[llvm] [RISCV] Improve cleanup phase of RISCV Insert VSETVLI pass (PR #67144)
via llvm-commits
llvm-commits at lists.llvm.org
Sat Sep 23 05:10:40 PDT 2023
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-risc-v
<details>
<summary>Changes</summary>
Make the pass better at eliminating redundant VSETVLIs in certain scenarios, as demonstrated in the updated tests.
---
Patch is 54.90 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/67144.diff
5 Files Affected:
- (modified) llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp (+56-55)
- (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll (+84-168)
- (modified) llvm/test/CodeGen/RISCV/rvv/vsetvli-insert-crossbb.ll (+1-2)
- (modified) llvm/test/CodeGen/RISCV/rvv/vsetvli-insert.ll (+2-6)
- (modified) llvm/test/CodeGen/RISCV/rvv/vsetvli-intrinsics.ll (+15-3)
``````````diff
diff --git a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
index 5584fa8d503dbe4..138c28e4effabf6 100644
--- a/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp
@@ -1426,43 +1426,6 @@ static bool isNonZeroAVL(const MachineOperand &MO) {
return 0 != MO.getImm();
}
-// Return true if we can mutate PrevMI to match MI without changing any the
-// fields which would be observed.
-static bool canMutatePriorConfig(const MachineInstr &PrevMI,
- const MachineInstr &MI,
- const DemandedFields &Used) {
- // If the VL values aren't equal, return false if either a) the former is
- // demanded, or b) we can't rewrite the former to be the later for
- // implementation reasons.
- if (!isVLPreservingConfig(MI)) {
- if (Used.VLAny)
- return false;
-
- // We don't bother to handle the equally zero case here as it's largely
- // uninteresting.
- if (Used.VLZeroness) {
- if (isVLPreservingConfig(PrevMI))
- return false;
- if (!isNonZeroAVL(MI.getOperand(1)) ||
- !isNonZeroAVL(PrevMI.getOperand(1)))
- return false;
- }
-
- // TODO: Track whether the register is defined between
- // PrevMI and MI.
- if (MI.getOperand(1).isReg() &&
- RISCV::X0 != MI.getOperand(1).getReg())
- return false;
- }
-
- if (!PrevMI.getOperand(2).isImm() || !MI.getOperand(2).isImm())
- return false;
-
- auto PriorVType = PrevMI.getOperand(2).getImm();
- auto VType = MI.getOperand(2).getImm();
- return areCompatibleVTYPEs(PriorVType, VType, Used);
-}
-
void RISCVInsertVSETVLI::doLocalPostpass(MachineBasicBlock &MBB) {
MachineInstr *NextMI = nullptr;
// We can have arbitrary code in successors, so VL and VTYPE
@@ -1475,36 +1438,74 @@ void RISCVInsertVSETVLI::doLocalPostpass(MachineBasicBlock &MBB) {
if (!isVectorConfigInstr(MI)) {
doUnion(Used, getDemanded(MI, MRI));
+ // We can't handle the case when the source AVL
+ // register of *NextMI is defined after MI
+ if (NextMI && NextMI->getOperand(1).isReg() && MI.getOperand(0).isReg() &&
+ MI.getOperand(0).getReg() == NextMI->getOperand(1).getReg())
+ Used.demandVL();
continue;
}
Register VRegDef = MI.getOperand(0).getReg();
if (VRegDef != RISCV::X0 &&
- !(VRegDef.isVirtual() && MRI->use_nodbg_empty(VRegDef)))
- Used.demandVL();
+ !(VRegDef.isVirtual() && MRI->use_nodbg_empty(VRegDef))) {
+ for (MachineInstr &UserMI : MRI->use_nodbg_instructions(VRegDef)) {
+ // ignore uses by *NextMI since we will deal with them explicitly
+ if (!NextMI || !UserMI.isIdenticalTo(*NextMI)) {
+ Used.demandVL();
+ break;
+ }
+ }
+ }
if (NextMI) {
- if (!Used.usedVL() && !Used.usedVTYPE()) {
- ToDelete.push_back(&MI);
- // Leave NextMI unchanged
- continue;
- } else if (canMutatePriorConfig(MI, *NextMI, Used)) {
- if (!isVLPreservingConfig(*NextMI)) {
- MI.getOperand(0).setReg(NextMI->getOperand(0).getReg());
- MI.getOperand(0).setIsDead(false);
- if (NextMI->getOperand(1).isImm())
- MI.getOperand(1).ChangeToImmediate(NextMI->getOperand(1).getImm());
- else
- MI.getOperand(1).ChangeToRegister(NextMI->getOperand(1).getReg(), false);
- MI.setDesc(NextMI->getDesc());
+ auto NextVRegDef = NextMI->getOperand(0).getReg();
+ auto &AVLOp = MI.getOperand(1), &NextAVLOp = NextMI->getOperand(1);
+ auto &VLTypeOp = MI.getOperand(2), &NextVLTypeOp = NextMI->getOperand(2);
+ if (areCompatibleVTYPEs(VLTypeOp.getImm(), NextVLTypeOp.getImm(), Used)) {
+ bool VLUsed =
+ Used.VLAny || (Used.VLZeroness &&
+ (!isNonZeroAVL(AVLOp) || !isNonZeroAVL(NextAVLOp)));
+ if (isVLPreservingConfig(*NextMI) || !VLUsed) {
+ // set VLType of MI to that of *NextMI
+ VLTypeOp.setImm(NextVLTypeOp.getImm());
+ // remove *NextMI and potentially replace its uses with MI
+ ToDelete.push_back(NextMI);
+ if (NextVRegDef != RISCV::X0 && !MRI->use_nodbg_empty(NextVRegDef)) {
+ if (VRegDef != RISCV::X0)
+ MRI->replaceRegWith(NextVRegDef, VRegDef);
+ else {
+ MI.getOperand(0).ChangeToRegister(
+ MRI->createVirtualRegister(&RISCV::GPRRegClass), true);
+ MRI->replaceRegWith(NextVRegDef, MI.getOperand(0).getReg());
+ }
+ }
+ // is the AVL op of *NextMI a non-zero reg storing the result of MI?
+ bool KeepAVL = NextAVLOp.isReg() && NextAVLOp.getReg() == VRegDef &&
+ VRegDef != RISCV::X0;
+ // set AVL of MI to that of *NextMI whenever it makes sense to do so
+ if (!isVLPreservingConfig(*NextMI) && !KeepAVL) {
+ if (NextAVLOp.isImm())
+ AVLOp.ChangeToImmediate(NextAVLOp.getImm());
+ else {
+ AVLOp.ChangeToRegister(NextAVLOp.getReg(), false);
+ // ensure that *MI did not inadvertently become VL-preserving
+ if (VRegDef == RISCV::X0 && AVLOp.getReg() == RISCV::X0)
+ MI.getOperand(0).ChangeToRegister(
+ MRI->createVirtualRegister(&RISCV::GPRRegClass), true);
+ }
+ MI.setDesc(NextMI->getDesc());
+ }
}
- MI.getOperand(2).setImm(NextMI->getOperand(2).getImm());
- ToDelete.push_back(NextMI);
- // fallthrough
}
}
NextMI = &MI;
Used = getDemanded(MI, MRI);
+ // ignore such uses since we can handle them explicitly
+ if (VRegDef == RISCV::X0) {
+ Used.VLAny = false;
+ Used.VLZeroness = false;
+ }
}
for (auto *MI : ToDelete)
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
index 49724cbc4418252..1bb197c84799b83 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-masked-gather.ll
@@ -2530,9 +2530,8 @@ define <8 x i32> @mgather_baseidx_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8 x i1>
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 2, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 1
; RV64ZVE32F-NEXT: .LBB35_4: # %else2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e8, mf2, ta, ma
@@ -2557,9 +2556,8 @@ define <8 x i32> @mgather_baseidx_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8 x i1>
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vsetivli zero, 6, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 5
; RV64ZVE32F-NEXT: .LBB35_9: # %else14
; RV64ZVE32F-NEXT: vsetivli zero, 2, e8, mf4, ta, ma
@@ -2577,9 +2575,8 @@ define <8 x i32> @mgather_baseidx_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8 x i1>
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 3, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 2
; RV64ZVE32F-NEXT: andi a2, a1, 8
; RV64ZVE32F-NEXT: beqz a2, .LBB35_6
@@ -2590,9 +2587,8 @@ define <8 x i32> @mgather_baseidx_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8 x i1>
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 3
; RV64ZVE32F-NEXT: andi a2, a1, 16
; RV64ZVE32F-NEXT: beqz a2, .LBB35_7
@@ -2613,9 +2609,8 @@ define <8 x i32> @mgather_baseidx_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8 x i1>
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 7, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 6
; RV64ZVE32F-NEXT: andi a1, a1, -128
; RV64ZVE32F-NEXT: beqz a1, .LBB35_11
@@ -2626,9 +2621,8 @@ define <8 x i32> @mgather_baseidx_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8 x i1>
; RV64ZVE32F-NEXT: slli a1, a1, 2
; RV64ZVE32F-NEXT: add a0, a0, a1
; RV64ZVE32F-NEXT: lw a0, 0(a0)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 7
; RV64ZVE32F-NEXT: vmv2r.v v8, v10
; RV64ZVE32F-NEXT: ret
@@ -2680,9 +2674,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 2, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 1
; RV64ZVE32F-NEXT: .LBB36_4: # %else2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e8, mf2, ta, ma
@@ -2707,9 +2700,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vsetivli zero, 6, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 5
; RV64ZVE32F-NEXT: .LBB36_9: # %else14
; RV64ZVE32F-NEXT: vsetivli zero, 2, e8, mf4, ta, ma
@@ -2727,9 +2719,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 3, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 2
; RV64ZVE32F-NEXT: andi a2, a1, 8
; RV64ZVE32F-NEXT: beqz a2, .LBB36_6
@@ -2740,9 +2731,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 3
; RV64ZVE32F-NEXT: andi a2, a1, 16
; RV64ZVE32F-NEXT: beqz a2, .LBB36_7
@@ -2763,9 +2753,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 7, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 6
; RV64ZVE32F-NEXT: andi a1, a1, -128
; RV64ZVE32F-NEXT: beqz a1, .LBB36_11
@@ -2776,9 +2765,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a1, a1, 2
; RV64ZVE32F-NEXT: add a0, a0, a1
; RV64ZVE32F-NEXT: lw a0, 0(a0)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 7
; RV64ZVE32F-NEXT: vmv2r.v v8, v10
; RV64ZVE32F-NEXT: ret
@@ -2834,9 +2822,8 @@ define <8 x i32> @mgather_baseidx_zext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 2, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 1
; RV64ZVE32F-NEXT: .LBB37_4: # %else2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e8, mf2, ta, ma
@@ -2862,9 +2849,8 @@ define <8 x i32> @mgather_baseidx_zext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vsetivli zero, 6, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 5
; RV64ZVE32F-NEXT: .LBB37_9: # %else14
; RV64ZVE32F-NEXT: vsetivli zero, 2, e8, mf4, ta, ma
@@ -2883,9 +2869,8 @@ define <8 x i32> @mgather_baseidx_zext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 3, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 2
; RV64ZVE32F-NEXT: andi a2, a1, 8
; RV64ZVE32F-NEXT: beqz a2, .LBB37_6
@@ -2897,9 +2882,8 @@ define <8 x i32> @mgather_baseidx_zext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 3
; RV64ZVE32F-NEXT: andi a2, a1, 16
; RV64ZVE32F-NEXT: beqz a2, .LBB37_7
@@ -2922,9 +2906,8 @@ define <8 x i32> @mgather_baseidx_zext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 7, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 6
; RV64ZVE32F-NEXT: andi a1, a1, -128
; RV64ZVE32F-NEXT: beqz a1, .LBB37_11
@@ -2936,9 +2919,8 @@ define <8 x i32> @mgather_baseidx_zext_v8i8_v8i32(ptr %base, <8 x i8> %idxs, <8
; RV64ZVE32F-NEXT: slli a1, a1, 2
; RV64ZVE32F-NEXT: add a0, a0, a1
; RV64ZVE32F-NEXT: lw a0, 0(a0)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 7
; RV64ZVE32F-NEXT: vmv2r.v v8, v10
; RV64ZVE32F-NEXT: ret
@@ -2992,9 +2974,8 @@ define <8 x i32> @mgather_baseidx_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <8 x i
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 2, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 1
; RV64ZVE32F-NEXT: .LBB38_4: # %else2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e16, m1, ta, ma
@@ -3019,9 +3000,8 @@ define <8 x i32> @mgather_baseidx_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <8 x i
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vsetivli zero, 6, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 5
; RV64ZVE32F-NEXT: .LBB38_9: # %else14
; RV64ZVE32F-NEXT: vsetivli zero, 2, e16, mf2, ta, ma
@@ -3039,9 +3019,8 @@ define <8 x i32> @mgather_baseidx_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <8 x i
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 3, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 2
; RV64ZVE32F-NEXT: andi a2, a1, 8
; RV64ZVE32F-NEXT: beqz a2, .LBB38_6
@@ -3052,9 +3031,8 @@ define <8 x i32> @mgather_baseidx_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <8 x i
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 3
; RV64ZVE32F-NEXT: andi a2, a1, 16
; RV64ZVE32F-NEXT: beqz a2, .LBB38_7
@@ -3075,9 +3053,8 @@ define <8 x i32> @mgather_baseidx_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <8 x i
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 7, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 6
; RV64ZVE32F-NEXT: andi a1, a1, -128
; RV64ZVE32F-NEXT: beqz a1, .LBB38_11
@@ -3088,9 +3065,8 @@ define <8 x i32> @mgather_baseidx_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <8 x i
; RV64ZVE32F-NEXT: slli a1, a1, 2
; RV64ZVE32F-NEXT: add a0, a0, a1
; RV64ZVE32F-NEXT: lw a0, 0(a0)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vsetivli zero, 8, e32, m2, ta, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a0
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 7
; RV64ZVE32F-NEXT: vmv2r.v v8, v10
; RV64ZVE32F-NEXT: ret
@@ -3143,9 +3119,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vsetivli zero, 2, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v12, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v12, 1
; RV64ZVE32F-NEXT: .LBB39_4: # %else2
; RV64ZVE32F-NEXT: vsetivli zero, 4, e16, m1, ta, ma
@@ -3170,9 +3145,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV64ZVE32F-NEXT: vsetvli zero, zero, e32, m1, ta, ma
-; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vsetivli zero, 6, e32, m2, tu, ma
+; RV64ZVE32F-NEXT: vmv.s.x v8, a2
; RV64ZVE32F-NEXT: vslideup.vi v10, v8, 5
; RV64ZVE32F-NEXT: .LBB39_9: # %else14
; RV64ZVE32F-NEXT: vsetivli zero, 2, e16, mf2, ta, ma
@@ -3190,9 +3164,8 @@ define <8 x i32> @mgather_baseidx_sext_v8i16_v8i32(ptr %base, <8 x i16> %idxs, <
; RV64ZVE32F-NEXT: slli a2, a2, 2
; RV64ZVE32F-NEXT: add a2, a0, a2
; RV64ZVE32F-NEXT: lw a2, 0(a2)
-; RV...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/67144
More information about the llvm-commits
mailing list