[clang] [llvm] LV][LAA] Vectorize math lib calls with mem write-only attribute (PR #78432)

[via cfe-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-analysis

Author: Paschalis Mpeis (paschalis-mpeis)

<details>
<summary>Changes</summary>

Teach LAA to consider safe specific math lib calls which are known to
 have set the memory write-only attribute. Those attributes are set to
 calls by `inferNonMandatoryLibFuncAttrs`, in BuildLibCalls.cpp, and the
 current ones are `modf`/`modff` and `frexp`/`frexpf`.
    
 This happens only when the calls are found through TLI to have
 vectorized counterparts.

---
Full diff: https://github.com/llvm/llvm-project/pull/78432.diff


2 Files Affected:

- (added) clang/test/CodeGen/aarch64-veclib-function-calls-linear-ptrs.c (+54) 
- (modified) llvm/lib/Analysis/LoopAccessAnalysis.cpp (+19) 


``````````diff
diff --git a/clang/test/CodeGen/aarch64-veclib-function-calls-linear-ptrs.c b/clang/test/CodeGen/aarch64-veclib-function-calls-linear-ptrs.c
new file mode 100644
index 000000000000000..957b3f5cb235d31
--- /dev/null
+++ b/clang/test/CodeGen/aarch64-veclib-function-calls-linear-ptrs.c
@@ -0,0 +1,54 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --filter "call.*(frexp|modf)" --version 4
+// RUN: %clang --target=aarch64-linux-gnu -march=armv8-a+sve -O3 -mllvm -vector-library=ArmPL -mllvm -force-vector-interleave=1 -mllvm -prefer-predicate-over-epilogue=predicate-dont-vectorize -emit-llvm -S -o - %s | FileCheck %s
+
+// REQUIRES: aarch64-registered-target
+
+/*
+Testing vectorization of math functions that have the attribute write-only to
+memory set. Given they have vectorized counterparts, they should be able to
+vectorize.
+*/
+
+// The following define is required to access some math functions.
+#define _GNU_SOURCE
+#include <math.h>
+
+// frexp/frexpf have no TLI mappings yet.
+
+// CHECK-LABEL: define dso_local void @frexp_f64(
+// CHECK-SAME: ptr nocapture noundef readonly [[IN:%.*]], ptr nocapture noundef writeonly [[OUT1:%.*]], ptr nocapture noundef writeonly [[OUT2:%.*]], i32 noundef [[N:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
+// CHECK:    [[CALL:%.*]] = tail call double @frexp(double noundef [[TMP0:%.*]], ptr noundef [[ADD_PTR:%.*]]) #[[ATTR5:[0-9]+]]
+//
+void frexp_f64(double *in, double *out1, int *out2, int N) {
+  for (int i = 0; i < N; ++i)
+    *out1 = frexp(in[i], out2+i);
+}
+
+// CHECK-LABEL: define dso_local void @frexp_f32(
+// CHECK-SAME: ptr nocapture noundef readonly [[IN:%.*]], ptr nocapture noundef writeonly [[OUT1:%.*]], ptr nocapture noundef writeonly [[OUT2:%.*]], i32 noundef [[N:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK:    [[CALL:%.*]] = tail call float @frexpf(float noundef [[TMP0:%.*]], ptr noundef [[ADD_PTR:%.*]]) #[[ATTR5]]
+//
+void frexp_f32(float *in, float *out1, int *out2, int N) {
+  for (int i = 0; i < N; ++i)
+    *out1 = frexpf(in[i], out2+i);
+}
+
+// CHECK-LABEL: define dso_local void @modf_f64(
+// CHECK-SAME: ptr nocapture noundef readonly [[IN:%.*]], ptr nocapture noundef writeonly [[OUT1:%.*]], ptr nocapture noundef writeonly [[OUT2:%.*]], i32 noundef [[N:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK:    [[TMP11:%.*]] = tail call <vscale x 2 x double> @armpl_svmodf_f64_x(<vscale x 2 x double> [[WIDE_MASKED_LOAD:%.*]], ptr [[TMP10:%.*]], <vscale x 2 x i1> [[ACTIVE_LANE_MASK:%.*]])
+// CHECK:    [[CALL:%.*]] = tail call double @modf(double noundef [[TMP14:%.*]], ptr noundef [[ADD_PTR:%.*]]) #[[ATTR6:[0-9]+]]
+//
+void modf_f64(double *in, double *out1, double *out2, int N) {
+  for (int i = 0; i < N; ++i)
+      out1[i] = modf(in[i], out2+i);
+}
+
+// CHECK-LABEL: define dso_local void @modf_f32(
+// CHECK-SAME: ptr nocapture noundef readonly [[IN:%.*]], ptr nocapture noundef writeonly [[OUT1:%.*]], ptr nocapture noundef writeonly [[OUT2:%.*]], i32 noundef [[N:%.*]]) local_unnamed_addr #[[ATTR0]] {
+// CHECK:    [[TMP11:%.*]] = tail call <vscale x 4 x float> @armpl_svmodf_f32_x(<vscale x 4 x float> [[WIDE_MASKED_LOAD:%.*]], ptr [[TMP10:%.*]], <vscale x 4 x i1> [[ACTIVE_LANE_MASK:%.*]])
+// CHECK:    [[CALL:%.*]] = tail call float @modff(float noundef [[TMP14:%.*]], ptr noundef [[ADD_PTR:%.*]]) #[[ATTR7:[0-9]+]]
+//
+void modf_f32(float *in, float *out1, float *out2, int N) {
+  for (int i = 0; i < N; ++i)
+      out1[i] = modff(in[i], out2+i);
+}
diff --git a/llvm/lib/Analysis/LoopAccessAnalysis.cpp b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
index aed60cc5a3f5ef0..0c8b4e51fcf5c16 100644
--- a/llvm/lib/Analysis/LoopAccessAnalysis.cpp
+++ b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
@@ -2274,6 +2274,20 @@ bool LoopAccessInfo::canAnalyzeLoop() {
   return true;
 }
 
+/// Returns whether \p I is a known math library call that has memory write-only
+/// attribute set.
+static bool isMathLibCallMemWriteOnly(const TargetLibraryInfo *TLI,
+                                      const Instruction &I) {
+  auto *Call = dyn_cast<CallInst>(&I);
+  if (!Call)
+    return false;
+
+  LibFunc Func;
+  TLI->getLibFunc(*Call, Func);
+  return Func == LibFunc::LibFunc_modf || Func == LibFunc::LibFunc_modff ||
+         Func == LibFunc::LibFunc_frexp || Func == LibFunc::LibFunc_frexpf;
+}
+
 void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
                                  const TargetLibraryInfo *TLI,
                                  DominatorTree *DT) {
@@ -2364,6 +2378,11 @@ void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
 
       // Save 'store' instructions. Abort if other instructions write to memory.
       if (I.mayWriteToMemory()) {
+        // We can safety handle math functions that have vectorized
+        // counterparts and have the memory write-only attribute set.
+        if (isMathLibCallMemWriteOnly(TLI, I))
+          continue;
+
         auto *St = dyn_cast<StoreInst>(&I);
         if (!St) {
           recordAnalysis("CantVectorizeInstruction", St)

``````````

</details>


https://github.com/llvm/llvm-project/pull/78432