[GlobalISel]: Simplify udiv lowering by determining known zeros

[AtariDreams]

No description provided.

[llvmbot]

@llvm/pr-subscribers-llvm-globalisel

Author: AtariDreams (AtariDreams)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/89678.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp (+3-1)
• (modified) llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll (+26)

diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
index c5ee354f13b7b4..2b300b63399427 100644
--- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
@@ -5069,6 +5069,8 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
   const unsigned EltBits = ScalarTy.getScalarSizeInBits();
   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
   LLT ScalarShiftAmtTy = ShiftAmtTy.getScalarType();
+  unsigned KnownLeadingZeros =
+      KB ? KB->getKnownBits(LHS).countMinLeadingZeros() : 0;
   auto &MIB = Builder;
 
   bool UseNPQ = false;
@@ -5087,7 +5089,7 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
     // TODO: Use undef values for divisor of 1.
     if (!Divisor.isOne()) {
       UnsignedDivisionByConstantInfo magics =
-          UnsignedDivisionByConstantInfo::get(Divisor);
+          UnsignedDivisionByConstantInfo::get(Divisor, KnownLeadingZeros);
 
       Magic = std::move(magics.Magic);
 
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
index 9a525151ca328b..c97a00ccdd4557 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
@@ -243,3 +243,29 @@ define <8 x i16> @pr38477(<8 x i16> %a0) {
   %1 = udiv <8 x i16> %a0, <i16 1, i16 119, i16 73, i16 -111, i16 -3, i16 118, i16 32, i16 31>
   ret <8 x i16> %1
 }
+
+define i32 @udiv_div_by_180(i32 %x)
+; SDAG-LABEL: udiv_div_by_180:
+; SDAG:       // %bb.0:
+; SDAG-NEXT:    mov w8, #5826 // =0x16c2
+; SDAG-NEXT:    and w9, w0, #0xff
+; SDAG-NEXT:    movk w8, #364, lsl #16
+; SDAG-NEXT:    umull x8, w9, w8
+; SDAG-NEXT:    lsr x0, x8, #32
+; SDAG-NEXT:    // kill: def $w0 killed $w0 killed $x0
+; SDAG-NEXT:    ret
+;
+; GISEL-LABEL: udiv_div_by_180:
+; GISEL:       // %bb.0:
+; GISEL-NEXT:    uxtb w8, w0
+; GISEL-NEXT:    mov w9, #5826 // =0x16c2
+; GISEL-NEXT:    movk w9, #364, lsl #16
+; GISEL-NEXT:    umull x8, w8, w9
+; GISEL-NEXT:    lsr x0, x8, #32
+; GISEL-NEXT:    // kill: def $w0 killed $w0 killed $x0
+; GISEL-NEXT:    ret
+{
+  %truncate = and i32 %x, 255
+  %udiv = udiv i32 %truncate, 180
+  ret i32 %udiv
+}

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: AtariDreams (AtariDreams)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/89678.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp (+3-1)
• (modified) llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll (+26)

diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
index c5ee354f13b7b4..2b300b63399427 100644
--- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp
@@ -5069,6 +5069,8 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
   const unsigned EltBits = ScalarTy.getScalarSizeInBits();
   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
   LLT ScalarShiftAmtTy = ShiftAmtTy.getScalarType();
+  unsigned KnownLeadingZeros =
+      KB ? KB->getKnownBits(LHS).countMinLeadingZeros() : 0;
   auto &MIB = Builder;
 
   bool UseNPQ = false;
@@ -5087,7 +5089,7 @@ MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
     // TODO: Use undef values for divisor of 1.
     if (!Divisor.isOne()) {
       UnsignedDivisionByConstantInfo magics =
-          UnsignedDivisionByConstantInfo::get(Divisor);
+          UnsignedDivisionByConstantInfo::get(Divisor, KnownLeadingZeros);
 
       Magic = std::move(magics.Magic);
 
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
index 9a525151ca328b..c97a00ccdd4557 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/combine-udiv.ll
@@ -243,3 +243,29 @@ define <8 x i16> @pr38477(<8 x i16> %a0) {
   %1 = udiv <8 x i16> %a0, <i16 1, i16 119, i16 73, i16 -111, i16 -3, i16 118, i16 32, i16 31>
   ret <8 x i16> %1
 }
+
+define i32 @udiv_div_by_180(i32 %x)
+; SDAG-LABEL: udiv_div_by_180:
+; SDAG:       // %bb.0:
+; SDAG-NEXT:    mov w8, #5826 // =0x16c2
+; SDAG-NEXT:    and w9, w0, #0xff
+; SDAG-NEXT:    movk w8, #364, lsl #16
+; SDAG-NEXT:    umull x8, w9, w8
+; SDAG-NEXT:    lsr x0, x8, #32
+; SDAG-NEXT:    // kill: def $w0 killed $w0 killed $x0
+; SDAG-NEXT:    ret
+;
+; GISEL-LABEL: udiv_div_by_180:
+; GISEL:       // %bb.0:
+; GISEL-NEXT:    uxtb w8, w0
+; GISEL-NEXT:    mov w9, #5826 // =0x16c2
+; GISEL-NEXT:    movk w9, #364, lsl #16
+; GISEL-NEXT:    umull x8, w8, w9
+; GISEL-NEXT:    lsr x0, x8, #32
+; GISEL-NEXT:    // kill: def $w0 killed $w0 killed $x0
+; GISEL-NEXT:    ret
+{
+  %truncate = and i32 %x, 255
+  %udiv = udiv i32 %truncate, 180
+  ret i32 %udiv
+}

[jayfoad]

LGTM

[jayfoad]

@AtariDreams why do you keep force-pushing? It makes it hard for reviewers to know what has changed since the last time they looked at the PR.

[AtariDreams]

@AtariDreams why do you keep force-pushing? It makes it hard for reviewers to know what has changed since the last time they looked at the PR.

Failed tests and I reran them to ensure everything is good.

[inclyc]

Failed tests and I reran them to ensure everything is good.

A simple way to do such things is to utilize git commit --fixup, and use git rebase -i main --autosquash. (The rebasing should be performed right after the change is ready to be merged, and perform only once.)