8346664: C2: Optimize mask check with constant offset #22856
Conversation
Extends the optimization of masked sums introduced in openjdk#6697 to cover constant values, which currently break the optimization. Such constant values arise in an expression of the following form, for example from MemorySegmentImpl#isAlignedForElement: (base + (index + 1) << 8) & 255 => MulNode (base + (index << 8 + 256)) & 255 => AddNode ((base + index << 8) + 256) & 255 Currently, "256" is not being recognized as a shifted value. This PR enables: ((base + index << 8) + 256) & 255 => MulNode (base + index << 8) & 255 => MulNode (PR openjdk#6697) base & 255
|
Hi @mernst-github, welcome to this OpenJDK project and thanks for contributing! We do not recognize you as Contributor and need to ensure you have signed the Oracle Contributor Agreement (OCA). If you have not signed the OCA, please follow the instructions. Please fill in your GitHub username in the "Username" field of the application. Once you have signed the OCA, please let us know by writing If you already are an OpenJDK Author, Committer or Reviewer, please click here to open a new issue so that we can record that fact. Please use "Add GitHub user mernst-github" as summary for the issue. If you are contributing this work on behalf of your employer and your employer has signed the OCA, please let us know by writing |
|
❗ This change is not yet ready to be integrated. |
|
@mernst-github The following label will be automatically applied to this pull request:
When this pull request is ready to be reviewed, an "RFR" email will be sent to the corresponding mailing list. If you would like to change these labels, use the /label pull request command. |
mernst-github
changed the title
| @@ -2052,94 +2052,88 @@ const Type* RotateRightNode::Value(PhaseGVN* phase) const { | |||
| } | |||
| } | |||
|
|
|||
| // Given an expression (AndX shift mask) or (AndX mask shift), | |||
| // Returns a lower bound of the number of trailing zeros in expr. | |||
| jint MulNode::AndIL_min_trailing_zeros(PhaseGVN* phase, Node* expr, BasicType bt) { | |||
There was a problem hiding this comment.
This could be a static function, I don't see much value in it being a method in MulNode.
| // Returns a lower bound of the number of trailing zeros in expr. | ||
| jint MulNode::AndIL_min_trailing_zeros(PhaseGVN* phase, Node* expr, BasicType bt) { | ||
| expr = expr->uncast(); | ||
| if (expr == nullptr) { |
There was a problem hiding this comment.
This should not be nullptr, you can safely remove it.
|
|
||
| if (type->is_con()) { | ||
| long con = type->get_con_as_long(type->basic_type()); | ||
| return con == 0L ? 0 : count_trailing_zeros(con); |
There was a problem hiding this comment.
For the sake of consistency, we should return the type width for con == 0, you can obtain this by type2aelementbytes(bt) * 8
|
|
||
| if (expr->Opcode() == Op_ConvI2L) { | ||
| expr = expr->in(1); | ||
| if (expr == nullptr) { |
There was a problem hiding this comment.
This cannot be nullptr, you can safely remove it, the same for expr->uncast() below. In general, the only case when the input of a ConvI2L (and other nodes) not being an int is when it is top, which means it is empty. A.k.a unreachable code.
| if (expr == nullptr) { | ||
| return 0; | ||
| } | ||
| type = phase->type(expr)->isa_int(); |
There was a problem hiding this comment.
You are trying to look through a ConvI2L, I think for the sake of consistency, you can reassign bt to T_INT at this point.
| return 0; | ||
| } | ||
| const TypeInt* rhs_t = phase->type(rhs)->isa_int(); | ||
| if (!rhs_t || !rhs_t->is_con()) { |
There was a problem hiding this comment.
We are trying to avoid implicit conversion to bool, you can use an explicit rhs_t != nullptr here.
| if (!rhs_t || !rhs_t->is_con()) { | ||
| return 0; | ||
| } | ||
| return rhs_t->get_con() & ((type->isa_int() ? BitsPerJavaInteger : BitsPerJavaLong) - 1); |
There was a problem hiding this comment.
If you reassign bt, you can do type2aelementbytes(bt), which IMO is clearer.
| bool MulNode::AndIL_shift_and_mask_is_always_zero(PhaseGVN* phase, Node* shift, Node* mask, BasicType bt, bool check_reverse) { | ||
| if (mask == nullptr || shift == nullptr) { | ||
| // mask M, we check for both operand orders. | ||
| bool MulNode::AndIL_is_always_zero(PhaseGVN* phase, Node* expr, Node* mask, BasicType bt, bool check_reverse) { |
There was a problem hiding this comment.
Actually you cannot conclude that ((x + y) & m) == 0 iff (x & m) == 0 when (y & m) == 0 because the addition x + y can carry some bit into the positions at which m is set. Consider this example for illustration:
(0b1010 + 0b0010) & 0b0100 == 0b1100 & 0b0100 == 0b0100 != 0
even when
0b1010 & 0b0100 == 0
0b0010 & 0b0100 == 0
The most trivial sufficient condition we are using here is that the lowest bit set of y is larger than the highest bit set of m. Because then adding y into x does not carry any bit into the result that is set in m but not set in x. This method can be a static function, too IMO.
| @@ -120,7 +120,7 @@ public static void checkShiftNonConstMaskLong(long res) { | |||
| @IR(counts = { IRNode.AND_I, "1" }) | |||
| @IR(failOn = { IRNode.ADD_I, IRNode.LSHIFT_I }) | |||
| public static int addShiftMaskInt(int i, int j) { | |||
| return (j + (i << 2)) & 3; // transformed to: return j & 3; | |||
| return (j + ((i + 1) << 2)) & 3; // transformed to: return j & 3; | |||
There was a problem hiding this comment.
I would prefer you adding other test cases instead of modifying existing ones.
Fixes JDK-8346664: extends the optimization of masked sums introduced in #6697 to cover constant values, which currently break the optimization.
Such constant values arise in an expression of the following form, for example from
MemorySegmentImpl#isAlignedForElement:Currently,
256is not being recognized as a shifted value. This PR enables further reduction:Implementation notes:
Progress
Error
Issue
Reviewing
Using
gitCheckout this PR locally:
$ git fetch https://git.openjdk.org/jdk.git pull/22856/head:pull/22856$ git checkout pull/22856Update a local copy of the PR:
$ git checkout pull/22856$ git pull https://git.openjdk.org/jdk.git pull/22856/headUsing Skara CLI tools
Checkout this PR locally:
$ git pr checkout 22856View PR using the GUI difftool:
$ git pr show -t 22856Using diff file
Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/22856.diff