pcc wrote:
I'm fixing the code generation for the test cases that I'm adding
(inhibit-zext-constant-hoist.ll) which were all extracted from a build of a
large internal program built with CFI. Previously f1 looked like this where
align was hoisted:
```
f1: # @f1
.cfi_startproc
# %bb.0:
movl $__typeid__ZTS1S_align, %eax
movzbl %al, %ecx
movb $64, %al
subb %cl, %al
movzbl %al, %eax
testl %edi, %edi
je .LBB0_3
# %bb.1:
movq (%rsi), %r8
movl $__typeid__ZTS1S_global_addr, %edx
movq %r8, %rdi
subq %rdx, %rdi
movq %rdi, %rdx
# kill: def $cl killed $cl killed $rcx
shrq %cl, %rdx
movl %eax, %ecx
shlq %cl, %rdi
orq %rdx, %rdi
cmpq $__typeid__ZTS1S_size_m1@ABS8, %rdi
jbe .LBB0_4
.LBB0_2:
ud1l 2(%eax), %eax
.LBB0_3:
movq (%rdx), %r8
movl $__typeid__ZTS1S_global_addr, %esi
movq %r8, %rdi
subq %rsi, %rdi
movq %rdi, %rsi
# kill: def $cl killed $cl killed $rcx
shrq %cl, %rsi
movl %eax, %ecx
shlq %cl, %rdi
orq %rsi, %rdi
cmpq $__typeid__ZTS1S_size_m1@ABS8, %rdi
movq %rdx, %rsi
ja .LBB0_2
.LBB0_4:
movq %rsi, %rdi
jmpq *(%r8) # TAILCALL
.Lfunc_end0:
.size f1, .Lfunc_end0-f1
.cfi_endproc
```
Now f1 looks like this:
```
f1: # @f1
.cfi_startproc
# %bb.0:
testl %edi, %edi
je .LBB0_3
# %bb.1:
movq (%rsi), %rax
movl $__typeid__ZTS1S_global_addr, %ecx
movq %rax, %rdx
subq %rcx, %rdx
rorq $__typeid__ZTS1S_align, %rdx
cmpq $__typeid__ZTS1S_size_m1@ABS8, %rdx
jbe .LBB0_4
.LBB0_2:
ud1l 2(%eax), %eax
.LBB0_3:
movq (%rdx), %rax
movl $__typeid__ZTS1S_global_addr, %ecx
movq %rax, %rsi
subq %rcx, %rsi
rorq $__typeid__ZTS1S_align, %rsi
cmpq $__typeid__ZTS1S_size_m1@ABS8, %rsi
movq %rdx, %rsi
ja .LBB0_2
.LBB0_4:
movq %rsi, %rdi
jmpq *(%rax) # TAILCALL
.Lfunc_end0:
.size f1, .Lfunc_end0-f1
.cfi_endproc
```
The other cases look similar before my change.
The poor codegen issue was introduced (I believe) by
https://github.com/llvm/llvm-project/pull/71040 which removed the zext
ConstantExprs that LowerTypeTests was using to keep everything in the same
basic block for matching.
I think that because it's the zext being hoisted and not the shifts it wouldn't
make a difference to use fshl/fshr but I can check.
https://github.com/llvm/llvm-project/pull/141326
_______________________________________________
llvm-branch-commits mailing list
llvm-branch-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
