diff options
| -rw-r--r-- | src/passes/OptimizeInstructions.cpp | 34 | ||||
| -rw-r--r-- | test/passes/optimize-instructions.txt | 70 | ||||
| -rw-r--r-- | test/passes/optimize-instructions.wast | 75 |
3 files changed, 179 insertions, 0 deletions
diff --git a/src/passes/OptimizeInstructions.cpp b/src/passes/OptimizeInstructions.cpp index e0c7217a5..1dd05dd0a 100644 --- a/src/passes/OptimizeInstructions.cpp +++ b/src/passes/OptimizeInstructions.cpp @@ -516,6 +516,12 @@ struct OptimizeInstructions : public WalkerPass<PostWalker<OptimizeInstructions, return ret; } } + // for or, we can potentially combine + if (binary->op == OrInt32) { + if (auto* ret = combineOr(binary)) { + return ret; + } + } // relation/comparisons allow for math optimizations if (binary->isRelational()) { if (auto* ret = optimizeRelational(binary)) { @@ -979,6 +985,34 @@ private: } } + // We can combine `or` operations, e.g. + // (x > y) | (x == y) ==> x >= y + Expression* combineOr(Binary* binary) { + assert(binary->op == OrInt32); + if (auto* left = binary->left->dynCast<Binary>()) { + if (auto* right = binary->right->dynCast<Binary>()) { + if (left->op != right->op && + ExpressionAnalyzer::equal(left->left, right->left) && + ExpressionAnalyzer::equal(left->right, right->right) && + !EffectAnalyzer(getPassOptions(), left->left).hasSideEffects() && + !EffectAnalyzer(getPassOptions(), left->right).hasSideEffects()) { + switch (left->op) { + // (x > y) | (x == y) ==> x >= y + case EqInt32: { + if (right->op == GtSInt32) { + left->op = GeSInt32; + return left; + } + break; + } + default: {} + } + } + } + } + return nullptr; + } + // fold constant factors into the offset void optimizeMemoryAccess(Expression*& ptr, Address& offset) { // ptr may be a const, but it isn't worth folding that in (we still have a const); in fact, diff --git a/test/passes/optimize-instructions.txt b/test/passes/optimize-instructions.txt index 7d82ae8b1..0e8654e48 100644 --- a/test/passes/optimize-instructions.txt +++ b/test/passes/optimize-instructions.txt @@ -3223,6 +3223,76 @@ (i32.const 2) ) ) + (func $pre-combine-or (; 75 ;) (type $5) (param $x i32) (param $y i32) + (drop + (i32.ge_s + (get_local $x) + (get_local $y) + ) + ) + (drop + (i32.ge_s + (get_local $x) + (get_local $y) + ) + ) + (drop + (i32.or + (i32.eq + (get_local $x) + (i32.const 1) + ) + (i32.gt_s + (get_local $x) + (get_local $y) + ) + ) + ) + (drop + (i32.or + (i32.eq + (get_local $x) + (get_local $y) + ) + (i32.gt_s + (get_local $x) + (i32.const 1) + ) + ) + ) + (drop + (i32.or + (i32.gt_s + (call $ne0) + (get_local $y) + ) + (i32.eq + (call $ne0) + (get_local $y) + ) + ) + ) + (drop + (i32.or + (i32.gt_s + (get_local $y) + (call $ne0) + ) + (i32.eq + (call $ne0) + (get_local $y) + ) + ) + ) + ) + (func $combine-or (; 76 ;) (type $5) (param $x i32) (param $y i32) + (drop + (i32.ge_s + (get_local $x) + (get_local $y) + ) + ) + ) ) (module (type $0 (func)) diff --git a/test/passes/optimize-instructions.wast b/test/passes/optimize-instructions.wast index 234b3913b..be9761a2b 100644 --- a/test/passes/optimize-instructions.wast +++ b/test/passes/optimize-instructions.wast @@ -3663,6 +3663,81 @@ (i32.const 2) ) ) + (func $pre-combine-or (param $x i32) (param $y i32) + (drop (i32.or + (i32.gt_s + (get_local $x) + (get_local $y) + ) + (i32.eq + (get_local $y) ;; ordering should not stop us + (get_local $x) + ) + )) + (drop (i32.or + (i32.eq ;; ordering should not stop us + (get_local $y) + (get_local $x) + ) + (i32.gt_s + (get_local $x) + (get_local $y) + ) + )) + (drop (i32.or + (i32.gt_s + (get_local $x) + (get_local $y) + ) + (i32.eq + (get_local $x) + (i32.const 1) ;; not equal + ) + )) + (drop (i32.or + (i32.gt_s + (get_local $x) + (i32.const 1) ;; not equal + ) + (i32.eq + (get_local $x) + (get_local $y) + ) + )) + (drop (i32.or + (i32.gt_s + (call $ne0) ;; side effects + (get_local $y) + ) + (i32.eq + (call $ne0) + (get_local $y) + ) + )) + (drop (i32.or + (i32.gt_s + (get_local $y) + (call $ne0) ;; side effects + ) + (i32.eq + (get_local $y) + (call $ne0) + ) + )) + ) + (func $combine-or (param $x i32) (param $y i32) + (drop (i32.or + (i32.gt_s + (get_local $x) + (get_local $y) + ) + (i32.eq + (get_local $x) + (get_local $y) + ) + )) + ;; TODO: more stuff here + ) ) (module (import "env" "memory" (memory $0 (shared 256 256))) |