[llvm] [ARM] Only change mask if demanded bits says we can optimize (PR #165106)
via llvm-commits
llvm-commits at lists.llvm.org
Sun Oct 26 16:51:14 PDT 2025
https://github.com/AZero13 updated https://github.com/llvm/llvm-project/pull/165106
>From 50be66956ac8612bc2d4c80573641ed8733e4cf0 Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Sat, 25 Oct 2025 12:27:59 -0400
Subject: [PATCH 1/3] [ARM] Only change mask if demanded bits says we can
optimize
Also enable a switch to turn off enable-logical-imm.
---
llvm/lib/Target/ARM/ARMISelLowering.cpp | 94 ++++++++---
llvm/test/CodeGen/ARM/funnel-shift-rot.ll | 5 +-
...st-and-by-const-from-lshr-in-eqcmp-zero.ll | 56 ++++---
llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll | 158 +++++++++++++-----
llvm/test/CodeGen/ARM/va_arg.ll | 2 +-
llvm/test/CodeGen/Thumb/bic_imm.ll | 6 +-
llvm/test/CodeGen/Thumb/branch-to-return.ll | 2 +-
llvm/test/CodeGen/Thumb2/active_lane_mask.ll | 2 +-
8 files changed, 228 insertions(+), 97 deletions(-)
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 313ae3d68fb83..41a0ca1271cf5 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -118,6 +118,7 @@ using namespace llvm;
#define DEBUG_TYPE "arm-isel"
STATISTIC(NumTailCalls, "Number of tail calls");
+STATISTIC(NumOptimizedImms, "Number of times immediates were optimized");
STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
STATISTIC(NumConstpoolPromoted,
@@ -142,6 +143,12 @@ static cl::opt<unsigned> ConstpoolPromotionMaxTotal(
cl::desc("Maximum size of ALL constants to promote into a constant pool"),
cl::init(128));
+static cl::opt<bool>
+ EnableOptimizeLogicalImm("arm-enable-logical-imm", cl::Hidden,
+ cl::desc("Enable ARM logical imm instruction "
+ "optimization"),
+ cl::init(true));
+
cl::opt<unsigned>
MVEMaxSupportedInterleaveFactor("mve-max-interleave-factor", cl::Hidden,
cl::desc("Maximum interleave factor for MVE VLDn to generate."),
@@ -20138,6 +20145,16 @@ void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
}
}
+static bool isLegalLogicalImmediate(unsigned Imm,
+ const ARMSubtarget *Subtarget) {
+ if (!Subtarget->isThumb())
+ return ARM_AM::getSOImmVal(Imm) != -1;
+ if (Subtarget->isThumb2())
+ return ARM_AM::getT2SOImmVal(Imm) != -1;
+ // Thumb1 only has 8-bit unsigned immediate.
+ return Imm <= 255;
+}
+
bool ARMTargetLowering::targetShrinkDemandedConstant(
SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
TargetLoweringOpt &TLO) const {
@@ -20146,8 +20163,7 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
if (!TLO.LegalOps)
return false;
- // Only optimize AND for now.
- if (Op.getOpcode() != ISD::AND)
+ if (!EnableOptimizeLogicalImm)
return false;
EVT VT = Op.getValueType();
@@ -20158,6 +20174,14 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
assert(VT == MVT::i32 && "Unexpected integer type");
+ // Exit early if we demand all bits.
+ if (DemandedBits.popcount() == 32)
+ return false;
+
+ // Only optimize AND for now.
+ if (Op.getOpcode() != ISD::AND)
+ return false;
+
// Make sure the RHS really is a constant.
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
if (!C)
@@ -20165,21 +20189,13 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
unsigned Mask = C->getZExtValue();
+ if (Mask == 0 || Mask == ~0U)
+ return false;
+
unsigned Demanded = DemandedBits.getZExtValue();
unsigned ShrunkMask = Mask & Demanded;
unsigned ExpandedMask = Mask | ~Demanded;
- // If the mask is all zeros, let the target-independent code replace the
- // result with zero.
- if (ShrunkMask == 0)
- return false;
-
- // If the mask is all ones, erase the AND. (Currently, the target-independent
- // code won't do this, so we have to do it explicitly to avoid an infinite
- // loop in obscure cases.)
- if (ExpandedMask == ~0U)
- return TLO.CombineTo(Op, Op.getOperand(0));
-
auto IsLegalMask = [ShrunkMask, ExpandedMask](unsigned Mask) -> bool {
return (ShrunkMask & Mask) == ShrunkMask && (~ExpandedMask & Mask) == 0;
};
@@ -20192,30 +20208,56 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
return TLO.CombineTo(Op, NewOp);
};
- // Prefer uxtb mask.
- if (IsLegalMask(0xFF))
- return UseMask(0xFF);
+ // If the mask is all zeros, let the target-independent code replace the
+ // result with zero.
+ if (ShrunkMask == 0) {
+ ++NumOptimizedImms;
+ return UseMask(ShrunkMask);
+ }
+
+ // If the mask is all ones, erase the AND. (Currently, the target-independent
+ // code won't do this, so we have to do it explicitly to avoid an infinite
+ // loop in obscure cases.)
+ if (ExpandedMask == ~0U) {
+ ++NumOptimizedImms;
+ return UseMask(ExpandedMask);
+ }
+
+ // If thumb, check for uxth and uxtb masks first and foremost.
+ if (Subtarget->isThumb1Only() && Subtarget->hasV6Ops()) {
+ if (IsLegalMask(0xFF)) {
+ ++NumOptimizedImms;
+ return UseMask(0xFF);
+ }
- // Prefer uxth mask.
- if (IsLegalMask(0xFFFF))
- return UseMask(0xFFFF);
+ if (IsLegalMask(0xFFFF)) {
+ ++NumOptimizedImms;
+ return UseMask(0xFFFF);
+ }
+ }
- // [1, 255] is Thumb1 movs+ands, legal immediate for ARM/Thumb2.
- // FIXME: Prefer a contiguous sequence of bits for other optimizations.
- if (ShrunkMask < 256)
+ // Don't optimize if it is legal already.
+ if (isLegalLogicalImmediate(Mask, Subtarget))
+ return false;
+
+ if (isLegalLogicalImmediate(~Mask, Subtarget))
+ return UseMask(Mask); // FIXME: Returning false causes infinite loop.
+
+ if (isLegalLogicalImmediate(ShrunkMask, Subtarget)) {
+ ++NumOptimizedImms;
return UseMask(ShrunkMask);
+ }
- // [-256, -2] is Thumb1 movs+bics, legal immediate for ARM/Thumb2.
- // FIXME: Prefer a contiguous sequence of bits for other optimizations.
- if ((int)ExpandedMask <= -2 && (int)ExpandedMask >= -256)
+ if (isLegalLogicalImmediate(~ExpandedMask, Subtarget)) {
+ ++NumOptimizedImms;
return UseMask(ExpandedMask);
+ }
// Potential improvements:
//
// We could try to recognize lsls+lsrs or lsrs+lsls pairs here.
// We could try to prefer Thumb1 immediates which can be lowered to a
// two-instruction sequence.
- // We could try to recognize more legal ARM/Thumb2 immediates here.
return false;
}
diff --git a/llvm/test/CodeGen/ARM/funnel-shift-rot.ll b/llvm/test/CodeGen/ARM/funnel-shift-rot.ll
index a1b6847d623d0..6f34a5fd00314 100644
--- a/llvm/test/CodeGen/ARM/funnel-shift-rot.ll
+++ b/llvm/test/CodeGen/ARM/funnel-shift-rot.ll
@@ -19,7 +19,7 @@ declare <4 x i32> @llvm.fshr.v4i32(<4 x i32>, <4 x i32>, <4 x i32>)
define i8 @rotl_i8_const_shift(i8 %x) {
; CHECK-LABEL: rotl_i8_const_shift:
; CHECK: @ %bb.0:
-; CHECK-NEXT: uxtb r1, r0
+; CHECK-NEXT: and r1, r0, #224
; CHECK-NEXT: lsl r0, r0, #3
; CHECK-NEXT: orr r0, r0, r1, lsr #5
; CHECK-NEXT: bx lr
@@ -161,8 +161,7 @@ define <4 x i32> @rotl_v4i32_rotl_const_shift(<4 x i32> %x) {
define i8 @rotr_i8_const_shift(i8 %x) {
; CHECK-LABEL: rotr_i8_const_shift:
; CHECK: @ %bb.0:
-; CHECK-NEXT: uxtb r1, r0
-; CHECK-NEXT: lsr r1, r1, #3
+; CHECK-NEXT: ubfx r1, r0, #3, #5
; CHECK-NEXT: orr r0, r1, r0, lsl #5
; CHECK-NEXT: bx lr
%f = call i8 @llvm.fshr.i8(i8 %x, i8 %x, i8 3)
diff --git a/llvm/test/CodeGen/ARM/hoist-and-by-const-from-lshr-in-eqcmp-zero.ll b/llvm/test/CodeGen/ARM/hoist-and-by-const-from-lshr-in-eqcmp-zero.ll
index 7cc623fb0a616..a21ac8944d7ad 100644
--- a/llvm/test/CodeGen/ARM/hoist-and-by-const-from-lshr-in-eqcmp-zero.ll
+++ b/llvm/test/CodeGen/ARM/hoist-and-by-const-from-lshr-in-eqcmp-zero.ll
@@ -21,9 +21,9 @@ define i1 @scalar_i8_signbit_eq(i8 %x, i8 %y) nounwind {
; ARM-LABEL: scalar_i8_signbit_eq:
; ARM: @ %bb.0:
; ARM-NEXT: uxtb r1, r1
-; ARM-NEXT: lsl r0, r0, r1
+; ARM-NEXT: mov r2, #128
+; ARM-NEXT: and r0, r2, r0, lsl r1
; ARM-NEXT: mov r1, #1
-; ARM-NEXT: uxtb r0, r0
; ARM-NEXT: eor r0, r1, r0, lsr #7
; ARM-NEXT: bx lr
;
@@ -42,7 +42,7 @@ define i1 @scalar_i8_signbit_eq(i8 %x, i8 %y) nounwind {
; THUMB78-NEXT: uxtb r1, r1
; THUMB78-NEXT: lsls r0, r1
; THUMB78-NEXT: movs r1, #1
-; THUMB78-NEXT: uxtb r0, r0
+; THUMB78-NEXT: and r0, r0, #128
; THUMB78-NEXT: eor.w r0, r1, r0, lsr #7
; THUMB78-NEXT: bx lr
%t0 = lshr i8 128, %y
@@ -122,9 +122,9 @@ define i1 @scalar_i16_signbit_eq(i16 %x, i16 %y) nounwind {
; ARM-LABEL: scalar_i16_signbit_eq:
; ARM: @ %bb.0:
; ARM-NEXT: uxth r1, r1
-; ARM-NEXT: lsl r0, r0, r1
+; ARM-NEXT: mov r2, #32768
+; ARM-NEXT: and r0, r2, r0, lsl r1
; ARM-NEXT: mov r1, #1
-; ARM-NEXT: uxth r0, r0
; ARM-NEXT: eor r0, r1, r0, lsr #15
; ARM-NEXT: bx lr
;
@@ -144,7 +144,7 @@ define i1 @scalar_i16_signbit_eq(i16 %x, i16 %y) nounwind {
; THUMB78-NEXT: uxth r1, r1
; THUMB78-NEXT: lsls r0, r1
; THUMB78-NEXT: movs r1, #1
-; THUMB78-NEXT: uxth r0, r0
+; THUMB78-NEXT: and r0, r0, #32768
; THUMB78-NEXT: eor.w r0, r1, r0, lsr #15
; THUMB78-NEXT: bx lr
%t0 = lshr i16 32768, %y
@@ -862,21 +862,35 @@ define <4 x i1> @vec_4xi32_nonsplat_undef2_eq(<4 x i32> %x, <4 x i32> %y) nounwi
;------------------------------------------------------------------------------;
define i1 @scalar_i8_signbit_ne(i8 %x, i8 %y) nounwind {
-; ARM-LABEL: scalar_i8_signbit_ne:
-; ARM: @ %bb.0:
-; ARM-NEXT: uxtb r1, r1
-; ARM-NEXT: lsl r0, r0, r1
-; ARM-NEXT: uxtb r0, r0
-; ARM-NEXT: lsr r0, r0, #7
-; ARM-NEXT: bx lr
+; ARM6-LABEL: scalar_i8_signbit_ne:
+; ARM6: @ %bb.0:
+; ARM6-NEXT: uxtb r1, r1
+; ARM6-NEXT: mov r2, #128
+; ARM6-NEXT: and r0, r2, r0, lsl r1
+; ARM6-NEXT: lsr r0, r0, #7
+; ARM6-NEXT: bx lr
;
-; THUMB-LABEL: scalar_i8_signbit_ne:
-; THUMB: @ %bb.0:
-; THUMB-NEXT: uxtb r1, r1
-; THUMB-NEXT: lsls r0, r1
-; THUMB-NEXT: uxtb r0, r0
-; THUMB-NEXT: lsrs r0, r0, #7
-; THUMB-NEXT: bx lr
+; ARM78-LABEL: scalar_i8_signbit_ne:
+; ARM78: @ %bb.0:
+; ARM78-NEXT: uxtb r1, r1
+; ARM78-NEXT: lsl r0, r0, r1
+; ARM78-NEXT: ubfx r0, r0, #7, #1
+; ARM78-NEXT: bx lr
+;
+; THUMB6-LABEL: scalar_i8_signbit_ne:
+; THUMB6: @ %bb.0:
+; THUMB6-NEXT: uxtb r1, r1
+; THUMB6-NEXT: lsls r0, r1
+; THUMB6-NEXT: uxtb r0, r0
+; THUMB6-NEXT: lsrs r0, r0, #7
+; THUMB6-NEXT: bx lr
+;
+; THUMB78-LABEL: scalar_i8_signbit_ne:
+; THUMB78: @ %bb.0:
+; THUMB78-NEXT: uxtb r1, r1
+; THUMB78-NEXT: lsls r0, r1
+; THUMB78-NEXT: ubfx r0, r0, #7, #1
+; THUMB78-NEXT: bx lr
%t0 = lshr i8 128, %y
%t1 = and i8 %t0, %x
%res = icmp ne i8 %t1, 0 ; we are perfectly happy with 'ne' predicate
@@ -1051,3 +1065,5 @@ define i1 @scalar_i8_signbit_eq_with_nonzero(i8 %x, i8 %y) nounwind {
%res = icmp eq i8 %t1, 1 ; should be comparing with 0
ret i1 %res
}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; THUMB: {{.*}}
diff --git a/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll b/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll
index 4b0419577cdf0..7f4315ffcaa6c 100644
--- a/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll
+++ b/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll
@@ -1,3 +1,4 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc -mtriple=thumbv8 %s -o - | FileCheck %s --check-prefixes=CHECK,T2
; RUN: llc -mtriple=thumbv8m.main %s -o - | FileCheck %s --check-prefixes=CHECK,T2
; RUN: llc -mtriple=thumbv8m.base %s -o - | FileCheck %s --check-prefixes=CHECK,T1
@@ -13,11 +14,23 @@
; Test sdiv i16
define dso_local signext i16 @f0(i16 signext %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f0
-; CHECK: movs r1, #2
-; CHECK-NEXT: sdiv r0, r0, r1
-; CHECK-NEXT: sxth r0, r0
-; CHECK-NEXT: bx lr
+; CHECK-LABEL: f0:
+; CHECK: @ %bb.0: @ %entry
+; CHECK-NEXT: movs r1, #2
+; CHECK-NEXT: sdiv r0, r0, r1
+; CHECK-NEXT: sxth r0, r0
+; CHECK-NEXT: bx lr
+;
+; V6M-LABEL: f0:
+; V6M: @ %bb.0: @ %entry
+; V6M-NEXT: movs r1, #255
+; V6M-NEXT: lsls r1, r1, #8
+; V6M-NEXT: ands r1, r0
+; V6M-NEXT: lsrs r1, r1, #15
+; V6M-NEXT: adds r0, r0, r1
+; V6M-NEXT: sxth r0, r0
+; V6M-NEXT: asrs r0, r0, #1
+; V6M-NEXT: bx lr
entry:
%0 = sdiv i16 %F, 2
@@ -26,10 +39,19 @@ entry:
; Same as above, but now with i32
define dso_local i32 @f1(i32 %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f1
-; CHECK: movs r1, #4
-; CHECK-NEXT: sdiv r0, r0, r1
-; CHECK-NEXT: bx lr
+; CHECK-LABEL: f1:
+; CHECK: @ %bb.0: @ %entry
+; CHECK-NEXT: movs r1, #4
+; CHECK-NEXT: sdiv r0, r0, r1
+; CHECK-NEXT: bx lr
+;
+; V6M-LABEL: f1:
+; V6M: @ %bb.0: @ %entry
+; V6M-NEXT: asrs r1, r0, #31
+; V6M-NEXT: lsrs r1, r1, #30
+; V6M-NEXT: adds r0, r0, r1
+; V6M-NEXT: asrs r0, r0, #2
+; V6M-NEXT: bx lr
entry:
%div = sdiv i32 %F, 4
@@ -38,10 +60,18 @@ entry:
; The immediate is not a power of 2, so we expect a sdiv.
define dso_local i32 @f2(i32 %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f2
-; CHECK: movs r1, #5
-; CHECK-NEXT: sdiv r0, r0, r1
-; CHECK-NEXT: bx lr
+; CHECK-LABEL: f2:
+; CHECK: @ %bb.0: @ %entry
+; CHECK-NEXT: movs r1, #5
+; CHECK-NEXT: sdiv r0, r0, r1
+; CHECK-NEXT: bx lr
+;
+; V6M-LABEL: f2:
+; V6M: @ %bb.0: @ %entry
+; V6M-NEXT: push {r7, lr}
+; V6M-NEXT: movs r1, #5
+; V6M-NEXT: bl __divsi3
+; V6M-NEXT: pop {r7, pc}
entry:
%div = sdiv i32 %F, 5
@@ -51,8 +81,28 @@ entry:
; Try a larger power of 2 immediate: immediates larger than
; 128 don't give any code size savings.
define dso_local i32 @f3(i32 %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f3
-; CHECK-NOT: sdiv
+; T2-LABEL: f3:
+; T2: @ %bb.0: @ %entry
+; T2-NEXT: asrs r1, r0, #31
+; T2-NEXT: add.w r0, r0, r1, lsr #24
+; T2-NEXT: asrs r0, r0, #8
+; T2-NEXT: bx lr
+;
+; T1-LABEL: f3:
+; T1: @ %bb.0: @ %entry
+; T1-NEXT: asrs r1, r0, #31
+; T1-NEXT: lsrs r1, r1, #24
+; T1-NEXT: adds r0, r0, r1
+; T1-NEXT: asrs r0, r0, #8
+; T1-NEXT: bx lr
+;
+; V6M-LABEL: f3:
+; V6M: @ %bb.0: @ %entry
+; V6M-NEXT: asrs r1, r0, #31
+; V6M-NEXT: lsrs r1, r1, #24
+; V6M-NEXT: adds r0, r0, r1
+; V6M-NEXT: asrs r0, r0, #8
+; V6M-NEXT: bx lr
entry:
%div = sdiv i32 %F, 256
ret i32 %div
@@ -65,20 +115,35 @@ attributes #0 = { minsize norecurse nounwind optsize readnone }
; the sdiv to sdiv, but to the faster instruction sequence.
define dso_local signext i16 @f4(i16 signext %F) {
-; T2-LABEL: f4
-; T2: uxth r1, r0
-; T2-NEXT: add.w r0, r0, r1, lsr #15
-; T2-NEXT: sxth r0, r0
-; T2-NEXT: asrs r0, r0, #1
-; T2-NEXT: bx lr
-
-; T1-LABEL: f4
-; T1: uxth r1, r0
-; T1-NEXT: lsrs r1, r1, #15
-; T1-NEXT: adds r0, r0, r1
-; T1-NEXT: sxth r0, r0
-; T1-NEXT: asrs r0, r0, #1
-; T1-NEXT: bx lr
+; T2-LABEL: f4:
+; T2: @ %bb.0: @ %entry
+; T2-NEXT: and r1, r0, #32768
+; T2-NEXT: add.w r0, r0, r1, lsr #15
+; T2-NEXT: sxth r0, r0
+; T2-NEXT: asrs r0, r0, #1
+; T2-NEXT: bx lr
+;
+; T1-LABEL: f4:
+; T1: @ %bb.0: @ %entry
+; T1-NEXT: movw r1, #65280
+; T1-NEXT: ands r1, r0
+; T1-NEXT: lsrs r1, r1, #15
+; T1-NEXT: adds r0, r0, r1
+; T1-NEXT: sxth r0, r0
+; T1-NEXT: asrs r0, r0, #1
+; T1-NEXT: bx lr
+;
+; V6M-LABEL: f4:
+; V6M: @ %bb.0: @ %entry
+; V6M-NEXT: movs r1, #255
+; V6M-NEXT: lsls r1, r1, #8
+; V6M-NEXT: ands r1, r0
+; V6M-NEXT: lsrs r1, r1, #15
+; V6M-NEXT: adds r0, r0, r1
+; V6M-NEXT: sxth r0, r0
+; V6M-NEXT: asrs r0, r0, #1
+; V6M-NEXT: bx lr
+
entry:
%0 = sdiv i16 %F, 2
@@ -86,18 +151,29 @@ entry:
}
define dso_local i32 @f5(i32 %F) {
-; T2-LABEL: f5
-; T2: asrs r1, r0, #31
-; T2-NEXT: add.w r0, r0, r1, lsr #30
-; T2-NEXT: asrs r0, r0, #2
-; T2-NEXT: bx lr
-
-; T1-LABEL: f5
-; T1: asrs r1, r0, #31
-; T1-NEXT: lsrs r1, r1, #30
-; T1-NEXT: adds r0, r0, r1
-; T1-NEXT: asrs r0, r0, #2
-; T1-NEXT: bx lr
+; T2-LABEL: f5:
+; T2: @ %bb.0: @ %entry
+; T2-NEXT: asrs r1, r0, #31
+; T2-NEXT: add.w r0, r0, r1, lsr #30
+; T2-NEXT: asrs r0, r0, #2
+; T2-NEXT: bx lr
+;
+; T1-LABEL: f5:
+; T1: @ %bb.0: @ %entry
+; T1-NEXT: asrs r1, r0, #31
+; T1-NEXT: lsrs r1, r1, #30
+; T1-NEXT: adds r0, r0, r1
+; T1-NEXT: asrs r0, r0, #2
+; T1-NEXT: bx lr
+;
+; V6M-LABEL: f5:
+; V6M: @ %bb.0: @ %entry
+; V6M-NEXT: asrs r1, r0, #31
+; V6M-NEXT: lsrs r1, r1, #30
+; V6M-NEXT: adds r0, r0, r1
+; V6M-NEXT: asrs r0, r0, #2
+; V6M-NEXT: bx lr
+
entry:
%div = sdiv i32 %F, 4
diff --git a/llvm/test/CodeGen/ARM/va_arg.ll b/llvm/test/CodeGen/ARM/va_arg.ll
index 41556b9fa2dec..7228b82b19c5a 100644
--- a/llvm/test/CodeGen/ARM/va_arg.ll
+++ b/llvm/test/CodeGen/ARM/va_arg.ll
@@ -35,7 +35,7 @@ define double @test2(i32 %a, ptr %b, ...) nounwind optsize {
; CHECK-NEXT: add r0, sp, #4
; CHECK-NEXT: stmib sp, {r2, r3}
; CHECK-NEXT: add r0, r0, #11
-; CHECK-NEXT: bic r0, r0, #3
+; CHECK-NEXT: bic r0, r0, #7
; CHECK-NEXT: str r2, [r1]
; CHECK-NEXT: add r1, r0, #8
; CHECK-NEXT: str r1, [sp]
diff --git a/llvm/test/CodeGen/Thumb/bic_imm.ll b/llvm/test/CodeGen/Thumb/bic_imm.ll
index 741b2cf8db2e3..d1730c21250a8 100644
--- a/llvm/test/CodeGen/Thumb/bic_imm.ll
+++ b/llvm/test/CodeGen/Thumb/bic_imm.ll
@@ -82,7 +82,7 @@ define void @truncated_neg256(i16 %a, ptr %p) {
;
; CHECK-T2-LABEL: truncated_neg256:
; CHECK-T2: @ %bb.0:
-; CHECK-T2-NEXT: bic r0, r0, #255
+; CHECK-T2-NEXT: and r0, r0, #65280
; CHECK-T2-NEXT: strh r0, [r1]
; CHECK-T2-NEXT: bx lr
%and = and i16 %a, -256
@@ -90,7 +90,6 @@ define void @truncated_neg256(i16 %a, ptr %p) {
ret void
}
-; FIXME: Thumb2 supports "bic r0, r0, #510"
define void @truncated_neg511(i16 %a, ptr %p) {
; CHECK-T1-LABEL: truncated_neg511:
; CHECK-T1: @ %bb.0:
@@ -105,8 +104,7 @@ define void @truncated_neg511(i16 %a, ptr %p) {
;
; CHECK-T2-LABEL: truncated_neg511:
; CHECK-T2: @ %bb.0:
-; CHECK-T2-NEXT: movw r2, #65025
-; CHECK-T2-NEXT: ands r0, r2
+; CHECK-T2-NEXT: bic r0, r0, #510
; CHECK-T2-NEXT: strh r0, [r1]
; CHECK-T2-NEXT: bx lr
%and = and i16 %a, -511
diff --git a/llvm/test/CodeGen/Thumb/branch-to-return.ll b/llvm/test/CodeGen/Thumb/branch-to-return.ll
index 11e8add6f215b..b190a11aeab17 100644
--- a/llvm/test/CodeGen/Thumb/branch-to-return.ll
+++ b/llvm/test/CodeGen/Thumb/branch-to-return.ll
@@ -26,7 +26,7 @@ define i32 @foo(ptr %x, i32 %n) {
; CHECK-NEXT: ldr.w r0, [r12]
; CHECK-NEXT: .LBB0_6: @ %for.body.preheader1
; CHECK-NEXT: subs r3, r1, r3
-; CHECK-NEXT: mvn r2, #12
+; CHECK-NEXT: mvn r2, #15
; CHECK-NEXT: and.w r1, r2, r1, lsl #2
; CHECK-NEXT: add r1, r12
; CHECK-NEXT: .LBB0_7: @ %for.body
diff --git a/llvm/test/CodeGen/Thumb2/active_lane_mask.ll b/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
index bcd92f81911b2..b75f1ff742bee 100644
--- a/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
+++ b/llvm/test/CodeGen/Thumb2/active_lane_mask.ll
@@ -283,7 +283,7 @@ define void @test_width2(ptr nocapture readnone %x, ptr nocapture %y, i8 zeroext
; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
; CHECK-NEXT: adds r0, r2, #1
; CHECK-NEXT: movs r3, #1
-; CHECK-NEXT: bic r0, r0, #1
+; CHECK-NEXT: and r0, r0, #510
; CHECK-NEXT: subs r0, #2
; CHECK-NEXT: add.w r0, r3, r0, lsr #1
; CHECK-NEXT: dls lr, r0
>From 5a0106a032eea44e166e380af73a2ab8bddc8a67 Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Sun, 26 Oct 2025 19:50:28 -0400
Subject: [PATCH 2/3] s
---
llvm/test/CodeGen/Thumb2/bf16-instructions.ll | 38 +++++++++----------
llvm/test/CodeGen/Thumb2/shift_parts.ll | 4 +-
2 files changed, 20 insertions(+), 22 deletions(-)
diff --git a/llvm/test/CodeGen/Thumb2/bf16-instructions.ll b/llvm/test/CodeGen/Thumb2/bf16-instructions.ll
index 313d237d54b35..3d0802a7e9ff7 100644
--- a/llvm/test/CodeGen/Thumb2/bf16-instructions.ll
+++ b/llvm/test/CodeGen/Thumb2/bf16-instructions.ll
@@ -2104,12 +2104,11 @@ define bfloat @test_copysign(bfloat %a, bfloat %b) {
;
; CHECK-FP-LABEL: test_copysign:
; CHECK-FP: @ %bb.0:
-; CHECK-FP-NEXT: vmov r0, s0
-; CHECK-FP-NEXT: vmov r1, s1
-; CHECK-FP-NEXT: and r1, r1, #32768
-; CHECK-FP-NEXT: bfc r0, #15, #17
-; CHECK-FP-NEXT: add r0, r1
-; CHECK-FP-NEXT: vmov.f16 s0, r0
+; CHECK-FP-NEXT: vmov r0, s1
+; CHECK-FP-NEXT: vmov r1, s0
+; CHECK-FP-NEXT: lsrs r0, r0, #15
+; CHECK-FP-NEXT: bfi r1, r0, #15, #1
+; CHECK-FP-NEXT: vmov.f16 s0, r1
; CHECK-FP-NEXT: vmov.f16 r0, s0
; CHECK-FP-NEXT: vmov s0, r0
; CHECK-FP-NEXT: bx lr
@@ -2128,11 +2127,11 @@ define bfloat @test_copysign_f32(bfloat %a, float %b) {
;
; CHECK-FP-LABEL: test_copysign_f32:
; CHECK-FP: @ %bb.0:
-; CHECK-FP-NEXT: vmov r0, s0
-; CHECK-FP-NEXT: vmov r1, s1
-; CHECK-FP-NEXT: and r1, r1, #-2147483648
-; CHECK-FP-NEXT: bfc r0, #15, #17
-; CHECK-FP-NEXT: orr.w r0, r0, r1, lsr #16
+; CHECK-FP-NEXT: vmov r0, s1
+; CHECK-FP-NEXT: vmov r1, s0
+; CHECK-FP-NEXT: and r0, r0, #-2147483648
+; CHECK-FP-NEXT: bic r1, r1, #32768
+; CHECK-FP-NEXT: orr.w r0, r1, r0, lsr #16
; CHECK-FP-NEXT: vmov.f16 s0, r0
; CHECK-FP-NEXT: vmov.f16 r0, s0
; CHECK-FP-NEXT: vmov s0, r0
@@ -2153,10 +2152,10 @@ define bfloat @test_copysign_f64(bfloat %a, double %b) {
;
; CHECK-FP-LABEL: test_copysign_f64:
; CHECK-FP: @ %bb.0:
+; CHECK-FP-NEXT: vmov r0, r1, d1
; CHECK-FP-NEXT: vmov r0, s0
-; CHECK-FP-NEXT: vmov r1, r2, d1
-; CHECK-FP-NEXT: and r1, r2, #-2147483648
-; CHECK-FP-NEXT: bfc r0, #15, #17
+; CHECK-FP-NEXT: and r1, r1, #-2147483648
+; CHECK-FP-NEXT: bic r0, r0, #32768
; CHECK-FP-NEXT: orr.w r0, r0, r1, lsr #16
; CHECK-FP-NEXT: vmov.f16 s0, r0
; CHECK-FP-NEXT: vmov.f16 r0, s0
@@ -2180,12 +2179,11 @@ define float @test_copysign_extended(bfloat %a, bfloat %b) {
;
; CHECK-FP-LABEL: test_copysign_extended:
; CHECK-FP: @ %bb.0:
-; CHECK-FP-NEXT: vmov r0, s0
-; CHECK-FP-NEXT: vmov r1, s1
-; CHECK-FP-NEXT: and r1, r1, #32768
-; CHECK-FP-NEXT: bfc r0, #15, #17
-; CHECK-FP-NEXT: add r0, r1
-; CHECK-FP-NEXT: lsls r0, r0, #16
+; CHECK-FP-NEXT: vmov r0, s1
+; CHECK-FP-NEXT: vmov r1, s0
+; CHECK-FP-NEXT: lsrs r0, r0, #15
+; CHECK-FP-NEXT: bfi r1, r0, #15, #1
+; CHECK-FP-NEXT: lsls r0, r1, #16
; CHECK-FP-NEXT: vmov s0, r0
; CHECK-FP-NEXT: bx lr
%r = call bfloat @llvm.copysign.f16(bfloat %a, bfloat %b)
diff --git a/llvm/test/CodeGen/Thumb2/shift_parts.ll b/llvm/test/CodeGen/Thumb2/shift_parts.ll
index b4ac405d82ed5..060c628013050 100644
--- a/llvm/test/CodeGen/Thumb2/shift_parts.ll
+++ b/llvm/test/CodeGen/Thumb2/shift_parts.ll
@@ -457,7 +457,7 @@ entry:
define i32 @ashr_demand_bottommask2(i64 %x) {
; CHECK-LABEL: ashr_demand_bottommask2:
; CHECK: @ %bb.0: @ %entry
-; CHECK-NEXT: mvn r0, #2
+; CHECK-NEXT: mvn r0, #3
; CHECK-NEXT: and.w r0, r0, r1, lsl #1
; CHECK-NEXT: bx lr
entry:
@@ -470,7 +470,7 @@ entry:
define i32 @lshr_demand_bottommask2(i64 %x) {
; CHECK-LABEL: lshr_demand_bottommask2:
; CHECK: @ %bb.0: @ %entry
-; CHECK-NEXT: mvn r0, #2
+; CHECK-NEXT: mvn r0, #3
; CHECK-NEXT: and.w r0, r0, r1, lsl #1
; CHECK-NEXT: bx lr
entry:
>From 147c9b8b29121a6e3930d1fd5aaf2a2f2abb0315 Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Sun, 26 Oct 2025 19:51:03 -0400
Subject: [PATCH 3/3] Update sdiv-pow2-thumb-size.ll
---
llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll | 158 +++++-------------
1 file changed, 41 insertions(+), 117 deletions(-)
diff --git a/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll b/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll
index 7f4315ffcaa6c..4b0419577cdf0 100644
--- a/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll
+++ b/llvm/test/CodeGen/ARM/sdiv-pow2-thumb-size.ll
@@ -1,4 +1,3 @@
-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc -mtriple=thumbv8 %s -o - | FileCheck %s --check-prefixes=CHECK,T2
; RUN: llc -mtriple=thumbv8m.main %s -o - | FileCheck %s --check-prefixes=CHECK,T2
; RUN: llc -mtriple=thumbv8m.base %s -o - | FileCheck %s --check-prefixes=CHECK,T1
@@ -14,23 +13,11 @@
; Test sdiv i16
define dso_local signext i16 @f0(i16 signext %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f0:
-; CHECK: @ %bb.0: @ %entry
-; CHECK-NEXT: movs r1, #2
-; CHECK-NEXT: sdiv r0, r0, r1
-; CHECK-NEXT: sxth r0, r0
-; CHECK-NEXT: bx lr
-;
-; V6M-LABEL: f0:
-; V6M: @ %bb.0: @ %entry
-; V6M-NEXT: movs r1, #255
-; V6M-NEXT: lsls r1, r1, #8
-; V6M-NEXT: ands r1, r0
-; V6M-NEXT: lsrs r1, r1, #15
-; V6M-NEXT: adds r0, r0, r1
-; V6M-NEXT: sxth r0, r0
-; V6M-NEXT: asrs r0, r0, #1
-; V6M-NEXT: bx lr
+; CHECK-LABEL: f0
+; CHECK: movs r1, #2
+; CHECK-NEXT: sdiv r0, r0, r1
+; CHECK-NEXT: sxth r0, r0
+; CHECK-NEXT: bx lr
entry:
%0 = sdiv i16 %F, 2
@@ -39,19 +26,10 @@ entry:
; Same as above, but now with i32
define dso_local i32 @f1(i32 %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f1:
-; CHECK: @ %bb.0: @ %entry
-; CHECK-NEXT: movs r1, #4
-; CHECK-NEXT: sdiv r0, r0, r1
-; CHECK-NEXT: bx lr
-;
-; V6M-LABEL: f1:
-; V6M: @ %bb.0: @ %entry
-; V6M-NEXT: asrs r1, r0, #31
-; V6M-NEXT: lsrs r1, r1, #30
-; V6M-NEXT: adds r0, r0, r1
-; V6M-NEXT: asrs r0, r0, #2
-; V6M-NEXT: bx lr
+; CHECK-LABEL: f1
+; CHECK: movs r1, #4
+; CHECK-NEXT: sdiv r0, r0, r1
+; CHECK-NEXT: bx lr
entry:
%div = sdiv i32 %F, 4
@@ -60,18 +38,10 @@ entry:
; The immediate is not a power of 2, so we expect a sdiv.
define dso_local i32 @f2(i32 %F) local_unnamed_addr #0 {
-; CHECK-LABEL: f2:
-; CHECK: @ %bb.0: @ %entry
-; CHECK-NEXT: movs r1, #5
-; CHECK-NEXT: sdiv r0, r0, r1
-; CHECK-NEXT: bx lr
-;
-; V6M-LABEL: f2:
-; V6M: @ %bb.0: @ %entry
-; V6M-NEXT: push {r7, lr}
-; V6M-NEXT: movs r1, #5
-; V6M-NEXT: bl __divsi3
-; V6M-NEXT: pop {r7, pc}
+; CHECK-LABEL: f2
+; CHECK: movs r1, #5
+; CHECK-NEXT: sdiv r0, r0, r1
+; CHECK-NEXT: bx lr
entry:
%div = sdiv i32 %F, 5
@@ -81,28 +51,8 @@ entry:
; Try a larger power of 2 immediate: immediates larger than
; 128 don't give any code size savings.
define dso_local i32 @f3(i32 %F) local_unnamed_addr #0 {
-; T2-LABEL: f3:
-; T2: @ %bb.0: @ %entry
-; T2-NEXT: asrs r1, r0, #31
-; T2-NEXT: add.w r0, r0, r1, lsr #24
-; T2-NEXT: asrs r0, r0, #8
-; T2-NEXT: bx lr
-;
-; T1-LABEL: f3:
-; T1: @ %bb.0: @ %entry
-; T1-NEXT: asrs r1, r0, #31
-; T1-NEXT: lsrs r1, r1, #24
-; T1-NEXT: adds r0, r0, r1
-; T1-NEXT: asrs r0, r0, #8
-; T1-NEXT: bx lr
-;
-; V6M-LABEL: f3:
-; V6M: @ %bb.0: @ %entry
-; V6M-NEXT: asrs r1, r0, #31
-; V6M-NEXT: lsrs r1, r1, #24
-; V6M-NEXT: adds r0, r0, r1
-; V6M-NEXT: asrs r0, r0, #8
-; V6M-NEXT: bx lr
+; CHECK-LABEL: f3
+; CHECK-NOT: sdiv
entry:
%div = sdiv i32 %F, 256
ret i32 %div
@@ -115,35 +65,20 @@ attributes #0 = { minsize norecurse nounwind optsize readnone }
; the sdiv to sdiv, but to the faster instruction sequence.
define dso_local signext i16 @f4(i16 signext %F) {
-; T2-LABEL: f4:
-; T2: @ %bb.0: @ %entry
-; T2-NEXT: and r1, r0, #32768
-; T2-NEXT: add.w r0, r0, r1, lsr #15
-; T2-NEXT: sxth r0, r0
-; T2-NEXT: asrs r0, r0, #1
-; T2-NEXT: bx lr
-;
-; T1-LABEL: f4:
-; T1: @ %bb.0: @ %entry
-; T1-NEXT: movw r1, #65280
-; T1-NEXT: ands r1, r0
-; T1-NEXT: lsrs r1, r1, #15
-; T1-NEXT: adds r0, r0, r1
-; T1-NEXT: sxth r0, r0
-; T1-NEXT: asrs r0, r0, #1
-; T1-NEXT: bx lr
-;
-; V6M-LABEL: f4:
-; V6M: @ %bb.0: @ %entry
-; V6M-NEXT: movs r1, #255
-; V6M-NEXT: lsls r1, r1, #8
-; V6M-NEXT: ands r1, r0
-; V6M-NEXT: lsrs r1, r1, #15
-; V6M-NEXT: adds r0, r0, r1
-; V6M-NEXT: sxth r0, r0
-; V6M-NEXT: asrs r0, r0, #1
-; V6M-NEXT: bx lr
-
+; T2-LABEL: f4
+; T2: uxth r1, r0
+; T2-NEXT: add.w r0, r0, r1, lsr #15
+; T2-NEXT: sxth r0, r0
+; T2-NEXT: asrs r0, r0, #1
+; T2-NEXT: bx lr
+
+; T1-LABEL: f4
+; T1: uxth r1, r0
+; T1-NEXT: lsrs r1, r1, #15
+; T1-NEXT: adds r0, r0, r1
+; T1-NEXT: sxth r0, r0
+; T1-NEXT: asrs r0, r0, #1
+; T1-NEXT: bx lr
entry:
%0 = sdiv i16 %F, 2
@@ -151,29 +86,18 @@ entry:
}
define dso_local i32 @f5(i32 %F) {
-; T2-LABEL: f5:
-; T2: @ %bb.0: @ %entry
-; T2-NEXT: asrs r1, r0, #31
-; T2-NEXT: add.w r0, r0, r1, lsr #30
-; T2-NEXT: asrs r0, r0, #2
-; T2-NEXT: bx lr
-;
-; T1-LABEL: f5:
-; T1: @ %bb.0: @ %entry
-; T1-NEXT: asrs r1, r0, #31
-; T1-NEXT: lsrs r1, r1, #30
-; T1-NEXT: adds r0, r0, r1
-; T1-NEXT: asrs r0, r0, #2
-; T1-NEXT: bx lr
-;
-; V6M-LABEL: f5:
-; V6M: @ %bb.0: @ %entry
-; V6M-NEXT: asrs r1, r0, #31
-; V6M-NEXT: lsrs r1, r1, #30
-; V6M-NEXT: adds r0, r0, r1
-; V6M-NEXT: asrs r0, r0, #2
-; V6M-NEXT: bx lr
-
+; T2-LABEL: f5
+; T2: asrs r1, r0, #31
+; T2-NEXT: add.w r0, r0, r1, lsr #30
+; T2-NEXT: asrs r0, r0, #2
+; T2-NEXT: bx lr
+
+; T1-LABEL: f5
+; T1: asrs r1, r0, #31
+; T1-NEXT: lsrs r1, r1, #30
+; T1-NEXT: adds r0, r0, r1
+; T1-NEXT: asrs r0, r0, #2
+; T1-NEXT: bx lr
entry:
%div = sdiv i32 %F, 4
More information about the llvm-commits
mailing list