Merge pull request #15639 from jacobly0/signed-mod

llvm/cbe: fix signed `@mod`/`@divFloor` computations

2 files changed, 44 insertions, 28 deletions

diff --git a/src/codegen/c.zig b/src/codegen/c.zig
index 7da99de5c1..86b74b1429 100644
--- a/src/codegen/c.zig
+++ b/src/codegen/c.zig
@@ -6508,23 +6508,16 @@ fn airSplat(f: *Function, inst: Air.Inst.Index) !CValue {
 
     const inst_ty = f.air.typeOfIndex(inst);
     const inst_scalar_ty = inst_ty.scalarType();
-    const inst_scalar_cty = try f.typeToIndex(inst_scalar_ty, .complete);
-    const need_memcpy = f.indexToCType(inst_scalar_cty).tag() == .array;
 
     const writer = f.object.writer();
     const local = try f.allocLocal(inst, inst_ty);
     const v = try Vectorize.start(f, inst, writer, inst_ty);
-    if (need_memcpy) try writer.writeAll("memcpy(&");
+    const a = try Assignment.init(f, inst_scalar_ty);
     try f.writeCValue(writer, local, .Other);
     try v.elem(f, writer);
-    try writer.writeAll(if (need_memcpy) ", &" else " = ");
+    try a.assign(f, writer);
     try f.writeCValue(writer, operand, .Other);
-    if (need_memcpy) {
-        try writer.writeAll(", sizeof(");
-        try f.renderCType(writer, inst_scalar_cty);
-        try writer.writeAll("))");
-    }
-    try writer.writeAll(";\n");
+    try a.end(f, writer);
     try v.end(f, inst, writer);
 
     return local;
diff --git a/src/codegen/llvm.zig b/src/codegen/llvm.zig
index 2189663338..991ac04573 100644
--- a/src/codegen/llvm.zig
+++ b/src/codegen/llvm.zig
@@ -7215,20 +7215,28 @@ pub const FuncGen = struct {
             return self.buildFloatOp(.floor, inst_ty, 1, .{result});
         }
         if (scalar_ty.isSignedInt()) {
-            // const d = @divTrunc(a, b);
-            // const r = @rem(a, b);
-            // return if (r == 0) d else d - ((a < 0) ^ (b < 0));
-            const result_llvm_ty = try self.dg.lowerType(inst_ty);
-            const zero = result_llvm_ty.constNull();
-            const div_trunc = self.builder.buildSDiv(lhs, rhs, "");
+            const target = self.dg.module.getTarget();
+            const inst_llvm_ty = try self.dg.lowerType(inst_ty);
+            const scalar_bit_size_minus_one = scalar_ty.bitSize(target) - 1;
+            const bit_size_minus_one = if (inst_ty.zigTypeTag() == .Vector) const_vector: {
+                const vec_len = inst_ty.vectorLen();
+                const scalar_llvm_ty = try self.dg.lowerType(scalar_ty);
+
+                const shifts = try self.gpa.alloc(*llvm.Value, vec_len);
+                defer self.gpa.free(shifts);
+
+                @memset(shifts, scalar_llvm_ty.constInt(scalar_bit_size_minus_one, .False));
+                break :const_vector llvm.constVector(shifts.ptr, vec_len);
+            } else inst_llvm_ty.constInt(scalar_bit_size_minus_one, .False);
+
+            const div = self.builder.buildSDiv(lhs, rhs, "");
             const rem = self.builder.buildSRem(lhs, rhs, "");
-            const rem_eq_0 = self.builder.buildICmp(.EQ, rem, zero, "");
-            const a_lt_0 = self.builder.buildICmp(.SLT, lhs, zero, "");
-            const b_lt_0 = self.builder.buildICmp(.SLT, rhs, zero, "");
-            const a_b_xor = self.builder.buildXor(a_lt_0, b_lt_0, "");
-            const a_b_xor_ext = self.builder.buildZExt(a_b_xor, div_trunc.typeOf(), "");
-            const d_sub_xor = self.builder.buildSub(div_trunc, a_b_xor_ext, "");
-            return self.builder.buildSelect(rem_eq_0, div_trunc, d_sub_xor, "");
+            const div_sign = self.builder.buildXor(lhs, rhs, "");
+            const div_sign_mask = self.builder.buildAShr(div_sign, bit_size_minus_one, "");
+            const zero = inst_llvm_ty.constNull();
+            const rem_nonzero = self.builder.buildICmp(.NE, rem, zero, "");
+            const correction = self.builder.buildSelect(rem_nonzero, div_sign_mask, zero, "");
+            return self.builder.buildNSWAdd(div, correction, "");
         }
         return self.builder.buildUDiv(lhs, rhs, "");
     }
@@ -7280,12 +7288,27 @@ pub const FuncGen = struct {
             return self.builder.buildSelect(ltz, c, a, "");
         }
         if (scalar_ty.isSignedInt()) {
-            const a = self.builder.buildSRem(lhs, rhs, "");
-            const b = self.builder.buildNSWAdd(a, rhs, "");
-            const c = self.builder.buildSRem(b, rhs, "");
+            const target = self.dg.module.getTarget();
+            const scalar_bit_size_minus_one = scalar_ty.bitSize(target) - 1;
+            const bit_size_minus_one = if (inst_ty.zigTypeTag() == .Vector) const_vector: {
+                const vec_len = inst_ty.vectorLen();
+                const scalar_llvm_ty = try self.dg.lowerType(scalar_ty);
+
+                const shifts = try self.gpa.alloc(*llvm.Value, vec_len);
+                defer self.gpa.free(shifts);
+
+                @memset(shifts, scalar_llvm_ty.constInt(scalar_bit_size_minus_one, .False));
+                break :const_vector llvm.constVector(shifts.ptr, vec_len);
+            } else inst_llvm_ty.constInt(scalar_bit_size_minus_one, .False);
+
+            const rem = self.builder.buildSRem(lhs, rhs, "");
+            const div_sign = self.builder.buildXor(lhs, rhs, "");
+            const div_sign_mask = self.builder.buildAShr(div_sign, bit_size_minus_one, "");
+            const rhs_masked = self.builder.buildAnd(rhs, div_sign_mask, "");
             const zero = inst_llvm_ty.constNull();
-            const ltz = self.builder.buildICmp(.SLT, lhs, zero, "");
-            return self.builder.buildSelect(ltz, c, a, "");
+            const rem_nonzero = self.builder.buildICmp(.NE, rem, zero, "");
+            const correction = self.builder.buildSelect(rem_nonzero, rhs_masked, zero, "");
+            return self.builder.buildNSWAdd(rem, correction, "");
         }
         return self.builder.buildURem(lhs, rhs, "");
     }