From 93ea8e708d5fc3dfe01de8d5b6d1061fef91b5d4 Mon Sep 17 00:00:00 2001 From: Christian Hagedorn Date: Tue, 15 Oct 2024 15:28:47 +0000 Subject: [PATCH] 8330045: Enhance array handling Co-authored-by: Christian Hagedorn Co-authored-by: Emanuel Peter Co-authored-by: Francisco Ferrari Bihurriet Co-authored-by: Martin Balao Reviewed-by: rhalade, ahgross, thartmann, epeter, adinn, roland --- src/hotspot/share/opto/subnode.cpp | 101 +++++++++++++----- .../compiler/c2/gvn/TestBoolNodeGVN.java | 59 ++++++---- 2 files changed, 113 insertions(+), 47 deletions(-) diff --git a/src/hotspot/share/opto/subnode.cpp b/src/hotspot/share/opto/subnode.cpp index 445eb168214..f1eca4682bd 100644 --- a/src/hotspot/share/opto/subnode.cpp +++ b/src/hotspot/share/opto/subnode.cpp @@ -1623,27 +1623,17 @@ Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) { return new BoolNode( ncmp, _test.negate() ); } - // Change ((x & (m - 1)) u< m) into (m > 0) - // This is the off-by-one variant of ((x & m) u<= m) - if (cop == Op_CmpU && - _test._test == BoolTest::lt && - cmp1_op == Op_AndI) { - Node* l = cmp1->in(1); - Node* r = cmp1->in(2); - for (int repeat = 0; repeat < 2; repeat++) { - bool match = r->Opcode() == Op_AddI && r->in(2)->find_int_con(0) == -1 && - r->in(1) == cmp2; - if (match) { - // arraylength known to be non-negative, so a (arraylength != 0) is sufficient, - // but to be compatible with the array range check pattern, use (arraylength u> 0) - Node* ncmp = cmp2->Opcode() == Op_LoadRange - ? phase->transform(new CmpUNode(cmp2, phase->intcon(0))) - : phase->transform(new CmpINode(cmp2, phase->intcon(0))); - return new BoolNode(ncmp, BoolTest::gt); - } else { - // commute and try again - l = cmp1->in(2); - r = cmp1->in(1); + // Transform: "((x & (m - 1)) u 0)" + // This is case [CMPU_MASK] which is further described at the method comment of BoolNode::Value_cmpu_and_mask(). + if (cop == Op_CmpU && _test._test == BoolTest::lt && cmp1_op == Op_AndI) { + Node* m = cmp2; // RHS: m + for (int add_idx = 1; add_idx <= 2; add_idx++) { // LHS: "(m + (-1)) & x" or "x & (m + (-1))"? + Node* maybe_m_minus_1 = cmp1->in(add_idx); + if (maybe_m_minus_1->Opcode() == Op_AddI && + maybe_m_minus_1->in(2)->find_int_con(0) == -1 && + maybe_m_minus_1->in(1) == m) { + Node* m_cmpu_0 = phase->transform(new CmpUNode(m, phase->intcon(0))); + return new BoolNode(m_cmpu_0, BoolTest::gt); } } } @@ -1809,9 +1799,57 @@ Node *BoolNode::Ideal(PhaseGVN *phase, bool can_reshape) { // } } -//------------------------------Value------------------------------------------ -// Change ((x & m) u<= m) or ((m & x) u<= m) to always true -// Same with ((x & m) u< m+1) and ((m & x) u< m+1) +// We use the following Lemmas/insights for the following two transformations (1) and (2): +// x & y <=u y, for any x and y (Lemma 1, masking always results in a smaller unsigned number) +// y u 0) +// This is the off-by-one variant of the above. +// +// We now prove that this replacement is correct. This is the same as proving +// "m >u 0" if and only if "x & (m - 1) u 0 <=> x & (m - 1) m >u 0": +// We prove this by contradiction: +// Assume m <=u 0 which is equivalent to m == 0: +// and thus +// x & (m - 1) u 0 => x & (m - 1) u 0, no underflow of "m - 1" +// +// +// Note that the signed version of "m > 0": +// m > 0 <=> x & (m - 1) 0 +// is false which is a contradiction. +// +// (1a) and (1b) is covered by this method since we can directly return a true value as type while (2) is covered +// in BoolNode::Ideal since we create a new non-constant node (see [CMPU_MASK]). const Type* BoolNode::Value_cmpu_and_mask(PhaseValues* phase) const { Node* cmp = in(1); if (cmp != nullptr && cmp->Opcode() == Op_CmpU) { @@ -1819,14 +1857,21 @@ const Type* BoolNode::Value_cmpu_and_mask(PhaseValues* phase) const { Node* cmp2 = cmp->in(2); if (cmp1->Opcode() == Op_AndI) { - Node* bound = nullptr; + Node* m = nullptr; if (_test._test == BoolTest::le) { - bound = cmp2; + // (1a) "((x & m) <=u m)", cmp2 = m + m = cmp2; } else if (_test._test == BoolTest::lt && cmp2->Opcode() == Op_AddI && cmp2->in(2)->find_int_con(0) == 1) { - bound = cmp2->in(1); + // (1b) "(x & m) in(1); + const TypeInt* rhs_m_type = phase->type(rhs_m)->isa_int(); + if (rhs_m_type->_lo > -1 || rhs_m_type->_hi < -1) { + // Exclude any case where m == -1 is possible. + m = rhs_m; + } } - if (cmp1->in(2) == bound || cmp1->in(1) == bound) { + if (cmp1->in(2) == m || cmp1->in(1) == m) { return TypeInt::ONE; } } diff --git a/test/hotspot/jtreg/compiler/c2/gvn/TestBoolNodeGVN.java b/test/hotspot/jtreg/compiler/c2/gvn/TestBoolNodeGVN.java index 3f2fe7ecf05..384dc6f8f11 100644 --- a/test/hotspot/jtreg/compiler/c2/gvn/TestBoolNodeGVN.java +++ b/test/hotspot/jtreg/compiler/c2/gvn/TestBoolNodeGVN.java @@ -56,27 +56,47 @@ public class TestBoolNodeGVN { @Test @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) @IR(failOn = IRNode.CMP_U, - phase = CompilePhase.AFTER_PARSING, - applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) public static boolean testShouldReplaceCpmUCase2(int x, int m) { return !(Integer.compareUnsigned((m & x), m) > 0); } @Test - @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) + @Arguments(values = {Argument.DEFAULT, Argument.RANDOM_EACH}) @IR(failOn = IRNode.CMP_U, - phase = CompilePhase.AFTER_PARSING, - applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) public static boolean testShouldReplaceCpmUCase3(int x, int m) { + m = Math.max(0, m); + return Integer.compareUnsigned((x & m), m + 1) < 0; + } + + @Test + @Arguments(values = {Argument.DEFAULT, Argument.RANDOM_EACH}) + @IR(failOn = IRNode.CMP_U, + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + public static boolean testShouldReplaceCpmUCase4(int x, int m) { + m = Math.max(0, m); + return Integer.compareUnsigned((m & x), m + 1) < 0; + } + + @Test + @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) + @IR(counts = {IRNode.CMP_U, "1"}, // m could be -1 and thus optimization cannot be applied + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + public static boolean testShouldNotReplaceCpmUCase1(int x, int m) { return Integer.compareUnsigned((x & m), m + 1) < 0; } @Test @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) - @IR(failOn = IRNode.CMP_U, - phase = CompilePhase.AFTER_PARSING, - applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) - public static boolean testShouldReplaceCpmUCase4(int x, int m) { + @IR(counts = {IRNode.CMP_U, "1"}, // m could be -1 and thus optimization cannot be applied + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + public static boolean testShouldNotReplaceCpmUCase2(int x, int m) { return Integer.compareUnsigned((m & x), m + 1) < 0; } @@ -92,8 +112,8 @@ public class TestBoolNodeGVN { @Test @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) @IR(counts = {IRNode.CMP_U, "1"}, - phase = CompilePhase.AFTER_PARSING, - applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) public static boolean testShouldHaveCpmUCase2(int x, int m) { return !(Integer.compareUnsigned((m & x), m - 1) > 0); } @@ -101,8 +121,8 @@ public class TestBoolNodeGVN { @Test @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) @IR(counts = {IRNode.CMP_U, "1"}, - phase = CompilePhase.AFTER_PARSING, - applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) public static boolean testShouldHaveCpmUCase3(int x, int m) { return Integer.compareUnsigned((x & m), m + 2) < 0; } @@ -110,22 +130,23 @@ public class TestBoolNodeGVN { @Test @Arguments(values = {Argument.DEFAULT, Argument.DEFAULT}) @IR(counts = {IRNode.CMP_U, "1"}, - phase = CompilePhase.AFTER_PARSING, - applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) + phase = CompilePhase.AFTER_PARSING, + applyIfPlatformOr = {"x64", "true", "aarch64", "true", "riscv64", "true"}) public static boolean testShouldHaveCpmUCase4(int x, int m) { return Integer.compareUnsigned((m & x), m + 2) < 0; } private static void testCorrectness() { int[] values = { - 0, 1, 5, 8, 16, 42, 100, new Random().nextInt(0, Integer.MAX_VALUE), Integer.MAX_VALUE + -100, -42, -16, -8, -5, -1, 0, 1, 5, 8, 16, 42, 100, + new Random().nextInt(), Integer.MAX_VALUE, Integer.MIN_VALUE }; for (int x : values) { for (int m : values) { - if (!testShouldReplaceCpmUCase1(x, m) | - !testShouldReplaceCpmUCase2(x, m) | - !testShouldReplaceCpmUCase3(x, m) | + if (!testShouldReplaceCpmUCase1(x, m) || + !testShouldReplaceCpmUCase2(x, m) || + !testShouldReplaceCpmUCase3(x, m) || !testShouldReplaceCpmUCase4(x, m)) { throw new RuntimeException("Bad result for x = " + x + " and m = " + m + ", expected always true"); }