PreISelIntrinsicLowering: Lower llvm.log to a loop if scalable vec arg

[steplong]

Similar to ab976a1, but for llvm.log.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Stephen Long (steplong)

Changes

Similar to ab976a1, but for llvm.log.

---

Full diff: https://github.com/llvm/llvm-project/pull/129744.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/PreISelIntrinsicLowering.cpp (+1)
• (added) llvm/test/Transforms/PreISelIntrinsicLowering/AArch64/expand-log.ll (+43)

diff --git a/llvm/lib/CodeGen/PreISelIntrinsicLowering.cpp b/llvm/lib/CodeGen/PreISelIntrinsicLowering.cpp
index 27fa0b43d74f6..50ccce877e4c1 100644
--- a/llvm/lib/CodeGen/PreISelIntrinsicLowering.cpp
+++ b/llvm/lib/CodeGen/PreISelIntrinsicLowering.cpp
@@ -560,6 +560,7 @@ bool PreISelIntrinsicLowering::lowerIntrinsics(Module &M) const {
       break;
     case Intrinsic::exp:
     case Intrinsic::exp2:
+    case Intrinsic::log:
       Changed |= forEachCall(F, [&](CallInst *CI) {
         Type *Ty = CI->getArgOperand(0)->getType();
         if (!isa<ScalableVectorType>(Ty))
diff --git a/llvm/test/Transforms/PreISelIntrinsicLowering/AArch64/expand-log.ll b/llvm/test/Transforms/PreISelIntrinsicLowering/AArch64/expand-log.ll
new file mode 100644
index 0000000000000..424729f498aa6
--- /dev/null
+++ b/llvm/test/Transforms/PreISelIntrinsicLowering/AArch64/expand-log.ll
@@ -0,0 +1,43 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=pre-isel-intrinsic-lowering -S < %s | FileCheck %s
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32"
+target triple = "aarch64"
+
+define <vscale x 4 x float> @scalable_vec_log(<vscale x 4 x float> %input) {
+; CHECK-LABEL: define <vscale x 4 x float> @scalable_vec_log(
+; CHECK-SAME: <vscale x 4 x float> [[INPUT:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP2:%.*]] = mul i64 [[TMP1]], 4
+; CHECK-NEXT:    br label %[[BB3:.*]]
+; CHECK:       [[BB3]]:
+; CHECK-NEXT:    [[TMP4:%.*]] = phi i64 [ 0, [[TMP0:%.*]] ], [ [[TMP9:%.*]], %[[BB3]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = phi <vscale x 4 x float> [ [[INPUT]], [[TMP0]] ], [ [[TMP8:%.*]], %[[BB3]] ]
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <vscale x 4 x float> [[TMP5]], i64 [[TMP4]]
+; CHECK-NEXT:    [[TMP7:%.*]] = call float @llvm.log.f32(float [[TMP6]])
+; CHECK-NEXT:    [[TMP8]] = insertelement <vscale x 4 x float> [[TMP5]], float [[TMP7]], i64 [[TMP4]]
+; CHECK-NEXT:    [[TMP9]] = add i64 [[TMP4]], 1
+; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[TMP9]], [[TMP2]]
+; CHECK-NEXT:    br i1 [[TMP10]], label %[[BB11:.*]], label %[[BB3]]
+; CHECK:       [[BB11]]:
+; CHECK-NEXT:    ret <vscale x 4 x float> [[TMP8]]
+;
+  %output = call <vscale x 4 x float> @llvm.log.nxv4f32(<vscale x 4 x float> %input)
+  ret <vscale x 4 x float> %output
+}
+
+define <4 x float> @fixed_vec_log(<4 x float> %input) {
+; CHECK-LABEL: define <4 x float> @fixed_vec_log(
+; CHECK-SAME: <4 x float> [[INPUT:%.*]]) {
+; CHECK-NEXT:    [[OUTPUT:%.*]] = call <4 x float> @llvm.log.v4f32(<4 x float> [[INPUT]])
+; CHECK-NEXT:    ret <4 x float> [[OUTPUT]]
+;
+  %output = call <4 x float> @llvm.log.v4f32(<4 x float> %input)
+  ret <4 x float> %output
+}
+
+declare <4 x float> @llvm.log.v4f32(<4 x float>) #0
+declare <vscale x 4 x float> @llvm.log.nxv4f32(<vscale x 4 x float>) #0
+
+; CHECK: attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
+; CHECK-NEXT: attributes #1 = { nocallback nofree nosync nounwind willreturn memory(none) }
+attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }

[kartcq]

Hi @arsenm Could you please review this change.
Also this needs to be cherrypicked to 21.x branch.
Kindly do the needful.

[steplong]

Hi @arsenm, could you review this?

[arsenm]

/cherry-pick 19ada02

[llvmbot]

/pull-request #153046