[llvm] 478b22d - [CGP] despeculateCountZeros - Don't create is-zero branch if cttz/ctlz source is known non-zero

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2021-07-24T13:11:49+01:00
New Revision: 478b22d95aee2b818008a7b92b8606175767d0de

URL: https://github.com/llvm/llvm-project/commit/478b22d95aee2b818008a7b92b8606175767d0de
DIFF: https://github.com/llvm/llvm-project/commit/478b22d95aee2b818008a7b92b8606175767d0de.diff

LOG: [CGP] despeculateCountZeros - Don't create is-zero branch if cttz/ctlz source is known non-zero

If value tracking can confirm that the cttz/ctlz source is known non-zero then we don't need to create a branch (which DAG will struggle to recover from).

Differential Revision: https://reviews.llvm.org/D106685

Added: 
    

Modified: 
    llvm/lib/CodeGen/CodeGenPrepare.cpp
    llvm/test/CodeGen/X86/clz.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp
index 1c0932ee40e53..77ce3d2fb5633 100644
--- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -2041,6 +2041,10 @@ static bool despeculateCountZeros(IntrinsicInst *CountZeros,
   if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits())
     return false;
 
+  // Bail if the value is never zero.
+  if (llvm::isKnownNonZero(CountZeros->getOperand(0), *DL))
+    return false;
+
   // The intrinsic will be sunk behind a compare against zero and branch.
   BasicBlock *StartBlock = CountZeros->getParent();
   BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false");

diff  --git a/llvm/test/CodeGen/X86/clz.ll b/llvm/test/CodeGen/X86/clz.ll
index 0c12135c127ff..755f6b43d0975 100644
--- a/llvm/test/CodeGen/X86/clz.ll
+++ b/llvm/test/CodeGen/X86/clz.ll
@@ -704,33 +704,21 @@ define i64 @cttz_i64_zero_test(i64 %n) {
 ; Don't generate the cmovne when the source is known non-zero (and bsr would
 ; not set ZF).
 ; rdar://9490949
-; FIXME: The compare and branch are produced late in IR (by CodeGenPrepare), and
-;        codegen doesn't know how to delete the movl and je.
 define i32 @ctlz_i32_fold_cmov(i32 %n) {
 ; X86-LABEL: ctlz_i32_fold_cmov:
 ; X86:       # %bb.0:
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    orl $1, %eax
-; X86-NEXT:    je .LBB16_1
-; X86-NEXT:  # %bb.2: # %cond.false
 ; X86-NEXT:    bsrl %eax, %eax
 ; X86-NEXT:    xorl $31, %eax
 ; X86-NEXT:    retl
-; X86-NEXT:  .LBB16_1:
-; X86-NEXT:    movl $32, %eax
-; X86-NEXT:    retl
 ;
 ; X64-LABEL: ctlz_i32_fold_cmov:
 ; X64:       # %bb.0:
 ; X64-NEXT:    orl $1, %edi
-; X64-NEXT:    je .LBB16_1
-; X64-NEXT:  # %bb.2: # %cond.false
 ; X64-NEXT:    bsrl %edi, %eax
 ; X64-NEXT:    xorl $31, %eax
 ; X64-NEXT:    retq
-; X64-NEXT:  .LBB16_1:
-; X64-NEXT:    movl $32, %eax
-; X64-NEXT:    retq
 ;
 ; X86-CLZ-LABEL: ctlz_i32_fold_cmov:
 ; X86-CLZ:       # %bb.0:
@@ -953,14 +941,9 @@ define i64 @ctlz_i64_zero_test_knownneverzero(i64 %n) {
 ; X64-LABEL: ctlz_i64_zero_test_knownneverzero:
 ; X64:       # %bb.0:
 ; X64-NEXT:    orq $1, %rdi
-; X64-NEXT:    je .LBB21_1
-; X64-NEXT:  # %bb.2: # %cond.false
 ; X64-NEXT:    bsrq %rdi, %rax
 ; X64-NEXT:    xorq $63, %rax
 ; X64-NEXT:    retq
-; X64-NEXT:  .LBB21_1:
-; X64-NEXT:    movl $64, %eax
-; X64-NEXT:    retq
 ;
 ; X86-CLZ-LABEL: ctlz_i64_zero_test_knownneverzero:
 ; X86-CLZ:       # %bb.0:
@@ -1026,13 +1009,8 @@ define i64 @cttz_i64_zero_test_knownneverzero(i64 %n) {
 ; X64:       # %bb.0:
 ; X64-NEXT:    movabsq $-9223372036854775808, %rax # imm = 0x8000000000000000
 ; X64-NEXT:    orq %rdi, %rax
-; X64-NEXT:    je .LBB22_1
-; X64-NEXT:  # %bb.2: # %cond.false
 ; X64-NEXT:    bsfq %rax, %rax
 ; X64-NEXT:    retq
-; X64-NEXT:  .LBB22_1:
-; X64-NEXT:    movl $64, %eax
-; X64-NEXT:    retq
 ;
 ; X86-CLZ-LABEL: cttz_i64_zero_test_knownneverzero:
 ; X86-CLZ:       # %bb.0: