[InstCombine] Canonicalize `and(zext(A), B)` into `select A, B & 1, 0` #66740

dtcxzyw · 2023-09-19T06:15:06Z

This patch canonicalizes the pattern and(zext(A), B) into select A, B & 1, 0. Thus, we can reuse transforms select B == even, B & 1, 0 -> 0 and select B == odd, B & 1, 0 -> zext(B == odd) in InstCombine.
It is an alternative to #66676.
Alive2: https://alive2.llvm.org/ce/z/598phE
Fixes #66733.
Fixes #66606.
Fixes #28612.

… & 1, 0`

llvmbot · 2023-09-19T06:16:17Z

@llvm/pr-subscribers-llvm-transforms

Changes

This patch canonicalizes the pattern and(zext(A), B) into select A, B & 1, 0. Thus, we can reuse transforms select B == even, B & 1, 0 -> 0 and select B == odd, B & 1, 0 -> zext(B == odd) in InstCombine.
It is an alternative to #66676.
Alive2: https://alive2.llvm.org/ce/z/598phE
Fixes #66606.

Full diff: https://github.com/llvm/llvm-project/pull/66740.diff

6 Files Affected:

(modified) llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp (+6)
(modified) llvm/test/Transforms/InstCombine/and-or-icmps.ll (+5-7)
(modified) llvm/test/Transforms/InstCombine/and.ll (+10-18)
(modified) llvm/test/Transforms/InstCombine/icmp.ll (+16-15)
(modified) llvm/test/Transforms/InstCombine/narrow.ll (+1-2)
(modified) llvm/test/Transforms/InstCombine/zext-or-icmp.ll (+5-5)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
index ca7dfa82ab5a5d7..7fa432e1fdc1790 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp
@@ -2654,6 +2654,12 @@ Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) {
       A->getType()->isIntOrIntVectorTy(1))
     return SelectInst::Create(A, Constant::getNullValue(Ty), B);
 
+  // and(zext(A), B) -> A ? (B & 1) : 0
+  if (match(&I, m_c_And(m_OneUse(m_ZExt(m_Value(A))), m_Value(B))) &&
+      A->getType()->isIntOrIntVectorTy(1))
+    return SelectInst::Create(A, Builder.CreateAnd(B, ConstantInt::get(Ty, 1)),
+                              Constant::getNullValue(Ty));
+
   // (iN X s>> (N-1)) & Y --> (X s< 0) ? Y : 0 -- with optional sext
   if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf(
                             m_AShr(m_Value(X), m_APIntAllowUndef(C)))),
diff --git a/llvm/test/Transforms/InstCombine/and-or-icmps.ll b/llvm/test/Transforms/InstCombine/and-or-icmps.ll
index 065dbf261e131bf..881a9b7ff129dbb 100644
--- a/llvm/test/Transforms/InstCombine/and-or-icmps.ll
+++ b/llvm/test/Transforms/InstCombine/and-or-icmps.ll
@@ -2769,9 +2769,9 @@ define i64 @icmp_slt_0_and_icmp_sgt_neg1_i64(i64 %x) {
 define i64 @icmp_slt_0_and_icmp_sge_neg1_i64_fail(i64 %x) {
 ; CHECK-LABEL: @icmp_slt_0_and_icmp_sge_neg1_i64_fail(
 ; CHECK-NEXT:    [[A:%.*]] = icmp sgt i64 [[X:%.*]], -2
-; CHECK-NEXT:    [[B:%.*]] = zext i1 [[A]] to i64
 ; CHECK-NEXT:    [[C:%.*]] = lshr i64 [[X]], 62
-; CHECK-NEXT:    [[D:%.*]] = and i64 [[C]], [[B]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[C]], 1
+; CHECK-NEXT:    [[D:%.*]] = select i1 [[A]], i64 [[TMP1]], i64 0
 ; CHECK-NEXT:    ret i64 [[D]]
 ;
   %A = icmp sge i64 %x, -1
@@ -2871,9 +2871,8 @@ define i32 @icmp_slt_0_and_icmp_sge_neg2_i32_multiuse1(i32 %x) {
 define i32 @icmp_slt_0_and_icmp_sge_neg2_i32_multiuse2(i32 %x) {
 ; CHECK-LABEL: @icmp_slt_0_and_icmp_sge_neg2_i32_multiuse2(
 ; CHECK-NEXT:    [[A:%.*]] = icmp sgt i32 [[X:%.*]], -3
-; CHECK-NEXT:    [[B:%.*]] = zext i1 [[A]] to i32
 ; CHECK-NEXT:    [[C:%.*]] = lshr i32 [[X]], 31
-; CHECK-NEXT:    [[D:%.*]] = and i32 [[C]], [[B]]
+; CHECK-NEXT:    [[D:%.*]] = select i1 [[A]], i32 [[C]], i32 0
 ; CHECK-NEXT:    call void @use32(i32 [[C]])
 ; CHECK-NEXT:    ret i32 [[D]]
 ;
@@ -2923,10 +2922,9 @@ define i32 @icmp_slt_0_or_icmp_eq_100_i32_multiuse_fail1(i32 %x) {
 
 define i32 @icmp_x_slt_0_and_icmp_y_ne_neg2_i32_multiuse_fail2(i32 %x, i32 %y) {
 ; CHECK-LABEL: @icmp_x_slt_0_and_icmp_y_ne_neg2_i32_multiuse_fail2(
-; CHECK-NEXT:    [[A:%.*]] = icmp ne i32 [[X:%.*]], -2
-; CHECK-NEXT:    [[B:%.*]] = zext i1 [[A]] to i32
+; CHECK-NEXT:    [[A_NOT:%.*]] = icmp eq i32 [[X:%.*]], -2
 ; CHECK-NEXT:    [[C:%.*]] = lshr i32 [[Y:%.*]], 31
-; CHECK-NEXT:    [[D:%.*]] = and i32 [[C]], [[B]]
+; CHECK-NEXT:    [[D:%.*]] = select i1 [[A_NOT]], i32 0, i32 [[C]]
 ; CHECK-NEXT:    call void @use32(i32 [[C]])
 ; CHECK-NEXT:    ret i32 [[D]]
 ;
diff --git a/llvm/test/Transforms/InstCombine/and.ll b/llvm/test/Transforms/InstCombine/and.ll
index 7ec4b0dbb0a12f0..9aa0df3062263cc 100644
--- a/llvm/test/Transforms/InstCombine/and.ll
+++ b/llvm/test/Transforms/InstCombine/and.ll
@@ -2443,8 +2443,8 @@ define i64 @test_and_or_constexpr_infloop() {
 
 define i32 @and_zext(i32 %a, i1 %b) {
 ; CHECK-LABEL: @and_zext(
-; CHECK-NEXT:    [[MASK:%.*]] = zext i1 [[B:%.*]] to i32
-; CHECK-NEXT:    [[R:%.*]] = and i32 [[MASK]], [[A:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[A:%.*]], 1
+; CHECK-NEXT:    [[R:%.*]] = select i1 [[B:%.*]], i32 [[TMP1]], i32 0
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %mask = zext i1 %b to i32
@@ -2454,8 +2454,8 @@ define i32 @and_zext(i32 %a, i1 %b) {
 
 define i32 @and_zext_commuted(i32 %a, i1 %b) {
 ; CHECK-LABEL: @and_zext_commuted(
-; CHECK-NEXT:    [[MASK:%.*]] = zext i1 [[B:%.*]] to i32
-; CHECK-NEXT:    [[R:%.*]] = and i32 [[MASK]], [[A:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[A:%.*]], 1
+; CHECK-NEXT:    [[R:%.*]] = select i1 [[B:%.*]], i32 [[TMP1]], i32 0
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %mask = zext i1 %b to i32
@@ -2478,8 +2478,8 @@ define i32 @and_zext_multiuse(i32 %a, i1 %b) {
 
 define <2 x i32> @and_zext_vec(<2 x i32> %a, <2 x i1> %b) {
 ; CHECK-LABEL: @and_zext_vec(
-; CHECK-NEXT:    [[MASK:%.*]] = zext <2 x i1> [[B:%.*]] to <2 x i32>
-; CHECK-NEXT:    [[R:%.*]] = and <2 x i32> [[MASK]], [[A:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i32> [[A:%.*]], <i32 1, i32 1>
+; CHECK-NEXT:    [[R:%.*]] = select <2 x i1> [[B:%.*]], <2 x i32> [[TMP1]], <2 x i32> zeroinitializer
 ; CHECK-NEXT:    ret <2 x i32> [[R]]
 ;
   %mask = zext <2 x i1> %b to <2 x i32>
@@ -2490,10 +2490,7 @@ define <2 x i32> @and_zext_vec(<2 x i32> %a, <2 x i1> %b) {
 ; tests from PR66606
 define i32 @and_zext_eq_even(i32 %a) {
 ; CHECK-LABEL: @and_zext_eq_even(
-; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[A:%.*]], 2
-; CHECK-NEXT:    [[NOT:%.*]] = zext i1 [[COND]] to i32
-; CHECK-NEXT:    [[R:%.*]] = and i32 [[NOT]], [[A]]
-; CHECK-NEXT:    ret i32 [[R]]
+; CHECK-NEXT:    ret i32 0
 ;
   %cond = icmp eq i32 %a, 2
   %not = zext i1 %cond to i32
@@ -2503,10 +2500,7 @@ define i32 @and_zext_eq_even(i32 %a) {
 
 define i32 @and_zext_eq_even_commuted(i32 %a) {
 ; CHECK-LABEL: @and_zext_eq_even_commuted(
-; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[A:%.*]], 2
-; CHECK-NEXT:    [[NOT:%.*]] = zext i1 [[COND]] to i32
-; CHECK-NEXT:    [[R:%.*]] = and i32 [[NOT]], [[A]]
-; CHECK-NEXT:    ret i32 [[R]]
+; CHECK-NEXT:    ret i32 0
 ;
   %cond = icmp eq i32 %a, 2
   %not = zext i1 %cond to i32
@@ -2517,8 +2511,7 @@ define i32 @and_zext_eq_even_commuted(i32 %a) {
 define i32 @and_zext_eq_odd(i32 %a) {
 ; CHECK-LABEL: @and_zext_eq_odd(
 ; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[A:%.*]], 3
-; CHECK-NEXT:    [[NOT:%.*]] = zext i1 [[COND]] to i32
-; CHECK-NEXT:    [[R:%.*]] = and i32 [[NOT]], [[A]]
+; CHECK-NEXT:    [[R:%.*]] = zext i1 [[COND]] to i32
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %cond = icmp eq i32 %a, 3
@@ -2530,8 +2523,7 @@ define i32 @and_zext_eq_odd(i32 %a) {
 define i32 @and_zext_eq_odd_commuted(i32 %a) {
 ; CHECK-LABEL: @and_zext_eq_odd_commuted(
 ; CHECK-NEXT:    [[COND:%.*]] = icmp eq i32 [[A:%.*]], 3
-; CHECK-NEXT:    [[NOT:%.*]] = zext i1 [[COND]] to i32
-; CHECK-NEXT:    [[R:%.*]] = and i32 [[NOT]], [[A]]
+; CHECK-NEXT:    [[R:%.*]] = zext i1 [[COND]] to i32
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %cond = icmp eq i32 %a, 3
diff --git a/llvm/test/Transforms/InstCombine/icmp.ll b/llvm/test/Transforms/InstCombine/icmp.ll
index 666d85d6cc32103..a5c05278f4232e8 100644
--- a/llvm/test/Transforms/InstCombine/icmp.ll
+++ b/llvm/test/Transforms/InstCombine/icmp.ll
@@ -4441,9 +4441,9 @@ define i1 @redundant_sign_bit_count_ugt_31_30(i32 %x) {
 define i1 @zext_bool_and_eq0(i1 %x, i8 %y) {
 ; CHECK-LABEL: @zext_bool_and_eq0(
 ; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[Y:%.*]], 1
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne i8 [[TMP1]], 0
-; CHECK-NEXT:    [[TMP3:%.*]] = and i1 [[TMP2]], [[X:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = xor i1 [[TMP3]], true
+; CHECK-NEXT:    [[R1:%.*]] = icmp eq i8 [[TMP1]], 0
+; CHECK-NEXT:    [[NOT_X:%.*]] = xor i1 [[X:%.*]], true
+; CHECK-NEXT:    [[R:%.*]] = select i1 [[NOT_X]], i1 true, i1 [[R1]]
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %zx = zext i1 %x to i8
@@ -4454,9 +4454,10 @@ define i1 @zext_bool_and_eq0(i1 %x, i8 %y) {
 
 define <2 x i1> @zext_bool_and_eq0_commute(<2 x i1> %x, <2 x i8> %p) {
 ; CHECK-LABEL: @zext_bool_and_eq0_commute(
-; CHECK-NEXT:    [[TMP1:%.*]] = trunc <2 x i8> [[P:%.*]] to <2 x i1>
-; CHECK-NEXT:    [[TMP2:%.*]] = and <2 x i1> [[TMP1]], [[X:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i1> [[TMP2]], <i1 true, i1 true>
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i8> [[P:%.*]], <i8 1, i8 1>
+; CHECK-NEXT:    [[R1:%.*]] = icmp eq <2 x i8> [[TMP1]], zeroinitializer
+; CHECK-NEXT:    [[NOT_X:%.*]] = xor <2 x i1> [[X:%.*]], <i1 true, i1 true>
+; CHECK-NEXT:    [[R:%.*]] = select <2 x i1> [[NOT_X]], <2 x i1> <i1 true, i1 true>, <2 x i1> [[R1]]
 ; CHECK-NEXT:    ret <2 x i1> [[R]]
 ;
   %y = mul <2 x i8> %p, %p ; thwart complexity-based canonicalization
@@ -4469,8 +4470,8 @@ define <2 x i1> @zext_bool_and_eq0_commute(<2 x i1> %x, <2 x i8> %p) {
 define i1 @zext_bool_and_ne0(i1 %x, i8 %y) {
 ; CHECK-LABEL: @zext_bool_and_ne0(
 ; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[Y:%.*]], 1
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne i8 [[TMP1]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[TMP2]], [[X:%.*]]
+; CHECK-NEXT:    [[R1:%.*]] = icmp ne i8 [[TMP1]], 0
+; CHECK-NEXT:    [[R:%.*]] = select i1 [[X:%.*]], i1 [[R1]], i1 false
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %zx = zext i1 %x to i8
@@ -4482,9 +4483,9 @@ define i1 @zext_bool_and_ne0(i1 %x, i8 %y) {
 define i1 @zext_bool_and_ne1(i1 %x, i8 %y) {
 ; CHECK-LABEL: @zext_bool_and_ne1(
 ; CHECK-NEXT:    [[TMP1:%.*]] = and i8 [[Y:%.*]], 1
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne i8 [[TMP1]], 0
-; CHECK-NEXT:    [[TMP3:%.*]] = and i1 [[TMP2]], [[X:%.*]]
-; CHECK-NEXT:    [[R:%.*]] = xor i1 [[TMP3]], true
+; CHECK-NEXT:    [[R1:%.*]] = icmp eq i8 [[TMP1]], 0
+; CHECK-NEXT:    [[NOT_X:%.*]] = xor i1 [[X:%.*]], true
+; CHECK-NEXT:    [[R:%.*]] = select i1 [[NOT_X]], i1 true, i1 [[R1]]
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %zx = zext i1 %x to i8
@@ -4495,8 +4496,8 @@ define i1 @zext_bool_and_ne1(i1 %x, i8 %y) {
 
 define <2 x i1> @zext_bool_and_eq1(<2 x i1> %x, <2 x i8> %y) {
 ; CHECK-LABEL: @zext_bool_and_eq1(
-; CHECK-NEXT:    [[TMP1:%.*]] = trunc <2 x i8> [[Y:%.*]] to <2 x i1>
-; CHECK-NEXT:    [[R:%.*]] = and <2 x i1> [[TMP1]], [[X:%.*]]
+; CHECK-NEXT:    [[R1:%.*]] = trunc <2 x i8> [[Y:%.*]] to <2 x i1>
+; CHECK-NEXT:    [[R:%.*]] = select <2 x i1> [[X:%.*]], <2 x i1> [[R1]], <2 x i1> zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[R]]
 ;
   %zx = zext <2 x i1> %x to <2 x i8>
@@ -4541,8 +4542,8 @@ define i1 @zext_bool_and_eq0_use(i1 %x, i64 %y) {
 
 define i1 @zext_bool_and_ne0_use(i1 %x, i64 %y) {
 ; CHECK-LABEL: @zext_bool_and_ne0_use(
-; CHECK-NEXT:    [[ZX:%.*]] = zext i1 [[X:%.*]] to i64
-; CHECK-NEXT:    [[A:%.*]] = and i64 [[ZX]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[Y:%.*]], 1
+; CHECK-NEXT:    [[A:%.*]] = select i1 [[X:%.*]], i64 [[TMP1]], i64 0
 ; CHECK-NEXT:    call void @use_i64(i64 [[A]])
 ; CHECK-NEXT:    [[R:%.*]] = icmp ne i64 [[A]], 0
 ; CHECK-NEXT:    ret i1 [[R]]
diff --git a/llvm/test/Transforms/InstCombine/narrow.ll b/llvm/test/Transforms/InstCombine/narrow.ll
index 9e4670b51ad139a..781974d33bf1157 100644
--- a/llvm/test/Transforms/InstCombine/narrow.ll
+++ b/llvm/test/Transforms/InstCombine/narrow.ll
@@ -155,8 +155,7 @@ define i1 @searchArray2(i32 %hay, ptr %haystack) {
 ; CHECK-NEXT:    [[IDX:%.*]] = getelementptr i32, ptr [[HAYSTACK:%.*]], i64 [[INDVAR]]
 ; CHECK-NEXT:    [[LD:%.*]] = load i32, ptr [[IDX]], align 4
 ; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[LD]], [[HAY:%.*]]
-; CHECK-NEXT:    [[ZEXT:%.*]] = zext i1 [[CMP1]] to i8
-; CHECK-NEXT:    [[AND]] = and i8 [[FOUND]], [[ZEXT]]
+; CHECK-NEXT:    [[AND]] = select i1 [[CMP1]], i8 [[FOUND]], i8 0
 ; CHECK-NEXT:    [[INDVAR_NEXT]] = add i64 [[INDVAR]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVAR_NEXT]], 1000
 ; CHECK-NEXT:    br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[LOOP]]
diff --git a/llvm/test/Transforms/InstCombine/zext-or-icmp.ll b/llvm/test/Transforms/InstCombine/zext-or-icmp.ll
index 74d0f2f21aec8e0..dcb8c081f154aa0 100644
--- a/llvm/test/Transforms/InstCombine/zext-or-icmp.ll
+++ b/llvm/test/Transforms/InstCombine/zext-or-icmp.ll
@@ -173,15 +173,15 @@ define i8 @PR49475_infloop(i32 %t0, i16 %insert, i64 %e, i8 %i162) {
 ; CHECK-NEXT:    [[B:%.*]] = icmp eq i32 [[T0:%.*]], 0
 ; CHECK-NEXT:    [[B2:%.*]] = icmp eq i16 [[INSERT:%.*]], 0
 ; CHECK-NEXT:    [[T1:%.*]] = or i1 [[B]], [[B2]]
-; CHECK-NEXT:    [[EXT:%.*]] = zext i1 [[T1]] to i32
-; CHECK-NEXT:    [[AND:%.*]] = and i32 [[EXT]], [[T0]]
-; CHECK-NEXT:    [[TMP1:%.*]] = or i32 [[AND]], 140
-; CHECK-NEXT:    [[XOR1:%.*]] = zext i32 [[TMP1]] to i64
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[T0]], 1
+; CHECK-NEXT:    [[TMP2:%.*]] = or i32 [[TMP1]], 140
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[TMP2]] to i64
+; CHECK-NEXT:    [[XOR1:%.*]] = select i1 [[T1]], i64 [[TMP3]], i64 140
 ; CHECK-NEXT:    [[CONV16:%.*]] = sext i8 [[I162:%.*]] to i64
 ; CHECK-NEXT:    [[SUB17:%.*]] = sub i64 [[CONV16]], [[E:%.*]]
 ; CHECK-NEXT:    [[SEXT:%.*]] = shl i64 [[SUB17]], 32
 ; CHECK-NEXT:    [[CONV18:%.*]] = ashr exact i64 [[SEXT]], 32
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sle i64 [[CONV18]], [[XOR1]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sge i64 [[XOR1]], [[CONV18]]
 ; CHECK-NEXT:    [[CONV19:%.*]] = zext i1 [[CMP]] to i16
 ; CHECK-NEXT:    [[OR21:%.*]] = or i16 [[CONV19]], [[INSERT]]
 ; CHECK-NEXT:    [[TOBOOL23_NOT:%.*]] = icmp eq i16 [[OR21]], 0

dianqk · 2023-09-19T06:51:36Z

We should use rebase instead of merge.
And I recommend force push to change the code before the review starts.

dtcxzyw · 2023-09-19T07:05:56Z

We should use rebase instead of merge. And I recommend force push to change the code before the review starts.

I am confused with the LLVM GitHub user guide. I witnessed someone mess up by doing rebase + force push.
Examples: #65853 #65543 #65535

nikic · 2023-09-19T08:51:19Z

I think this is the right canonicalization at the IR level, as select allows better analysis than zext(icmp). However, this seems to be universally worse for the backend: https://llvm.godbolt.org/z/Yh7brfc8b So I think we would need an SDAG undo transform. Interestingly, this already happens for riscv64, so maybe the undo transform already exists but is not enabled for all target.

@goldsteinn What do you think?

sftlbcn · 2023-09-19T12:16:56Z

this also fixes #28612

goldsteinn · 2023-09-19T15:03:02Z

I think this is the right canonicalization at the IR level, as select allows better analysis than zext(icmp). However, this seems to be universally worse for the backend: https://llvm.godbolt.org/z/Yh7brfc8b So I think we would need an SDAG undo transform. Interestingly, this already happens for riscv64, so maybe the undo transform already exists but is not enabled for all target.

@goldsteinn What do you think?

Yeah, I think I agree, I looked a bit around DAGCombiner and can't find where risv does this but pretty sure its in the RISCV target codes as commenting out the entirety of visitSelect, visitSelect_CC, and visitAndLike doesn't change the riscv codegen.

I'd say since we also do:

  %zx = sext i1 %x to i32
  %r = and i32 %zx, %y

to

  %r = select i1 %x, i32 %masked, i32 0

This makes sense in that context.

I'll throw up a patch shortly to cover this case in the backend.

Edit: its a bit more ambiguous if the i1 is coming from setcc: https://llvm.godbolt.org/z/dxEs4KrMs
Edit2: Its also a bit more challenging to fix this up in the DAGCombiner b.c booleans are not handled the same way across targets.

goldsteinn · 2023-09-19T17:12:45Z

Created: #66793 which fixes the backend regression this can cause.

nikic · 2023-09-20T05:58:51Z

You might want to check whether this also fixes #66740 and if so add a test.

nikic

LGTM

… & 1, 0` Fix tests for PR66733.

dtcxzyw · 2023-09-21T08:53:08Z

You might want to check whether this also fixes #66740 and if so add a test.

Confirmed. I have added a pre-commit test for #66733 in 1a73a6b.

dtcxzyw · 2023-09-21T08:58:16Z

To avoid backend regressions, I will merge this PR after #66793.

llvm#66740) This patch canonicalizes the pattern `and(zext(A), B)` into `select A, B & 1, 0`. Thus, we can reuse transforms `select B == even, B & 1, 0 -> 0` and `select B == odd, B & 1, 0 -> zext(B == odd)` in `InstCombine`. It is an alternative to llvm#66676. Alive2: https://alive2.llvm.org/ce/z/598phE Fixes llvm#66733. Fixes llvm#66606. Fixes llvm#28612.

…B & 1, 0` (llvm#66740)" This reverts commit a7f962c.

dtcxzyw added 3 commits September 19, 2023 10:19

[InstCombine] Canonicalize and(zext(A), B) into select A, B & 1, 0

85f7911

Merge branch 'main' into canonicalize-and-zext-bool

8198bea

fixup! [InstCombine] Canonicalize and(zext(A), B) into `select A, B…

ca389a9

… & 1, 0`

dtcxzyw requested review from nikic and goldsteinn September 19, 2023 06:15

llvmbot added the llvm:transforms label Sep 19, 2023

dtcxzyw mentioned this pull request Sep 19, 2023

[InstSimplify] Fold and A, (zext (icmp eq A, 0)) into 0 #66676

Closed

nikic mentioned this pull request Sep 20, 2023

(a == 0) & (((short)a) >= 0) is not always optimized to just a == 0 #66733

Closed

nikic approved these changes Sep 20, 2023

View reviewed changes

dtcxzyw added 2 commits September 21, 2023 16:26

Merge branch 'main' into canonicalize-and-zext-bool

4fc50e3

fixup! [InstCombine] Canonicalize and(zext(A), B) into `select A, B…

422b262

… & 1, 0` Fix tests for PR66733.

Merge branch 'main' into canonicalize-and-zext-bool

c8630d7

dtcxzyw merged commit a7f962c into llvm:main Sep 28, 2023

dtcxzyw deleted the canonicalize-and-zext-bool branch September 28, 2023 18:52

dtcxzyw mentioned this pull request Jul 16, 2024

[InstCombine] Improve select simplification based on known bits #97289

Open

cyyself added a commit to cyyself/llvm-project that referenced this pull request Jul 17, 2024

Revert "[InstCombine] Canonicalize and(zext(A), B) into `select A, …

e63ee38

…B & 1, 0` (llvm#66740)" This reverts commit a7f962c.

cyyself added a commit to cyyself/llvm-project that referenced this pull request Jul 18, 2024

Revert "[InstCombine] Canonicalize and(zext(A), B) into `select A, …

78a42e0

…B & 1, 0` (llvm#66740)" This reverts commit a7f962c.

cyyself added a commit to cyyself/llvm-project that referenced this pull request Jul 18, 2024

Revert "[InstCombine] Canonicalize and(zext(A), B) into `select A, …

8b16eec

…B & 1, 0` (llvm#66740)" This reverts commit a7f962c.

cyyself mentioned this pull request Jul 22, 2024

[InstCombine] Performance degradation caused by more branch after commit b5a9361 and #66740 #99835

Open

cyyself added a commit to cyyself/llvm-project that referenced this pull request Jul 24, 2024

Revert "[InstCombine] Canonicalize and(zext(A), B) into `select A, …

6537238

…B & 1, 0` (llvm#66740)" This reverts commit a7f962c.

cyyself added a commit to cyyself/llvm-project that referenced this pull request Aug 14, 2024

Revert "[InstCombine] Canonicalize and(zext(A), B) into `select A, …

4fef138

…B & 1, 0` (llvm#66740)" This reverts commit a7f962c.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[InstCombine] Canonicalize `and(zext(A), B)` into `select A, B & 1, 0` #66740

[InstCombine] Canonicalize `and(zext(A), B)` into `select A, B & 1, 0` #66740

Uh oh!

dtcxzyw commented Sep 19, 2023 •

edited

Loading

Uh oh!

llvmbot commented Sep 19, 2023

Uh oh!

dianqk commented Sep 19, 2023 •

edited

Loading

Uh oh!

dtcxzyw commented Sep 19, 2023

Uh oh!

nikic commented Sep 19, 2023

Uh oh!

sftlbcn commented Sep 19, 2023

Uh oh!

goldsteinn commented Sep 19, 2023 •

edited

Loading

Uh oh!

goldsteinn commented Sep 19, 2023

Uh oh!

nikic commented Sep 20, 2023

Uh oh!

nikic left a comment

Uh oh!

dtcxzyw commented Sep 21, 2023 •

edited

Loading

Uh oh!

dtcxzyw commented Sep 21, 2023

Uh oh!

Uh oh!

[InstCombine] Canonicalize and(zext(A), B) into select A, B & 1, 0 #66740

[InstCombine] Canonicalize and(zext(A), B) into select A, B & 1, 0 #66740

Uh oh!

Conversation

dtcxzyw commented Sep 19, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Sep 19, 2023

Uh oh!

dianqk commented Sep 19, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dtcxzyw commented Sep 19, 2023

Uh oh!

nikic commented Sep 19, 2023

Uh oh!

sftlbcn commented Sep 19, 2023

Uh oh!

goldsteinn commented Sep 19, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

goldsteinn commented Sep 19, 2023

Uh oh!

nikic commented Sep 20, 2023

Uh oh!

nikic left a comment

Choose a reason for hiding this comment

Uh oh!

dtcxzyw commented Sep 21, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dtcxzyw commented Sep 21, 2023

Uh oh!

Uh oh!

[InstCombine] Canonicalize `and(zext(A), B)` into `select A, B & 1, 0` #66740

[InstCombine] Canonicalize `and(zext(A), B)` into `select A, B & 1, 0` #66740

dtcxzyw commented Sep 19, 2023 •

edited

Loading

dianqk commented Sep 19, 2023 •

edited

Loading

goldsteinn commented Sep 19, 2023 •

edited

Loading

dtcxzyw commented Sep 21, 2023 •

edited

Loading