[AMDGPU] Fix error in #88512.

[PeddleSpam]

Fixes error in GlobalISel CTLZ lowering caused by #88512.

[llvmbot]

@llvm/pr-subscribers-llvm-globalisel

@llvm/pr-subscribers-backend-amdgpu

Author: Leon Clark (PeddleSpam)

Changes

• Reapply "[ctx_profile] Integration test (#92456)"
• [Github] Revert accidental changes to dependabot config
• Fix: remove wrongly pushed etime-function.mlir at toplevel (#92634)
• [MCAsmParser] .macro/.rept/.irp/.irpc: remove excess \n after expansion
• [flang][OpenMP] Re-enable tests when building OpenMP as a runtime (#89046)
• [flang][OpenMP] Try to unify induction var privatization for OMP regions. (#91116)
• [MCAsmParser] Improve .rept/.irp tests
• [clang][ThreadSafety] Skip past implicit cast in translateAttrExpr
• [clang][NFC] Further improvements to const-correctness
• [GlobalIsel] Combine select to integer min max more (#92570)
• [X86][CodeGen] Support flags copy lowering for CCMP/CTEST (#91849)
• [mlir] Add operator<< for printing Block (#92550)
• [flang][cuf] Add attr gen dependency to fix #92635
• [nfc][ctx_profile] Fix printf - related -Wformat-pedantic
• [NVPTX] support immediate values in st.param instructions (#91523)
• [VPlan] Remove unused removeLastOperand (NFC).
• [dsymutil] Use operator==(StringRef, StringRef) (NFC)
• [DWARFLinker] Use an implicit conversion of SmallString to StringRef (NFC)
• [DXIL] Use consistent SmallVector parameters
• [DAG] Use copysign in frem power-2 fold. (#91751)
• [VectorCombine] Don't transform single shuffles in shuffleToIdentity
• update_test_checks: match IR basic block labels (#88979)
• [ThinLTO]Sort imported GUIDs before cache key update (#92622)
• [nfc][InstrFDO]Encapsulate header writes in a class member function (#90142)
• Reformat
• Quick fix for a waning in clang_rt.ctx_profile [-Wgnu-anonymous-struct]
• [NewPM][AMDGPU] Add CodeGenPassBuilder (#91040)
• [gn build] Port b4ba3fe
• [GISel][RISCV] Legalize G_CONSTANT_FOLD_BARRIER (#89960)
• [VectorCombine] Additional extend tests for shuffleToIdentity. NFC
• [DAG] canCreateUndefOrPoison - merge INSERT_VECTOR_ELT/EXTRACT_VECTOR_ELT cases. NFC.
• [ctx_profile] Pass lib path into test
• [DAG] canCreateUndefOrPoison - only compute extract/index vector elt index knownbits when not poison
• [DAG] visitAVG - rewrite "fold (avgfloor x, 0) -> x >> 1" to use SDPatternMatch
• [DAG] visitABD - rewrite "(abs x, 0)" folds to use SDPatternMatch
• Revert "[Bounds-Safety] Temporarily relax a counted_by attribute restriction on flexible array members"
• Revert "[BoundsSafety] Allow 'counted_by' attribute on pointers in structs in C (#90786)"
• Revert "[Bounds-Safety] Fix pragma-attribute-supported-attributes-list.test"
• [Clang][CodeGen] Start migrating away from assuming the Default AS is 0 (#88182)
• [CodeGen][SDAG] Skip preferred extend at O0 (#92643)
• [CodeGen][SDAG] Track returntwice in lowering info (#92640)
• [llvm] Add KnownBits implementations for avgFloor and avgCeil (#86445)
• SimplifyLibCalls: Permit pow(2, x) -> ldexp(1, x) fold for vectors (#92532)
• [VPlan] Simplify (X && Y) || (X && !Y) -> X. (#89386)
• HLSL availability diagnostics design doc (#92207)
• [DOCS] ORCv2.rst Typo (#89482)
• [Clang][HLSL] Add environment parameter to availability attribute (#89809)
• ValueTracking: Correct undef handling for constant FP vectors (#92557)
• [BOLT] Fix preserved offset in fixDoubleJumps (#92485)
• [AMDGPU] Use LSH for lowering ctlz_zero_undef.i8/i16 (#88512)
• [TableGen] Avoid std::string copy. NFC
• Update llvm-bugs.yml (#77243)
• [llvm] Use operator==(StringRef, StringRef) (NFC) (#92705)
• [clang-format][NFC] Clean up SortIncludesTest.cpp
• [mlir] Use operator==(StringRef, StringRef) (NFC) (#92706)
• [CallPromotionUtils]Implement conditional indirect call promotion with vtable-based comparison (#81378)
• [clang] Use operator==(StringRef, StringRef) (NFC) (#92708)
• [SDAG][X86] Extend SplitVecOp_VSETCC for STRICT_FSETCC. (#92509)
• [llvm] Use StringRef::contains (NFC) (#92710)
• [Serialization] Read the initializer for interesting static variables before consuming it (#92353)
• [BOLT][NFC] Don't assign YAML profile to functions with no CFG (#92487)
• [InstCombine] Fold pointer adding in integer to arithmetic add (#91596)
• [AMDGPU] Use removeFnAttrFromReachable in lower-module-lds pass. (#92686)
• [AMDGPU] Fix kernarg preloading crash with some types and alignments (#91625)
• [ThinLTO] Populate declaration import status except for distributed ThinLTO under a default-off new option (#88024)
• [NFC] Remove unused ASTWriter::getTypeID
• [SCEV] Don't use non-deterministic constant folding for trip counts (#90942)
• Revert "[ThinLTO] Populate declaration import status except for distributed ThinLTO under a default-off new option" (#92715)
• [llvm] Use SmallString::str (NFC) (#92712)
• [AMDGPU] Only set Info.memVT when not later overridden (#92670)
• [MC] Make UseAssemblerInfoForParsing mostly true
• MIPS: Support '%w' token in inline asm template for MSA (#91920)
• Clang/MIPS: Add +fp64 if MSA and no explicit -mfp option (#91949)
• MIPS/Clang: Use FP32 by default if CPU is mips1 (#92122)
• [ELF] Support high address DW_EH_sdata4 for ELFCLASS32
• [PowerPC]perform bitcast lowering only at 64 bit
• [LoongArch] Select {DIV,MOD}.{W,WU} instruction to eliminate explicit sign extension (#92205)
• [Clang] Fix __is_array returning true for zero-sized arrays (#86652)
• [OpenCL] Add cl_khr_kernel_clock builtins (#91950)
• [clang][ExtractAPI] Remove symbols defined in categories to external types unless requested (#92522)
• [RISCV][CostModel] Remove cost of icmp inst in icmp+select with SFB. (#91158)
• [DebugInfo][GVNSink] Fix #77415: GVNSink fails to optimize LLVM IR with debug info (#77602)
• [AArch64] Add PreTest for optimizing MOV to ORR
• [Driver][PS5] Set visibility option defaults (#92091)
• [AArch64] Optimize MOV to ORR when load symmetric constants (#86249)
• [Coverage] Rework !SystemHeadersCoverage (#91446)
• [lldb][Windows] Fixed LibcxxChronoTimePointSecondsSummaryProvider() (#92701)
• [ConstantFolding] Canonicalize constexpr GEPs to i8 (#89872)
• InstSimplify: increase shufflevector test coverage (#92407)
• [flang][HLFIR] Adapt SimplifyHLFIRIntrinsics to run on all top level ops (#92573)
• movimm-expand-ldst.mir (d3d6565) requires asserts
• [SLP] NFC. Use TreeEntry::getOperand if setOperandsInOrder is called (#92727)
• [MLIR][OpenMP] NFC: Split OpenMP dialect definitions (#91741)
• [mlir][irdl] Fix missing verifier in irdl.parametric (#92700)
• [VPlan] Add commutative binary OR matcher, use in transform. (#92539)
• [CloneFunction] Remove check that is no longer necessary (#92577)
• [ValueTracking] Fix incorrect inferrence about the signbit of sqrt (#92510)
• [LAA] Add tests with invariant accesses using vector types.
• [clang] CTAD alias: Fix missing template arg packs during the transformation (#92535)
• [TableGen] HasOneUse builtin predicate on PatFrags (#91578)
• [clang] Make PS template DLL attribute propagation the same as MSVC (#92549)
• [DebugInfo][NaryReassociate] Fix missing debug location updates (#92545)
• [clang] Use SmallString::str (NFC) (#92717)
• [libcxx] locale.cpp: Move build_name helper into unnamed namespace (#92461)
• [Offload] Remove unused version script for plugins
• [AMDGPU] Fix error in #88512.

---

Full diff: https://github.com/llvm/llvm-project/pull/92770.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp (+1-1)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index 15a4b6796880f..3523fcc7dbd50 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -4168,7 +4168,7 @@ bool AMDGPULegalizerInfo::legalizeCTLZ_ZERO_UNDEF(MachineInstr &MI,
 
   auto ShiftAmt = B.buildConstant(S32, 32u - NumBits);
   auto Extend = B.buildAnyExt(S32, {Src}).getReg(0u);
-  auto Shift = B.buildLShr(S32, {Extend}, ShiftAmt);
+  auto Shift = B.buildShl(S32, {Extend}, ShiftAmt);
   auto Ctlz = B.buildInstr(AMDGPU::G_AMDGPU_FFBH_U32, {S32}, {Shift});
   B.buildTrunc(Dst, Ctlz);
   MI.eraseFromParent();

[jayfoad]

This ought to require an update to the tests too.