JIT: Strip shift amount masking via VN to prevent CSE interference by Copilot · Pull Request #126801 · dotnet/runtime

Copilot · 2026-04-11T22:49:03Z

Description

Roslyn emits shl(x, and(y, 31)) for shift operations. On architectures where hardware already masks the shift amount, the AND is redundant. Currently LowerShift strips it, but that runs after CSE. When multiple shifts share the same variable shift amount, CSE hoists (shift & 31) into a temp, and LowerShift no longer recognizes the pattern:

uint ShiftAndCSE(uint foo, int shift)
{
    uint res = (foo >> shift);
    res <<= shift;
    return res;
}

; Before: CSE'd mask not removed
and      r8d, 31
shrx     eax, edx, r8d
shlx     eax, eax, r8d

; After: VN sees through the mask, no AND emitted
shrx     eax, edi, esi
shlx     eax, eax, esi

Changes

src/coreclr/jit/valuenum.cpp: In fgValueNumberTree, when computing the VN for shift operations (GT_LSH, GT_RSH, GT_RSZ), if the shift amount is AND(y, mask) where the mask preserves all relevant shift bits (0x1F for 32-bit, 0x3F for 64-bit), use y's VN instead of the AND's VN. This lets CSE see through the redundant mask. Guarded by the same target set as LowerShift (TARGET_XARCH || TARGET_ARM64 || TARGET_LOONGARCH64 || TARGET_RISCV64). Mirrors the existing LowerShift mask check logic.
src/tests/JIT/opt/Shifts/ShiftMaskCSE.cs: New test covering the CSE+shift interaction pattern — multiple shifts sharing a variable shift amount across 32-bit, 64-bit, signed, and unsigned variants.

The existing LowerShift optimization is intentionally kept as a safety net for AND masks introduced by later phases. Rotation recognition in morph (fgRecognizeAndMorphBitwiseRotation) is unaffected — it handles both masked and unmasked shift amounts.

When Roslyn emits `x << y`, it generates `shl(x, and(y, 31))`. On architectures where the hardware already masks the shift amount (x86, x64, ARM64, LoongArch64, RISC-V64), this AND is redundant. Previously, the AND was only stripped late in LowerShift, after CSE had already run. If two shifts share the same variable shift amount, CSE would hoist the `(shift & 31)` expression, preventing LowerShift from recognizing the pattern. Now we strip the mask early during import, before CSE runs. This prevents the AND from being CSE'd and avoids generating unnecessary masking instructions. Agent-Logs-Url: https://github.com/dotnet/runtime/sessions/0d8e29d4-bfec-465e-bd94-77a38f880f35 Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Agent-Logs-Url: https://github.com/dotnet/runtime/sessions/0d8e29d4-bfec-465e-bd94-77a38f880f35 Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

EgorBo · 2026-04-12T15:30:35Z

@copilot please adjust RangeOps::ShiftRight and ShiftLeft as well.

Since the importer now strips the AND mask from shift amounts, the range checker may see shift amounts with wider ranges than [0..31]. Add ClampShiftAmount to apply the implicit hardware masking so range analysis is not regressed. Agent-Logs-Url: https://github.com/dotnet/runtime/sessions/1c1b5e88-5f5e-4f48-a357-0bf2704a210b Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot · 2026-04-12T15:56:01Z

@copilot please adjust RangeOps::ShiftRight and ShiftLeft as well.

Done in 2772485. Added ClampShiftAmount helper that applies the implicit hardware masking (& 31) to the shift amount range, and wired it into both ShiftRight and ShiftLeft. This prevents range analysis from regressing when the importer strips the explicit AND mask.

EgorBo · 2026-04-12T18:14:53Z

@copilot revert all changes and only make VN changes - if shift's arg is AND that we can ignore - use its op1 VN effectively ignoring AND.

tannergooding · 2026-04-12T18:18:09Z

src/coreclr/jit/importer.cpp

+#if defined(TARGET_XARCH) || defined(TARGET_ARM64) || defined(TARGET_LOONGARCH64) || defined(TARGET_RISCV64)
+                // Shift instructions on these targets mask the shift amount,
+                // so strip the redundant AND mask (e.g., & 31 for 32-bit or & 63 for 64-bit)
+                // that Roslyn emits. Doing this early prevents CSE from hoisting
+                // the mask and hiding it from later optimizations.
+                {
+                    size_t mask = 0x1f;
+#ifdef TARGET_64BIT
+                    if (varTypeIsLong(type))
+                    {
+                        mask = 0x3f;
+                    }
+#endif
+                    while (op2->OperIs(GT_AND) && op2->gtGetOp2()->IsCnsIntOrI())
+                    {
+                        if ((static_cast<size_t>(op2->gtGetOp2()->AsIntCon()->IconValue()) & mask) != mask)
+                        {
+                            break;
+                        }
+                        op2 = op2->gtGetOp1();
+                    }
+                }
+#endif
+
+                op1 = gtNewOperNode(oper, type, op1, op2);


@EgorBo, this rather feels like something we should just handle gtFoldExpr. It should make the logic simpler and would kick in anywhere such a pattern is found.

Basically just look at the rhs of the shift operation to see if its GT_AND and see if it is x & cns where cns is one of the well known masks. In such a case, just fold away the and

regardless of where I put it, it currently produces regressions, I assume some of them might be range check analysis related - not sure, needs an investigation. I want to check if we can simply ignore its existence on the VN level

Revert the importer.cpp shift mask stripping and rangecheck.h ClampShiftAmount changes per reviewer feedback. Instead, strip redundant AND masks from shift amounts at VN level: in fgValueNumberTree, when computing VN for shift operations, if the shift amount is AND(y, mask) where the mask preserves all relevant shift bits, use y's VN instead. This lets CSE see through the redundant mask without modifying the tree structure. Agent-Logs-Url: https://github.com/dotnet/runtime/sessions/c731e39e-f338-4d79-b29a-a13762165d21 Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot · 2026-04-12T18:57:13Z

@copilot revert all changes and only make VN changes - if shift's arg is AND that we can ignore - use its op1 VN effectively ignoring AND.

Done in 8d83b8a. Reverted all importer.cpp and rangecheck.h changes. The optimization is now VN-only: in fgValueNumberTree, when computing VN for shift operations (GT_LSH/GT_RSH/GT_RSZ), if the shift amount is AND(y, mask) where the mask preserves all relevant shift bits, the shift's VN uses y's VN instead of the AND's VN.

Initial plan

bf069e7

Copilot AI assigned Copilot and EgorBo Apr 11, 2026

Copilot AI requested review from Copilot and removed request for Copilot April 11, 2026 22:49

Copilot started work on behalf of EgorBo April 11, 2026 22:49 View session

Copilot AI linked an issue Apr 11, 2026 that may be closed by this pull request

JIT: Strip out masking of shiftAmount way earlier when possible #126799

Open

github-actions bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Apr 11, 2026

Copilot AI requested review from Copilot and removed request for Copilot April 11, 2026 23:36

Extract maskOp local variable for readability in shift mask stripping

3091ef0

Agent-Logs-Url: https://github.com/dotnet/runtime/sessions/0d8e29d4-bfec-465e-bd94-77a38f880f35 Co-authored-by: EgorBo <523221+EgorBo@users.noreply.github.com>

Copilot AI requested review from Copilot and removed request for Copilot April 11, 2026 23:38

Copilot AI changed the title ~~[WIP] Optimize shiftAmount masking in JIT import process~~ JIT: Strip shift amount masking in importer to prevent CSE interference Apr 11, 2026

Copilot AI requested a review from EgorBo April 11, 2026 23:42

Copilot finished work on behalf of EgorBo April 11, 2026 23:42

Copilot started work on behalf of EgorBo April 12, 2026 15:30 View session

Copilot AI requested review from Copilot and removed request for Copilot April 12, 2026 15:55

Copilot finished work on behalf of EgorBo April 12, 2026 15:57

Copilot started work on behalf of EgorBo April 12, 2026 18:15 View session

tannergooding reviewed Apr 12, 2026

View reviewed changes

Copilot AI requested review from Copilot and removed request for Copilot April 12, 2026 18:55

Copilot AI changed the title ~~JIT: Strip shift amount masking in importer to prevent CSE interference~~ JIT: Strip shift amount masking via VN to prevent CSE interference Apr 12, 2026

Copilot finished work on behalf of EgorBo April 12, 2026 18:58

This was referenced Apr 12, 2026

"We stopped hearing from agent Azure Pipelines 32. Verify the agent machine is running and has a healthy network connection" dotnet/dnceng#1886

Open

Unable to pull image from mcr.microsoft.com #117164

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

JIT: Strip shift amount masking via VN to prevent CSE interference#126801

JIT: Strip shift amount masking via VN to prevent CSE interference#126801
Copilot wants to merge 5 commits intomainfrom
copilot/jit-strip-masking-shiftamount

Copilot AI commented Apr 11, 2026 •

edited

Loading

Uh oh!

EgorBo commented Apr 12, 2026

Uh oh!

Copilot AI commented Apr 12, 2026

Uh oh!

EgorBo commented Apr 12, 2026

Uh oh!

tannergooding Apr 12, 2026 •

edited

Loading

Uh oh!

EgorBo Apr 12, 2026

Uh oh!

Copilot AI commented Apr 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

Copilot AI commented Apr 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Changes

Uh oh!

EgorBo commented Apr 12, 2026

Uh oh!

Copilot AI commented Apr 12, 2026

Uh oh!

EgorBo commented Apr 12, 2026

Uh oh!

tannergooding Apr 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

EgorBo Apr 12, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI commented Apr 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Copilot AI commented Apr 11, 2026 •

edited

Loading

tannergooding Apr 12, 2026 •

edited

Loading