[llvm] ea9d639 - Fix reordering of instructions during VirtRegRewriter unbundling

[Carl Ritson via llvm-commits]

Author: Carl Ritson
Date: 2020-10-06T09:43:02+09:00
New Revision: ea9d6392f4a375b5183be02ac93eb3a75c169360

URL: https://github.com/llvm/llvm-project/commit/ea9d6392f4a375b5183be02ac93eb3a75c169360
DIFF: https://github.com/llvm/llvm-project/commit/ea9d6392f4a375b5183be02ac93eb3a75c169360.diff

LOG: Fix reordering of instructions during VirtRegRewriter unbundling

When unbundling COPY bundles in VirtRegRewriter the start of the
bundle is not correctly referenced in the unbundling loop.

The effect of this is that unbundled instructions are sometimes
inserted out-of-order, particular in cases where multiple
reordering have been applied to avoid clobbering dependencies.
The resulting instruction sequence clobbers dependencies.

Reviewed By: foad

Differential Revision: https://reviews.llvm.org/D88821

Added: 
    

Modified: 
    llvm/lib/CodeGen/VirtRegMap.cpp
    llvm/test/CodeGen/AMDGPU/splitkit-copy-bundle.mir

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/VirtRegMap.cpp b/llvm/lib/CodeGen/VirtRegMap.cpp
index 4837cf11bc33..024fedb2eee2 100644
--- a/llvm/lib/CodeGen/VirtRegMap.cpp
+++ b/llvm/lib/CodeGen/VirtRegMap.cpp
@@ -452,7 +452,7 @@ void VirtRegRewriter::expandCopyBundle(MachineInstr &MI) const {
       // instruction, the bundle will have been completely undone.
       if (BundledMI != BundleStart) {
         BundledMI->removeFromBundle();
-        MBB.insert(FirstMI, BundledMI);
+        MBB.insert(BundleStart, BundledMI);
       } else if (BundledMI->isBundledWithSucc()) {
         BundledMI->unbundleFromSucc();
         BundleStart = &*std::next(BundledMI->getIterator());

diff  --git a/llvm/test/CodeGen/AMDGPU/splitkit-copy-bundle.mir b/llvm/test/CodeGen/AMDGPU/splitkit-copy-bundle.mir
index 423eb928b481..05668eeeef7b 100644
--- a/llvm/test/CodeGen/AMDGPU/splitkit-copy-bundle.mir
+++ b/llvm/test/CodeGen/AMDGPU/splitkit-copy-bundle.mir
@@ -298,11 +298,11 @@ body:             |
     ; VR: SI_SPILL_S512_SAVE killed renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27, %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32 :: (store 64 into %stack.0, align 4, addrspace 5)
     ; VR: S_NOP 0, implicit-def $sgpr8, implicit-def $sgpr12, implicit-def $sgpr16, implicit-def $sgpr20, implicit-def $sgpr24, implicit-def $sgpr28, implicit-def $sgpr32, implicit-def $sgpr36, implicit-def $sgpr40, implicit-def $sgpr44, implicit-def $sgpr48, implicit-def $sgpr52, implicit-def $sgpr56, implicit-def $sgpr60, implicit-def $sgpr64, implicit-def $sgpr68, implicit-def $sgpr72, implicit-def $sgpr74, implicit-def $sgpr78, implicit-def $sgpr82, implicit-def $sgpr86, implicit-def $sgpr90, implicit-def $sgpr94, implicit-def $sgpr98
     ; VR: renamable $sgpr12_sgpr13_sgpr14_sgpr15_sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27 = SI_SPILL_S512_RESTORE %stack.0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3, implicit $sgpr32 :: (load 64 from %stack.0, align 4, addrspace 5)
-    ; VR: renamable $sgpr16 = COPY killed renamable $sgpr20
+    ; VR: renamable $sgpr12_sgpr13 = COPY killed renamable $sgpr16_sgpr17
     ; VR: renamable $sgpr15 = COPY killed renamable $sgpr19
     ; VR: renamable $sgpr18_sgpr19 = COPY killed renamable $sgpr22_sgpr23
+    ; VR: renamable $sgpr16 = COPY killed renamable $sgpr20
     ; VR: renamable $sgpr21 = COPY killed renamable $sgpr25
-    ; VR: renamable $sgpr12_sgpr13 = COPY renamable $sgpr16_sgpr17
     ; VR: renamable $sgpr22 = COPY killed renamable $sgpr26
     ; VR: renamable $sgpr4_sgpr5_sgpr6_sgpr7 = IMPLICIT_DEF
     ; VR: renamable $sgpr8 = S_BUFFER_LOAD_DWORD_SGPR renamable $sgpr4_sgpr5_sgpr6_sgpr7, killed renamable $sgpr12, 0, 0 :: (dereferenceable invariant load 4)