[llvm] c40877d - [RISCV] Attach an implicit source operand on vector copies (#126155)

[via llvm-commits]

Author: Min-Yih Hsu
Date: 2025-02-08T16:25:27-08:00
New Revision: c40877d095eaa03d64e614723a69f1d68717f32a

URL: https://github.com/llvm/llvm-project/commit/c40877d095eaa03d64e614723a69f1d68717f32a
DIFF: https://github.com/llvm/llvm-project/commit/c40877d095eaa03d64e614723a69f1d68717f32a.diff

LOG: [RISCV] Attach an implicit source operand on vector copies (#126155)

Somtimes when we're breaking up a large vector copy into several smaller
ones, not every single smaller source registers are initialized at the
time when the original COPY happens, and the verifier will not be
pleased when seeing the smaller copies reading from an undef register.
This patch is a workaround for the said issue by attaching an implicit
read of the source operand on the newly generated copies.

This is tested by llvm/test/CodeGen/RISCV/rvv/vector-interleave.ll which
would have crashed the compiler without this fix when
LLVM_EXPENSIVE_CHECK is enabled. Original context:
https://github.com/llvm/llvm-project/pull/124825#issuecomment-2639097531

---------

Co-authored-by: Craig Topper <craig.topper at sifive.com>

Added: 
    llvm/test/CodeGen/RISCV/postra-copy-expand.mir

Modified: 
    llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
index 773319ba908c87a..1ec299e3c8cc0a6 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp
@@ -437,6 +437,11 @@ void RISCVInstrInfo::copyPhysRegVector(
       MIB.addReg(RISCV::VL, RegState::Implicit);
       MIB.addReg(RISCV::VTYPE, RegState::Implicit);
     }
+    // Add an implicit read of the original source to silence the verifier
+    // in the cases where some of the smaller VRs we're copying from might be
+    // undef, caused by the fact that the original, larger source VR might not
+    // be fully initialized at the time this COPY happens.
+    MIB.addReg(SrcReg, RegState::Implicit);
 
     // If we are copying reversely, we should decrease the encoding.
     SrcEncoding += (ReversedCopy ? -NumCopied : NumCopied);

diff  --git a/llvm/test/CodeGen/RISCV/postra-copy-expand.mir b/llvm/test/CodeGen/RISCV/postra-copy-expand.mir
new file mode 100644
index 000000000000000..e5b85659a0340c8
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/postra-copy-expand.mir
@@ -0,0 +1,24 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=riscv64 -mattr=+v -run-pass=postrapseudos %s -o - | FileCheck %s
+
+---
+name:            copy
+isSSA:           false
+noVRegs:         true
+liveins:
+  - { reg: '$v0', virtual-reg: '' }
+body:             |
+  bb.0:
+    liveins: $v0
+
+    ; CHECK-LABEL: name: copy
+    ; CHECK: liveins: $v0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: $v20m2 = VMV2R_V $v14m2, implicit $vtype, implicit $v14_v15_v16_v17_v18
+    ; CHECK-NEXT: $v22m2 = VMV2R_V $v16m2, implicit $vtype, implicit $v14_v15_v16_v17_v18
+    ; CHECK-NEXT: $v24 = VMV1R_V $v18, implicit $vtype, implicit $v14_v15_v16_v17_v18, implicit $vtype
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    renamable $v20_v21_v22_v23_v24 = COPY renamable $v14_v15_v16_v17_v18, implicit $vtype
+    PseudoRET implicit $v0
+
+...