check_cfc: dash g generate different code with machine-cp

[llvmbot]

Bugzilla Link	44422
Version	trunk
OS	Linux
Depends On	#43779
Blocks	#37076
Reporter	LLVM Bugzilla Contributor
CC	@jmorse,@pogo59

Extended Description

Debug Info should have no effect on codegen, using check_cfc to scan llvm_test_suite, find a bug here:

---

[Reproduce]
$HOME/llvm-project/clang/utils/check_cfc/clang++ $HOME/test-suite/MultiSource/Benchmarks/Bullet/btConvexConvexAlgorithm.cpp -w -c -O2 -I $HOME/test-suite/MultiSource/Benchmarks/Bullet/include -o tmp.ll

[result]
Check CFC, checking: dash_g_no_change, dash_s_no_change
/home/chris/test-suite/MultiSource/Benchmarks/Bullet/btConvexConvexAlgorithm.cpp Code difference detected with -g
--- /tmp/tmpKmXkNI.o

+++ /tmp/tmp7GmGG6.o

@@ -254,6 +254,6 @@

  39c:	f3 41 0f 10 74 85 28 	movss  0x28(%r13,%rax,4),%xmm6
  3a3:	f3 45 0f 10 74 8d 28 	movss  0x28(%r13,%rcx,4),%xmm14
  3aa:	75 06                	jne    3b2 <_ZN23btConvexConvexAlgorithm16processCollisionEP17btCollisionObjectS1_RK16btDispatcherInfoP16btManifoldResult+0x1e2>

• 3ac:	0f 8b 7f 11 00 00    	jnp    1531 <_ZN23btConvexConvexAlgorithm16processCollisionEP17btCollisionObjectS1_RK16btDispatcherInfoP16btManifoldResult+0x1361>
  

• 3ac:	0f 8b 82 11 00 00    	jnp    1534 <_ZN23btConvexConvexAlgorithm16processCollisionEP17btCollisionObjectS1_RK16btDispatcherInfoP16btManifoldResult+0x1364>
  3b2:	0f 28 cd             	movaps %xmm5,%xmm1
  3b5:	f3 0f 59 cf          	mulss  %xmm7,%xmm1
  

*** Diff truncated ***

[llvmbot]

Analysis about this issue:

[Simply reproduce]
clang++ $HOME/test-suite/MultiSource/Benchmarks/Bullet/btConvexConvexAlgorithm.cpp -w -c -mllvm -opt-bisect-limit=6173 -O2 -I $HOME/test-suite/MultiSource/Benchmarks/Bullet/include -o out1.o 2>log1
objdump -d -j .text out1.o > out1.obj
clang++ -g $HOME/test-suite/MultiSource/Benchmarks/Bullet/btConvexConvexAlgorithm.cpp -w -c -mllvm -opt-bisect-limit=6179 -O2 -I $HOME/test-suite/MultiSource/Benchmarks/Bullet/include -o out2.o 2>log2
objdump -d -j .text out2.o > out2.obj
colordiff out1.obj out2.obj

[Root cause]
the DBG_VALUE break (669) Machine Copy Propagation PASS to handle BackwardCopyPropagateBlock() for LEA64 and COPY.

Case1: (with no debug value)

BB:
renamable $rbp = nuw LEA64r %stack.4, 1, $noreg, 64, $noreg
renamable $r8 = COPY killed renamable $rbp

Result:
MCP: Replacing $rbp
with $r8
renamable $r8 = nuw LEA64r %stack.4.input, 1, $noreg, 64, $noreg

Case2: (-g: with debug value)

BB:
renamable $rbp = nuw LEA64r %stack.4, 1, $noreg, 64, $noreg
DBG_VALUE
renamable $r8 = COPY killed renamable $rbp

Result: (MCP will not replace $rbp with $r8)
renamable $rbp = nuw LEA64r %stack.4, 1, $noreg, 64, $noreg
DBG_VALUE
renamable $r8 = COPY killed renamable $rbp

[llvmbot]

[Root cause]
the DBG_VALUE break (669) Machine Copy Propagation PASS to handle
BackwardCopyPropagateBlock() for LEA64 and COPY.

Correct "669" with "6173":
BISECT: running pass (6173) Machine Copy Propagation Pass on function (_ZN23btConvexConvexAlgorithm16processCollisionEP17btCollisionObjectS1_RK16btDispatcherInfoP16btManifoldResult)

[llvmbot]

mentioned in issue #43779