[llvm] [ARM] Only change mask if demanded bits says we can optimize (PR #165106)
via llvm-commits
llvm-commits at lists.llvm.org
Sat Oct 25 17:51:25 PDT 2025
https://github.com/AZero13 updated https://github.com/llvm/llvm-project/pull/165106
>From f812585ecfc0a13f0643d351f7bf056a81d8e3bd Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Sat, 25 Oct 2025 12:27:59 -0400
Subject: [PATCH 1/2] [ARM] Only change mask if demanded bits says we can
optimize
---
llvm/lib/Target/ARM/ARMISelLowering.cpp | 44 ++++++++++----
llvm/test/CodeGen/ARM/funnel-shift-rot.ll | 5 +-
...st-and-by-const-from-lshr-in-eqcmp-zero.ll | 57 ++++++++++++-------
llvm/test/CodeGen/ARM/va_arg.ll | 2 +-
llvm/test/CodeGen/Thumb/branch-to-return.ll | 2 +-
llvm/test/CodeGen/Thumb2/active_lane_mask.ll | 2 +-
6 files changed, 74 insertions(+), 38 deletions(-)
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 313ae3d68fb83..2ca81332ea7c9 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -20138,6 +20138,17 @@ void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
}
}
+static bool isLegalLogicalImmediate(unsigned Imm,
+ const ARMSubtarget *Subtarget) {
+ // Handle special cases first
+ 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 {
@@ -20179,7 +20190,6 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
// 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;
};
@@ -20192,22 +20202,33 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
return TLO.CombineTo(Op, NewOp);
};
- // Prefer uxtb mask.
+ // If thumb, check for uxth and uxtb masks.
+ if (Subtarget->isThumb1Only()) {
+ if (IsLegalMask(0xFF))
+ return UseMask(0xFF);
+
+ if (IsLegalMask(0xFF00))
+ return UseMask(0xFF00);
+ }
+
+ // Don't optimize if it is legal already.
+ if (isLegalLogicalImmediate(Mask, Subtarget))
+ return false;
+
+ // bic
+ if (isLegalLogicalImmediate(~Mask, Subtarget))
+ return false;
+
if (IsLegalMask(0xFF))
return UseMask(0xFF);
- // Prefer uxth mask.
- if (IsLegalMask(0xFFFF))
- return UseMask(0xFFFF);
+ if (IsLegalMask(0xFF00))
+ return UseMask(0xFF00);
- // [1, 255] is Thumb1 movs+ands, legal immediate for ARM/Thumb2.
- // FIXME: Prefer a contiguous sequence of bits for other optimizations.
- if (ShrunkMask < 256)
+ if (isLegalLogicalImmediate(ShrunkMask, Subtarget))
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))
return UseMask(ExpandedMask);
// Potential improvements:
@@ -20215,7 +20236,6 @@ bool ARMTargetLowering::targetShrinkDemandedConstant(
// 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..36feb8c4f1f4b 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 +862,36 @@ 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: movs r1, #128
+; THUMB6-NEXT: ands r1, r0
+; THUMB6-NEXT: lsrs r0, r1, #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 +1066,5 @@ 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/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/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 361381a134d232e3e4484ab641ca2d7ee6d16c16 Mon Sep 17 00:00:00 2001
From: AZero13 <gfunni234 at gmail.com>
Date: Sat, 25 Oct 2025 20:50:59 -0400
Subject: [PATCH 2/2] Update hoist-and-by-const-from-lshr-in-eqcmp-zero.ll
---
.../ARM/hoist-and-by-const-from-lshr-in-eqcmp-zero.ll | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)
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 36feb8c4f1f4b..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
@@ -881,9 +881,8 @@ define i1 @scalar_i8_signbit_ne(i8 %x, i8 %y) nounwind {
; THUMB6: @ %bb.0:
; THUMB6-NEXT: uxtb r1, r1
; THUMB6-NEXT: lsls r0, r1
-; THUMB6-NEXT: movs r1, #128
-; THUMB6-NEXT: ands r1, r0
-; THUMB6-NEXT: lsrs r0, r1, #7
+; THUMB6-NEXT: uxtb r0, r0
+; THUMB6-NEXT: lsrs r0, r0, #7
; THUMB6-NEXT: bx lr
;
; THUMB78-LABEL: scalar_i8_signbit_ne:
More information about the llvm-commits
mailing list