Skip to content

Commit

Permalink
[X86] preferABDSToABSWithNSW - use ABDS for i32/i64 if we have CMOV
Browse files Browse the repository at this point in the history 
Now that we have better ABDS lowering, prefer cmov(sub(x,y),sub(y,x)) to cmov(abs(sub(x,y)),sub(x,y)) to improve ILP
  • Loading branch information
@RKSimon
RKSimon committed Sep 5, 2024
1 parent fa38527 commit 1104056
Show file tree
Hide file tree
Showing 3 changed files with 45 additions and 35 deletions.
2 changes: 1 addition & 1 deletion llvm/lib/Target/X86/X86ISelLowering.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -58210,7 +58210,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
}

bool X86TargetLowering::preferABDSToABSWithNSW(EVT VT) const {
return false;
return Subtarget.canUseCMOV() && (VT == MVT::i32 || VT == MVT::i64);
}

// Prefer (non-AVX512) vector TRUNCATE(SIGN_EXTEND_INREG(X)) to use of PACKSS.
Expand Down
36 changes: 22 additions & 14 deletions llvm/test/CodeGen/X86/abds-neg.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1027,19 +1027,22 @@ define i16 @abd_subnsw_i16_undef(i16 %a, i16 %b) nounwind {
define i32 @abd_subnsw_i32(i32 %a, i32 %b) nounwind {
; X86-LABEL: abd_subnsw_i32:
; X86: # %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: subl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: movl %ecx, %eax
; X86-NEXT: movl %ecx, %edx
; X86-NEXT: subl %eax, %edx
; X86-NEXT: subl %ecx, %eax
; X86-NEXT: cmovll %edx, %eax
; X86-NEXT: negl %eax
; X86-NEXT: cmovnsl %ecx, %eax
; X86-NEXT: retl
;
; X64-LABEL: abd_subnsw_i32:
; X64: # %bb.0:
; X64-NEXT: subl %esi, %edi
; X64-NEXT: movl %edi, %eax
; X64-NEXT: subl %esi, %eax
; X64-NEXT: subl %edi, %esi
; X64-NEXT: cmovgel %esi, %eax
; X64-NEXT: negl %eax
; X64-NEXT: cmovnsl %edi, %eax
; X64-NEXT: retq
%sub = sub nsw i32 %a, %b
%abs = call i32 @llvm.abs.i32(i32 %sub, i1 false)
Expand All @@ -1050,19 +1053,22 @@ define i32 @abd_subnsw_i32(i32 %a, i32 %b) nounwind {
define i32 @abd_subnsw_i32_undef(i32 %a, i32 %b) nounwind {
; X86-LABEL: abd_subnsw_i32_undef:
; X86: # %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: subl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: movl %ecx, %eax
; X86-NEXT: movl %ecx, %edx
; X86-NEXT: subl %eax, %edx
; X86-NEXT: subl %ecx, %eax
; X86-NEXT: cmovll %edx, %eax
; X86-NEXT: negl %eax
; X86-NEXT: cmovnsl %ecx, %eax
; X86-NEXT: retl
;
; X64-LABEL: abd_subnsw_i32_undef:
; X64: # %bb.0:
; X64-NEXT: subl %esi, %edi
; X64-NEXT: movl %edi, %eax
; X64-NEXT: subl %esi, %eax
; X64-NEXT: subl %edi, %esi
; X64-NEXT: cmovgel %esi, %eax
; X64-NEXT: negl %eax
; X64-NEXT: cmovnsl %edi, %eax
; X64-NEXT: retq
%sub = sub nsw i32 %a, %b
%abs = call i32 @llvm.abs.i32(i32 %sub, i1 true)
Expand Down Expand Up @@ -1090,10 +1096,11 @@ define i64 @abd_subnsw_i64(i64 %a, i64 %b) nounwind {
;
; X64-LABEL: abd_subnsw_i64:
; X64: # %bb.0:
; X64-NEXT: subq %rsi, %rdi
; X64-NEXT: movq %rdi, %rax
; X64-NEXT: subq %rsi, %rax
; X64-NEXT: subq %rdi, %rsi
; X64-NEXT: cmovgeq %rsi, %rax
; X64-NEXT: negq %rax
; X64-NEXT: cmovnsq %rdi, %rax
; X64-NEXT: retq
%sub = sub nsw i64 %a, %b
%abs = call i64 @llvm.abs.i64(i64 %sub, i1 false)
Expand Down Expand Up @@ -1121,10 +1128,11 @@ define i64 @abd_subnsw_i64_undef(i64 %a, i64 %b) nounwind {
;
; X64-LABEL: abd_subnsw_i64_undef:
; X64: # %bb.0:
; X64-NEXT: subq %rsi, %rdi
; X64-NEXT: movq %rdi, %rax
; X64-NEXT: subq %rsi, %rax
; X64-NEXT: subq %rdi, %rsi
; X64-NEXT: cmovgeq %rsi, %rax
; X64-NEXT: negq %rax
; X64-NEXT: cmovnsq %rdi, %rax
; X64-NEXT: retq
%sub = sub nsw i64 %a, %b
%abs = call i64 @llvm.abs.i64(i64 %sub, i1 true)
Expand Down
42 changes: 22 additions & 20 deletions llvm/test/CodeGen/X86/abds.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -928,19 +928,20 @@ define i16 @abd_subnsw_i16_undef(i16 %a, i16 %b) nounwind {
define i32 @abd_subnsw_i32(i32 %a, i32 %b) nounwind {
; X86-LABEL: abd_subnsw_i32:
; X86: # %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: subl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: movl %ecx, %eax
; X86-NEXT: negl %eax
; X86-NEXT: cmovsl %ecx, %eax
; X86-NEXT: movl %ecx, %edx
; X86-NEXT: subl %eax, %edx
; X86-NEXT: subl %ecx, %eax
; X86-NEXT: cmovll %edx, %eax
; X86-NEXT: retl
;
; X64-LABEL: abd_subnsw_i32:
; X64: # %bb.0:
; X64-NEXT: subl %esi, %edi
; X64-NEXT: movl %edi, %eax
; X64-NEXT: negl %eax
; X64-NEXT: cmovsl %edi, %eax
; X64-NEXT: subl %esi, %eax
; X64-NEXT: subl %edi, %esi
; X64-NEXT: cmovgel %esi, %eax
; X64-NEXT: retq
%sub = sub nsw i32 %a, %b
%abs = call i32 @llvm.abs.i32(i32 %sub, i1 false)
Expand All @@ -950,19 +951,20 @@ define i32 @abd_subnsw_i32(i32 %a, i32 %b) nounwind {
define i32 @abd_subnsw_i32_undef(i32 %a, i32 %b) nounwind {
; X86-LABEL: abd_subnsw_i32_undef:
; X86: # %bb.0:
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: subl {{[0-9]+}}(%esp), %ecx
; X86-NEXT: movl %ecx, %eax
; X86-NEXT: negl %eax
; X86-NEXT: cmovsl %ecx, %eax
; X86-NEXT: movl %ecx, %edx
; X86-NEXT: subl %eax, %edx
; X86-NEXT: subl %ecx, %eax
; X86-NEXT: cmovll %edx, %eax
; X86-NEXT: retl
;
; X64-LABEL: abd_subnsw_i32_undef:
; X64: # %bb.0:
; X64-NEXT: subl %esi, %edi
; X64-NEXT: movl %edi, %eax
; X64-NEXT: negl %eax
; X64-NEXT: cmovsl %edi, %eax
; X64-NEXT: subl %esi, %eax
; X64-NEXT: subl %edi, %esi
; X64-NEXT: cmovgel %esi, %eax
; X64-NEXT: retq
%sub = sub nsw i32 %a, %b
%abs = call i32 @llvm.abs.i32(i32 %sub, i1 true)
Expand All @@ -986,10 +988,10 @@ define i64 @abd_subnsw_i64(i64 %a, i64 %b) nounwind {
;
; X64-LABEL: abd_subnsw_i64:
; X64: # %bb.0:
; X64-NEXT: subq %rsi, %rdi
; X64-NEXT: movq %rdi, %rax
; X64-NEXT: negq %rax
; X64-NEXT: cmovsq %rdi, %rax
; X64-NEXT: subq %rsi, %rax
; X64-NEXT: subq %rdi, %rsi
; X64-NEXT: cmovgeq %rsi, %rax
; X64-NEXT: retq
%sub = sub nsw i64 %a, %b
%abs = call i64 @llvm.abs.i64(i64 %sub, i1 false)
Expand All @@ -1013,10 +1015,10 @@ define i64 @abd_subnsw_i64_undef(i64 %a, i64 %b) nounwind {
;
; X64-LABEL: abd_subnsw_i64_undef:
; X64: # %bb.0:
; X64-NEXT: subq %rsi, %rdi
; X64-NEXT: movq %rdi, %rax
; X64-NEXT: negq %rax
; X64-NEXT: cmovsq %rdi, %rax
; X64-NEXT: subq %rsi, %rax
; X64-NEXT: subq %rdi, %rsi
; X64-NEXT: cmovgeq %rsi, %rax
; X64-NEXT: retq
%sub = sub nsw i64 %a, %b
%abs = call i64 @llvm.abs.i64(i64 %sub, i1 true)
Expand Down

0 comments on commit 1104056

Please sign in to comment.