[llvm] [RISCV] AddEdge between first mask producer and user of V0 (PR #146855)
via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 3 03:51:35 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-risc-v
Author: Liao Chunyu (ChunyuLiao)
<details>
<summary>Changes</summary>
If two mask virtual registers have overlapping live ranges, may a move(vmv* v0, vx) will be generated to save the second mask. By moving the first mask's producer after the mask's use, the spill can be eliminated, and the move can be removed.
Try to remove vmv1r: https://gcc.godbolt.org/z/zbsWvfWYW
Before this patch, the loop body:
sub a4, a2, a3
sh3add a5, a3, a1
vsetvli a4, a4, e64, m2, ta, ma
vle64.v v12, (a5)
sh3add a5, a3, a0
vadd.vi v14, v12, -9
vmsle.vi v0, v12, -1
vmsltu.vx v8, v14, a6
vmerge.vim v14, v10, 9, v0
vmv1r.v v0, v8
vmerge.vvm v8, v14, v12, v0
add a3, a3, a4
vse64.v v8, (a5)
bne a3, a2, .LBB0_5
After this patch:
sub a4, a2, a3
sh3add a5, a3, a1
vsetvli a4, a4, e64, m2, ta, ma
vle64.v v10, (a5)
vmsle.vi v0, v10, -1
vadd.vi v12, v10, -9
vmerge.vim v14, v8, 9, v0
vmsltu.vx v0, v12, a6
sh3add a5, a3, a0
vmerge.vvm v10, v14, v10, v0
add a3, a3, a4
vse64.v v10, (a5)
bne a3, a2, .LBB0_5
---
Full diff: https://github.com/llvm/llvm-project/pull/146855.diff
2 Files Affected:
- (modified) llvm/lib/Target/RISCV/RISCVVectorMaskDAGMutation.cpp (+23-1)
- (modified) llvm/test/CodeGen/RISCV/rvv/vscale-vw-web-simplification.ll (+72-80)
``````````diff
diff --git a/llvm/lib/Target/RISCV/RISCVVectorMaskDAGMutation.cpp b/llvm/lib/Target/RISCV/RISCVVectorMaskDAGMutation.cpp
index be54a8c95a978..96430fb2cd1e6 100644
--- a/llvm/lib/Target/RISCV/RISCVVectorMaskDAGMutation.cpp
+++ b/llvm/lib/Target/RISCV/RISCVVectorMaskDAGMutation.cpp
@@ -68,11 +68,33 @@ class RISCVVectorMaskDAGMutation : public ScheduleDAGMutation {
void apply(ScheduleDAGInstrs *DAG) override {
SUnit *NearestUseV0SU = nullptr;
+ SmallVector<SUnit *, 8> DefMask;
for (SUnit &SU : DAG->SUnits) {
const MachineInstr *MI = SU.getInstr();
- if (MI->findRegisterUseOperand(RISCV::V0, TRI))
+ if (RISCVII::getLMul(MI->getDesc().TSFlags) != RISCVVType::LMUL_8 &&
+ isSoleUseCopyToV0(SU))
+ DefMask.push_back(&SU);
+
+ if (MI->findRegisterUseOperand(RISCV::V0, TRI)) {
NearestUseV0SU = &SU;
+ if (DefMask.size() > 1 && !MI->isCopy()) {
+ // Copy may not be a real use, so skip it here.
+ SUnit *FirstDefV0SU = DefMask[0];
+ SUnit *SecondDefV0SU = DefMask[1];
+ Register FirstVReg = FirstDefV0SU->getInstr()->getOperand(0).getReg();
+ Register SecondVReg =
+ SecondDefV0SU->getInstr()->getOperand(0).getReg();
+ LiveIntervals *LIS = static_cast<ScheduleDAGMILive *>(DAG)->getLIS();
+ LiveInterval &FirstLI = LIS->getInterval(FirstVReg);
+ LiveInterval &SecondLI = LIS->getInterval(SecondVReg);
+ if (FirstLI.overlaps(SecondLI))
+ DAG->addEdge(FirstDefV0SU, SDep(&SU, SDep::Artificial));
+ }
+ if (DefMask.size() > 0)
+ DefMask.erase(DefMask.begin());
+ }
+
if (NearestUseV0SU && NearestUseV0SU != &SU && isSoleUseCopyToV0(SU) &&
// For LMUL=8 cases, there will be more possibilities to spill.
// FIXME: We should use RegPressureTracker to do fine-grained
diff --git a/llvm/test/CodeGen/RISCV/rvv/vscale-vw-web-simplification.ll b/llvm/test/CodeGen/RISCV/rvv/vscale-vw-web-simplification.ll
index 206838917d004..28d6e631d524d 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vscale-vw-web-simplification.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vscale-vw-web-simplification.ll
@@ -153,20 +153,19 @@ define <vscale x 2 x i32> @vwop_vscale_sext_i1i32_multiple_users(ptr %x, ptr %y,
; NO_FOLDING: # %bb.0:
; NO_FOLDING-NEXT: vsetvli a3, zero, e32, m1, ta, mu
; NO_FOLDING-NEXT: vlm.v v8, (a0)
-; NO_FOLDING-NEXT: vlm.v v9, (a1)
-; NO_FOLDING-NEXT: vlm.v v10, (a2)
-; NO_FOLDING-NEXT: vmv.v.i v11, 0
+; NO_FOLDING-NEXT: vmv.v.i v10, 0
; NO_FOLDING-NEXT: vmv.v.v v0, v8
-; NO_FOLDING-NEXT: vmerge.vim v12, v11, -1, v0
+; NO_FOLDING-NEXT: vmerge.vim v11, v10, -1, v0
+; NO_FOLDING-NEXT: vlm.v v0, (a1)
+; NO_FOLDING-NEXT: vlm.v v9, (a2)
+; NO_FOLDING-NEXT: vmerge.vim v12, v10, -1, v0
; NO_FOLDING-NEXT: vmv.v.v v0, v9
-; NO_FOLDING-NEXT: vmerge.vim v9, v11, -1, v0
-; NO_FOLDING-NEXT: vmv.v.v v0, v10
-; NO_FOLDING-NEXT: vmerge.vim v10, v11, -1, v0
-; NO_FOLDING-NEXT: vmul.vv v9, v12, v9
-; NO_FOLDING-NEXT: vsub.vv v11, v12, v10
+; NO_FOLDING-NEXT: vmerge.vim v9, v10, -1, v0
+; NO_FOLDING-NEXT: vmul.vv v10, v11, v12
+; NO_FOLDING-NEXT: vsub.vv v11, v11, v9
; NO_FOLDING-NEXT: vmv.v.v v0, v8
-; NO_FOLDING-NEXT: vadd.vi v10, v10, -1, v0.t
-; NO_FOLDING-NEXT: vor.vv v8, v9, v10
+; NO_FOLDING-NEXT: vadd.vi v9, v9, -1, v0.t
+; NO_FOLDING-NEXT: vor.vv v8, v10, v9
; NO_FOLDING-NEXT: vor.vv v8, v8, v11
; NO_FOLDING-NEXT: ret
;
@@ -174,20 +173,19 @@ define <vscale x 2 x i32> @vwop_vscale_sext_i1i32_multiple_users(ptr %x, ptr %y,
; FOLDING: # %bb.0:
; FOLDING-NEXT: vsetvli a3, zero, e32, m1, ta, mu
; FOLDING-NEXT: vlm.v v8, (a0)
-; FOLDING-NEXT: vlm.v v9, (a1)
-; FOLDING-NEXT: vlm.v v10, (a2)
-; FOLDING-NEXT: vmv.v.i v11, 0
+; FOLDING-NEXT: vmv.v.i v10, 0
; FOLDING-NEXT: vmv.v.v v0, v8
-; FOLDING-NEXT: vmerge.vim v12, v11, -1, v0
+; FOLDING-NEXT: vmerge.vim v11, v10, -1, v0
+; FOLDING-NEXT: vlm.v v0, (a1)
+; FOLDING-NEXT: vlm.v v9, (a2)
+; FOLDING-NEXT: vmerge.vim v12, v10, -1, v0
; FOLDING-NEXT: vmv.v.v v0, v9
-; FOLDING-NEXT: vmerge.vim v9, v11, -1, v0
-; FOLDING-NEXT: vmv.v.v v0, v10
-; FOLDING-NEXT: vmerge.vim v10, v11, -1, v0
-; FOLDING-NEXT: vmul.vv v9, v12, v9
-; FOLDING-NEXT: vsub.vv v11, v12, v10
+; FOLDING-NEXT: vmerge.vim v9, v10, -1, v0
+; FOLDING-NEXT: vmul.vv v10, v11, v12
+; FOLDING-NEXT: vsub.vv v11, v11, v9
; FOLDING-NEXT: vmv.v.v v0, v8
-; FOLDING-NEXT: vadd.vi v10, v10, -1, v0.t
-; FOLDING-NEXT: vor.vv v8, v9, v10
+; FOLDING-NEXT: vadd.vi v9, v9, -1, v0.t
+; FOLDING-NEXT: vor.vv v8, v10, v9
; FOLDING-NEXT: vor.vv v8, v8, v11
; FOLDING-NEXT: ret
%a = load <vscale x 2 x i1>, ptr %x
@@ -209,20 +207,19 @@ define <vscale x 2 x i8> @vwop_vscale_sext_i1i8_multiple_users(ptr %x, ptr %y, p
; NO_FOLDING: # %bb.0:
; NO_FOLDING-NEXT: vsetvli a3, zero, e8, mf4, ta, mu
; NO_FOLDING-NEXT: vlm.v v8, (a0)
-; NO_FOLDING-NEXT: vlm.v v9, (a1)
-; NO_FOLDING-NEXT: vlm.v v10, (a2)
-; NO_FOLDING-NEXT: vmv.v.i v11, 0
+; NO_FOLDING-NEXT: vmv.v.i v10, 0
; NO_FOLDING-NEXT: vmv1r.v v0, v8
-; NO_FOLDING-NEXT: vmerge.vim v12, v11, -1, v0
+; NO_FOLDING-NEXT: vmerge.vim v11, v10, -1, v0
+; NO_FOLDING-NEXT: vlm.v v0, (a1)
+; NO_FOLDING-NEXT: vlm.v v9, (a2)
+; NO_FOLDING-NEXT: vmerge.vim v12, v10, -1, v0
; NO_FOLDING-NEXT: vmv1r.v v0, v9
-; NO_FOLDING-NEXT: vmerge.vim v9, v11, -1, v0
-; NO_FOLDING-NEXT: vmv1r.v v0, v10
-; NO_FOLDING-NEXT: vmerge.vim v10, v11, -1, v0
-; NO_FOLDING-NEXT: vmul.vv v9, v12, v9
-; NO_FOLDING-NEXT: vsub.vv v11, v12, v10
+; NO_FOLDING-NEXT: vmerge.vim v9, v10, -1, v0
+; NO_FOLDING-NEXT: vmul.vv v10, v11, v12
+; NO_FOLDING-NEXT: vsub.vv v11, v11, v9
; NO_FOLDING-NEXT: vmv1r.v v0, v8
-; NO_FOLDING-NEXT: vadd.vi v10, v10, -1, v0.t
-; NO_FOLDING-NEXT: vor.vv v8, v9, v10
+; NO_FOLDING-NEXT: vadd.vi v9, v9, -1, v0.t
+; NO_FOLDING-NEXT: vor.vv v8, v10, v9
; NO_FOLDING-NEXT: vor.vv v8, v8, v11
; NO_FOLDING-NEXT: ret
;
@@ -230,20 +227,19 @@ define <vscale x 2 x i8> @vwop_vscale_sext_i1i8_multiple_users(ptr %x, ptr %y, p
; FOLDING: # %bb.0:
; FOLDING-NEXT: vsetvli a3, zero, e8, mf4, ta, mu
; FOLDING-NEXT: vlm.v v8, (a0)
-; FOLDING-NEXT: vlm.v v9, (a1)
-; FOLDING-NEXT: vlm.v v10, (a2)
-; FOLDING-NEXT: vmv.v.i v11, 0
+; FOLDING-NEXT: vmv.v.i v10, 0
; FOLDING-NEXT: vmv1r.v v0, v8
-; FOLDING-NEXT: vmerge.vim v12, v11, -1, v0
+; FOLDING-NEXT: vmerge.vim v11, v10, -1, v0
+; FOLDING-NEXT: vlm.v v0, (a1)
+; FOLDING-NEXT: vlm.v v9, (a2)
+; FOLDING-NEXT: vmerge.vim v12, v10, -1, v0
; FOLDING-NEXT: vmv1r.v v0, v9
-; FOLDING-NEXT: vmerge.vim v9, v11, -1, v0
-; FOLDING-NEXT: vmv1r.v v0, v10
-; FOLDING-NEXT: vmerge.vim v10, v11, -1, v0
-; FOLDING-NEXT: vmul.vv v9, v12, v9
-; FOLDING-NEXT: vsub.vv v11, v12, v10
+; FOLDING-NEXT: vmerge.vim v9, v10, -1, v0
+; FOLDING-NEXT: vmul.vv v10, v11, v12
+; FOLDING-NEXT: vsub.vv v11, v11, v9
; FOLDING-NEXT: vmv1r.v v0, v8
-; FOLDING-NEXT: vadd.vi v10, v10, -1, v0.t
-; FOLDING-NEXT: vor.vv v8, v9, v10
+; FOLDING-NEXT: vadd.vi v9, v9, -1, v0.t
+; FOLDING-NEXT: vor.vv v8, v10, v9
; FOLDING-NEXT: vor.vv v8, v8, v11
; FOLDING-NEXT: ret
%a = load <vscale x 2 x i1>, ptr %x
@@ -445,16 +441,15 @@ define <vscale x 2 x i32> @vwop_vscale_zext_i1i32_multiple_users(ptr %x, ptr %y,
; NO_FOLDING-NEXT: vsetvli a3, zero, e32, m1, ta, mu
; NO_FOLDING-NEXT: vlm.v v0, (a0)
; NO_FOLDING-NEXT: vlm.v v8, (a2)
-; NO_FOLDING-NEXT: vlm.v v9, (a1)
-; NO_FOLDING-NEXT: vmv.v.i v10, 0
-; NO_FOLDING-NEXT: vmerge.vim v11, v10, 1, v0
+; NO_FOLDING-NEXT: vmv.v.i v9, 0
+; NO_FOLDING-NEXT: vmerge.vim v10, v9, 1, v0
; NO_FOLDING-NEXT: vmv.v.v v0, v8
-; NO_FOLDING-NEXT: vmerge.vim v8, v10, 1, v0
-; NO_FOLDING-NEXT: vadd.vv v10, v11, v8
-; NO_FOLDING-NEXT: vsub.vv v8, v11, v8
-; NO_FOLDING-NEXT: vmv.v.v v0, v9
-; NO_FOLDING-NEXT: vor.vv v10, v10, v11, v0.t
-; NO_FOLDING-NEXT: vor.vv v8, v10, v8
+; NO_FOLDING-NEXT: vmerge.vim v8, v9, 1, v0
+; NO_FOLDING-NEXT: vlm.v v0, (a1)
+; NO_FOLDING-NEXT: vadd.vv v9, v10, v8
+; NO_FOLDING-NEXT: vsub.vv v8, v10, v8
+; NO_FOLDING-NEXT: vor.vv v9, v9, v10, v0.t
+; NO_FOLDING-NEXT: vor.vv v8, v9, v8
; NO_FOLDING-NEXT: ret
;
; FOLDING-LABEL: vwop_vscale_zext_i1i32_multiple_users:
@@ -462,16 +457,15 @@ define <vscale x 2 x i32> @vwop_vscale_zext_i1i32_multiple_users(ptr %x, ptr %y,
; FOLDING-NEXT: vsetvli a3, zero, e32, m1, ta, mu
; FOLDING-NEXT: vlm.v v0, (a0)
; FOLDING-NEXT: vlm.v v8, (a2)
-; FOLDING-NEXT: vlm.v v9, (a1)
-; FOLDING-NEXT: vmv.v.i v10, 0
-; FOLDING-NEXT: vmerge.vim v11, v10, 1, v0
+; FOLDING-NEXT: vmv.v.i v9, 0
+; FOLDING-NEXT: vmerge.vim v10, v9, 1, v0
; FOLDING-NEXT: vmv.v.v v0, v8
-; FOLDING-NEXT: vmerge.vim v8, v10, 1, v0
-; FOLDING-NEXT: vadd.vv v10, v11, v8
-; FOLDING-NEXT: vsub.vv v8, v11, v8
-; FOLDING-NEXT: vmv.v.v v0, v9
-; FOLDING-NEXT: vor.vv v10, v10, v11, v0.t
-; FOLDING-NEXT: vor.vv v8, v10, v8
+; FOLDING-NEXT: vmerge.vim v8, v9, 1, v0
+; FOLDING-NEXT: vlm.v v0, (a1)
+; FOLDING-NEXT: vadd.vv v9, v10, v8
+; FOLDING-NEXT: vsub.vv v8, v10, v8
+; FOLDING-NEXT: vor.vv v9, v9, v10, v0.t
+; FOLDING-NEXT: vor.vv v8, v9, v8
; FOLDING-NEXT: ret
%a = load <vscale x 2 x i1>, ptr %x
%b = load <vscale x 2 x i1>, ptr %y
@@ -493,16 +487,15 @@ define <vscale x 2 x i8> @vwop_vscale_zext_i1i8_multiple_users(ptr %x, ptr %y, p
; NO_FOLDING-NEXT: vsetvli a3, zero, e8, mf4, ta, mu
; NO_FOLDING-NEXT: vlm.v v0, (a0)
; NO_FOLDING-NEXT: vlm.v v8, (a2)
-; NO_FOLDING-NEXT: vlm.v v9, (a1)
-; NO_FOLDING-NEXT: vmv.v.i v10, 0
-; NO_FOLDING-NEXT: vmerge.vim v11, v10, 1, v0
+; NO_FOLDING-NEXT: vmv.v.i v9, 0
+; NO_FOLDING-NEXT: vmerge.vim v10, v9, 1, v0
; NO_FOLDING-NEXT: vmv1r.v v0, v8
-; NO_FOLDING-NEXT: vmerge.vim v8, v10, 1, v0
-; NO_FOLDING-NEXT: vadd.vv v10, v11, v8
-; NO_FOLDING-NEXT: vsub.vv v8, v11, v8
-; NO_FOLDING-NEXT: vmv1r.v v0, v9
-; NO_FOLDING-NEXT: vor.vv v10, v10, v11, v0.t
-; NO_FOLDING-NEXT: vor.vv v8, v10, v8
+; NO_FOLDING-NEXT: vmerge.vim v8, v9, 1, v0
+; NO_FOLDING-NEXT: vlm.v v0, (a1)
+; NO_FOLDING-NEXT: vadd.vv v9, v10, v8
+; NO_FOLDING-NEXT: vsub.vv v8, v10, v8
+; NO_FOLDING-NEXT: vor.vv v9, v9, v10, v0.t
+; NO_FOLDING-NEXT: vor.vv v8, v9, v8
; NO_FOLDING-NEXT: ret
;
; FOLDING-LABEL: vwop_vscale_zext_i1i8_multiple_users:
@@ -510,16 +503,15 @@ define <vscale x 2 x i8> @vwop_vscale_zext_i1i8_multiple_users(ptr %x, ptr %y, p
; FOLDING-NEXT: vsetvli a3, zero, e8, mf4, ta, mu
; FOLDING-NEXT: vlm.v v0, (a0)
; FOLDING-NEXT: vlm.v v8, (a2)
-; FOLDING-NEXT: vlm.v v9, (a1)
-; FOLDING-NEXT: vmv.v.i v10, 0
-; FOLDING-NEXT: vmerge.vim v11, v10, 1, v0
+; FOLDING-NEXT: vmv.v.i v9, 0
+; FOLDING-NEXT: vmerge.vim v10, v9, 1, v0
; FOLDING-NEXT: vmv1r.v v0, v8
-; FOLDING-NEXT: vmerge.vim v8, v10, 1, v0
-; FOLDING-NEXT: vadd.vv v10, v11, v8
-; FOLDING-NEXT: vsub.vv v8, v11, v8
-; FOLDING-NEXT: vmv1r.v v0, v9
-; FOLDING-NEXT: vor.vv v10, v10, v11, v0.t
-; FOLDING-NEXT: vor.vv v8, v10, v8
+; FOLDING-NEXT: vmerge.vim v8, v9, 1, v0
+; FOLDING-NEXT: vlm.v v0, (a1)
+; FOLDING-NEXT: vadd.vv v9, v10, v8
+; FOLDING-NEXT: vsub.vv v8, v10, v8
+; FOLDING-NEXT: vor.vv v9, v9, v10, v0.t
+; FOLDING-NEXT: vor.vv v8, v9, v8
; FOLDING-NEXT: ret
%a = load <vscale x 2 x i1>, ptr %x
%b = load <vscale x 2 x i1>, ptr %y
``````````
</details>
https://github.com/llvm/llvm-project/pull/146855
More information about the llvm-commits
mailing list