|
| 1 | +#include "mlir/IR/PatternMatch.h" |
| 2 | +#include "mlir/Transforms/WalkPatternRewriteDriver.h" |
| 3 | + |
| 4 | +#include "src/enzyme_ad/jax/Passes/Passes.h" |
| 5 | + |
| 6 | +#include "stablehlo/dialect/StablehloOps.h" |
| 7 | + |
| 8 | +namespace mlir { |
| 9 | +namespace enzyme { |
| 10 | +#define GEN_PASS_DEF_WHILELOOPOUTSIDEVALUESADDTOARGUMENTLISTPASS |
| 11 | +#include "src/enzyme_ad/jax/Passes/Passes.h.inc" |
| 12 | +} // namespace enzyme |
| 13 | +} // namespace mlir |
| 14 | + |
| 15 | +using namespace mlir; |
| 16 | +using namespace mlir::stablehlo; |
| 17 | + |
| 18 | +namespace { |
| 19 | + |
| 20 | +static bool definedOutside(Value v, Operation *op) { |
| 21 | + return !op->isAncestor(v.getParentBlock()->getParentOp()); |
| 22 | +} |
| 23 | + |
| 24 | +struct SHLOWhileOpUpdateArgumentListPattern final |
| 25 | + : public OpRewritePattern<stablehlo::WhileOp> { |
| 26 | + using OpRewritePattern::OpRewritePattern; |
| 27 | + |
| 28 | + LogicalResult matchAndRewrite(stablehlo::WhileOp whileOp, |
| 29 | + PatternRewriter &rewriter) const override { |
| 30 | + // Collect values used inside cond/body that are defined outside the WhileOp |
| 31 | + SmallVector<Value, 4> extraValues; |
| 32 | + SmallPtrSet<Value, 8> seen; |
| 33 | + |
| 34 | + auto collectExternal = [&](Region ®ion) { |
| 35 | + region.walk([&](Operation *op) { |
| 36 | + for (OpOperand &operand : op->getOpOperands()) { |
| 37 | + Value v = operand.get(); |
| 38 | + if (!v) |
| 39 | + continue; |
| 40 | + if (definedOutside(v, whileOp) && !seen.contains(v)) { |
| 41 | + seen.insert(v); |
| 42 | + extraValues.push_back(v); |
| 43 | + } |
| 44 | + } |
| 45 | + }); |
| 46 | + }; |
| 47 | + |
| 48 | + collectExternal(whileOp.getCond()); |
| 49 | + collectExternal(whileOp.getBody()); |
| 50 | + |
| 51 | + if (extraValues.empty()) |
| 52 | + return failure(); |
| 53 | + |
| 54 | + // Build new operand list = existing operands + external values |
| 55 | + SmallVector<Value, 8> newOperands(whileOp.getOperands().begin(), |
| 56 | + whileOp.getOperands().end()); |
| 57 | + for (Value v : extraValues) |
| 58 | + newOperands.push_back(v); |
| 59 | + |
| 60 | + SmallVector<Type, 8> newResultTypes; |
| 61 | + newResultTypes.reserve(newOperands.size()); |
| 62 | + for (Value v : newOperands) |
| 63 | + newResultTypes.push_back(v.getType()); |
| 64 | + |
| 65 | + auto newWhile = stablehlo::WhileOp::create(rewriter, whileOp.getLoc(), |
| 66 | + newResultTypes, newOperands); |
| 67 | + |
| 68 | + rewriter.inlineRegionBefore(whileOp.getCond(), newWhile.getCond(), |
| 69 | + newWhile.getCond().end()); |
| 70 | + rewriter.inlineRegionBefore(whileOp.getBody(), newWhile.getBody(), |
| 71 | + newWhile.getBody().end()); |
| 72 | + |
| 73 | + // Append block arguments for the extra values |
| 74 | + Block &condBlock = newWhile.getCond().front(); |
| 75 | + Block &bodyBlock = newWhile.getBody().front(); |
| 76 | + |
| 77 | + unsigned origArgCount = whileOp.getNumOperands(); |
| 78 | + SmallVector<BlockArgument, 8> addedCondArgs, addedBodyArgs; |
| 79 | + addedCondArgs.reserve(extraValues.size()); |
| 80 | + addedBodyArgs.reserve(extraValues.size()); |
| 81 | + for (Value v : extraValues) { |
| 82 | + addedCondArgs.push_back(condBlock.addArgument(v.getType(), v.getLoc())); |
| 83 | + addedBodyArgs.push_back(bodyBlock.addArgument(v.getType(), v.getLoc())); |
| 84 | + } |
| 85 | + |
| 86 | + // Remap uses of external values inside the regions to the new block args |
| 87 | + auto remapRegionUses = [&](Region ®ion, ArrayRef<Value> externals, |
| 88 | + ArrayRef<BlockArgument> args) { |
| 89 | + region.walk([&](Operation *op) { |
| 90 | + for (OpOperand &operand : op->getOpOperands()) { |
| 91 | + Value v = operand.get(); |
| 92 | + for (auto [ext, arg] : llvm::zip(externals, args)) { |
| 93 | + if (v == ext) { |
| 94 | + operand.set(arg); |
| 95 | + break; |
| 96 | + } |
| 97 | + } |
| 98 | + } |
| 99 | + }); |
| 100 | + }; |
| 101 | + |
| 102 | + remapRegionUses(newWhile.getCond(), extraValues, addedCondArgs); |
| 103 | + remapRegionUses(newWhile.getBody(), extraValues, addedBodyArgs); |
| 104 | + |
| 105 | + Operation *terminator = bodyBlock.getTerminator(); |
| 106 | + if (!terminator) { |
| 107 | + return rewriter.notifyMatchFailure(whileOp, "missing body terminator"); |
| 108 | + } |
| 109 | + |
| 110 | + auto retOp = dyn_cast<stablehlo::ReturnOp>(terminator); |
| 111 | + assert(retOp && "expected stablehlo::ReturnOp"); |
| 112 | + |
| 113 | + SmallVector<Value, 8> newRetVals(retOp.getOperands().begin(), |
| 114 | + retOp.getOperands().end()); |
| 115 | + for (BlockArgument arg : addedBodyArgs) |
| 116 | + newRetVals.push_back(arg); |
| 117 | + |
| 118 | + rewriter.setInsertionPoint(terminator); |
| 119 | + rewriter.replaceOpWithNewOp<stablehlo::ReturnOp>(terminator, newRetVals); |
| 120 | + |
| 121 | + for (unsigned i = 0; i < origArgCount; ++i) |
| 122 | + rewriter.replaceAllUsesWith(whileOp.getResult(i), newWhile.getResult(i)); |
| 123 | + rewriter.eraseOp(whileOp); |
| 124 | + return success(); |
| 125 | + } |
| 126 | +}; |
| 127 | + |
| 128 | +struct WhileLoopOutsideValuesAddToArgumentListPass |
| 129 | + : public enzyme::impl::WhileLoopOutsideValuesAddToArgumentListPassBase< |
| 130 | + WhileLoopOutsideValuesAddToArgumentListPass> { |
| 131 | + using WhileLoopOutsideValuesAddToArgumentListPassBase:: |
| 132 | + WhileLoopOutsideValuesAddToArgumentListPassBase; |
| 133 | + |
| 134 | + void runOnOperation() override { |
| 135 | + RewritePatternSet patterns(&getContext()); |
| 136 | + patterns.add<SHLOWhileOpUpdateArgumentListPattern>(patterns.getContext()); |
| 137 | + walkAndApplyPatterns(getOperation(), std::move(patterns)); |
| 138 | + } |
| 139 | +}; |
| 140 | + |
| 141 | +} // namespace |
0 commit comments