[RISCV] Improve cleanup phase of RISCV Insert VSETVLI pass #67144
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
simeonkr
changed the title
|
@llvm/pr-subscribers-backend-risc-v ChangesMake 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:
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]
|
topperc
reviewed
Sep 24, 2023
| @@ -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() && | |||
There was a problem hiding this comment.
MI here can be any instruction right? What if it has more than 1 def?
There was a problem hiding this comment.
Or it might not even be a def. It could be a store instruction that uses the same register?
There was a problem hiding this comment.
To address this, I've tried to make the check as general as possible: 22d8abb.
topperc
reviewed
Sep 24, 2023
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
simeonkr
force-pushed
the
upstream-vsetvli-cleanup
branch
from
15e2a7f
to
22d8abb
Compare
wangpc-pp
reviewed
Sep 26, 2023
| @@ -329,9 +329,9 @@ entry: | |||
| define double @test17(i64 %avl, <vscale x 1 x double> %a, <vscale x 1 x double> %b) nounwind { | |||
| ; CHECK-LABEL: test17: | |||
| ; CHECK: # %bb.0: # %entry | |||
| ; CHECK-NEXT: vsetvli a0, a0, e64, m1, ta, ma | |||
| ; CHECK-NEXT: vfmv.f.s fa5, v8 | |||
| ; CHECK-NEXT: vsetvli a0, a0, e32, mf2, ta, ma | |||
There was a problem hiding this comment.
This can be removed too? It just sets vl, vtype will be changed by vsetvli below.
There was a problem hiding this comment.
This was due to a use by an instruction which was marked for deletion. Unfortunately, to account for such cases I needed to make ToDelete a SmallPtrSet so that I could quickly discard instructions added to this set: e8f2fbe
artagnon
reviewed
Sep 29, 2023
|
Gentle ping |
yetingk
reviewed
Oct 11, 2023
yetingk
reviewed
Oct 11, 2023
added 5 commits
October 12, 2023 11:13
Make the pass better at eliminating redundant VSETVLIs in certain scenarios, as demonstrated in the updated tests.
simeonkr
force-pushed
the
upstream-vsetvli-cleanup
branch
from
702b488
to
1c1b326
Compare
zhongjuzhe
reviewed
Oct 13, 2023
| @@ -834,8 +834,7 @@ define <vscale x 2 x i32> @pre_lmul(<vscale x 2 x i32> %x, <vscale x 2 x i32> %y | |||
| ; CHECK-LABEL: pre_lmul: | |||
| ; CHECK: # %bb.0: # %entry | |||
| ; CHECK-NEXT: andi a0, a0, 1 | |||
| ; CHECK-NEXT: vsetvli a1, zero, e64, m1, ta, ma | |||
| ; CHECK-NEXT: vsetvli zero, a1, e32, m1, ta, ma | |||
| ; CHECK-NEXT: vsetvli a1, zero, e32, m1, ta, ma | |||
| ; CHECK-NEXT: vadd.vv v8, v8, v9 | |||
There was a problem hiding this comment.
It seems odd.
According to the LLVM IR, is should be processing VLMAX number of element of SEW = 64, LMUL = 1.
Suppose RVV VLEN = 128bit:
SEW = 64, LMUL = 1, the vsetvlmax output should be 2 elements.
SEW = 32, LMUL = 1, the vsetvlmax output should be 4 elements.
According to LLVM IR:
%vl = tail call i64 @llvm.riscv.vsetvlimax.i64(i64 3, i64 0)
%a = call <vscale x 2 x i32> @llvm.riscv.vadd.nxv2i32(<vscale x 2 x i32> undef, <vscale x 2 x i32> %x, <vscale x 2 x i32> %y, i64 %vl)
%vl output should be 2 elements, then later "vadd" should be processing 2 elements.
Here the new assembly after this patch is:
; CHECK-NEXT: vsetvli a1, zero, e32, m1, ta, ma
; CHECK-NEXT: vadd.vv v8, v8, v9
The "vadd.vv" is processing 4 elements.
There was a problem hiding this comment.
What is the purpose of vsetvli zero, a1, e32, m1, ta, ma?
There was a problem hiding this comment.
Good catch! I think it is a wrong optimization here.
What is the purpose of
vsetvli zero, a1, e32, m1, ta, ma?
It sets vl to a1 (result of %vl = tail call i64 @llvm.riscv.vsetvlimax.i64(i64 3, i64 0)) and changes vtype to e32, m1, which matches semantics of LLVM IR.
There was a problem hiding this comment.
Should the optimization therefore be the single instruction vsetvli a1, zero, e64, m1, ta, ma?
There was a problem hiding this comment.
I don't think so.
We need vl to be VLMAX of e64, m1 and vtype to be e32, m1 (so the final vl will be half of the VLMAX of e32, m1). Optimization may not be feasible here.
added 2 commits
October 19, 2023 16:39
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Make the pass better at eliminating redundant VSETVLIs in certain scenarios, as demonstrated in the updated tests.