[llvm] 0289dad - [InstCombine] Add folds for (icmp eq/ne (and (add/sub/xor A, P2), P2), 0/P2)
Noah Goldstein via llvm-commits
llvm-commits at lists.llvm.org
Fri Oct 27 15:36:48 PDT 2023
Author: Noah Goldstein
Date: 2023-10-27T17:36:30-05:00
New Revision: 0289dad538fa2fdc1a82a26f5d19f94fbd20d949
URL: https://github.com/llvm/llvm-project/commit/0289dad538fa2fdc1a82a26f5d19f94fbd20d949
DIFF: https://github.com/llvm/llvm-project/commit/0289dad538fa2fdc1a82a26f5d19f94fbd20d949.diff
LOG: [InstCombine] Add folds for (icmp eq/ne (and (add/sub/xor A, P2), P2), 0/P2)
- `(icmp eq/ne (and (add/sub/xor X, P2), P2), P2)`
-> `(icmp eq/ne (and X, P2), 0)`
- `(icmp eq/ne (and (add/sub/xor X, P2), P2), 0)`
-> `(icmp eq/ne (and X, P2), P2)`
Folds like this come up with reasonable regularity in odd/even loops.
Proofs: https://alive2.llvm.org/ce/z/45pq2x
Closes #67836
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
llvm/test/Transforms/InstCombine/icmp-and-add-sub-xor-p2.ll
llvm/test/Transforms/InstCombine/pr25342.ll
llvm/test/Transforms/PGOProfile/cspgo_profile_summary.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index a09c9b48be9d5b2..d254c4706e0aacd 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -5483,6 +5483,32 @@ Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
m_CombineAnd(m_Value(B), m_Unless(m_ImmConstant())))))
return new ICmpInst(Pred, Builder.CreateXor(A, B), Cst);
+ {
+ // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
+ auto m_Matcher =
+ m_CombineOr(m_CombineOr(m_c_Add(m_Value(B), m_Deferred(A)),
+ m_c_Xor(m_Value(B), m_Deferred(A))),
+ m_Sub(m_Value(B), m_Deferred(A)));
+ std::optional<bool> IsZero = std::nullopt;
+ if (match(&I, m_c_ICmp(PredUnused, m_OneUse(m_c_And(m_Value(A), m_Matcher)),
+ m_Deferred(A))))
+ IsZero = false;
+ // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
+ else if (match(&I,
+ m_ICmp(PredUnused, m_OneUse(m_c_And(m_Value(A), m_Matcher)),
+ m_Zero())))
+ IsZero = true;
+
+ if (IsZero && isKnownToBeAPowerOfTwo(A, /* OrZero */ true, /*Depth*/ 0, &I))
+ // (icmp eq/ne (and (add/sub/xor X, P2), P2), P2)
+ // -> (icmp eq/ne (and X, P2), 0)
+ // (icmp eq/ne (and (add/sub/xor X, P2), P2), 0)
+ // -> (icmp eq/ne (and X, P2), P2)
+ return new ICmpInst(Pred, Builder.CreateAnd(B, A),
+ *IsZero ? A
+ : ConstantInt::getNullValue(A->getType()));
+ }
+
return nullptr;
}
diff --git a/llvm/test/Transforms/InstCombine/icmp-and-add-sub-xor-p2.ll b/llvm/test/Transforms/InstCombine/icmp-and-add-sub-xor-p2.ll
index b9796e448630196..c8a3dfcd68cd469 100644
--- a/llvm/test/Transforms/InstCombine/icmp-and-add-sub-xor-p2.ll
+++ b/llvm/test/Transforms/InstCombine/icmp-and-add-sub-xor-p2.ll
@@ -9,8 +9,8 @@ define i1 @src_add_eq_p2(i8 %x, i8 %yy) {
; CHECK-NEXT: [[Y:%.*]] = and i8 [[NY]], [[YY]]
; CHECK-NEXT: [[X1:%.*]] = add i8 [[Y]], [[X:%.*]]
; CHECK-NEXT: call void @use.i8(i8 [[X1]])
-; CHECK-NEXT: [[V:%.*]] = and i8 [[X1]], [[Y]]
-; CHECK-NEXT: [[R:%.*]] = icmp eq i8 [[V]], [[Y]]
+; CHECK-NEXT: [[TMP1:%.*]] = and i8 [[Y]], [[X]]
+; CHECK-NEXT: [[R:%.*]] = icmp eq i8 [[TMP1]], 0
; CHECK-NEXT: ret i1 [[R]]
;
%ny = sub i8 0, %yy
@@ -49,8 +49,8 @@ define i1 @src_xor_ne_zero(i8 %x, i8 %yy) {
; CHECK-NEXT: [[Y:%.*]] = and i8 [[NY]], [[YY]]
; CHECK-NEXT: [[X1:%.*]] = xor i8 [[Y]], [[X:%.*]]
; CHECK-NEXT: call void @use.i8(i8 [[X1]])
-; CHECK-NEXT: [[V:%.*]] = and i8 [[X1]], [[Y]]
-; CHECK-NEXT: [[R:%.*]] = icmp ne i8 [[V]], 0
+; CHECK-NEXT: [[TMP1:%.*]] = and i8 [[Y]], [[X]]
+; CHECK-NEXT: [[R:%.*]] = icmp ne i8 [[TMP1]], [[Y]]
; CHECK-NEXT: ret i1 [[R]]
;
%ny = sub i8 0, %yy
@@ -89,8 +89,8 @@ define <2 x i1> @src_sub_ne_p2(<2 x i8> %x, <2 x i8> %yy) {
; CHECK-NEXT: [[Y:%.*]] = and <2 x i8> [[NY]], [[YY]]
; CHECK-NEXT: [[X1:%.*]] = sub <2 x i8> [[X:%.*]], [[Y]]
; CHECK-NEXT: call void @use.v2i8(<2 x i8> [[X1]])
-; CHECK-NEXT: [[V:%.*]] = and <2 x i8> [[X1]], [[Y]]
-; CHECK-NEXT: [[R:%.*]] = icmp ne <2 x i8> [[V]], [[Y]]
+; CHECK-NEXT: [[TMP1:%.*]] = and <2 x i8> [[Y]], [[X]]
+; CHECK-NEXT: [[R:%.*]] = icmp ne <2 x i8> [[TMP1]], zeroinitializer
; CHECK-NEXT: ret <2 x i1> [[R]]
;
%ny = sub <2 x i8> zeroinitializer, %yy
@@ -107,8 +107,8 @@ define <2 x i1> @src_sub_eq_zero(<2 x i8> %x, <2 x i8> %yy) {
; CHECK-NEXT: [[Y:%.*]] = shl <2 x i8> <i8 1, i8 2>, [[YY:%.*]]
; CHECK-NEXT: [[X1:%.*]] = sub <2 x i8> [[X:%.*]], [[Y]]
; CHECK-NEXT: call void @use.v2i8(<2 x i8> [[X1]])
-; CHECK-NEXT: [[V:%.*]] = and <2 x i8> [[X1]], [[Y]]
-; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[V]], zeroinitializer
+; CHECK-NEXT: [[TMP1:%.*]] = and <2 x i8> [[Y]], [[X]]
+; CHECK-NEXT: [[R:%.*]] = icmp eq <2 x i8> [[TMP1]], [[Y]]
; CHECK-NEXT: ret <2 x i1> [[R]]
;
%y = shl <2 x i8> <i8 1, i8 2>, %yy
diff --git a/llvm/test/Transforms/InstCombine/pr25342.ll b/llvm/test/Transforms/InstCombine/pr25342.ll
index e1a6822e7908474..2f85f99c4ce003e 100644
--- a/llvm/test/Transforms/InstCombine/pr25342.ll
+++ b/llvm/test/Transforms/InstCombine/pr25342.ll
@@ -78,7 +78,7 @@ define void @multi_phi(i32 signext %n) {
; CHECK-NEXT: entry:
; CHECK-NEXT: br label [[FOR_COND:%.*]]
; CHECK: for.cond:
-; CHECK-NEXT: [[TMP0:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[TMP6:%.*]], [[ODD_BB:%.*]] ]
+; CHECK-NEXT: [[TMP0:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[TMP7:%.*]], [[ODD_BB:%.*]] ]
; CHECK-NEXT: [[I_0:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC:%.*]], [[ODD_BB]] ]
; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[I_0]], [[N:%.*]]
; CHECK-NEXT: br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_END:%.*]]
@@ -92,14 +92,14 @@ define void @multi_phi(i32 signext %n) {
; CHECK-NEXT: [[SUB_I:%.*]] = fsub float [[MUL_I]], [[MUL4_I]]
; CHECK-NEXT: [[ADD_I:%.*]] = fadd float [[SUB_I]], [[TMP0]]
; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[I_0]], 1
-; CHECK-NEXT: [[BIT0:%.*]] = and i32 [[INC]], 1
-; CHECK-NEXT: [[EVEN_NOT_NOT:%.*]] = icmp eq i32 [[BIT0]], 0
-; CHECK-NEXT: br i1 [[EVEN_NOT_NOT]], label [[EVEN_BB:%.*]], label [[ODD_BB]]
+; CHECK-NEXT: [[TMP5:%.*]] = and i32 [[I_0]], 1
+; CHECK-NEXT: [[EVEN_NOT_NOT_NOT:%.*]] = icmp eq i32 [[TMP5]], 0
+; CHECK-NEXT: br i1 [[EVEN_NOT_NOT_NOT]], label [[ODD_BB]], label [[EVEN_BB:%.*]]
; CHECK: even.bb:
-; CHECK-NEXT: [[TMP5:%.*]] = fadd float [[SUB_I]], [[ADD_I]]
+; CHECK-NEXT: [[TMP6:%.*]] = fadd float [[SUB_I]], [[ADD_I]]
; CHECK-NEXT: br label [[ODD_BB]]
; CHECK: odd.bb:
-; CHECK-NEXT: [[TMP6]] = phi float [ [[ADD_I]], [[FOR_BODY]] ], [ [[TMP5]], [[EVEN_BB]] ]
+; CHECK-NEXT: [[TMP7]] = phi float [ [[ADD_I]], [[FOR_BODY]] ], [ [[TMP6]], [[EVEN_BB]] ]
; CHECK-NEXT: br label [[FOR_COND]]
; CHECK: for.end:
; CHECK-NEXT: store float [[TMP0]], ptr @dd, align 4
diff --git a/llvm/test/Transforms/PGOProfile/cspgo_profile_summary.ll b/llvm/test/Transforms/PGOProfile/cspgo_profile_summary.ll
index b4a27ee65613365..7040bac6a4c43f0 100644
--- a/llvm/test/Transforms/PGOProfile/cspgo_profile_summary.ll
+++ b/llvm/test/Transforms/PGOProfile/cspgo_profile_summary.ll
@@ -104,7 +104,7 @@ for.end:
; CSPGOSUMMARY-LABEL: @foo
; CSPGOSUMMARY: %even.odd.i = select i1 %tobool.i{{[0-9]*}}, ptr @even, ptr @odd
; CSPGOSUMMARY-SAME: !prof ![[BW_CSPGO_BAR]]
-; CSPGOSUMMARY: %even.odd.i2 = select i1 %tobool.i{{[0-9]*}}, ptr @even, ptr @odd
+; CSPGOSUMMARY: %even.odd.i2 = select i1 %tobool.i{{[0-9]*}}, ptr @odd, ptr @even
; CSPGOSUMMARY-SAME: !prof ![[BW_CSPGO_BAR]]
declare dso_local i32 @bar_m(i32)
More information about the llvm-commits
mailing list