[llvm-branch-commits] [llvm] ValueTracking: x - floor(x) cannot introduce overflow (PR #189003)

[via llvm-branch-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

<details>
<summary>Changes</summary>

This returns a value with an absolute value less than 1 so it
should be possible to propagate no-infs.

---
Full diff: https://github.com/llvm/llvm-project/pull/189003.diff


2 Files Affected:

- (modified) llvm/lib/Analysis/ValueTracking.cpp (+9-1) 
- (modified) llvm/test/Transforms/Attributor/nofpclass-fmul.ll (+42) 


``````````diff
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index eca4b98b21226..7f6b0c27d7518 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -4977,7 +4977,15 @@ static constexpr KnownFPClass::MinMaxKind getMinMaxKind(Intrinsic::ID IID) {
 /// \return true if this is a floating point value that is known to have a
 /// magnitude smaller than 1. i.e., fabs(X) <= 1.0
 static bool isAbsoluteValueLessEqualOne(const Value *V) {
-  // TODO: Handle frexp and x - floor(x)?
+  // TODO: Handle frexp
+  // TODO: Other rounding intrinsics?
+
+  // fabs(x - floor(x)) <= 1
+  const Value *SubFloorX;
+  if (match(V, m_FSub(m_Value(SubFloorX),
+                      m_Intrinsic<Intrinsic::floor>(m_Deferred(SubFloorX)))))
+    return true;
+
   return match(V, m_Intrinsic<Intrinsic::amdgcn_trig_preop>(m_Value())) ||
          match(V, m_Intrinsic<Intrinsic::amdgcn_fract>(m_Value()));
 }
diff --git a/llvm/test/Transforms/Attributor/nofpclass-fmul.ll b/llvm/test/Transforms/Attributor/nofpclass-fmul.ll
index db2cb9ed9c2ba..1b1cf60a33500 100644
--- a/llvm/test/Transforms/Attributor/nofpclass-fmul.ll
+++ b/llvm/test/Transforms/Attributor/nofpclass-fmul.ll
@@ -910,3 +910,45 @@ define float @ret_fmul__not_inf__neg1.5(float nofpclass(inf) %x) {
   %mul = fmul float %x, -1.5
   ret float %mul
 }
+
+define float @ret_fmul__not_inf__fsub_floor_pat(float nofpclass(inf) %x, float %y) {
+; CHECK-LABEL: define nofpclass(inf) float @ret_fmul__not_inf__fsub_floor_pat(
+; CHECK-SAME: float nofpclass(inf) [[X:%.*]], float [[Y:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[FLOOR_Y:%.*]] = call float @llvm.floor.f32(float [[Y]]) #[[ATTR2]]
+; CHECK-NEXT:    [[Y_SUB_FLOOR_Y:%.*]] = fsub float [[Y]], [[FLOOR_Y]]
+; CHECK-NEXT:    [[MUL:%.*]] = fmul float [[X]], [[Y_SUB_FLOOR_Y]]
+; CHECK-NEXT:    ret float [[MUL]]
+;
+  %floor.y = call float @llvm.floor.f32(float %y)
+  %y.sub.floor.y = fsub float %y, %floor.y
+  %mul = fmul float %x, %y.sub.floor.y
+  ret float %mul
+}
+
+define float @ret_fmul__not_inf__fsub_floor_pat_commute(float nofpclass(inf) %x, float %y) {
+; CHECK-LABEL: define nofpclass(inf) float @ret_fmul__not_inf__fsub_floor_pat_commute(
+; CHECK-SAME: float nofpclass(inf) [[X:%.*]], float [[Y:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[FLOOR_Y:%.*]] = call float @llvm.floor.f32(float [[Y]]) #[[ATTR2]]
+; CHECK-NEXT:    [[Y_SUB_FLOOR_Y:%.*]] = fsub float [[Y]], [[FLOOR_Y]]
+; CHECK-NEXT:    [[MUL:%.*]] = fmul float [[Y_SUB_FLOOR_Y]], [[X]]
+; CHECK-NEXT:    ret float [[MUL]]
+;
+  %floor.y = call float @llvm.floor.f32(float %y)
+  %y.sub.floor.y = fsub float %y, %floor.y
+  %mul = fmul float %y.sub.floor.y, %x
+  ret float %mul
+}
+
+define float @ret_fmul__not_inf__fsub_floor_pat_wrong_floor_val(float nofpclass(inf) %x, float %y, float %z) {
+; CHECK-LABEL: define float @ret_fmul__not_inf__fsub_floor_pat_wrong_floor_val(
+; CHECK-SAME: float nofpclass(inf) [[X:%.*]], float [[Y:%.*]], float [[Z:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[FLOOR_Y:%.*]] = call float @llvm.floor.f32(float [[Z]]) #[[ATTR2]]
+; CHECK-NEXT:    [[Y_SUB_FLOOR_Y:%.*]] = fsub float [[Y]], [[FLOOR_Y]]
+; CHECK-NEXT:    [[MUL:%.*]] = fmul float [[X]], [[Y_SUB_FLOOR_Y]]
+; CHECK-NEXT:    ret float [[MUL]]
+;
+  %floor.y = call float @llvm.floor.f32(float %z)
+  %y.sub.floor.y = fsub float %y, %floor.y
+  %mul = fmul float %x, %y.sub.floor.y
+  ret float %mul
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/189003