16,17,64,22222333] ++
++ [1 bsl P || P <- lists:seq(8, 12) ++ lists:seq(26, 36)],
++
+ %% Generate random pairs of smalls.
+- Pairs0 = [{rand:uniform(MaxSmall) * rand_sign(),
++ Pairs0 = [{rand:uniform(MaxSmall),
+ rand:uniform(MaxSmall) * rand_sign()} ||
+- _ <- lists:seq(1, 75)],
+-
+- Pairs1 = [{rand:uniform(MaxSmall), N} ||
+- N <- [-4,-3,-2,-1,1,2,3,5,17,63,64,1111,22222]] ++ Pairs0,
++ _ <- lists:seq(1, 50)],
++ Pairs1 = [{rand:uniform(MaxSmall), N} || N <- Divisors] ++ Pairs0,
++ Pairs2 = [{N, M} || N <- lists:seq(0, 7), M <- [-2,-1,1,2,3,4]] ++ Pairs1,
++ Pairs3 = [{abs(M) * (rand:uniform(10)+1) + rand:uniform(1000), M} ||
++ M <- Divisors] ++ Pairs2,
+
+ %% Generate pairs of numbers whose product are bignums.
+ [{rand:uniform(MaxSmall),1 bsl Pow} ||
+- Pow <- lists:seq(NumBitsMaxSmall - 4, NumBitsMaxSmall - 1)] ++ Pairs1.
++ Pow <- lists:seq(NumBitsMaxSmall - 4, NumBitsMaxSmall - 1)] ++ Pairs3.
+
+ rand_sign() ->
+ case rand:uniform() < 0.2 of
+@@ -572,6 +576,22 @@ gen_div_function({Name,{A,B}}) ->
+ put(prevent_div_rem_fusion, Q),
+ R = X rem Y,
+ {Q, R};
++ '@Name@'(pos_integer1, X, fixed) when is_integer(X), 0 =< X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(pos_integer2, X, fixed) when is_integer(X), 0 =< X, X < _@APlusOne@ ->
++ Y = _@B@,
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(pos_integer3, X, fixed) when is_integer(X), 0 =< X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ put(prevent_div_rem_fusion, Q),
++ R = X rem Y,
++ {Q, R};
+ '@Name@'(number0, X, Y) when -_@APlusOne@ < X, X < _@APlusOne@,
+ -_@BPlusOne@ < Y, Y < _@BPlusOne@ ->
+ Q = X div Y,
+@@ -593,6 +613,108 @@ gen_div_function({Name,{A,B}}) ->
+ Q = X div Y,
+ {Q, R};
+ '@Name@'(number4, X, fixed) when -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ put(prevent_div_rem_fusion, Q),
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(any0, X, fixed) ->
++ Y = _@B@,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(any1, X, fixed) ->
++ Y = _@B@,
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(any2, X, fixed) ->
++ Y = _@B@,
++ Q = X div Y,
++ put(prevent_div_rem_fusion, Q),
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X0, Y0, integer0) ->
++ Q = X0 div Y0,
++ R = X0 rem Y0,
++ if X0 > 0, Y0 > 0 ->
++ <<X:_@NumBitsA@>> = <<X0:_@NumBitsA@>>,
++ <<Y:_@NumBitsB@>> = <<Y0:_@NumBitsB@>>,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ true ->
++ {Q, R}
++ end;
++ '@Name@'(X, fixed, integer1) when is_integer(X), -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, fixed, integer2) when is_integer(X), -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(X, fixed, integer3) when is_integer(X), -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ put(prevent_div_rem_fusion, Q),
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, fixed, pos_integer1) when is_integer(X), 0 =< X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, fixed, pos_integer2) when is_integer(X), 0 =< X, X < _@APlusOne@ ->
++ Y = _@B@,
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(X, fixed, pos_integer3) when is_integer(X), 0 =< X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ put(prevent_div_rem_fusion, Q),
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, Y, number0) when -_@APlusOne@ < X, X < _@APlusOne@,
++ -_@BPlusOne@ < Y, Y < _@BPlusOne@ ->
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, Y, number1) when -_@APlusOne@ < X, X < _@APlusOne@,
++ -_@BPlusOne@ < Y, Y < _@BPlusOne@ ->
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(X, fixed, number2) when -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, fixed, number3) when -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(X, fixed, number4) when -_@APlusOne@ < X, X < _@APlusOne@ ->
++ Y = _@B@,
++ Q = X div Y,
++ put(prevent_div_rem_fusion, Q),
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, fixed, any0) ->
++ Y = _@B@,
++ Q = X div Y,
++ R = X rem Y,
++ {Q, R};
++ '@Name@'(X, fixed, any1) ->
++ Y = _@B@,
++ R = X rem Y,
++ Q = X div Y,
++ {Q, R};
++ '@Name@'(X, fixed, any2) ->
+ Y = _@B@,
+ Q = X div Y,
+ put(prevent_div_rem_fusion, Q),
+@@ -602,16 +724,60 @@ gen_div_function({Name,{A,B}}) ->
+ test_division([{Name,{A,B}}|T], Mod) ->
+ F = fun Mod:Name/3,
+ try
+- Res0 = {A div B, A rem B},
+- Res0 = F(integer0, A, B),
+- Res0 = F(integer1, A, fixed),
+- Res0 = F(integer2, A, fixed),
+- Res0 = F(integer3, A, fixed),
+- Res0 = F(number0, A, B),
+- Res0 = F(number1, A, B),
+- Res0 = F(number2, A, fixed),
+- Res0 = F(number3, A, fixed),
+- Res0 = F(number4, A, fixed)
++ PosRes = {A div B, A rem B},
++ NegRes = {-A div B, -A rem B},
++
++ PosRes = F(integer0, A, B),
++ PosRes = F(integer1, A, fixed),
++ PosRes = F(integer2, A, fixed),
++ PosRes = F(integer3, A, fixed),
++ PosRes = F(pos_integer1, A, fixed),
++ PosRes = F(pos_integer2, A, fixed),
++ PosRes = F(pos_integer3, A, fixed),
++ PosRes = F(number0, A, B),
++ PosRes = F(number1, A, B),
++ PosRes = F(number2, A, fixed),
++ PosRes = F(number3, A, fixed),
++ PosRes = F(number4, A, fixed),
++ PosRes = F(any0, A, fixed),
++ PosRes = F(any1, A, fixed),
++ PosRes = F(any2, A, fixed),
++
++ PosRes = F(A, B, integer0),
++ PosRes = F(A, fixed, integer1),
++ PosRes = F(A, fixed, integer2),
++ PosRes = F(A, fixed, integer3),
++ PosRes = F(A, fixed, pos_integer1),
++ PosRes = F(A, fixed, pos_integer2),
++ PosRes = F(A, fixed, pos_integer3),
++ PosRes = F(A, B, number0),
++ PosRes = F(A, B, number1),
++ PosRes = F(A, fixed, number2),
++ PosRes = F(A, fixed, number3),
++ PosRes = F(A, fixed, number4),
++ PosRes = F(A, fixed, any0),
++ PosRes = F(A, fixed, any1),
++ PosRes = F(A, fixed, any2),
++
++ NegRes = F(integer0, -A, B),
++ NegRes = F(integer1, -A, fixed),
++ NegRes = F(integer2, -A, fixed),
++ NegRes = F(integer3, -A, fixed),
++ NegRes = F(number0, -A, B),
++ NegRes = F(number1, -A, B),
++ NegRes = F(number2, -A, fixed),
++ NegRes = F(number3, -A, fixed),
++ NegRes = F(number4, -A, fixed),
++
++ NegRes = F(-A, B, integer0),
++ NegRes = F(-A, fixed, integer1),
++ NegRes = F(-A, fixed, integer2),
++ NegRes = F(-A, fixed, integer3),
++ NegRes = F(-A, B, number0),
++ NegRes = F(-A, B, number1),
++ NegRes = F(-A, fixed, number2),
++ NegRes = F(-A, fixed, number3),
++ NegRes = F(-A, fixed, number4)
+ catch
+ C:R:Stk ->
+ io:format("~p failed. numbers: ~p ~p\n", [Name,A,B]),
+--
+2.34.1
+
--
2.34.1

0 comments on commit b1c2382

Please sign in to comment.