[llvm] [AMDGPU] Implement computeKnownBitsForTargetNode for amdgcn_ubfe and amdgcn_sbfe Intrinsics. (PR #200560)
Maxim Kuznetsov via llvm-commits
llvm-commits at lists.llvm.org
Sat May 30 02:48:46 PDT 2026
https://github.com/lakreite created https://github.com/llvm/llvm-project/pull/200560
For ubfe, the upper (32 - width) bits of the result are always zero.
For sbfe, the upper bits are sign-extended from sign bit.
For zero width result is always zero.
In both cases known bits from the source operand are shifted down by offset before being trunced, extended and returned.
Partially addresses https://github.com/llvm/llvm-project/issues/128390.
>From 375ad0615b2f261b235ac13115bdfd5b027d4ee6 Mon Sep 17 00:00:00 2001
From: Max Kuznetsov <lakreite at gmail.com>
Date: Sat, 30 May 2026 12:35:11 +0300
Subject: [PATCH] [AMDGPU] Implement computeKnownBitsForTargetNode for
amdgcn_ubfe and amdgcn_sbfe intrinsics.
---
llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 29 ++++++++++++
.../CodeGen/AMDGPU/compute-known-bits-bfe.ll | 46 +++++++++++++++++++
2 files changed, 75 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/compute-known-bits-bfe.ll
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 11f30f6210939..794791758212f 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -6147,6 +6147,35 @@ void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
Known.Zero.setHighBits(llvm::countl_zero(MaxValue));
break;
}
+ case Intrinsic::amdgcn_ubfe:
+ case Intrinsic::amdgcn_sbfe: {
+ bool IsSigned = (IID == Intrinsic::amdgcn_sbfe);
+ SDValue Src = Op.getOperand(1);
+ auto *OffsetNode = dyn_cast<ConstantSDNode>(Op.getOperand(2));
+ auto *WidthNode = dyn_cast<ConstantSDNode>(Op.getOperand(3));
+ if (!OffsetNode || !WidthNode)
+ break;
+
+ KnownBits SrcKnown = DAG.computeKnownBits(Src, Depth + 1);
+
+ unsigned Offset = OffsetNode->getZExtValue() & 31;
+ uint64_t Width = std::min(WidthNode->getZExtValue(),
+ 32 - static_cast<uint64_t>(Offset));
+
+ SrcKnown.Zero.ashrInPlace(Offset);
+ SrcKnown.One.ashrInPlace(Offset);
+
+ SrcKnown.One &= APInt::getLowBitsSet(32, Width);
+
+ if (Width == 0)
+ SrcKnown.Zero |= APInt::getAllOnes(32);
+ else
+ SrcKnown = IsSigned ? SrcKnown.trunc(Width).sext(32)
+ : SrcKnown.trunc(Width).zext(32);
+
+ Known = SrcKnown;
+ break;
+ }
default:
break;
}
diff --git a/llvm/test/CodeGen/AMDGPU/compute-known-bits-bfe.ll b/llvm/test/CodeGen/AMDGPU/compute-known-bits-bfe.ll
new file mode 100644
index 0000000000000..1a8e7442450cb
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/compute-known-bits-bfe.ll
@@ -0,0 +1,46 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GCN %s
+
+define i32 @sbfe_sext(i32 %src) nounwind {
+; GCN-LABEL: sbfe_sext:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %mask = xor i32 128, -1
+ %and = and i32 %src, %mask
+ %sbfe = call i32 @llvm.amdgcn.sbfe.i32(i32 %and, i32 4, i32 4)
+ %res = and i32 %sbfe, 240
+ ret i32 %res
+}
+
+define i32 @sbfe_width_zero(i32 %src) nounwind {
+; GCN-LABEL: sbfe_width_zero:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %res = call i32 @llvm.amdgcn.sbfe.i32(i32 %src, i32 4, i32 0)
+ ret i32 %res
+}
+
+define i32 @ubfe_sext(i32 %src) nounwind {
+; GCN-LABEL: ubfe_sext:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %sbfe = call i32 @llvm.amdgcn.ubfe.i32(i32 %src, i32 4, i32 4)
+ %res = and i32 %sbfe, 240
+ ret i32 %res
+}
+
+define i32 @ubfe_width_zero(i32 %src) nounwind {
+; GCN-LABEL: ubfe_width_zero:
+; GCN: ; %bb.0:
+; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT: v_mov_b32_e32 v0, 0
+; GCN-NEXT: s_setpc_b64 s[30:31]
+ %res = call i32 @llvm.amdgcn.ubfe.i32(i32 %src, i32 4, i32 0)
+ ret i32 %res
+}
More information about the llvm-commits
mailing list