[InstCombine] Relax the one-use constraints for icmp pred (binop X, Z), (binop Y, Z)

[dtcxzyw]

This patch relaxes the one-use constraints for icmp pred (binop X, Z), (binop Y, Z). It will enable more optimizations with pointer arithmetic.
One example in boost::match_results::set_size:

declare void @use(i64)
define i1 @src(ptr %a1, ptr %a2, ptr %add.ptr.i66, i64 %sub.ptr.rhs.cast.i) {
  %sub.ptr.lhs.cast.i = ptrtoint ptr %a1 to i64
  %sub.ptr.rhs.cast.i = ptrtoint ptr %a2 to i64
  %sub.ptr.sub.i = sub i64 %sub.ptr.lhs.cast.i, %sub.ptr.rhs.cast.i
  %sub.ptr.div.i = sdiv exact i64 %sub.ptr.sub.i, 24
  call void @use(i64 %sub.ptr.div.i)
  %sub.ptr.lhs.cast.i.i = ptrtoint ptr %add.ptr.i66 to i64
  %sub.ptr.sub.i.i = sub i64 %sub.ptr.lhs.cast.i.i, %sub.ptr.rhs.cast.i
  %sub.ptr.div.i.i = sdiv exact i64 %sub.ptr.sub.i.i, 24
  %cmp.i.not.i.i = icmp eq i64 %sub.ptr.div.i.i, %sub.ptr.div.i
  ret i1 %cmp.i.not.i.i
}
define i1 @tgt(ptr %a1, ptr %a2, ptr %add.ptr.i66, i64 %sub.ptr.rhs.cast.i) {
  %sub.ptr.lhs.cast.i = ptrtoint ptr %a1 to i64
  %sub.ptr.rhs.cast.i = ptrtoint ptr %a2 to i64
  %sub.ptr.sub.i = sub i64 %sub.ptr.lhs.cast.i, %sub.ptr.rhs.cast.i
  %sub.ptr.div.i = sdiv exact i64 %sub.ptr.sub.i, 24
  call void @use(i64 %sub.ptr.div.i)
  %cmp.i.not.i.i = icmp eq i64 %sub.ptr.sub.i.i, %sub.ptr.sub.i
  ret i1 %cmp.i.not.i.i
}

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Yingwei Zheng (dtcxzyw)

Changes

This patch relaxes the one-use constraints for icmp pred (binop X, Z), (binop Y, Z). It will enable more optimizations with pointer arithmetic.
One example in boost::match_results::set_size:

declare void @<!-- -->use(i64)
define i1 @<!-- -->src(ptr %a1, ptr %a2, ptr %add.ptr.i66, i64 %sub.ptr.rhs.cast.i) {
  %sub.ptr.lhs.cast.i = ptrtoint ptr %a1 to i64
  %sub.ptr.rhs.cast.i = ptrtoint ptr %a2 to i64
  %sub.ptr.sub.i = sub i64 %sub.ptr.lhs.cast.i, %sub.ptr.rhs.cast.i
  %sub.ptr.div.i = sdiv exact i64 %sub.ptr.sub.i, 24
  call void @<!-- -->use(i64 %sub.ptr.div.i)
  %sub.ptr.lhs.cast.i.i = ptrtoint ptr %add.ptr.i66 to i64
  %sub.ptr.sub.i.i = sub i64 %sub.ptr.lhs.cast.i.i, %sub.ptr.rhs.cast.i
  %sub.ptr.div.i.i = sdiv exact i64 %sub.ptr.sub.i.i, 24
  %cmp.i.not.i.i = icmp eq i64 %sub.ptr.div.i.i, %sub.ptr.div.i
  ret i1 %cmp.i.not.i.i
}
define i1 @<!-- -->tgt(ptr %a1, ptr %a2, ptr %add.ptr.i66, i64 %sub.ptr.rhs.cast.i) {
  %sub.ptr.lhs.cast.i = ptrtoint ptr %a1 to i64
  %sub.ptr.rhs.cast.i = ptrtoint ptr %a2 to i64
  %sub.ptr.sub.i = sub i64 %sub.ptr.lhs.cast.i, %sub.ptr.rhs.cast.i
  %sub.ptr.div.i = sdiv exact i64 %sub.ptr.sub.i, 24
  call void @<!-- -->use(i64 %sub.ptr.div.i)
  %cmp.i.not.i.i = icmp eq i64 %sub.ptr.sub.i.i, %sub.ptr.sub.i
  ret i1 %cmp.i.not.i.i
}

---

Full diff: https://github.com/llvm/llvm-project/pull/76384.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp (+2-2)
• (modified) llvm/test/Transforms/InstCombine/icmp.ll (+17)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 289976718e52f3..0afb415a482411 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -4911,8 +4911,8 @@ Instruction *InstCombinerImpl::foldICmpBinOp(ICmpInst &I,
     }
   }
 
-  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() && BO0->hasOneUse() &&
-      BO1->hasOneUse() && BO0->getOperand(1) == BO1->getOperand(1)) {
+  if (BO0 && BO1 && BO0->getOpcode() == BO1->getOpcode() &&
+      BO0->getOperand(1) == BO1->getOperand(1)) {
     switch (BO0->getOpcode()) {
     default:
       break;
diff --git a/llvm/test/Transforms/InstCombine/icmp.ll b/llvm/test/Transforms/InstCombine/icmp.ll
index 1c7bb36f0d34c0..6d881be56dc2b4 100644
--- a/llvm/test/Transforms/InstCombine/icmp.ll
+++ b/llvm/test/Transforms/InstCombine/icmp.ll
@@ -815,6 +815,23 @@ define i1 @test46(i32 %X, i32 %Y, i32 %Z) {
   ret i1 %C
 }
 
+define i1 @test46_multiuse(i32 %X, i32 %Y, i32 %Z) {
+; CHECK-LABEL: @test46_multiuse(
+; CHECK-NEXT:    [[A:%.*]] = ashr exact i32 [[X:%.*]], [[Z:%.*]]
+; CHECK-NEXT:    call void @use_i32(i32 [[A]])
+; CHECK-NEXT:    [[B:%.*]] = ashr exact i32 [[Y:%.*]], [[Z]]
+; CHECK-NEXT:    call void @use_i32(i32 [[B]])
+; CHECK-NEXT:    [[C:%.*]] = icmp ult i32 [[X]], [[Y]]
+; CHECK-NEXT:    ret i1 [[C]]
+;
+  %A = ashr exact i32 %X, %Z
+  call void @use_i32(i32 %A)
+  %B = ashr exact i32 %Y, %Z
+  call void @use_i32(i32 %B)
+  %C = icmp ult i32 %A, %B
+  ret i1 %C
+}
+
 ; PR9343 #5
 define i1 @test47(i32 %X, i32 %Y, i32 %Z) {
 ; CHECK-LABEL: @test47(

[dtcxzyw]

This patch breaks the alias analysis and causes some regressions. We should do the simplification when at least one binop is one-use.

[dtcxzyw]

This patch breaks the alias analysis and causes some regressions. We should do the simplification when at least one binop is one-use.

Done. All regressions have been fixed except for dtcxzyw/llvm-opt-benchmark#49 (comment). I will post a patch for clang later.

[nikic]

LGTM