[AArch64] Replace AND with LSL#2 for LDR target (#34101) #89531

ParkHanbum · 2024-04-21T05:31:47Z

Currently, process of replacing bitwise operations consisting of
LSR/LSL with And is performed by DAGCombiner.

However, in certain cases, the AND generated by this process
can be removed.

Consider following case:

        lsr x8, x8, #56
        and x8, x8, #0xfc
        ldr w0, [x2, x8]
        ret

In this case, we can remove the AND by changing the target of LDR
to [X2, X8, LSL #2] and right-shifting amount change to 56 to 58.

after changed:

        lsr x8, x8, #58
        ldr w0, [x2, x8, lsl #2]
        ret

This patch checks to see if the SHIFTING + AND operation on load
target can be optimized and optimizes it if it can.

llvmbot · 2024-04-21T05:32:20Z

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-aarch64

Author: hanbeom (ParkHanbum)

Changes

Currently, process of replacing bitwise operations consisting of
LSR/LSL with And is performed by DAGCombiner.

However, in certain cases, the AND generated by this process
can be removed.

Consider following case:

        lsr x8, x8, #<!-- -->56
        and x8, x8, #<!-- -->0xfc
        ldr w0, [x2, x8]
        ret

In this case, we can remove the AND by changing the target of LDR
to [X2, X8, LSL #2] and right-shifting amount change to 56 to 58.

after changed:

        lsr x8, x8, #<!-- -->58
        ldr w0, [x2, x8, lsl #<!-- -->2]
        ret

This patch checks to see if the SHIFTING + AND operation on load
target can be optimized and optimizes it if it can.

Full diff: https://github.com/llvm/llvm-project/pull/89531.diff

2 Files Affected:

(modified) llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp (+62)
(added) llvm/test/CodeGen/AArch64/peephole-load.mir (+160)

diff --git a/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp b/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp
index 22da7ddef98a2a..1c331c88042317 100644
--- a/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp
+++ b/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp
@@ -128,6 +128,7 @@ struct AArch64MIPeepholeOpt : public MachineFunctionPass {
   bool visitINSviGPR(MachineInstr &MI, unsigned Opc);
   bool visitINSvi64lane(MachineInstr &MI);
   bool visitFMOVDr(MachineInstr &MI);
+  bool visitLOAD(MachineInstr &MI);
   bool runOnMachineFunction(MachineFunction &MF) override;
 
   StringRef getPassName() const override {
@@ -690,6 +691,64 @@ bool AArch64MIPeepholeOpt::visitFMOVDr(MachineInstr &MI) {
   return true;
 }
 
+bool AArch64MIPeepholeOpt::visitLOAD(MachineInstr &MI) {
+  Register LdOp2Reg = MI.getOperand(2).getReg();
+  unsigned RegSize = TRI->getRegSizeInBits(LdOp2Reg, *MRI);
+
+  // Consider:
+  // (ldr w, [x, (and x, (ubfm x, x, imms, immr), C1)])
+  // If bitmask C1 of And is all the bits remaining after
+  // bitshifting to UBFM minus last 2 bits, try to optimize.
+  // Optimize to:
+  // (ldr w, [x (ubfm x, x, imms, immr), lsl #2])
+  {
+    if (!MI.getOperand(4).isImm() || MI.getOperand(4).getImm() != 0)
+      return false;
+
+    MachineInstr *AndMI = MRI->getUniqueVRegDef(LdOp2Reg);
+    if (!AndMI || AndMI->getOpcode() != AArch64::ANDXri ||
+        !AndMI->getOperand(2).isImm())
+      return false;
+
+    uint64_t AndMask = AArch64_AM::decodeLogicalImmediate(
+        AndMI->getOperand(2).getImm(), RegSize);
+    MachineInstr *ShtMI = MRI->getUniqueVRegDef(AndMI->getOperand(1).getReg());
+    uint64_t Mask = 0;
+    if (!ShtMI || ShtMI->getOpcode() != AArch64::UBFMXri)
+      return false;
+    uint64_t imms = ShtMI->getOperand(2).getImm();
+    uint64_t immr = ShtMI->getOperand(3).getImm();
+    uint64_t new_imms = 0;
+    uint64_t new_immr = 0;
+    if (imms <= immr) {
+      if (immr != RegSize - 1)
+        return false;
+      Mask = ((uint64_t)1 << (RegSize - imms)) - 4;
+      new_imms = imms+2;
+      new_immr = immr;
+    } else {
+      // we only need to handle case lsl #1
+      if ((imms - immr != 1) || imms != RegSize - 1)
+        return false;
+      Mask = UINT64_MAX - 3;
+      new_imms = 1;
+      new_immr = imms;
+    }
+
+    // check this shifting can be treat as PreIndex Shifting.
+    if (AndMask == Mask) {
+      AndMI->eraseFromParent();
+      ShtMI->getOperand(2).setImm(new_imms);
+      ShtMI->getOperand(3).setImm(new_immr);
+      MI.getOperand(2).setReg(ShtMI->getOperand(0).getReg());
+      MI.getOperand(4).setImm(1);
+      return true;
+    }
+  }
+
+  return false;
+}
+
 bool AArch64MIPeepholeOpt::runOnMachineFunction(MachineFunction &MF) {
   if (skipFunction(MF.getFunction()))
     return false;
@@ -771,6 +830,9 @@ bool AArch64MIPeepholeOpt::runOnMachineFunction(MachineFunction &MF) {
       case AArch64::FMOVDr:
         Changed |= visitFMOVDr(MI);
         break;
+      case AArch64::LDRWroX:
+        Changed |= visitLOAD(MI);
+        break;
       }
     }
   }
diff --git a/llvm/test/CodeGen/AArch64/peephole-load.mir b/llvm/test/CodeGen/AArch64/peephole-load.mir
new file mode 100644
index 00000000000000..8c9555a48997b9
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/peephole-load.mir
@@ -0,0 +1,160 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -run-pass=aarch64-mi-peephole-opt -o - -mtriple=aarch64-unknown-linux -verify-machineinstrs %s | FileCheck %s
+
+---
+name: transform_lsr_and_ldr_to_lsr_ldr2
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x1, $x2
+    ; CHECK-LABEL: name: transform_lsr_and_ldr_to_lsr_ldr2
+    ; CHECK: liveins: $x0, $x1, $x2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[MADDXrrr:%[0-9]+]]:gpr64 = MADDXrrr [[COPY1]], [[COPY2]], $xzr
+    ; CHECK-NEXT: [[UBFMXri:%[0-9]+]]:gpr64 = UBFMXri killed [[MADDXrrr]], 58, 63
+    ; CHECK-NEXT: [[LDRWroX:%[0-9]+]]:gpr32 = LDRWroX [[COPY]], killed [[UBFMXri]], 0, 1
+    ; CHECK-NEXT: $w0 = COPY [[LDRWroX]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $w0
+    %2:gpr64common = COPY $x2
+    %1:gpr64 = COPY $x1
+    %0:gpr64 = COPY $x0
+    %3:gpr64 = MADDXrrr %1, %0, $xzr
+    %4:gpr64 = UBFMXri killed %3, 56, 63
+    %5:gpr64common = ANDXri killed %4, 8069
+    %6:gpr32 = LDRWroX %2, killed %5, 0, 0
+    $w0 = COPY %6
+    RET_ReallyLR implicit $w0
+...
+---
+name: transform_lsl1_and_ldr_to_lsr1_ldr2
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x1, $x2
+    ; CHECK-LABEL: name: transform_lsl1_and_ldr_to_lsr1_ldr2
+    ; CHECK: liveins: $x0, $x1, $x2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[MADDXrrr:%[0-9]+]]:gpr64 = MADDXrrr [[COPY1]], [[COPY2]], $xzr
+    ; CHECK-NEXT: [[UBFMXri:%[0-9]+]]:gpr64 = UBFMXri killed [[MADDXrrr]], 1, 63
+    ; CHECK-NEXT: [[LDRWroX:%[0-9]+]]:gpr32 = LDRWroX [[COPY]], killed [[UBFMXri]], 0, 1
+    ; CHECK-NEXT: $w0 = COPY [[LDRWroX]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $w0
+    %2:gpr64common = COPY $x2
+    %1:gpr64 = COPY $x1
+    %0:gpr64 = COPY $x0
+    %3:gpr64 = MADDXrrr %1, %0, $xzr
+    %4:gpr64 = UBFMXri killed %3, 63, 62
+    %5:gpr64common = ANDXri killed %4, 8125
+    %6:gpr32 = LDRWroX %2, killed %5, 0, 0
+    $w0 = COPY %6
+    RET_ReallyLR implicit $w0
+...
+---
+name: donot_transform_and_ldr
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x1, $x2
+    ; CHECK-LABEL: name: donot_transform_and_ldr
+    ; CHECK: liveins: $x0, $x1, $x2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[MADDXrrr:%[0-9]+]]:gpr64 = MADDXrrr [[COPY1]], [[COPY2]], $xzr
+    ; CHECK-NEXT: [[ANDXri:%[0-9]+]]:gpr64common = ANDXri killed [[MADDXrrr]], 8125
+    ; CHECK-NEXT: [[LDRWroX:%[0-9]+]]:gpr32 = LDRWroX [[COPY]], killed [[ANDXri]], 0, 0
+    ; CHECK-NEXT: $w0 = COPY [[LDRWroX]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $w0
+    %2:gpr64common = COPY $x2
+    %1:gpr64 = COPY $x1
+    %0:gpr64 = COPY $x0
+    %3:gpr64 = MADDXrrr %1, %0, $xzr
+    %4:gpr64common = ANDXri killed %3, 8125
+    %5:gpr32 = LDRWroX %2, killed %4, 0, 0
+    $w0 = COPY %5
+    RET_ReallyLR implicit $w0
+...
+---
+name: donot_transform_if_not_lsl
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x1, $x2
+    ; CHECK-LABEL: name: donot_transform_if_not_lsl
+    ; CHECK: liveins: $x0, $x1, $x2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[MADDXrrr:%[0-9]+]]:gpr64 = MADDXrrr [[COPY1]], [[COPY2]], $xzr
+    ; CHECK-NEXT: [[UBFMXri:%[0-9]+]]:gpr64 = UBFMXri killed [[MADDXrrr]], 64, 62
+    ; CHECK-NEXT: [[ANDXri:%[0-9]+]]:gpr64common = ANDXri killed [[UBFMXri]], 8125
+    ; CHECK-NEXT: [[LDRWroX:%[0-9]+]]:gpr32 = LDRWroX [[COPY]], killed [[ANDXri]], 0, 0
+    ; CHECK-NEXT: $w0 = COPY [[LDRWroX]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $w0
+    %2:gpr64common = COPY $x2
+    %1:gpr64 = COPY $x1
+    %0:gpr64 = COPY $x0
+    %3:gpr64 = MADDXrrr %1, %0, $xzr
+    %4:gpr64 = UBFMXri killed %3, 64, 62
+    %5:gpr64common = ANDXri killed %4, 8125
+    %6:gpr32 = LDRWroX %2, killed %5, 0, 0
+    $w0 = COPY %6
+    RET_ReallyLR implicit $w0
+...
+---
+name: donot_transform_if_not_lsr
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x1, $x2
+    ; CHECK-LABEL: name: donot_transform_if_not_lsr
+    ; CHECK: liveins: $x0, $x1, $x2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x1
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[MADDXrrr:%[0-9]+]]:gpr64 = MADDXrrr [[COPY1]], [[COPY2]], $xzr
+    ; CHECK-NEXT: [[UBFMXri:%[0-9]+]]:gpr64 = UBFMXri killed [[MADDXrrr]], 62, 62
+    ; CHECK-NEXT: [[ANDXri:%[0-9]+]]:gpr64common = ANDXri killed [[UBFMXri]], 8069
+    ; CHECK-NEXT: [[LDRWroX:%[0-9]+]]:gpr32 = LDRWroX [[COPY]], killed [[ANDXri]], 0, 0
+    ; CHECK-NEXT: $w0 = COPY [[LDRWroX]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $w0
+    %2:gpr64common = COPY $x2
+    %1:gpr64 = COPY $x1
+    %0:gpr64 = COPY $x0
+    %3:gpr64 = MADDXrrr %1, %0, $xzr
+    %4:gpr64 = UBFMXri killed %3, 62, 62
+    %5:gpr64common = ANDXri killed %4, 8069
+    %6:gpr32 = LDRWroX %2, killed %5, 0, 0
+    $w0 = COPY %6
+    RET_ReallyLR implicit $w0
+...
+---
+name: donot_transform_if_not_exist_and_and_lsl
+tracksRegLiveness: true
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x2
+    ; CHECK-LABEL: name: donot_transform_if_not_exist_and_and_lsl
+    ; CHECK: liveins: $x0, $x2
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY $x0
+    ; CHECK-NEXT: [[UBFMXri:%[0-9]+]]:gpr64 = UBFMXri [[COPY1]], 61, 60
+    ; CHECK-NEXT: [[LDRWroX:%[0-9]+]]:gpr32 = LDRWroX [[COPY]], killed [[UBFMXri]], 0, 0
+    ; CHECK-NEXT: $w0 = COPY [[LDRWroX]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $w0
+    %2:gpr64common = COPY $x2
+    %0:gpr64 = COPY $x0
+    %3:gpr64 = UBFMXri %0, 61, 60
+    %4:gpr32 = LDRWroX %2, killed %3, 0, 0
+    $w0 = COPY %4
+    RET_ReallyLR implicit $w0

github-actions · 2024-04-21T05:35:28Z

✅ With the latest revision this PR passed the C/C++ code formatter.

davemgreen · 2024-04-22T06:39:06Z

Could this be handled better in DAG combine? Maybe with something similar to reassociationCanBreakAddressingModePattern?

ParkHanbum · 2024-04-22T07:23:17Z

ok. I'll find way

ParkHanbum · 2024-04-22T13:55:33Z

@davemgreen
I wrote code to restore and back to shl in the visit load of DAGCombiner.

GEN Create new node: t26: i64 = shl t23, Constant:i64<2>
GEN : t26: i64 = shl t23, Constant:i64<2>

But in this case, DAG change shl to and again in visitSHL.

Combining: t26: i64 = shl t23, Constant:i64<2>
Creating constant: t27: i64 = Constant<54>
Creating constant: t28: i64 = Constant<-1>
Creating constant: t29: i64 = Constant<-72057594037927936>
Creating constant: t30: i64 = Constant<1020>
Creating new node: t31: i64 = srl t7, Constant:i64<54>
Creating new node: t32: i64 = and t31, Constant:i64<1020>
 ... into: t32: i64 = and t31, Constant:i64<1020>

Do you have any advice on where to implement this?

davemgreen · 2024-04-23T07:59:00Z

Hi. There is code in the DAG combiner to attempt to stop reassociations that would break the addressing mode.
https://github.com/llvm/llvm-project/blob/654846560c8892b32ea1e92c1fffc5e70bf0e75f/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp#L2690
I've not looked into this case in a lot of details, there is a chance the transform happens in a few different places, but could the same thing be done here?

ParkHanbum · 2024-04-24T12:44:36Z

@davemgreen Thanks, I think I'm almost there thanks to your advice, I was thinking I should do it in LOAD. it's my mistake.

Can I ask one more thing?

I can check the legality of the address via TLI.isLegalAddressingMode(), but what function in the TLI can I utilize to check if PreIndex is allowed in architecture?

My current implementation also enforces the following cases in RISCV

define signext i32 @srli_1_sh2add(ptr %0, i64 %1) {
; RV64I-LABEL: srli_1_sh2add:
; RV64I:       # %bb.0:
; RV64I-NEXT:    slli a1, a1, 1
; RV64I-NEXT:    andi a1, a1, -4
; RV64I-NEXT:    add a0, a0, a1
; RV64I-NEXT:    lw a0, 0(a0)
; RV64I-NEXT:    ret
;
; RV64ZBA-LABEL: srli_1_sh2add:
; RV64ZBA:       # %bb.0:
; RV64ZBA-NEXT:    srli a1, a1, 1
; RV64ZBA-NEXT:    sh2add a0, a1, a0
; RV64ZBA-NEXT:    lw a0, 0(a0)
; RV64ZBA-NEXT:    ret
  %3 = lshr i64 %1, 1
  %4 = getelementptr inbounds i32, ptr %0, i64 %3
  %5 = load i32, ptr %4, align 4
  ret i32 %5
}

before:

srli_1_sh2add:
# %bb.0:
    slli a1, a1, 1
    andi a1, a1, -4
    add a0, a0, a1
    lw a0, 0(a0)
    ret

after:

srli_1_sh2add:                          # @srli_1_sh2add
	.cfi_startproc
# %bb.0:
	srli	a1, a1, 1
	slli	a1, a1, 2
	add	a0, a0, a1
	lw	a0, 0(a0)
	ret

ParkHanbum · 2024-05-31T10:51:47Z

@davemgreen I think it works done. would you review again please?

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

justinfargnoli

LGTM, but please wait for @davemgreen's review.

davemgreen

Thanks, this does look like a good place for it.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

davemgreen

Thanks. LGTM

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

ParkHanbum · 2024-06-07T10:00:43Z

I tested this with test-suite in m1 macbook and here is that results:

current main

Failed Tests (1):
  test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-alloca-1.test

This

Failed Tests (2):
  test-suite :: SingleSource/Regression/C/gcc-c-torture/execute/GCC-C-execute-alloca-1.test
  test-suite :: SingleSource/UnitTests/AArch64/aarch64-acle-fmv-features.test

what different result at test aarch64-acle-fmv-features.test between main and this.

 22 bti
======main======
 23 simd
======this======
 23         FAIL
 24 simd
 25 fp

is this properly result? if not please let me know how to track which point was incorrect.

davemgreen · 2024-06-09T19:52:28Z

Does the test reliably fail with the new compiler? It sounds like it might be unrelated, as there have been a number of changes around FMV lately. The test looks like it fork()'s, so maybe something was going wrong with it? If it is a real failure cause by this patch then you might need to debug a bit to see what might be going wrong.

ParkHanbum · 2024-06-10T14:35:18Z

sorry, it was my mistake. I was tested it with my built llvm version.
I have tested it again with Apple clang version 15.0.0 and it passed.

RSilicon · 2024-07-16T16:37:10Z

Any updates on this?

ParkHanbum · 2024-07-17T01:18:50Z

nope

davemgreen

Sorry - I think I believed you had commit access. I've given this an extra test and it seems OK. If you can update the suggestion to fix a warning, we can get this in.

davemgreen · 2024-08-21T08:34:34Z

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

+          ShouldADD->getOpcode() == ISD::ADD && ShouldADD->hasOneUse()) {
+        if (auto ShouldLOAD = dyn_cast<LoadSDNode>(*ShouldADD->use_begin())) {
+          unsigned ByteVT = ShouldLOAD->getMemoryVT().getSizeInBits() / 8;
+          if ((1 << ShlAmt) == ByteVT &&


-> (1ULL << ShlAmt) == ByteVT

…x, c3)) when load

Currently, process of replacing bitwise operations consisting of `(shl (srl x, c1), c2)` with `And` is performed by `DAGCombiner`. However, in certain case like `(shl (srl, x, c1) 2)` is do not need to transform to `AND` if it was used to `Load` Target. Consider following case: ``` lsr x8, x8, llvm#56 and x8, x8, #0xfc ldr w0, [x2, x8] ret ``` In this case, we can remove the `AND` by changing the target of `LDR` to `[X2, X8, LSL llvm#2]` and right-shifting amount change to 56 to 58. after changed: ``` lsr x8, x8, llvm#58 ldr w0, [x2, x8, lsl llvm#2] ret ``` This patch checks to see if the `(shl (srl x, c1) 2)` operation on `load` target can be prevent transform to `And`.

davemgreen

Thanks (I didn't get notification for the update before)

ParkHanbum · 2024-08-25T09:48:48Z

@davemgreen never mind. I'm always thanks for your reviewing.

) Currently, process of replacing bitwise operations consisting of `LSR`/`LSL` with `And` is performed by `DAGCombiner`. However, in certain cases, the `AND` generated by this process can be removed. Consider following case: ``` lsr x8, x8, llvm#56 and x8, x8, #0xfc ldr w0, [x2, x8] ret ``` In this case, we can remove the `AND` by changing the target of `LDR` to `[X2, X8, LSL llvm#2]` and right-shifting amount change to 56 to 58. after changed: ``` lsr x8, x8, llvm#58 ldr w0, [x2, x8, lsl llvm#2] ret ``` This patch checks to see if the `SHIFTING` + `AND` operation on load target can be optimized and optimizes it if it can.

llvmbot added the backend:AArch64 label Apr 21, 2024

ParkHanbum force-pushed the i34101 branch from feebf12 to c8ee27e Compare April 21, 2024 13:41

ParkHanbum force-pushed the i34101 branch from c8ee27e to 378b159 Compare May 31, 2024 10:48

llvmbot added the llvm:SelectionDAG SelectionDAGISel as well label May 31, 2024

justinfargnoli reviewed May 31, 2024

View reviewed changes

ParkHanbum force-pushed the i34101 branch from 378b159 to 5f02043 Compare June 1, 2024 10:07

ParkHanbum requested a review from justinfargnoli June 1, 2024 10:08

justinfargnoli requested a review from davemgreen June 1, 2024 17:41

justinfargnoli assigned ParkHanbum Jun 1, 2024

justinfargnoli approved these changes Jun 1, 2024

View reviewed changes

davemgreen reviewed Jun 3, 2024

View reviewed changes

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp Outdated Show resolved Hide resolved

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp Outdated Show resolved Hide resolved

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp Outdated Show resolved Hide resolved

ParkHanbum force-pushed the i34101 branch from 5f02043 to 95479a5 Compare June 3, 2024 09:48

ParkHanbum requested a review from davemgreen June 3, 2024 09:49

ParkHanbum force-pushed the i34101 branch from 95479a5 to a9fa7ec Compare June 3, 2024 20:50

davemgreen approved these changes Jun 4, 2024

View reviewed changes

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp Outdated Show resolved Hide resolved

ParkHanbum force-pushed the i34101 branch 2 times, most recently from b518518 to b380592 Compare June 6, 2024 10:13

ParkHanbum requested a review from davemgreen June 7, 2024 13:37

davemgreen reviewed Aug 21, 2024

View reviewed changes

ParkHanbum added 2 commits August 21, 2024 21:37

[AArch64] Add tests for prevent (shl (srl x, c1), c2) -> (and (shift …

33d7398

…x, c3)) when load

ParkHanbum force-pushed the i34101 branch from b380592 to c956250 Compare August 21, 2024 13:16

ParkHanbum requested a review from davemgreen August 24, 2024 19:05

davemgreen approved these changes Aug 24, 2024

View reviewed changes

davemgreen merged commit 77fccb3 into llvm:main Aug 24, 2024
8 checks passed

davemgreen mentioned this pull request Sep 6, 2024

[Aarch64] lsr v, v, #56+ldr d0, [x9, x8, lsl #3] optimized to lsr v, v, #53+and v, v, #0x7f8+ldr d0, [x9, x8] #107526

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[AArch64] Replace AND with LSL#2 for LDR target (#34101) #89531

[AArch64] Replace AND with LSL#2 for LDR target (#34101) #89531

ParkHanbum commented Apr 21, 2024

llvmbot commented Apr 21, 2024 •

edited

Loading

github-actions bot commented Apr 21, 2024 •

edited

Loading

davemgreen commented Apr 22, 2024

ParkHanbum commented Apr 22, 2024

ParkHanbum commented Apr 22, 2024 •

edited

Loading

davemgreen commented Apr 23, 2024

ParkHanbum commented Apr 24, 2024 •

edited

Loading

ParkHanbum commented May 31, 2024

justinfargnoli left a comment

davemgreen left a comment

davemgreen left a comment

ParkHanbum commented Jun 7, 2024 •

edited

Loading

davemgreen commented Jun 9, 2024

ParkHanbum commented Jun 10, 2024 •

edited

Loading

RSilicon commented Jul 16, 2024

ParkHanbum commented Jul 17, 2024

davemgreen left a comment

davemgreen Aug 21, 2024

davemgreen left a comment

ParkHanbum commented Aug 25, 2024

[AArch64] Replace AND with LSL#2 for LDR target (#34101) #89531

[AArch64] Replace AND with LSL#2 for LDR target (#34101) #89531

Conversation

ParkHanbum commented Apr 21, 2024

llvmbot commented Apr 21, 2024 • edited Loading

github-actions bot commented Apr 21, 2024 • edited Loading

davemgreen commented Apr 22, 2024

ParkHanbum commented Apr 22, 2024

ParkHanbum commented Apr 22, 2024 • edited Loading

davemgreen commented Apr 23, 2024

ParkHanbum commented Apr 24, 2024 • edited Loading

ParkHanbum commented May 31, 2024

justinfargnoli left a comment

Choose a reason for hiding this comment

davemgreen left a comment

Choose a reason for hiding this comment

davemgreen left a comment

Choose a reason for hiding this comment

ParkHanbum commented Jun 7, 2024 • edited Loading

davemgreen commented Jun 9, 2024

ParkHanbum commented Jun 10, 2024 • edited Loading

RSilicon commented Jul 16, 2024

ParkHanbum commented Jul 17, 2024

davemgreen left a comment

Choose a reason for hiding this comment

davemgreen Aug 21, 2024

Choose a reason for hiding this comment

davemgreen left a comment

Choose a reason for hiding this comment

ParkHanbum commented Aug 25, 2024

llvmbot commented Apr 21, 2024 •

edited

Loading

github-actions bot commented Apr 21, 2024 •

edited

Loading

ParkHanbum commented Apr 22, 2024 •

edited

Loading

ParkHanbum commented Apr 24, 2024 •

edited

Loading

ParkHanbum commented Jun 7, 2024 •

edited

Loading

ParkHanbum commented Jun 10, 2024 •

edited

Loading